The Latest Research AI and Maschine Learning

Table of contents :
Untitled

Citation preview

WITH DISCUSSION QUESTIONS

THE LATEST RESEARCH AI AND MACHINE LEARNING

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

Introduction

“In the sphere of business, AI is poised have a transformational impact … the bottleneck now is in management, implementation, and business imagination.” ERIK BRYNJOLFSSON AND ANDREW MCAFEE, “The Business of Artificial Intelligence” There’s no longer any doubt that AI and machine learning will transform entire professions and industries in the coming years. And while the long-term reverberations of AI’s expanding role in society are fascinating and worrying, at HBR our authors are focused on how organizations can put it into use now. We all still wonder (and worry about) what business will look like in a decade, but machine intelligence is a technology with implications for your company today. In this collection, we’ve selected some of HBR’s most important research, analysis, and case studies on AI and machine learning from the past two years to help you understand what you and your colleagues need to know about machine intelligence, now. The first section, Understanding AI and

Machine Learning, explains the current business potential of AI, helps you identify opportunities in your organization, and provides an inside look at how leaders in the field of AI are deploying the technology in their own companies. The next section, How AI Is Being Used Now, presents current uses of AI ranging from back-office processes in health care to research on board selection to out-of-the-box marketing products for automotive retail. Finally, The Near Future looks ahead to what’s coming next in AI, with our authors predicting that we’ll see new ways of collaborating with machines and increasing automation in more industries and job functions.

The Latest Research: AI and Machine Learning provides you with today’s essential thinking on machine intelligence in business. Use it to kick off urgent conversations about how to implement AI initiatives in your organization. At the end of each article, you’ll find discussion questions to share with your team and help you apply what you’ve learned across your company.

©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTS RESERVED.

1

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

Table of Contents UNDERSTANDING AI AND MACHINE LEARING: WHAT THEY ARE AND HOW TO USE THEM 4

The Business of Artificial Intelligence By Erik Brynjolfsson and Andrew McAfee

14 Deep Learning Will Radically Change the Ways We Interact with Technology By Aditya Singh 24 Inside Facebook’s AI Workshop An Interview with Joaquin Candela by Scott Berinato 31 How to Spot a Machine Learning Opportunity, Even If You Aren’t a Data Scientist By Kathryn Hume 37 Is Your Company’s Data Actually Valuable in the AI Era? By Ajay Agrawal, Joshua Gans, and Avi Goldfarb 42 How AI Fits into Your Data Science Team Q&A with Hilary Mason

HOW AI IS BEING USED NOW: NEW PROCESSES, INSIGHTS, AND PRODUCTS 46 Artificial Intelligence for the Real World By Thomas H. Davenport and Rajeev Ronanki

80 How Harley-Davidson Used Artificial Intelligence to Increase New York Sales Leads by 2,390% By Brad Power 86 Machine Learning Can Help B2B Firms Learn More About Their Customers By Stephan Kudyba and Thomas H. Davenport 92 How We Built a Virtual Scheduling Assistant at Microsoft By Andrés Monroy-Hernández and Justin Cranshaw 98 How Companies Are Already Using AI By Satya Ramaswamy

THE NEAR FUTURE: CHANGES TO JOBS, THE ECONOMY, AND SOCIETY 106 Collaborative Intelligence: Humans and AI Are Joining Forces By H. James Wilson and Paul Daugherty 118 Robo-Advisers Are Coming to Consulting and Corporate Strategy By Thomas H. Davenport, Barry Libert, and Megan Beck 125 10 Promising AI Applications in Health Care By Brian Kalis, Matt Collier, and Richard Fu

57 How AI Is Taking the Scut Work Out of Health Care By Jonathan Bush

131 The First Wave of Corporate AI Is Doomed to Fail By Kartik Hosanagar and Apoorv Saxena

62 3 Ways Companies Are Building a Business Around AI By Quentin Hardy

137 AI Is the Future of Cybersecurity, for Better and for Worse By Roman V. Yampolskiy

69 What Changes When AI Is So Accessible That Everyone Can Use It? By H. James Wilson and Paul Daugherty

142 How AI Could Help the Public Sector By Emma Martinho-Truswell

74 Research: Could Machine Learning Help Companies Select Better Board Directors? By Isil Erel, Léa H. Stern, Chenhao Tan, and Michael S. Weisbach

©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTS RESERVED.

147 How Will AI Change Work? Here Are 5 Schools of Thought By Mark Knickrehm 153 Related Resources

22

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

UNDERSTANDING AI AND MACHINE LEARNING: WHAT THEY ARE AND HOW TO USE THEM

©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTS RESERVED.

3

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

THE BUSINESS OF ARTIFICIAL INTELLIGENCE WHAT IT CAN — AND CANNOT — DO FOR YOUR ORGANIZATION BY ERIK BRYNJOLFSSON AND ANDREW MCAFEE For more than 250 years the fundamental drivers of economic growth have been technological innovations. The most important of these are what economists call general-purpose technologies — a category that includes the steam engine, electricity, and the internal combustion engine. Each one catalyzed waves of complementary innovations and opportunities. The internal combustion engine, for example, gave rise to cars, trucks, airplanes, chain saws, and lawnmowers, along with big-box retailers, shopping centers, crossdocking warehouses, new supply chains, and, when you think about it, suburbs. Companies as diverse as Walmart, UPS, and Uber found ways to leverage the technology to create profitable new business models. ©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTS RESERVED.

4

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

ERIK BRYNJOLFSSON AND ANDREW MCAFEE



THE AUTHORS

ERIK BRYNJOLFSSON AND ANDREW MCAFEE The most important general-purpose technology of our era is artificial intelligence, particularly machine learning (ML) — that is, the machine’s ability to keep improving its performance without humans having to explain exactly how to accomplish all the tasks it’s given. Within just the past few years machine learning has become far more effective and widely available. We can now build systems that learn how to perform tasks on their own. Why is this such a big deal? Two reasons. First, we humans know more than we can tell: We can’t explain exactly how we’re able to do a lot of things — from recognizing a face to making a smart move in the ancient Asian strategy game of Go. Prior to ML, this inability to articulate our own knowledge meant that we couldn’t automate many tasks. Now we can. Second, ML systems are often excellent learners. They can achieve superhuman performance in a wide range of activities, including detecting fraud and diagnosing disease. Excellent digital learners are being deployed across the economy, and their impact will be profound. In the sphere of business, AI is poised have a transformational impact, on the scale of earlier general-purpose technologies. Although it is already in use in thousands of companies around the world, most big opportunities have not yet been tapped. The effects of AI will be magnified in the coming decade, as manufacturing, retailing, transportation, finance, health care, law, advertising, insurance, entertainment, education, and virtually every other industry transform their core processes and business models to take advantage of machine learning. The bottleneck now is in management, implementation, and business imagination. Like so many other new technologies, however, AI has generated lots of unrealistic expectations. We see business plans liberally sprinkled with references to machine learning, neural nets, and other forms of the technology, with little connection to its real capabilities. Simply calling a dating site “AI-powered,” for example doesn’t make it any more effective, but it might help with fundraising. This article will cut through the noise to describe the real potential of AI, its practical implications, and the barriers to its adoption.

Erik Brynjolfsson (@erikbryn) is the director of MIT’s Initiative on the Digital Economy, the Schussel Family Professor of Management Science at the MIT Sloan School of Management, and a research associate at NBER. His research examines the effects of information technologies on business strategy, productivity and performance, digital commerce, and intangible assets. At MIT he teaches courses on the economics of information and the Analytics Lab. Brynjolfsson was among the first researchers to measure IT’s productivity contributions and the complementary role of organizational capital and other intangibles. His research provided the first quantification of online product variety value, known as the “long tail,”

and developed pricing and bundling models for information goods. He earned his AB and his SM in applied mathematics and decision sciences at Harvard and his PhD in managerial economics at the Sloan School. Brynjolfsson is the author of several books, including, with Andrew McAfee, Machine, Platform, Crowd: Harnessing Our Digital Future (2017) and the New York Times best seller The Second Machine Age: Work, Progress, and Prosperity in a Time of Brilliant Technologies (2014). You can find his papers here. Andrew McAfee (@amcafee), a principal research scientist at MIT, studies how digital technologies are changing business, the economy, and society. With Erik Brynjolfsson he coauthored Machine, Platform, Crowd:

Harnessing Our Digital Future (2017) and The Second Machine Age: Work, Progress, and Prosperity in a Time of Brilliant Technologies (2014), which was a New York Times best seller and was shortlisted for the Financial Times/ McKinsey Business Book of the Year Award. McAfee writes academic papers, a blog for the Financial Times, and articles for publications including Harvard Business Review, the Economist, the Wall Street Journal, and the New York Times. He has talked about his work on Charlie Rose and 60 Minutes; at TED, Davos, and the Aspen Ideas Festival; and in front of many other audiences. McAfee was educated at Harvard and MIT, where he is a cofounder of the institute’s Initiative on the Digital Economy.

WHAT CAN AI DO TODAY?

The term artificial intelligence was coined in 1955 by John McCarthy, a math professor at Dartmouth who organized the seminal conference on the topic the following year. Ever since, perhaps in part because of its evocative name, the field has given rise to more than its share of fantastic claims and promises. In 1957 the economist Herbert Simon predicted that computers would beat humans at chess within 10 years. (It took 40.) In 1967 the cognitive scientist Marvin Minsky said, “Within a generation the problem of creating ‘artificial intelligence’ will be substantially solved.” Simon and Minsky were both intellectual giants, but they erred badly. Thus it’s understandable that dramatic claims about future breakthroughs meet with a certain amount of skepticism.

ALTHOUGH AI IS ALREADY IN USE IN THOUSANDS OF COMPANIES AROUND THE WORLD, MOST BIG OPPORTUNITIES HAVE NOT YET BEEN TAPPED. ©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTS RESERVED.

5

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

ERIK BRYNJOLFSSON AND ANDREW MCAFEE



Let’s start by exploring what AI is already doing and how quickly it is improving. The biggest advances have been in two broad areas: perception and cognition. In the former category some of the most practical advances have been made in relation to speech. Voice recognition is still far from perfect, but millions of people are now using it — think Siri, Alexa, and Google Assistant. The text you are now reading was originally dictated to a computer and transcribed with sufficient accuracy to make it faster than typing. A study by the Stanford computer scientist James Landay and colleagues found that speech recognition is now about three times as fast, on average, as typing on a cell phone. The error rate, once 8.5%, has dropped to 4.9%. What’s striking is that this substantial improvement has come not over the past 10 years but just since the summer of 2016. Image recognition, too, has improved dramatically. You may have noticed that Facebook and other apps now recognize many of your friends’ faces in posted photos and prompt you to tag them with their names. An app running on your smartphone will recognize virtually any bird in the wild. Image recognition is even replacing ID cards at corporate headquarters. Vision systems, such as those used in self-driving cars, formerly made a mistake when identifying a pedestrian as often as once in 30 frames (the cameras in these systems record about 30 frames a second); now they err less often than once in 30 million frames. The error rate for recognizing images from a large database called ImageNet, with several million photographs of common, obscure, or downright weird images, fell from higher than 30% in 2010 to about 4% in 2016 for the best systems. (See the exhibit “Puppy or Muffin?”) The speed of improvement has accelerated rapidly in recent years as a new approach, based on very large or “deep” neural nets, was adopted. The ML approach for vision systems is still far from flawless — but even people have trouble quickly recognizing puppies’ faces or, more embarrassingly, see their cute faces where none exist. The second type of major improvement has been in cognition and problem solving. Machines have already beaten the finest (human) players of poker and Go — achievements that experts had predicted would take at least another decade. Google’s DeepMind team has used ML systems to improve the cooling efficiency at data centers by more than 15%, even after they were optimized by human experts. Intelligent agents are being used by the cybersecurity company Deep Instinct to detect malware, and by PayPal to prevent money laundering. A system using IBM technology automates the claims process at an insurance company in Singapore, and a system from Lumidatum, a data

©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTS RESERVED.

PUPPY OR MUFFIN? PROGRESS IN IMAGE RECOGNITION Machines have made real strides in distinguishing among similar-looking categories of images.

KAREN ZACK/@TEENYBISCUIT

VISION ERROR RATE 30%

Algorithms

25 20 15 10

Humans

5 0

2010

2011

2012

2013

2014

SOURCE ELECTRONIC FRONTIER FOUNDATION

 HBR.ORG THE BIG IDEA 5

2015

2016

© HBR.ORG

6

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

ERIK BRYNJOLFSSON AND ANDREW MCAFEE



science platform firm, offers timely advice to improve customer support. Dozens of companies are using ML to decide which trades to execute on Wall Street, and more and more credit decisions are made with its help. Amazon employs ML to optimize inventory and improve product recommendations to customers. Infinite Analytics developed one ML system to predict whether a user would click on a particular ad, improving online ad placement for a global consumer packaged goods company, and another to improve customers’ search and discovery process at a Brazilian online retailer. The first system increased advertising ROI threefold, and the second resulted in a $125 million increase in annual revenue. Machine learning systems are not only replacing older algorithms in many applications, but are now superior at many tasks that were once done best by humans. Although the systems are far from perfect, their error rate — about 5% — on the ImageNet database is at or better than human-level performance. Voice recognition, too, even in noisy environments, is now nearly equal to human performance. Reaching this threshold opens up vast new possibilities for transforming the workplace and the economy. Once AI-based systems surpass human performance at a given task, they are much likelier to spread quickly. For instance, Aptonomy and Sanbot, makers respectively of drones and robots, are using improved vision systems to automate much of the work of security guards. The software company Affectiva, among others, is using them to recognize emotions such as joy, surprise, and anger in focus groups. And Enlitic is one of several deep-learning start-ups that use them to scan medical images to help diagnose cancer. These are impressive achievements, but the applicability of AI-based systems is still quite narrow. For instance, their remarkable performance on the ImageNet database, even with its millions of images, doesn’t always translate into similar success “in the wild,” where lighting conditions, angles, image resolution, and context may be very different. More fundamentally, we can marvel at a system that understands Chinese speech and translates it into English, but we don’t expect such a system to know what a particular Chinese character means — let alone where to eat in Beijing. If someone performs a task well, it’s natural to assume that the person has some competence in related tasks. But ML systems are trained to do specific tasks, and typically their knowledge does not generalize. The fallacy that a computer’s narrow understanding implies broader understanding is perhaps the biggest source of confusion, and exaggerated claims, about AI’s progress. We are far from machines that exhibit general intelligence across diverse domains.

©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTS RESERVED.

SUPERVISED LEARNING SYSTEMS

As two pioneers in the field, Tom Mitchell and Michael I. Jordan, have noted, most of the recent progress in machine learning involves mapping from a set of inputs to a set of outputs. Some examples: Input X

Output Y

Application

Voice recording

Transcript

Speech recognition

Historical market data

Future market data

Trading bots

Photograph

Caption

Image tagging

Drug chemical properties Treatment efficacy

Pharma R&D

Store transaction details

Is the transaction fraudulent? Fraud detection

Recipe ingredients

Customer reviews

Food recommendations

Purchase histories

Future purchase behavior

Customer retention

Car locations and speed

Traffic flow

Traffic lights

Faces

Names

Face recognition

UNDERSTANDING MACHINE LEARNING

The most important thing to understand about ML is that it represents a fundamentally different approach to creating software: The machine learns from examples, rather than being explicitly programmed for a particular outcome. This is an important break from previous practice. For most of the past 50 years, advances in information technology and its applications have focused on codifying existing knowledge and procedures and embedding them in machines. Indeed, the term “coding” denotes the painstaking process of transferring knowledge from developers’ heads into a form that machines can understand and execute. This approach has a fundamental weakness: Much of the knowledge we all have is tacit, meaning that we can’t fully explain it. It’s nearly impossible for us to write down instructions that would enable another person to learn how to ride a bike or to recognize a friend’s face. In other words, we all know more than we can tell. This fact is so important that it has a name: Polanyi’s Paradox, for the philosopher and polymath Michael Polanyi, who described it in 1964. Polanyi’s Paradox not only limits what we can tell one another but has historically placed a fundamental restriction on our ability to endow machines with intelligence. For a long time that

 HBR.ORG THE BIG IDEA 6

7

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

ERIK BRYNJOLFSSON AND ANDREW MCAFEE



limited the activities that machines could productively perform in the economy. Machine learning is overcoming those limits. In this second wave of the second machine age, machines built by humans are learning from examples and using structured feedback to solve on their own problems such as Polanyi’s classic one of recognizing a face.

DIFFERENT FLAVORS OF MACHINE LEARNING

Artificial intelligence and machine learning come in many flavors, but most of the successes in recent years have been in one category: supervised learning systems, in which the machine is given lots of examples of the correct answer to a particular problem. This process almost always involves mapping from a set of inputs, X, to a set of outputs, Y. For instance, the inputs might be pictures of various animals, and the correct outputs might be labels for those animals: dog, cat, horse. The inputs could also be waveforms from a sound recording and the outputs could be words: “yes,” “no,” “hello,” “good-bye.” (See the exhibit “Supervised Learning Systems.”) Successful systems often use a training set of data with thousands or even millions of examples, each of which has been labeled with the correct answer. The system can then be let loose to look at new examples. If the training has gone well, the system will predict answers with a high rate of accuracy. The algorithms that have driven much of this success depend on an approach called deep learning, which uses neural networks. Deep learning algorithms have a significant advantage over earlier generations of ML algorithms: They can make better use of much larger data sets. The old systems would improve as the number of examples in the training data grew, but only up to a point, after which additional data didn’t lead to better predictions. According to Andrew Ng, one of the giants

©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTS RESERVED.

of the field, deep neural nets don’t seem to level off in this way: More data leads to better and better predictions. Some very large systems are trained by using 36 million examples or more. Of course, working with extremely large data sets requires more and more processing power, which is one reason the very big systems are often run on supercomputers or specialized computer architectures. Any situation in which you have a lot of data on behavior and are trying to predict an outcome is a potential application for supervised learning systems. Jeff Wilke, who leads Amazon’s consumer business, says that supervised learning systems have largely replaced the memory-based filtering algorithms that were used to make personalized recommendations to customers. In other cases, classic algorithms for setting inventory levels and optimizing supply chains have been replaced by more efficient and robust systems based on machine learning. JPMorgan Chase introduced a system for reviewing commercial loan contracts; work that used to take loan officers 360,000 hours can now be done in a few seconds. And supervised learning systems are now being used to diagnose skin cancer. These are just a few examples. It’s comparatively straightforward to label a body of data and use it to train a supervised learner; that’s why supervised ML systems are more common than unsupervised ones, at least for now. Unsupervised learning systems seek to learn on their own. We humans are excellent unsupervised learners: We pick up most of our knowledge of the world (such as how to recognize a tree) with little or no labeled data. But it is exceedingly difficult to develop a successful machine learning system that works this way. If and when we learn to build robust unsupervised learners, exciting possibilities will open up. These machines could look at complex problems in fresh ways to

 HBR.ORG THE BIG IDEA 7

ABOVE: THIS IS WHAT IT MEANS TO WORK WITH ARTIFICIAL INTELLIGENCE. THE RESULTS ARE HUMAN AND NOT HUMAN. RECOGNIZABLE BUT ALSO UNEXPECTED. ARE THEY BEAUTIFUL? FRIGHTENING? DELIGHTFUL?

8

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

ERIK BRYNJOLFSSON AND ANDREW MCAFEE



help us discover patterns — in the spread of diseases, in price moves across securities in a market, in customers’ purchase behaviors, and so on — that we are currently unaware of. Such possibilities lead Yann LeCun, the head of AI research at Facebook and a professor at NYU, to compare supervised learning systems to the frosting on the cake and unsupervised learning to the cake itself. Another small but growing area within the field is reinforcement learning. This approach is embedded in systems that have mastered Atari video games and board games like Go. It is also helping to optimize data center power usage and to develop trading strategies for the stock market. Robots created by Kindred use machine learning to identify and sort objects they’ve never

encountered before, speeding up the “pick and place” process in distribution centers for consumer goods. In reinforcement learning systems the programmer specifies the current state of the system and the goal, lists allowable actions, and describes the elements of the environment that constrain the outcomes for each of those actions. Using the allowable actions, the system has to figure out how to get as close to the goal as possible. These systems work well when humans can specify the goal but not necessarily how to get there. For instance, Microsoft used reinforcement learning to select headlines for MSN.com news stories by “rewarding” the system with a higher score when more visitors clicked on the link. The system tried to maximize its score on the basis of the rules its designers gave it. Of

BELOW: LOOK DEEP, AND YOU’LL SEE THE HUMAN IN THE ALGORITHM. LOOK DEEPER, AND YOU’LL SEE THE ALGORITHM IN THE INTELLIGENCE.

DESIGNING AND IMPLEMENTING NEW COMBINATIONS OF TECHNOLOGIES, HUMAN SKILLS, AND CAPITAL ASSETS TO MEET CUSTOMERS’ NEEDS REQUIRES LARGE-SCALE CREATIVITY AND PLANNING. IT IS A TASK THAT MACHINES ARE NOT VERY GOOD AT.

©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTS RESERVED.

 HBR.ORG THE BIG IDEA 8

9

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

ERIK BRYNJOLFSSON AND ANDREW MCAFEE



course, this means that a reinforcement learning system will optimize for the goal you explicitly reward, not necessarily the goal you really care about (such as lifetime customer value), so specifying the goal correctly and clearly is critical.

PUTTING MACHINE LEARNING TO WORK

There are three pieces of good news for organizations looking to put ML to use today. First, AI skills are spreading quickly. The world still has not nearly enough data scientists and machine learning experts, but the demand for them is being met by online educational resources as well as by universities. The best of these, including Udacity, Coursera, and fast.ai, do much more than teach introductory concepts; they can actually get smart, motivated students to the point of being able to create industrial-grade ML deployments. In addition to training their own people, interested companies can use online talent platforms such as Upwork, Topcoder, and Kaggle to find ML experts with verifiable expertise. The second welcome development is that the necessary algorithms and hardware for modern AI can be bought or rented as needed. Google, Amazon, Microsoft, Salesforce, and other companies are making powerful ML infrastructure available via the cloud. The cutthroat competition among these rivals means that companies that want to experiment with or deploy ML will see more and more capabilities available at ever-lower prices over time. The final piece of good news, and probably the most underappreciated, is that you may not need all that much data to start making productive use of ML. The performance of most machine learning systems improves as they’re given more data to work with, so it seems logical to conclude that the company with the most data will win. That might be the case if “win” means “dominate the global market for a single application such as ad targeting or speech recognition.” But if success is defined instead as significantly improving performance, then sufficient data is often surprisingly easy to obtain. For example, Udacity cofounder Sebastian Thrun noticed that some of his salespeople were much more effective than others when replying to inbound queries in a chat room. Thrun and his graduate student Zayd Enam realized that their chat room logs were essentially a set of labeled training data — exactly what a supervised learning system needs. Interactions that led to a sale were labeled successes, and all others were labeled failures. Zayd used the data to predict what answers successful salespeople were likely to give in response to certain very common inquiries and then shared those predictions with the other salespeople to nudge them toward better performance. After 1,000 training cycles,

©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTS RESERVED.

the salespeople had increased their effectiveness by 54% and were able to serve twice as many customers at a time. The AI start-up WorkFusion takes a similar approach. It works with companies to bring higher levels of automation to back-office processes such as paying international invoices and settling large trades between financial institutions. The reason these processes haven’t been automated yet is that they’re complicated; relevant information isn’t always presented the same way every time (“How do we know what currency they’re talking about?”), and some interpretation and judgment are necessary. WorkFusion’s software watches in the background as people do their work and uses their actions as training data for the cognitive task of classification (“This invoice is in dollars. This one is in yen. This one is in euros…”). Once the system is confident enough in its classifications, it takes over the process. Machine learning is driving changes at three levels: tasks and occupations, business processes, and business models. An example of task-and-occupation redesign is the use of machine vision systems to identify potential cancer cells — freeing up radiologists to focus on truly critical cases, to communicate with patients, and to coordinate with other physicians. An example of process redesign is the reinvention of the workflow and layout of Amazon fulfillment centers after the introduction of robots and optimization algorithms based on machine learning. Similarly, business models need to be rethought to take advantage of ML systems that can intelligently recommend music or movies in a personalized way. Instead of selling songs à la carte on the basis of consumer choices, a better model might offer a subscription to a personalized station that predicted and played music a particular customer would like, even if the person had never heard it before. Note that machine learning systems hardly ever replace the entire job, process, or business model. Most often they complement human activities, which can make their work ever more valuable. The most effective rule for the new division of labor is rarely, if ever, “give all tasks to the machine.” Instead, if the successful completion of a process requires 10 steps, one or two of them may become automated while the rest become more valuable for humans to do. For instance, the chat room sales support system at Udacity didn’t try to build a bot that could take over all the conversations; rather, it advised human salespeople about how to improve their performance. The humans remained in charge but became vastly more effective and efficient. This approach is usually much more feasible than trying to design machines that can do everything humans can do. It often leads to better, more satisfying work for the

 HBR.ORG THE BIG IDEA 9

10

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

ERIK BRYNJOLFSSON AND ANDREW MCAFEE



WHILE ALL THE RISKS OF AI ARE VERY REAL, THE APPROPRIATE BENCHMARK IS NOT PERFECTION BUT THE BEST AVAILABLE ALTERNATIVE. people involved and ultimately to a better outcome for customers. Designing and implementing new combinations of technologies, human skills, and capital assets to meet customers’ needs requires large-scale creativity and planning. It is a task that machines are not very good at. That makes being an entrepreneur or a business manager one of society’s most rewarding jobs in the age of ML.

RISKS AND LIMITS

The second wave of the second machine age brings with it new risks. In particular, machine learning systems often have low “interpretability,” meaning that humans have difficulty figuring out how the systems reached their decisions. Deep neural networks may have hundreds of millions of connections, each of which contributes a small amount to the ultimate decision. As a result, these systems’ predictions tend to resist simple, clear explanation. Unlike humans, machines are not (yet!) good storytellers. They can’t always give a rationale for why a particular applicant was accepted or rejected for a job, or a particular medicine was recommended. Ironically, even as we have begun to overcome Polanyi’s Paradox, we’re facing a kind of reverse version: Machines know more than they can tell us. This creates three risks. First, the machines may have hidden biases, derived not from any intent of the designer but from the data provided to train the system. For instance, if a system learns which job applicants to accept for an interview by using a data set of decisions made by human recruiters in the past, it may inadvertently learn to perpetuate their racial, gender, ethnic, or other biases. Moreover, these biases may not appear as an explicit rule but, rather, be embedded in subtle interactions among the thousands of factors considered. A second risk is that, unlike traditional systems built on explicit logic rules, neural network systems deal with statistical truths rather than literal truths. That can make it difficult, if not impossible, to prove with complete certainty that the system will work in all cases — especially in situations that weren’t represented in the training data. Lack of verifiability can be a concern in mission-critical applications, such as controlling a nuclear power plant, or when life-or-death decisions are involved.

©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTS RESERVED.

Third, when the ML system does make errors, as it almost inevitably will, diagnosing and correcting exactly what’s going wrong can be difficult. The underlying structure that led to the solution can be unimaginably complex, and the solution may be far from optimal if the conditions under which the system was trained change. While all these risks are very real, the appropriate benchmark is not perfection but the best available alternative. After all, we humans, too, have biases, make mistakes, and have trouble explaining truthfully how we arrived at a particular decision. The advantage of machine-based systems is that they can be improved over time and will give consistent answers when presented with the same data. Does that mean there is no limit to what artificial intelligence and machine learning can do? Perception and cognition cover a great deal of territory — from driving a car to forecasting sales to deciding whom to hire or promote. We believe the chances are excellent that AI will soon reach superhuman levels of performance in most or all of these areas. So what won’t AI and ML be able to do? We sometimes hear “Artificial intelligence will never be good at assessing emotional, crafty, sly, inconsistent human beings — it’s too rigid and impersonal for that.” We don’t agree. ML systems like those at Affectiva are already at or beyond human-level performance in discerning a person’s emotional state on the basis of tone of voice or facial expression. Other systems can infer when even the world’s best poker players are bluffing well enough to beat them at the amazingly complex game Heads-up No-Limit Texas Hold’em. Reading people accurately is subtle work, but it’s not magic. It requires perception and cognition — exactly the areas in which ML is currently strong and getting stronger all the time. A great place to start a discussion of the limits of AI is with Pablo Picasso’s observation about computers: “But they are useless. They can only give you answers.” They’re actually far from useless, as ML’s recent triumphs show, but Picasso’s observation still provides insight. Computers are devices for answering questions, not for posing them. That means entrepreneurs, innovators, scientists, creators, and other kinds of people who figure out what problem or opportunity to tackle next, or what new territory to explore, will continue to be essential.

 HBR.ORG THE BIG IDEA 10

11

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

ERIK BRYNJOLFSSON AND ANDREW MCAFEE



Similarly, there’s a huge difference between passively assessing someone’s mental state or morale and actively working to change it. ML systems are getting quite good at the former but remain well behind us at the latter. We humans are a deeply social species; other humans, not machines, are best at tapping into social drives such as compassion, pride, solidarity, and shame in order to persuade, motivate, and inspire. In 2014 the TED Conference and the XPrize Foundation announced an award for “the first artificial intelligence to come to this stage and give a TED Talk compelling enough to win a standing ovation from the audience.” We doubt the award will be claimed anytime soon. We think the biggest and most important opportunities for human smarts in this new age of superpowerful ML lie at the intersection of two areas: figuring out what problems to work on next, and persuading a lot of people to tackle them and go along with the solutions. This is a decent definition of leadership, which is becoming much more important in the second machine age. The status quo of dividing up work between minds and machines is falling apart very quickly. Companies that stick with it are going to find themselves at an ever-greater competitive disadvantage compared with rivals who are willing and able to put ML to use in all the places where it is appropriate and who can figure out how to effectively integrate its capabilities with humanity’s. A time of tectonic change in the business world has begun, brought on by technological progress. As was the case with steam power and electricity, it’s not access to the new technologies themselves, or even to the best technologists, that separates winners from losers. Instead, it’s innovators who are open-minded enough to see past the status quo and envision very different approaches, and savvy enough to put them into place. One of machine learning’s greatest legacies may well be the creation of a new generation of business leaders. In our view, artificial intelligence, especially machine learning, is the most important general-purpose technology of our era. The impact of these innovations on business and the economy will be reflected not only in their direct contributions but also in their ability to enable and inspire complementary innovations. New products and processes are being made possible by better vision systems, speech recognition, intelligent problem solving, and many other capabilities that machine learning delivers. Some experts have gone even further. Gil Pratt, who now heads the Toyota Research Institute, has compared the current wave of AI technology to the Cambrian explosion 500 million years ago that birthed a tremendous variety of new life forms. Then as now, one of the key new capabilities was vision. When animals first gained

©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTS RESERVED.

this capability, it allowed them to explore the environment far more effectively; that catalyzed an enormous increase in the number of species, both predators and prey, and in the range of ecological niches that were filled. Today as well we expect to see a variety of new products, services, processes, and organizational forms and also numerous extinctions. There will certainly be some weird failures along with unexpected successes. Although it is hard to predict exactly which companies will dominate in the new environment, a general principle is clear: The most nimble and adaptable companies and executives will thrive. Organizations that can rapidly sense and respond to opportunities will seize the advantage in the AI-enabled landscape. So the successful strategy is to be willing to experiment and learn quickly. If managers aren’t ramping up experiments in the area of machine learning, they aren’t doing their job. Over the next decade, AI won’t replace managers, but managers who use AI will replace those who don’t. 

 HBR.ORG THE BIG IDEA 11

12

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

Discussion Questions

The Business of Artificial Intelligence BY ERIK BRYNJOLFSSON AND ANDREW MCAFEE

1. Consider where the AI bottlenecks are in your company. Is senior leadership willing to imagine your company’s transformation? Are existing structures too entrenched? Do you have enough data science capabilities? Is risk aversion or fear of failure keeping your company from piloting AI initiatives? 2. Think about the types of problems that people in your company see as candidates for machine learning initiatives. Can you articulate how they are similar or different from each other? (For instance, does one require image recognition, while another requires understanding spoken language?) Which possible initiatives might help your company develop expertise that it can apply to other AI challenges? 3. Is AI still in the “science fiction” or “something we’ll do in the future” or “wait and see” stage at your company? What processes seem as if they are the likeliest candidates for automation? What first steps can you take?

Over the next decade, AI won’t replace managers, but managers who use AI will replace those who don’t.

COPYRIGHT © 2018 HARVARD BUSINESS SCHOOL PUBLISHING ALL RIGHTS RESERVED. ©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTSCORPORATION. RESERVED.

134

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

REPRINT H03F8N PUBLISHED ON HBR.ORG JANUARY 30, 2017

ARTICLE TECHNOLOGY

Deep Learning Will Radically Change the Ways We Interact with Technology by Aditya Singh

©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTS RESERVED.

14

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

TECHNOLOGY

Deep Learning Will Radically Change the Ways We Interact with Technology

by Aditya Singh JANUARY 30, 2017

Even though heat and sound are both forms of energy, when you were a kid, you probably didn’t need to be told not to speak in thermal convection. And each time your children come across a stray animal, they likely don’t have to self-consciously rehearse a subroutine of zoological attributes to decide whether it’s a cat or a dog. Human beings come pre-loaded with the cognitive gear to simply

COPYRIGHT © 2017 HARVARD BUSINESS SCHOOL PUBLISHING ALL RIGHTS RESERVED. ©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTSCORPORATION. RESERVED.

152

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

perceive these distinctions. The differences appear so obvious, and knowing the differences comes so naturally to us, that we refer to it as common sense. Computers, in contrast, need step-by-step handholding—in the form of deterministic algorithms—to render even the most basic of judgments. Despite decades of unbroken gains in speed and processing capacity, machines can’t do what the average toddler does without even trying. That is—until now. Over the last half-dozen years, deep learning, a branch of artificial intelligence inspired by the structure of the human brain, has made enormous strides in giving machines the ability to intuit the physical world. At Facebook’s AI lab, they’ve built a deep learning system capable of answering simple questions to which it had never previously been exposed. The Echo, Amazon’s smart speaker, uses deep learning techniques. Three years ago, Microsoft’s chief research officer impressed attendees at a lecture in China with a demonstration of deep learning speech software that translated his spoken English into Chinese, then instantly delivered the translation using a simulation of his voice speaking Mandarin—with an error rate of just 7%. It now uses the technology to improve voice search on Windows mobile and Bing. The most powerful tech companies in the world have been quietly deploying deep learning to improve their products and services, and none has invested more than Google. It has “bet the company” on AI, says the New York Times, committing huge resources and scooping up many of the leading researchers in the field. And its efforts have borne fruit. A few years ago, a Google deep learning network was shown 10 million unlabeled images from YouTube, and proved to be nearly twice as accurate at identifying the objects in the images (cats, human faces, flowers, various species of fish, and thousands of others) as any previous method. When Google deployed deep learning on its Android voice search, errors dropped by 25% overnight. At the beginning of this year, another Google deep learning system defeated one of the best players of Go—the world’s most complex board game. This is only the beginning. I believe that over the next few years start-ups and the usual big tech suspects will use deep learning to upgrade a wide suite of existing applications, and to create new products and services. Entirely new business lines and markets will spring up, which will, in turn, give rise to still more innovation. Deep learning systems will become easier to use and more widely available. And I predict that deep learning will change the way people interact with technology as radically as operating systems transformed ordinary people’s access to computers.

Deep Learning Historically, computers performed tasks by being programmed with deterministic algorithms, which detailed every step that had to be taken. This worked well in many situations, from performing elaborate calculations to defeating chess grandmasters. But it hasn’t worked as well in situations where providing an explicit algorithm wasn’t possible—such as recognizing faces or emotions, or answering novel questions. Trying to approach those challenges by hand-coding the myriad attributes of a face or phoneme was too labor-intensive, and left machines unable to process data that didn’t fit within the explicit COPYRIGHT © 2017 HARVARD BUSINESS SCHOOL PUBLISHING ALL RIGHTS RESERVED. ©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTSCORPORATION. RESERVED.

163

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

parameters provided by the programmers. Think of the difference between modern voice-assistants like Siri or Alexa, which allow you to ask for things in various ways using natural language, vs. automated phone menu systems, which only perform if you use the specific set of non-negotiable words that they were programmed to understand. By contrast, deep learning-based systems make sense of data for themselves, without the need of an explicit algorithm. Loosely inspired by the human brain, these machines learn, in a very real sense, from their experience. And some are now about as good at object and speech recognition as people. So how does deep learning work? Deep learning systems are modeled after the neural networks in the neocortex of the human brain, where higher-level cognition occurs. In the brain, a neuron is a cell that transmits electrical or chemical information. When connected with other neurons, it forms a neural network. In machines, the neurons are virtual—basically bits of code running statistical regressions. String enough of these virtual neurons together and you get a virtual neural network. Think of every neuron in the network below as a simple statistical model: it takes in some inputs, and it passes along some output.

For a neural network to be useful, though, it requires training. To train a neural network, a set of virtual neurons are mapped out and assigned a random numerical “weight,” which determines how the neurons respond to new data (digitized objects or sounds). Like in any statistical or machine learning, the machine initially gets to see the correct answers, too. So if the network doesn’t accurately identify the input – doesn’t see a face in an image, for example — then the system adjusts

COPYRIGHT © 2017 HARVARD BUSINESS SCHOOL PUBLISHING ALL RIGHTS RESERVED. ©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTSCORPORATION. RESERVED.

174

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

the weights—i.e., how much attention each neuron paid to the data—in order to produce the right answer. Eventually, after sufficient training, the neural network will consistently recognize the correct patterns in speech or images. The idea of artificial neurons has been around for at least 60 years, when, in the 1950s, Frank Rosenblatt built a “perceptron” made of motors, dials, and light detectors, which he successfully trained to tell the difference between basic shapes. But early neural networks were extremely limited in the number of neurons they could simulate, which meant they couldn’t recognize complex patterns. Three developments in the last decade made deep learning viable. First, Geoffrey Hinton and other researchers at the University of Toronto developed a breakthrough method for software neurons to teach themselves by layering their training. (Hinton now splits his time between the University of Toronto and Google.) A first layer of neurons will learn how to distinguish basic features, say, an edge or a contour, by being blasted with millions of data points. Once the layer learns how to recognize these things accurately, it gets fed to the next layer, which trains itself to identify more complex features, say, a nose or an ear. Then that layer gets fed to another layer, which trains itself to recognize still greater levels of abstraction, and so on, layer after layer—hence the “deep” in deep learning—until the system can reliably recognize very complex phenomenon, like a human face.

COPYRIGHT © 2017 HARVARD BUSINESS SCHOOL PUBLISHING ALL RIGHTS RESERVED. ©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTSCORPORATION. RESERVED.

185

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

COPYRIGHT © 2017 HARVARD BUSINESS SCHOOL PUBLISHING ALL RIGHTS RESERVED. ©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTSCORPORATION. RESERVED.

196

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

The second development responsible for recent advancements in AI is the sheer amount of data that is now available. Rapid digitization has resulted in the production of large-scale data, and that data is oxygen for training deep learning systems. Children can pick something up after being shown how to do it just a few times. AI-powered machines, however, need to be exposed to countless examples. Deep learning is essentially a brute-force process for teaching machines how a thing is done or what a thing is. Show a deep learning neural network 19 million pictures of cats and probabilities emerge, inclinations are ruled out, and the software neurons eventually figure out what statistically significant factors equate to feline. It learns how to spot a cat. That’s why Big Data is so important— without it, deep learning just doesn’t work. Finally, a team at Stanford led by Andrew Ng (now at Baidu) made a breakthrough when they realized that graphics processing unit chips, or GPUs, which were invented for the visual processing demands of video games, could be repurposed for deep learning. Until recently, typical computer chips could only process one event at a time, but GPUs were designed for parallel computing. Using these chips to run neural networks, with their millions of connections, in parallel sped up the training and abilities of deep learning systems by several orders of magnitude. It made it possible for a machine to learn in a day something that had previously taken many weeks. The most advanced deep learning networks today are made up of millions of simulated neurons, with billions of connections between them, and can be trained through unsupervised learning. It is the most effective practical application of artificial intelligence that’s yet been devised. For some tasks, the best deep learning systems are pattern recognizers on par with people. And the technology is moving aggressively from the research lab into industry.

Deep Learning OS 1.0 As impressive as the gains from deep learning have been already, these are early days. If I analogize it to the personal computer, deep learning is in the green-and-black-DOS-screen stage of its evolution. A great deal of time and effort, at present, is being spent doing for deep learning—cleaning, labelling, and interpreting data, for example—rather than doing with deep learning. But in the next couple of years, start-ups and established companies will begin releasing commercial solutions for building production-ready deep learning applications. Making use of open-source frameworks such as TensorFlow, these solutions will dramatically reduce the effort, time, and costs of creating complex deep learning systems. Together they will constitute the building blocks of a deep learning operating system. A deep learning operating system will permit the widespread adoption of practical AI. In the same way that Windows and Mac OS allowed regular consumers to use computers and SaaS gave them access to the cloud, tech companies in the next few years will democratize deep learning. Eventually, a deep learning OS will allow people who aren’t computer scientists or natural language processing researchers to use deep learning to solve real-life problems, like detecting diseases instead of identifying cats.

COPYRIGHT © 2017 HARVARD BUSINESS SCHOOL PUBLISHING ALL RIGHTS RESERVED. ©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTSCORPORATION. RESERVED.

207

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

The first new companies making up the deep learning operating system will be working on solutions in data, software, hardware. Data. Getting good quality large scale data is the biggest barrier to adopting deep learning. But both service shops and software platforms will arise to deal with the data problem. Companies are already creating internal intelligent platforms that assist humans to label data quickly. Future data labeling platforms will be embedded in the design of the application, so that the data created by using a product will be captured for training purposes. And there will be new service-based companies that will outsource labeling to low-cost countries, as well as create labeled data through synthetic means. Software. There are two main areas here where I see innovation happening: 1) The design and programming of neural networks. Different deep learning architectures, such as CNNs and RNNs, support different types of applications (image, text, etc.). Some use a combination of neural network architectures. As for training, many applications will use a combination of machine learning algorithms, deep learning, reinforcement learning, or unsupervised learning for solving different sub-parts of the application. I predict that someone will build a machine learning design engine solution, which will examine an application, training data set, infrastructure resources, and so on, and recommend the right architecture and algorithms to be used. 2) A marketplace of reusable neural network modules. As described above, different layers in a neural network learn different concepts and then build on each other. This architecture naturally creates opportunity to share and reuse trained neural networks. A layer of virtual neurons that’s been trained to identify an edge, on its way up to recognizing the face of cat, could also be repurposed as the base layer for recognizing the face of a person. Already, Tensorflow, the most popular deep learning framework, supports reusing an entire subgraph component. Soon, the community of machine learning experts contributing open source modules will create the potential for deep learning versions of GitHub and StackOverflow. Hardware. Finding the optimal mix of GPUs, CPUs, cloud resources; determining the level of parallelization; and performing cost analyses are complex decisions for developers. This creates an opportunity for platform and service-based companies to recommend the right infrastructure for training tasks. Additionally, there will be companies that provide infrastructure services—such as orchestration, scale-out, management, and load balancing—on specialized hardware for deep learning. Moreover, I expect incumbents as well as start-ups to launch their own deep learningoptimized chips. These are just some of the possibilities. I’m certain there are many more lurking in other entrepreneurial minds, because the promise of this technology is immense. We are beginning to build machines that can learn for themselves and that have some semblance of sensible judgment. Palak Dalal (HBS ’17) contributed research and analysis to this piece. COPYRIGHT © 2017 HARVARD BUSINESS SCHOOL PUBLISHING ALL RIGHTS RESERVED. ©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTSCORPORATION. RESERVED.

218

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

Aditya Singh is a partner at Foundation Capital.

COPYRIGHT © 2017 HARVARD BUSINESS SCHOOL PUBLISHING ALL RIGHTS RESERVED. ©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTSCORPORATION. RESERVED.

229

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

Discussion Questions

Deep Learning Will Radically Change the Ways We Interact with Technology BY ADITYA SINGH

1. After reading these first two articles, do you have a better understanding of the distinctions between AI, machine learning, and deep learning? Could you explain them to a colleague? 2. Singh suggests that AI will soon change the way people interact with technology. One way is that voice recognition is becoming faster and more reliable than typing. How would your business be affected if speaking to a computer or devices became the norm rather than typing? In what ways would it improve processes? In what way might it complicate things?

Eventually, a deep learning OS will allow people who aren’t computer scientists or natural language processing researchers to use deep learning to solve real-life problems, like detecting diseases instead of identifying cats.

COPYRIGHT © 2018 HARVARD BUSINESS SCHOOL PUBLISHING ALL RIGHTS RESERVED. ©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTSCORPORATION. RESERVED.

235

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

ARTICLE INSIDE FACEBOOK’S AI WORKSHOP

At the social network behemoth, machine learning has become a platform for the platform. by Scott Berinato

“Look, a machine learning algorithm really is a lookup table, right? Where the key is the input, like an image, and

©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTS RESERVED.

the value is the label for the input, like ‘a horse.’ I have a bunch of examples of something. Pictures of horses. I give the algorithm as many as I can. ‘This is a horse. This is a horse. This isn’t a horse. This is a horse.’ And the algorithm keeps those in a table. Then, if a new example comes along — or if I tell it to watch for new examples — well, the algorithm just goes and looks at all those examples we fed it. Which rows in the table look similar? And how similar? It’s trying to decide, ‘Is this new thing a horse? I think so.’ If it’s right, the image gets put in the ‘This is a horse’ group, and if it’s wrong, it gets put in the ‘This isn’t a horse’ group. Next time, it has more data to look up. “One challenge is how do we decide how similar a new picture is to the ones stored in the table. One aspect of machine learning is to learn similarity GETTY IMAGES

W

ithin Facebook’s cavernous Building 20, about halfway between the lobby (panoramic views of the Ravenswood Slough) and the kitchen (hot breakfast, smoothies, gourmet coffee), in a small conference room called Lollapalooza, Joaquin Candela is trying to explain artificial intelligence to a layperson. Candela — bald, compact, thoughtful — runs Facebook’s Applied Machine Learning (AML) group — the engine room of AI at Facebook, which, increasingly, makes it the engine room of Facebook in general. After some verbal searching, he finally says:

functions. Another challenge is, What happens when your table grows really large? For every new image, you would need to make a zillion comparisons…. So another aspect of machine learning is to approximate a large stored table with a function instead of going through every image. The function knows how to roughly estimate what the corresponding value should be. That’s the essence of ML — to approximate a gigantic table with a function. This is what learning is about.” There’s more to it than that, obviously, but it’s a good starting point when talking about AI, because it makes it sound real, almost boring. Mechanical. So much of the conversation around AI is awash in mystical descriptions of its power and in reverence for its nearmagic capabilities. Candela doesn’t like that and tries to use more-prosaic terms. It’s powerful, yes, but not magical. It has limitations. During presentations, he’s fond of showing a slide with a wizard and a factory, telling audiences that Facebook thinks of AI like the latter, because “wizards don’t scale.” And that’s what Facebook has done with AI and machine learning: scaled it

24

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

SCOTT BERINATO



at a breakneck pace. A few years ago the company’s machine learning group numbered just a few and needed days to run an experiment. Now, Candela says, several hundred employees run thousands of experiments a day. AI is woven so intricately into the platform that it would be impossible to separate the products — your feed, your chat, your kid’s finsta — from the algorithms. Nearly everything users see and do is informed by AI and machine learning. Understanding how and why Facebook has so fully embraced AI can help any organization that’s ready to invest in an algorithmic future. It would be easy to assume that Facebook, with all its resources, would simply get the best talent and write the best algorithms — game over. But Candela took a different approach. Certainly the talent is strong, and the algorithms are good. Some of them are designed to “see” images or automatically filter them. Some understand conversations and can respond to them. Some translate between languages. Some try to predict what you’ll like and buy. But in some ways the algorithms are not his main focus. Instead, he’s been busy creating an AI workshop in which anyone in the company can use AI to achieve a goal. Basically, Candela built an AI platform for the platform. Whether you’re a deeply knowledgeable programmer or a complete newbie, you can take advantage of his wares. Here’s how he did it and what you can learn from it.

SOYUZ

Candela, a veteran of Microsoft Research, arrived at Facebook in 2012 to work in the company’s ads business. He and a handful of staffers inherited a ranking algorithm for better targeting users with ads. Candela describes the machine learning code he inherited as “robust but not the latest.” More than once

he compares it to Soyuz, the 1960s Soviet spacecraft. Basic but reliable. Gets the job done even if it’s not the newest, best thing. “It’ll get you up there and down. But it’s not the latest covnet [convolutional neural net] of the month.” You might assume, then, that the first thing Candela set out to do was to upgrade the algorithm. Get rid of Soyuz in favor of a space plane. It wasn’t. “To get more value, I can do three things,” he says. “I can improve the algorithm itself, make it more sophisticated. I can throw more and better data at the algorithm so that the existing code produces better results. And I can change the speed of experimentation to get more results faster. “We focused on data and speed, not on a better algorithm.” Candela describes this decision as “dramatic” and “hard.” Computer scientists, especially academic-minded ones, are rewarded for inventing new algorithms or improving existing ones. A better statistical model is the goal. Getting cited in a journal is validation. Wowing your peers gives you cred. It requires a shift in thinking to get those engineers to focus on business impact before optimal statistical model. He thinks many companies are making the mistake of structuring their efforts around building the best algorithms, or hiring developers who claim to have the best algorithms, because that’s how many AI developers think. But for a company, a good algorithm that improves the business is more valuable than vanguard statistical models. In truth, Candela says, real algorithmic breakthroughs are few and far between — two or three a year at best. If his team focused its energies there, it would take lots of effort to make marginal gains. He hammers these points home constantly: Figure out the impact on the business first. Know what you’re

©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTS RESERVED.

solving for. Know what business challenge you need to address. “You might look for the shiniest algorithm or the people who are telling you they have the most advanced algorithm. And you really should be looking for people who are most obsessed with getting any algorithm to do a job. That’s kind of a profound thing that I think is lost in a lot of the conversation. I had a conversation with our resident machine learning geek at our office, and we were just talking about different people doing AI. He said, ‘Nobody really thinks their algorithms are very good or whatever.’ It makes me think, maybe that’s fine. “I’m not saying don’t work on the algorithm at all. I’m saying that focusing on giving it more data and better data, and then experimenting faster, makes a lot more sense.” So rather than defining success as building the best natural language processing algorithm, he defines it as deploying one that will help users find a restaurant when they ask their friends, “Where can I get a good bite around here?” Instead of being thrilled that some computer vision algorithm is nearing pixel-perfect object recognition, he gets excited if that AI is good enough to notice that you post a lot of pictures of the beach and can help you buy a swimsuit. The strategy worked when he started at Facebook. Ad revenues rose. Candela’s profile rose. It was suggested that AML become a centralized function for all of Facebook. Candela said no. Twice. “I was concerned about the ‘If you build it, they will come’ phenomenon.” Just creating bits of artificial intelligence in the hope that people would see the value and adopt it wouldn’t work. But he did pick his spots. He collaborated with the feeds team while saying no to many other groups. Then he worked with the Messenger team. His team grew and took on more projects with other teams.

25

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

SCOTT BERINATO



H2

Candela goes to a whiteboard to describe how he built his AI factory inside Facebook. The key, he says, was figuring out where on the product development path AI fits. He draws something like the graph on this page (see the exhibit “Where AI Fits In at Facebook”). H3 — Horizon 3 or three years out from product — is the realm of R&D and science. Often, data scientists who work on AI think of themselves as here, improving algorithms and looking for new ways to get machines to learn. Candela didn’t put his team here for the reasons already mentioned. It’s too far from impact on the business. H1, approaching product delivery, is where the product teams live — the feeds team, the Instagram team, the ads team. AI doesn’t go here either, because it would be difficult to retrofit products this deeply developed. It would be like building a car and then deciding that it should be self-driving after you started to put it together. That leaves H2, between the science and the product, as the place AML lives at Facebook. AML is a conduit

Where AI Fits In at Facebook

PRODUCT READINESS

By 2015 Candela could see that his group would need to centralize, so he turned his attention to how he’d build such an operation. He was still worried about the “build it and they will come” phenomenon, so he focused less on how his team would be structured and more on how the group would connect to the rest of Facebook. “You build a factory that makes amazing widgets, and you forget to design the loading docks into your factory?” He laughs. “Well, enjoy your widgets.” Only then, about three years in, did Candela think about upgrading some of his algorithms. (Incidentally, even today, the emergency escape spacecraft attached to the International Space Station is a Soyuz.)

H3

H2

H1

R&D

AML

Product delivery

TIME SOURCE FACEBOOK

for transferring the science into the product. It does not do research for research’s sake, and it does not build and ship products. As the upward slope in the product’s readiness shows, it’s a dynamic space. Pointing to H2, Candela says, “This needs to feel uncomfortable all the time. The people you need to hire need to be okay with that, and they need to be incredibly selfless. Because if your work is successful, you spin it out. And you need to fail quite a bit. I’m comfortable with a 50% failure rate.” If the team is failing less, Candela suspects its members are too risk averse, or they’re taking on challenges that are sliding them closer to H1’s product focus. “Maybe we do something like that and it works, but it’s still a failure, because the product teams should be taking that on, not us. If you own a piece of technology that the ads team should operate themselves to generate value, give it to them, and then increase your level of ambition in the machine learning

©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTS RESERVED.

© HBR.ORG

space before something becomes product.” So Candela’s team is neither earning the glory of inventing new statistical models nor putting products out into the world. It’s a factory of specialists who translate others’ science for others’ products and fail half the time.

PUSH/PULL

All that being said, the lines between the three realms — H3, H2, and H1 — still aren’t crisp. In some cases Candela’s team does look at the science of machine learning, to solve specific problems. And sometimes it does help build the products. That was especially true as AML got off the ground, because many people in the business hadn’t yet been exposed to AI and what it could do for them. In one case AML built a translation algorithm. The team dipped into the research space to look at how existing translation algorithms worked and could be improved, because bad translations, which either don’t

26

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

SCOTT BERINATO



make sense or create a misleading interpretation, are in some ways worse than no translation. “Early on it was more push, more tenacity on our part,” Candela says. “But it was gentle tenacity. We weren’t going to throw something over the fence and tell the product team, ‘This is great, use it.’” That meant that his team helped write some product code. Doing a little bit of the science and a little bit of the product in addition to its core function was meant to inspire the product team members to see what AML could do for them. What the two teams built — a product that allowed community pages to instantly translate into several languages — worked. Other projects were similarly pushed out, and now the international team and other product groups at Facebook are pulling from AML, asking to use code in their products themselves. “Look, it’s nowhere near where I want it to be,” Candela says. “I’d like to have all the product leaders in the company get together quarterly for AI reviews. That will certainly happen. But the conversation in the past two years has completely changed. Now if I walk from one end of this building to the other and I bump into, I don’t know, the video team or the Messenger team, they’ll stop me and say, ‘Hey, we’re excited to try this. We think we can build a product on this.’ That didn’t happen before.” AML’s success, though, has created a new challenge for Candela. Now that everyone wants a piece of AML, the factory has to scale.

LAYER CAKE

Candela couldn’t scale just by saying yes to every project and adding bodies to get the work done. So he organized in other ways. First he subdivided his team according to the type of AI its members would focus on:

Applied Machine Learning Translation

TEXT Natural Language

a serious hardware array, so many experiments can be run simultaneously. (The system allows for more than 6 million predictions a second). All of this is to increase the velocity of running experiments on the data and scale.

AI/ML expertise required LESS

Self-serve AI

For non-technical users, e.g. LUMOS

Reusable engines

For developers outside of AML, e.g. CLUE Speech

AUDIO

ML algorithms

Generalizable by discipline

Deep learning framework Caffe2

AI backbone MORE

Camera

FBLearner Flow

Ease of use Computer Laura Vision

VISUAL

MORE

Self-serve AI

For non-technical users, e.g. LUMOS

Reusable engines

For developers outside of AML, e.g. CLUE

ML algorithms

Computer Photography

Generalizable by discipline

Deep learning framework Caffe2

SOURCE FACEBOOK

AI backbone

© HBR.ORG

LESS

This created common denominators so that one team — say, computer vision — could work on any machine learning application involving parsing images and reuse its work whenever possible. Next came a large-scale engineering effort to build Facebook’s own AI backbone, called FBLearner Flow. Here algorithms are deployed once and made reusable for anyone who may need them. The time-consuming parts of setting up and running experiments are automated, and past results are stored and made available and easily searchable. And the system runs through

©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTS RESERVED.

FBLearner Flow

Ability to build and customize AI LESS

Self-serve AI

For non-technical users, e.g. LUMOS

Reusable engines

For developers outside of AML, e.g. CLUE

ML algorithms

Generalizable by discipline

Deep learning framework Caffe2

AI backbone MORE

FBLearner Flow

SOURCE FACEBOOK

© HBR.ORG

27

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

SCOTT BERINATO



The system was also designed to accommodate many kinds of possible users. Candela believes that for AI to work, and to scale even further, he must help people outside AML do the work themselves. He created what he calls a layer cake of artificial intelligence. The bottom layers focus on AML’s work: refining the core engine (with a strong focus on optimizing performance, especially for mobile) and working with machine learning algorithms. The upper layers focus on tools that make it possible for those outside AML to exploit the algorithms with less AML involvement. “It’s all about what you expose to the user,” Candela says. In some cases he’s built systems that developers outside AML can take advantage of to build and run their own models.

REX

A good example of Candela’s team structure and the push/pull dynamic comes from some AI built to surface content on the basis of what you type. The natural-language machine learning team created an engine to understand conversational typing. This bit of intelligence first found its way into the Messenger chat client. AML developed the models while the product team developed use cases and “intents” — lingo for the types of tasks you want the engine to learn. For example, training natural language AI to recognize and reliably respond to a phrase like “I’m looking for the best…” is an intent. The first few such intents were deployed to Messenger through a product called M Suggestions. If you sent a chat to a friend that said “I’ll meet you there in 30 minutes,” M Suggestions might prompt you with an offer to hire a car. As the tools for building intent models developed and the product team became more conversant with

them, AML’s role diminished. Now the Messenger team has improved M Suggestions by building dozens more intents on its own. Still, this bit of natural language AI wasn’t built just for chat. It’s reusable. It was codified as CLUE, for “conversational learning understanding engine.” It found its way into more Facebook applications. It’s being adapted for status updates and feeds. Social recommendations — or social rex, as everyone calls them — are increasingly driven by AI. If you typed “I’m traveling to Omaha and I really want to find a good steak downtown,” AI might respond as if it were one of your friends, with a comment on your post, rex such as a list of steakhouses, and a tagged map of where they are relative to downtown. If your friend replied to you and said, “It also has some great vegetarian restaurants,” the algorithm might again reply with pertinent data. Social rex intents are not yet being developed without AML, but the goal is to have them move out of Candela’s group, just as M Suggestions did. In general, the idea is to make product teams AI-capable themselves. “We’ll teach you to fish,” Candela says, “and you go fish, and we’ll drag up the next thing. We’ll build a fishing boat. And once you’re using the fishing boat, I’m going to build a cannery, right?” At the moment, about 70% of the AI work on the backbone is done by people outside Candela’s team. That’s possible in part because of the interface with AI. In some cases, as with a tool called Lumos, machine learning can be used by nondevelopers.

HORSEBACK RIDING AND CEREAL BOXES

Lumos is computer vision AI, a tool that can comb through photos on Facebook or Instagram or other platforms and learn what they contain. You can train it to see anything. It has helped automate the discovery and banning

©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTS RESERVED.

of pornographic or violent content, IP appropriation (improper use of brands and logos), and other unwelcome content. It can also help identify things you like and do (to drive personalized advertising and recommendations), on the basis of photos in your feeds. I watch a demo in which engineers select “horseback riding” as our intent, the thing we’ll be looking for. The interface is simple: A few clicks, a couple of forms to fill out — What are you looking for? How much data do you want to look at? — and the algorithm gets to work finding pictures of horseback riding. Thumbnails start to fill the page. The algorithm has searched for horseback riding before, so it’s already quite good at finding it. My guess is that north of 80% of the images that pop up are indeed of horseback riding, and they show remarkable variety. Here’s one with someone posing at a standstill. Here’s one with the horse rearing. Here’s an equestrian jumping. The algorithm finds shapes and boundaries between shapes and builds on previous knowledge of what those interactions mean. It knows things about what combination of pixels is most likely a person, for example, and what’s a horse. It knows when it “sees” a person and a horse together with the person situated close above the horse. And it decides that this looks like horseback riding. We also find pictures that aren’t horseback riding — one is a person standing next to a horse; another is a person on a mule — and check those off as not matches. They’re framed in red, in case there’s any doubt. The algorithm internalizes that information — adds it to the lookup table — for use next time. A simple chart at the top of the page shows the algorithm’s accuracy and confidence over time. It’s always an S curve, slow to learn at first, then rapidly improving, then tapering off on how much more accurate it can get. It’s very good at seeing horseback riding.

28

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

SCOTT BERINATO



Other potentially valuable pictures are harder for AI to parse. “Receipts” is tricky to suss out because it can look to a computer just like type on a page; but there would be some interesting apps for AI that could recognize and “read” receipts. The engineers show how bowling alleys and escalators often confuse the algorithm because they have similar shapes and visual properties. I ask, “What about something like ‘food’?” This brings us to an important point about machine learning: It’s only as good as its training. We call up food as a topic to train. Indeed, we see lots of pictures of fruits and vegetables, a few of plates at restaurants. All food. We also see a cereal box. Is that food? Well, yes. Or no. It’s a box. But there’s food in it. When we buy it, we’re buying food, not the box. If I asked if there was any food in the cupboard, you wouldn’t say, “No, just a cereal box.” (Or, more pertinent to Facebook, if I posted a picture of a cereal box, should it think I’m posting about food or about a box?) As a picture, as a piece of data, it’s a box. Should we mark this as a match or a miss? Here’s part of the art of machine learning. When training algorithms, one needs to use clearly definable categories. Food is probably too general in some ways, and the algorithm will either improperly hit or miss on images because it’s hard to know what we mean when we say, “Show me pictures of food.” “Vegetable” is a better idea to train on. And when training, everyone must define terms in the same way. Imagine two people training the algorithm when one always marks cereal boxes as food, and the other marks them as not food. Now imagine that happening at scale, on terabytes of visual data. The same applies to natural language processing. Humans are very good

at interpreting text in context to find sophisticated meaning. For example, I may type, “Gee, I love that movie about the superheroes. It’s so, so original! I hope they make a hundred more of them.” My friends, who know me and know some of the mechanics of sarcasm, may readily understand that my meaning is the opposite of what I’m typing. Artificial intelligence is still learning how to decide the meaning of something like that. To figure out if I’m being sarcastic, it has to go much further than just learning how to parse grammar and vocabulary. It has to see what else I’ve said and posted, and try to find other clues that will tell it whether I really loved the movie and I want 100 more or I actually detested it — because getting that wrong is not good for a platform that wants to create affinities with me. If I was being sarcastic and my feed starts filling up with superhero movie ads, I’m probably not enjoying the experience.

NOT MAGIC

It’s details like these — showing where AI is still limited, and how humans have such a core role in training it, and how solving problems and creating value are more important than finding great models — that Candela is thinking about near the end of the day, when he’s talking about the mythic status AI has gained. He’s railing against what he perceives as laziness in those who find the idea of AI-as-magic-bullet appealing and don’t apply critical thinking to it. “What frustrates me,” he says, “is that everybody knows what a statistician is and what a data analyst can do. If I want to know ‘Hey, what age segment behaves in what way?’ I get the data analyst. “So when people skip that, and they come to us and say, ‘Hey, give me a machine learning algorithm that will do what we do,’ I’m like, ‘What is it that I look like? What problem are you trying

©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTS RESERVED.

to solve? What’s your goal? What are the trade-offs?’” Sometimes they’re surprised that there are trade-offs. “If that person doesn’t have answers to those questions, I’m thinking, ‘What the hell are you thinking AI is?’” They are thinking it’s magic. “But it’s not. That’s the part where I tell people, ‘You don’t need machine learning. You need to build a data science team that helps you think through a problem and apply the human litmus test. Sit with them. Look at your data. If you can’t tell what’s going on, if you don’t have any intuition, if you can’t build a very simple, rulebased system — like, Hey, if a person is younger than 20 and living in this geography, then do this thing — if you can’t do that, then I’m extremely nervous even talking about throwing AI at your problem.’ “I’m delighted when other executives come to me and start not from wanting to understand the technology but from a problem they have that they’ve thought very, very deeply about. And sometimes — often, in fact — a simple, good old rule-based system, if you have the right data, will get you 80% of the way to solving the problem.” “And guess what? It’s going to have the benefit that everybody understands it. Exhaust the human brain first.”  About the author: Scott Berinato is a senior editor at Harvard Business Review and the author of Good Charts: The HBR Guide to Making Smarter, More Persuasive Data Visualizations (2016).

29

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

Discussion Questions

Inside Facebook’s AI Workshop AN INTERVIEW WITH JOAQUIN CANDELA BY SCOTT BERINATO

1. Candela repeats three points on AI initiatives: “Figure out the impact on the business first. Know what you’re solving for. Know what business challenge you need to address.” Do these principles drive AI initiatives at your company? Which one of these principles does your company excel at, and why? Which one could use the most work, and how might you address it? 2. Candela describes a data problem in which one algorithm trainer is marking a cereal box as “food” and another is marking it as a “box.” Can you think of similar types of problems that might exist in your company’s data? What impact could these be having on your company’s processes or results?

During presentations, he’s fond of showing a slide with a wizard and a factory, telling audiences that Facebook thinks of AI like the latter, because “wizards don’t scale.”

COPYRIGHT © 2018 HARVARD BUSINESS SCHOOL PUBLISHING ALL RIGHTS RESERVED. ©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTSCORPORATION. RESERVED.

306

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

REPRINT H03Z0H PUBLISHED ON HBR.ORG OCTOBER 20, 2017

ARTICLE ANALYTICS

How to Spot a Machine Learning Opportunity, Even If You Aren’t a Data Scientist

by Kathryn Hume

©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTS RESERVED.

31

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

ANALYTICS

How to Spot a Machine Learning Opportunity, Even If You Aren’t a Data Scientist by Kathryn Hume OCTOBER 20, 2017

SYLVERARTS/ISTOCK

Artificial intelligence is no longer just a niche subfield of computer science. Tech giants have been using AI for years: Machine learning algorithms power Amazon product recommendations, Google Maps, and the content that Facebook, Instagram, and Twitter display in social media feeds. But

COPYRIGHT © 2017 HARVARD BUSINESS SCHOOL PUBLISHING ALL RIGHTS RESERVED. ©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTSCORPORATION. RESERVED.

322

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

William Gibson’s adage applies well to AI adoption: The future is already here, it’s just not evenly distributed. The average company faces many challenges in getting started with machine learning, including a shortage of data scientists. But just as important is a shortage of executives and nontechnical employees able to spot AI opportunities. And spotting those opportunities doesn’t require a PhD in statistics or even the ability to write code. (It will, spoiler alert, require a brief trip back to high school algebra.) Having an intuition for how machine learning algorithms work – even in the most general sense – is becoming an important business skill. Machine learning scientists can’t work in a vacuum; business stakeholders should help them identify problems worth solving and allocate subject matter experts to distill their knowledge into labels for data sets, provide feedback on output, and set the objectives for algorithmic success. As Andrew Ng has written: “Almost all of AI’s recent progress is through one type, in which some input data (A) is used to quickly generate some simple response (B).” But how does this work? Think back to high school math — I promise this will be brief — when you first learned the equation for a straight line: y = mx + b. Algebraic equations like this represent the relationship between two variables, x and y. In high school algebra, you’d be told what m and b are, be given an input value for x, and then be asked to plug them into the equation to solve for y. In this case, you start with the equation and then calculate particular values. Supervised learning reverses this process, solving for m and b, given a set of x’s and y’s. In supervised learning, you start with many particulars — the data — and infer the general equation. And the learning part means you can update the equation as you see more x’s and y’s, changing the slope of the line to better fit the data. The equation almost never identifies the relationship between each x and y with 100% accuracy, but the generalization is powerful because later on you can use it to do algebra on new data. Once you’ve found a slope that captures a relationship between x and y reliably, if you are given a new x value, you can make an educated guess about the corresponding value of y. As you might imagine, many exciting machine learning problems can’t be reduced to a simple equation like y = mx + b. But at their essence, supervised machine learning algorithms are also solving for complex versions of m, based on labeled values for x and y, so they can predict future y’s from future x’s. If you’ve ever taken a statistics course or worked with predictive analytics, this should all sound familiar: It’s the idea behind linear regression, one of the simpler forms of supervised learning. To return to Ng’s formulation, supervised learning requires you to have examples of both the input data and the response, both the x’s and the y’s. If you have both of those, supervised learning lets

COPYRIGHT © 2017 HARVARD BUSINESS SCHOOL PUBLISHING ALL RIGHTS RESERVED. ©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTSCORPORATION. RESERVED.

333

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

you come up with an equation that approximates that relationship, so in the future you can guess y values for any new value of x. So the question of how to identify AI opportunities starts with asking: What are some outcomes worth guessing? And do we have the data necessary to do supervised learning? For example, let’s say a data scientist is tasked with predicting real estate prices for a neighborhood. After analyzing the data, she finds that housing price (y) is tightly correlated to size of house (x). So, she’d use many data points containing both houses’ size and price, use statistics to estimate the slope (m), and then use the equation y = mx + b to predict the price for a given house based on its size. This is linear regression, and it remains incredibly powerful. Organizations use similar techniques to predict future product sales, investment portfolio risk, or customer churn. Again, the statistics behind different algorithms vary in complexity. Some techniques output simple point predictions (We think y will happen!) and others output a range of possible predictions with affiliated confidence rates (There’s a 70% chance y will happen, but if we change one assumption, our confidence falls to 60%). These are all examples of prediction problems, but supervised learning is also used for classification. Classification tasks clump data into buckets. Here a data scientist looks for features in data that are reliable proxies for categories she wants to separate: If data has feature x, it goes into bucket one; if not, it goes into bucket two. You can still think of this as using x’s to predict y’s, but in this case y isn’t a number but a type. Organizations use classification algorithms to filter spam, diagnose abnormalities on X-rays, identify relevant documents for a lawsuit, sort résumés for a job, or segment customers. But classification gains its true power when the number of classes increases. Classification can be extended beyond binary choices like “Is it spam or not?” to include lots of different buckets. Perception tasks, like training a computer to recognize objects in images, are also classification tasks, they just have many output classes (for example, the various animal species names) instead of just Bucket 1 and Bucket 2. This makes supervised learning systems look smarter than they are, as we assume their ability to learn concepts mirrors our own. In fact, they’re just bucketing data into buckets 1, 2, 3…n, according to the “m” learned for the function. So far, this all feels rather abstract. How can you bring it down to earth and learn how to identify these mathematical structures in your everyday work? There are a few ways you can determine whether a task presents a good supervised learning opportunity.

COPYRIGHT © 2017 HARVARD BUSINESS SCHOOL PUBLISHING ALL RIGHTS RESERVED. ©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTSCORPORATION. RESERVED.

344

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

First, write down what you do in your job. Break apart your activities into: things you do daily or regularly versus things you do sporadically; things that have become second nature versus things that require patient deliberation or lots of thought; and things that are part of a process versus things you do on your own. For those tasks that you perform regularly, on your own, and that feel automatic, identify how many others in your organization do similar tasks and how many people have done this historically. Examine the nature of the task. Does it include predicting something or bucketing something into categories? Ask yourself: If 10 colleagues in your organization performed the task, would they all agree on the answer? If humans can’t agree something is true or false, computers can’t reliably transform judgment calls into statistical patterns. How long have people in the organization been doing something similar to this task? If it’s been a long time, has the organization kept a record of successfully completed tasks? If yes, this could be used as a training data set for your supervised learning algorithm. If no, you may need to start collecting this data today, and then you can keep a human in the loop to train the algorithm over time. Next, sit down with a data science team and tell them about the task. Walk them through your thought process and tell them what aspects of information you focus on when you complete your task. This will help them determine if automation is feasible and tease out the aspects of the data that will be most predictive of the desired output. Ask yourself, if this were automated, how might that change the products we offer to our customers? Ask, what is the worst thing that could happen to the business if this were to be automated? And finally, ask, what is the worst thing that could happen to the business if the algorithm outputs the wrong answer or an answer with a 65% or 70% accuracy rate? What is the accuracy threshold the business requires to go ahead and automate this task? Succeeding with supervised learning entails a shift in the perspective on how work gets done. It entails using past work — all that human judgment and subject matter expertise — to create an algorithm that applies that expertise to future work. When used well, this makes employees more productive and creates new value. But it starts with identifying problems worth solving and thinking about them in terms of inputs and outputs, x’s and y’s.

Kathryn Hume is vice president of product and strategy at integrate.ai, a Toronto-based startup.

COPYRIGHT © 2017 HARVARD BUSINESS SCHOOL PUBLISHING ALL RIGHTS RESERVED. ©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTSCORPORATION. RESERVED.

355

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

Discussion Questions

How to Spot a Machine Learning Opportunity, Even If You Aren’t a Data Scientist BY KATHRYN HUME

1. Given Hume’s framework, beginning at the top of the previous page with “First, write down what you do in your job,” are there any tasks in your own job that you can potentially automate? Would you welcome the opportunity to automate these tasks? Why or why not? 2. How could your company implement this process for exploring automation potential throughout the organization? To which job functions or departments could your company apply supervised machine learning? How would people in your company respond to this type of analysis? Would you ask employees to look for automation opportunities in their own jobs?

Having an intuition for how machine-learning algorithms work—even in the most general sense—is becoming an important business skill.

COPYRIGHT © 2018 HARVARD BUSINESS SCHOOL PUBLISHING ALL RIGHTS RESERVED. ©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTSCORPORATION. RESERVED.

367

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

REPRINT H0442I PUBLISHED ON HBR.ORG JANUARY 17, 2018

ARTICLE DATA

Is Your Company’s Data Actually Valuable in the AI Era? by Ajay Agrawal, Joshua Gans and Avi Goldfarb

©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTS RESERVED.

37

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

DATA

Is Your Company’s Data Actually Valuable in the AI Era? by Ajay Agrawal, Joshua Gans and Avi Goldfarb JANUARY 17, 2018

Carmen Martínez Torrón /Hayon Thapaliya/Getty Images

COPYRIGHT © 2018 HARVARD BUSINESS SCHOOL PUBLISHING ALL RIGHTS RESERVED. ©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTSCORPORATION. RESERVED.

382

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

AI is coming. That is what we heard throughout 2017 and will likely continue to hear throughout this year. For established businesses that are not Google or Facebook, a natural question to ask is: What have we got that is going to allow us to survive this transition? In our experience, when business leaders ask this with respect to AI, the answer they are given is “data.” This view is confirmed by the business press. There are hundreds of articles claiming that “data is the new oil” — by which they mean it is a fuel that will drive the AI economy. If that is the case, then your company can consider itself lucky. You collected all this data, and then it turned out you were sitting on an oil reserve when AI happened to show up. But when you have that sort of luck, it is probably a good idea to ask “Are we really that lucky?” The “data is oil” analogy does have some truth to it. Like internal combustion engines with oil, AI needs data to run. AI takes in raw data and converts it into something useful for decision making. Want to know the weather tomorrow? Let’s use data on past weather. Want to know yogurt sales next week? Let’s use data on past yogurt sales. AIs are prediction machines driven by data. But does AI need your data? There is a tendency these days to see all data as potentially valuable for AI, but that isn’t really the case. Yes, data, like oil, is used day-to-day to operate your prediction machine. But the data you are sitting on now is likely not that data. Instead, the data you have now, which your company accumulated over time, is the type of data used to build the prediction machine — not operate it. The data you have now is training data. You use that data as input to train an algorithm. And you use that algorithm to generate predictions to inform actions. So, yes, that does mean your data is valuable. But it does not mean your business can survive the storm. Once your data is used to train a prediction machine, it is devalued. It is not useful anymore for that sort of prediction. And there are only so many predictions your data will be useful for. To continue the oil analogy, data can be burned. It is somewhat lost after use. Scientists know this. They spend years collecting data, but once it has produced research findings, it sits unused in a file drawer or on back-up disk. Your business may be sitting on an oil well, but it’s finite. It doesn’t guarantee you more in the AI economy than perhaps a more favorable liquidation value. Even to the extent that your data could be valuable, your ability to capture that value may be limited. How many other sources of comparable data exist? If you are one of many yogurt vendors, then your database containing the past 10 years of yogurt sales and related data (price, temperature, sales of related products like ice cream) will have less market value than if you are the only owner of that type of data. In other words, just as with oil, the greater the number of other suppliers of your type of data, the less value you can capture from your training data. The value of your training data is further influenced by the value generated through enhanced prediction accuracy. Your training data is more

COPYRIGHT © 2018 HARVARD BUSINESS SCHOOL PUBLISHING ALL RIGHTS RESERVED. ©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTSCORPORATION. RESERVED.

393

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

valuable if enhanced prediction accuracy can increase yogurt sales by $100 million rather than only $10 million. Moreover, the ongoing value of data usually comes from the actions you take in your day-to-day business — the new data you accrue each day. New data allows you to operate your prediction machine after it is trained. It also enables you to improve your prediction machine through learning. While 10 years of data on past yogurt sales is valuable for training an AI model to predict future yogurt sales, the actual predictions used to manage the supply chain require operational data on an ongoing basis. And this is the important point for today’s incumbent companies. An AI startup that acquires a trove of data on past yogurt sales can train an AI model to predict future sales. It can’t actually use its model to make decisions unless the startup obtains ongoing operational data to learn from. Unlike startups, large enterprises generate operational data every day. That’s an asset. The more operations, the more data. Furthermore, the owner of the operation can actually make use of the prediction. It can use the prediction to enhance its future operation. In the AI economy, the value of your accumulated data is limited to a one-time benefit from training your AI model. And the value of training data is, like oil or any other input, influenced by the overall supply — it’s less valuable when more people have it. In contrast, the value of your ongoing operational data is not limited to a one-time benefit, but rather provides a perpetual benefit for operating and further enhancing your prediction machine. So, despite all the talk about data being the new oil, your accumulated historical data isn’t the thing. However, it may be the thing that gets you to the thing. Its value for your future business prospects is low. But if you can find ways to generate a new, ongoing data stream that delivers a performance advantage in terms of your AI’s predictive power, that will give you sustainable leverage when AI arrives.

Ajay Agrawal is the Peter Munk Professor of Entrepreneurship at the University of Toronto’s Rotman School of Management and Research Associate at the National Bureau of Economic Research in Cambridge, MA. He is founder of the Creative Destruction Lab, co-founder of The Next AI, and co-founder of Kindred. He is the co-author of Prediction Machines: The Simple Economics of Artificial Intelligence (Harvard Business School Press, April 2018).

Joshua Gans is the Jeffrey S. Skoll Chair of Technical Innovation and Entrepreneurship at the University of Toronto’s Rotman School of Management and serves as chief economist in the Creative Destruction Lab. He is the co-author of Prediction Machines: The Simple Economics of Artificial Intelligence (Harvard Business School Press, April 2018). His book, The Disruption Dilemma, is published by MIT Press.

Avi Goldfarb is the Ellison Professor of Marketing at the Rotman School of Management, University of Toronto. He is also a Research Associate at the National Bureau of Economic Research, Chief Data Scientist at the Creative Destruction Lab, and Senior Editor at Marketing Science. He is the co-author of Prediction Machines: The Simple Economics of Artificial Intelligence (Harvard Business School Press, April 2018).

COPYRIGHT © 2018 HARVARD BUSINESS SCHOOL PUBLISHING ALL RIGHTS RESERVED. ©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTSCORPORATION. RESERVED.

404

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

Discussion Questions

Is Your Company’s Data Actually Valuable in the AI Era? BY AJAY AGRAWAL, JOSHUA GANS, AND AVI GOLDFARB

1. What kinds of data does your company have? Is it a trove of data on past sales or behaviors? Or is it ongoing operational data? Or both? 2. Looking at your company’s data in the context the authors describe, are your data assets more valuable or less valuable than you had previously thought? What might you do to increase the value of your data in the AI era?

Yes, data, like oil, is used day-to-day to operate your prediction machine. But the data you are sitting on now is likely not that data. Instead, the data you have now, which your company accumulated over time, is the type of data used to build the prediction machine—not operate it.

COPYRIGHT © 2018 HARVARD BUSINESS SCHOOL PUBLISHING ALL RIGHTS RESERVED. ©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTSCORPORATION. RESERVED.

418

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

Q&A: HILARY MASON HOW AI FITS INTO YOUR DATA SCIENCE TEAM

It helps to know the three things data scientists do.

HBR: AI is a hot topic right now. As a data scientist and a researcher, how do you think about the recent progress in your field?

©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTS RESERVED.

MASON: If we were having this conversation eight or 10 years ago, it would have been about big data — about whether we could even build the infrastructure to get all the data in one place and query it. Once you can do that, you can do analytics — which is essentially counting things to answer questions that have business value or product value. People could always count things in data, but the change we saw about eight years ago was that new software made doing it affordable and accessible for a wide variety of people who never could do it before. And that led to the rise of data science, which is about counting things cleverly, predicting things, and building models on data. Because that modeling was now so much cheaper, it was applied not just to very high value problems, like actuarial science, but to things that may seem fairly trivial, like recommendations, search results, and that kind of stuff. GETTY IMAGES

I

n their HBR Big Idea feature, Erik Brynjolfsson and Andrew McAfee argue that AI and machine learning will soon become “general-purpose technologies” as significant as electricity or the internal combustion engine. They represent a landmark change in our technical capabilities and will power the next wave of economic growth. But how will we put them into practice? Where in the organization will these new capabilities sit, and how will companies take advantage of them? To get a practical, on-the-ground view, HBR senior editor Walter Frick spoke with Hilary Mason, the founder of Fast Forward Labs, a machine intelligence research firm. Here are excerpts from their conversation.

Then we had machine learning, which is a set of tools inside data science that let you count things cleverly and incorporate feedback loops. We began using the models to get more data from the world and fed it back into those models so that they improved over time. Now today we talk about AI. The term itself is a little bit loose and has both a technical meaning and a marketing meaning, but it’s essentially about using machine learning — and specifically deep learning — to enable applications that are built on top of this stack. That means that you can’t do AI without machine learning. You also can’t do machine learning without analytics, and you can’t do analytics without data infrastructure. And so that’s how I see them all being related. How do machine learning and AI fit into companies’ existing data capabilities? Data science is used in multiple ways inside an organization, and a really common mistake I see people make in managing it is assuming that because it runs on one tech stack, it’s just one thing. But I’d break it down into three capabilities, all of which rely on the same technology. The first capability is understanding the business. That’s

42

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

HILARY MASON



analytics, or business intelligence — being able to ask questions and analyze information to make better decisions. It’s usually run out of the CFO or COO’s office. It’s not necessarily a technical domain. The second capability is product data science: building algorithms and systems — which may use machine learning and AI — that actually improve the product. This is where things like spam filters, recommendation systems, search algorithms, and data visualization come in. This capability usually sits under a line of business and is run out of product development or engineering. The last data capability is one that tends to get neglected or lumped in with product data science. It’s an R&D capability — using data to open up new product, new business, and new revenue opportunities. And are all three capabilities changed by machine learning and AI? Let’s take a moment and look more closely at what deep learning offers, since it’s central to a lot of what people now call AI and a big part of the progress in machine learning in recent years. First, deep learning makes data that was previously inaccessible to any kind of analysis accessible — you can actually find value in video and audio data, for example. The number of companies that have a large amount of that kind of data is still fairly small, but I do think it’s likely to increase over time. Even analytics is impacted by the ability to use image data rather than just text or structured data. Second, deep learning enables new approaches to solving very difficult data science problems — text summarization, for example. Deep learning allows you to create predictive models at a level of quality and sophistication that was previously out of reach. And so deep learning also enhances the product

function of data science because it can generate new product opportunities. For example, several companies are using deep learning very successfully in e-commerce recommendation systems. Then of course deep learning affects the R&D function by pushing the frontier of what is technically possible. So data science is about analytics, product development, and R&D. Is this a walk-before-you-run situation? Or should companies attempt all three at once? It’s a little bit of both. You’ll leave opportunities on the table if you pursue only one of these use cases. However, it really helps to get your infrastructure and analytics piece to be fairly solid before jumping into R&D. And in practice we see that people are much more comfortable investing in cost-saving initiatives before they invest in new revenue opportunities. It’s just more culturally acceptable. What other mistakes do you see companies making in their data science efforts? A big one involves process. We’ve noticed that people shoehorn this kind of stuff into the software-engineering process, and that doesn’t work. Developing data science systems is fundamentally different in several ways. At the outset of a data science project, you don’t know if it’s going to work. At the outset of a software-engineering project, you know it’s going to work. This means that software-engineering processes fail when they encounter uncertainty. By contrast, data science requires an experimental process that allows for uncertainty. Also, every company has its own cultural hurdle to get over. A lot of companies aren’t places where you can work on something that doesn’t succeed, so the poor data scientists who do the risky research projects end

©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTS RESERVED.

up getting penalized in their annual reviews because they worked on something for two months that didn’t pay off, even though they did great work. Data science requires having that cultural space to experiment and work on things that might fail. Companies need to understand that they’re investing in a portfolio of initiatives, some of which will eventually pay off, generating dramatically more value than incremental product improvements do. How do you navigate all the buzz around this topic, and how do you recommend executives do so? I remain a relentless optimist about the potential of what we’re now calling AI, but I’m also a pragmatist in the sense that I need to deliver systems that work to our clients, and that is quite a constraint. There are some folks running around making claims that are clearly exaggerated and ridiculous. In other cases things that a few years ago we would have called a regression analysis are now being called AI, just to enhance their value from a marketing perspective. So my advice is to keep in mind that there is no magic. At a conceptual level nothing here is out of reach of any executive’s understanding. And if someone is pitching you on an idea and says, “I don’t want to explain how it works, but it’s AI,” it’s really important to keep asking: How does it work? What data goes in? What patterns might be in the data that the system could be learning? And what comes out? Because what comes out of a deep learning system is generally just a previously unlabeled data point that now has a label, along with some confidence in that label, and that’s it. It’s not intelligent in the sense that you and I are — and we’re still a long, long way away from anything that looks like the kind of intelligence a human has. 

43

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

Discussion Questions

How AI Fits into Your Data Science Team Q&A WITH HILARY MASON

1. After reading this article, do you understand what Mason describes as the three capabilities of data science—analytics, product development, and R&D? Can you identify these functions at your company? Does your company view these as three distinct capabilities? 2. Does your data science team have a cultural space that allows failure? Is your company betting on a portfolio of data initiatives, only some of which will pay off? How could your data science group encourage and even celebrate failure as a means to innovation?

If someone is pitching you on an idea and says, “I don’t want to explain how it works, but it’s AI,” it’s really important to keep asking: How does it work? What data goes in? What patterns might be in the data that the system could be learning? And what comes out?

COPYRIGHT © 2018 HARVARD BUSINESS SCHOOL PUBLISHING ALL RIGHTS RESERVED. ©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTSCORPORATION. RESERVED.

449

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

HOW AI IS BEING USED NOW: NEW PROCESSES, INSIGHTS, AND PRODUCTS

©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTS RESERVED.

45

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

REPRINT R1801H PUBLISHED IN HBR JANUARY–FEBRUARY 2018

ARTICLE TECHNOLOGY

Artificial Intelligence for the Real World Don’t start with moon shots. by Thomas H. Davenport and Rajeev Ronanki

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or ©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTS RESERVED. 46 800-988-0886 for additional copies.

FEATURE ARTIFICIAL INTELLIGENCE FOR THE REAL WORLD

Artificial Intelligence for the Real World Don’t start with moon shots.

BY THOMAS H. DAVENPORT AND RAJEEV RONANKI

IMAGES BY JAMES WHEATON AND ANDREW NGUYEN

©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTS RESERVED. 47 This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

In 2013, the MD Anderson Cancer Center launched a “moon shot” project: diagnose and recommend treatment plans for certain forms of cancer using IBM’s Watson cognitive system. But in 2017, the project was put on hold after costs topped $62 million—and the system had yet to be used on patients. At the same time, the cancer center’s IT group was experimenting with using cognitive ©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTS RESERVED. 48 This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

FEATURE ARTIFICIAL INTELLIGENCE FOR THE REAL WORLD

IN BRIEF THE PROBLEM

Cognitive technologies are increasingly being used to solve business problems, but many of the most ambitious AI projects encounter setbacks or fail.

THE APPROACH

Companies should take an incremental rather than a transformative approach and focus on augmenting rather than replacing human capabilities.

THE PROCESS

To get the most out of AI, firms must understand which technologies perform what types of tasks, create a prioritized portfolio of projects based on business needs, and develop plans to scale up across the company.

technologies to do much less ambitious jobs, such as making hotel and restaurant recommendations for patients’ families, determining which patients needed help paying bills, and addressing staff IT problems. The results of these projects have been much more promising: The new systems have contributed to increased patient satisfaction, improved financial performance, and a decline in time spent on tedious data entry by the hospital’s care managers. Despite the setback on the moon shot, MD Anderson remains committed to using cognitive technology—that is, next-generation artificial intelligence—to enhance cancer treatment, and is currently developing a variety of new projects at its center of competency for cognitive computing. The contrast between the two approaches is relevant to anyone planning AI initiatives. Our survey of 250 executives who are familiar with their companies’ use of cognitive technology shows that three-quarters of them believe that AI will substantially transform their companies within three years. However, our study of 152 projects in almost as many companies also reveals that highly ambitious moon shots are less likely to be successful than “low-­ hanging fruit” projects that enhance business processes. This shouldn’t be surprising—such has been the case with the great majority of new technologies that companies have adopted in the past. But the hype surrounding artificial intelligence has been especially powerful, and some organizations have been seduced by it. In this article, we’ll look at the various categories of AI being employed and provide a framework for how companies should begin to build up their cognitive capabilities in the next several years to achieve their business objectives.

THREE TYPES OF AI

It is useful for companies to look at AI through the lens of business capabilities rather than technologies. Broadly speaking, AI can support three important business needs: automating business processes, gaining insight through data analysis, and engaging with customers and employees. (See the exhibit “Cognitive Projects by Type.”) Process automation. Of the 152 projects we studied, the most common type was the automation of digital and physical tasks—typically back-office administrative and financial activities—using robotic process automation technologies. RPA is more advanced than earlier business-process automation tools, because the “robots” (that is, code on a server) act like a human inputting and consuming information from multiple IT systems. Tasks include: • transferring data from e‑mail and call center systems into systems of record—for example, updating customer files with address changes or service additions;

• replacing lost credit or ATM cards, reaching into multiple systems to update records and handle customer communications; • reconciling failures to charge for services across billing systems by extracting information from multiple document types; and • “reading” legal and contractual documents to extract provisions using natural language processing. RPA is the least expensive and easiest to implement of the cognitive technologies we’ll discuss here, and typically brings a quick and high return on investment. (It’s also the least “smart” in the sense that these applications aren’t programmed to learn and improve, though developers are slowly adding more intelligence and learning capability.) It is particularly well suited to working across multiple back-end systems. At NASA, cost pressures led the agency to launch four RPA pilots in accounts payable and receivable, IT spending, and human resources—all managed by a shared services center. The four projects worked well—in the HR application, for example, 86% of transactions were completed without human intervention—and are being rolled out across the organization. NASA is now implementing more RPA bots, some with higher levels of intelligence. As Jim Walker, project leader for the shared services organization notes, “So far it’s not rocket science.” One might imagine that robotic process automation would quickly put people out of work. But across the 71 RPA projects we reviewed (47% of the total), replacing administrative employees was neither the primary objective nor a common outcome. Only a few projects led to reductions in head count, and in most cases, the tasks in question had already been shifted to outsourced workers. As technology improves, robotic automation projects are likely to lead to some job losses in the future, particularly in the offshore business-process outsourcing industry. If you can outsource a task, you can probably automate it. Cognitive insight. The second most common type of project in our study (38% of the total) used algorithms to detect patterns in vast volumes of data and interpret their meaning. Think of it as “analytics on steroids.” These machine-learning applications are being used to: • predict what a particular customer is likely to buy; • identify credit fraud in real time and detect insurance claims fraud; • analyze warranty data to identify safety or quality problems in automobiles and other manufactured products; • automate personalized targeting of digital ads; and • provide insurers with more-accurate and detailed actuarial modeling. Cognitive insights provided by machine learning differ from those available from traditional analytics

©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTS RESERVED. 49 This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

COGNITIVE PROJECTS BY TYPE We studied 152 cognitive technology projects and found that they fell into three categories. ROBOTICS & COGNITIVE AUTOMATION

COGNITIVE INSIGHT

COGNITIVE ENGAGEMENT

71 57 24

in three ways: They are usually much more data-­ intensive and detailed, the models typically are trained on some part of the data set, and the models get better—that is, their ability to use new data to make predictions or put things into categories improves over time. Versions of machine learning (deep learning, in particular, which attempts to mimic the activity in the human brain in order to recognize patterns) can perform feats such as recognizing images and speech. Machine learning can also make available new data for better analytics. While the activity of data curation has historically been quite labor-intensive, now machine learning can identify probabilistic matches— data that is likely to be associated with the same person or company but that appears in slightly different

formats—across databases. GE has used this technology to integrate supplier data and has saved $80 million in its first year by eliminating redundancies and negotiating contracts that were previously managed at the business unit level. Similarly, a large bank used this technology to extract data on terms from supplier contracts and match it with invoice numbers, identifying tens of millions of dollars in products and services not supplied. Deloitte’s audit practice is using cognitive insight to extract terms from contracts, which enables an audit to address a much higher proportion of documents, often 100%, without human auditors’ having to painstakingly read through them. Cognitive insight applications are typically used to improve performance on jobs only machines can do— tasks such as programmatic ad buying that involve

©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTS RESERVED. 50 This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

FEATURE ARTIFICIAL INTELLIGENCE FOR THE REAL WORLD

THE BUSINESS BENEFITS OF AI

We surveyed 250 executives who were familiar with their companies’ use of cognitive technologies to learn about their goals for AI initiatives. More than half said their primary goal was to make existing products better. Reducing head count was mentioned by only 22%. PERCENTAGE OF EXECUTIVES WHO CITE THE FOLLOWING AS BENEFITS OF AI

ENHANCE THE FEATURES, FUNCTIONS, AND PERFORMANCE OF OUR PRODUCTS

51%

MAKE BETTER DECISIONS

35%

CREATE NEW PRODUCTS

32%

OPTIMIZE INTERNAL BUSINESS OPERATIONS

36%

FREE UP WORKERS TO BE MORE CREATIVE BY AUTOMATING TASKS

36%

PURSUE NEW MARKETS

25%

CAPTURE AND APPLY SCARCE KNOWLEDGE WHERE NEEDED

25%

OPTIMIZE EXTERNAL PROCESSES LIKE MARKETING AND SALES

30%

REDUCE HEAD COUNT THROUGH AUTOMATION

22% SOURCE DELOITTE 2017

such high-speed data crunching and automation that they’ve long been beyond human ability—so they’re not generally a threat to human jobs. Cognitive engagement. Projects that engage employees and customers using natural language processing chatbots, intelligent agents, and machine learning were the least common type in our study (accounting for 16% of the total). This category includes: • intelligent agents that offer 24/7 customer service addressing a broad and growing array of issues from password requests to technical support questions— all in the customer’s natural language;

• internal sites for answering employee questions on topics including IT, employee benefits, and HR policy; • product and service recommendation systems for retailers that increase personalization, engagement, and sales—typically including rich language or images; and • health treatment recommendation systems that help providers create customized care plans that take into account individual patients’ health status and previous treatments. The companies in our study tended to use cognitive engagement technologies more to interact with employees than with customers. That may change as firms become more comfortable turning customer interactions over to machines. Vanguard, for example, is piloting an intelligent agent that helps its customer service staff answer frequently asked questions. The plan is to eventually allow customers to engage with the cognitive agent directly, rather than with the human customer-service agents. SEBank, in Sweden, and the medical technology giant Becton, Dickinson, in the United States, are using the lifelike intelligent-agent avatar Amelia to serve as an internal employee help desk for IT support. SEBank has recently made Amelia available to customers on a limited basis in order to test its performance and customer response. Companies tend to take a conservative approach to customer-facing cognitive engagement technologies largely because of their immaturity. Facebook, for example, found that its Messenger chatbots couldn’t answer 70% of customer requests without human intervention. As a result, Facebook and several other firms are restricting bot-based interfaces to certain topic domains or conversation types. Our research suggests that cognitive engagement apps are not currently threatening customer service or sales rep jobs. In most of the projects we studied, the goal was not to reduce head count but to handle growing numbers of employee and customer interactions without adding staff. Some organizations were planning to hand over routine communications to machines, while transitioning customer-support personnel to more-complex activities such as handling customer issues that escalate, conducting extended unstructured dialogues, or reaching out to customers before they call in with problems. As companies become more familiar with cognitive tools, they are experimenting with projects that combine elements from all three categories to reap the benefits of AI. An Italian insurer, for example, developed a “cognitive help desk” within its IT organization. The system engages with employees using deep-­ learning technology (part of the cognitive insights category) to search frequently asked questions and answers, previously resolved cases, and documentation to come up with solutions to employees’ problems. It uses a smart-­routing capability (business process automation) to forward the most complex problems to

©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTS RESERVED. 51 This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

human representatives, and it uses natural language processing to support user requests in Italian. Despite their rapidly expanding experience with cognitive tools, however, companies face significant obstacles in development and implementation. On the basis of our research, we’ve developed a four-step framework for integrating AI technologies that can help companies achieve their objectives, whether the projects are moon shoots or business-process enhancements.

1. UNDERSTANDING THE TECHNOLOGIES

Before embarking on an AI initiative, companies must understand which technologies perform what types of tasks, and the strengths and limitations of each. Rulebased expert systems and robotic process automation, for example, are transparent in how they do their work, but neither is capable of learning and improving. Deep learning, on the other hand, is great at learning from large volumes of labeled data, but it’s almost impossible to understand how it creates the models it does. This “black box” issue can be problematic in highly regulated industries such as financial services, in which regulators insist on knowing why decisions are made in a certain way. We encountered several organizations that wasted time and money pursuing the wrong technology for the job at hand. But if they’re armed with a good understanding of the different technologies, companies are better positioned to determine which might best address specific needs, which vendors to work with, and how quickly a system can be implemented. Acquiring this understanding requires ongoing research and education, usually within IT or an innovation group. In particular, companies will need to leverage the capabilities of key employees, such as data scientists, who have the statistical and big-data skills necessary to learn the nuts and bolts of these technologies. A main success factor is your people’s willingness to learn. Some will leap at the opportunity, while others will want to stick with tools they’re familiar with. Strive to have a high percentage of the former. If you don’t have data science or analytics capabilities in-house, you’ll probably have to build an ecosystem of external service providers in the near term. If you expect to be implementing longer-term AI projects, you will want to recruit expert in-house talent. Either way, having the right capabilities is essential to progress. Given the scarcity of cognitive technology talent, most organizations should establish a pool of resources—perhaps in a centralized function such as IT or strategy—and make experts available to high-­ priority projects throughout the organization. As needs and talent proliferate, it may make sense to dedicate groups to particular business functions or units, but even then a central coordinating function can be useful in managing projects and careers.

THE CHALLENGES OF AI

Executives in our survey identified several factors that can stall or derail AI initiatives, ranging from integration issues to scarcity of talent. PERCENTAGE WHO CITE THE FOLLOWING AS OBSTACLES

IT’S HARD TO INTEGRATE COGNITIVE PROJECTS WITH EXISTING PROCESSES AND SYSTEMS

47%

TECHNOLOGIES AND EXPERTISE ARE TOO EXPENSIVE

40%

MANAGERS DON’T UNDERSTAND COGNITIVE TECHNOLOGIES AND HOW THEY WORK

37%

WE CAN’T GET ENOUGH PEOPLE WITH EXPERTISE IN THE TECHNOLOGY

35%

TECHNOLOGIES ARE IMMATURE

31%

TECHNOLOGIES HAVE BEEN OVERSOLD IN THE MARKETPLACE

18% SOURCE DELOITTE 2017

2. CREATING A PORTFOLIO OF PROJECTS

The next step in launching an AI program is to systematically evaluate needs and capabilities and then develop a prioritized portfolio of projects. In the companies we studied, this was usually done in workshops or through small consulting engagements. We recommend that companies conduct assessments in three broad areas. Identifying the opportunities. The first assessment determines which areas of the business could benefit most from cognitive applications. Typically, they are parts of the company where “knowledge”— insight derived from data analysis or a collection of texts—is at a premium but for some reason is not available. • Bottlenecks. In some cases, the lack of cognitive insights is caused by a bottleneck in the flow of information; knowledge exists in the organization, but it is not optimally distributed. That’s often the case in health care, for example, where knowledge tends to be siloed within practices, departments, or academic medical centers.

©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTS RESERVED. 52 This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

FEATURE ARTIFICIAL INTELLIGENCE FOR THE REAL WORLD

• Scaling challenges. In other cases, knowledge exists, but the process for using it takes too long or is expensive to scale. Such is often the case with knowledge developed by financial advisers. That’s why many investment and wealth management firms now offer AI-supported “robo-advice” capabilities that provide clients with cost-effective guidance for routine financial issues. In the pharmaceutical industry, Pfizer is tackling the scaling problem by using IBM’s Watson to accelerate the laborious process of drug-discovery research in immuno-oncology, an emerging approach to cancer treatment that uses the body’s immune system to help fight cancer. Immunooncology drugs can take up to 12 years to bring to market. By combining a sweeping literature review with Pfizer’s own data, such as lab reports, Watson is helping researchers to surface relationships and find hidden patterns that should speed the identification of new drug targets, combination therapies for study, and patient selection strategies for this new class of drugs. • Inadequate firepower. Finally, a company may collect more data than its existing human or computer firepower can adequately analyze and apply. For example, a company may have massive amounts of data on consumers’ digital behavior but lack insight about what it means or how it can be strategically applied. To address this, companies are using machine learning to support tasks such as programmatic buying of personalized digital ads or, in the case of Cisco Systems and IBM, to create tens of thousands of “propensity models” for determining which customers are likely to buy which products. Determining the use cases. The second area of assessment evaluates the use cases in which cognitive applications would generate substantial value and contribute to business success. Start by asking key questions such as: How critical to your overall strategy is addressing the targeted problem? How difficult would it be to implement the proposed AI solution—both technically and organizationally? Would the benefits from launching the application be worth the effort? Next, prioritize the use cases according to which offer the most short- and long-term value, and which might ultimately be integrated into a broader platform or suite of cognitive capabilities to create competitive advantage. Selecting the technology. The third area to assess examines whether the AI tools being considered for each use case are truly up to the task. Chatbots and intelligent agents, for example, may frustrate some companies because most of them can’t yet match human problem solving beyond simple scripted cases (though they are improving rapidly). Other technologies, like robotic process automation that can streamline simple processes such as invoicing, may in fact slow down more-complex production systems. And

while deep learning visual recognition systems can recognize images in photos and videos, they require lots of labeled data and may be unable to make sense of a complex visual field. In time, cognitive technologies will transform how companies do business. Today, however, it’s wiser to take incremental steps with the currently available technology while planning for transformational change in the not-too-distant future. You may ultimately want to turn customer interactions over to bots, for example, but for now it’s probably more feasible—and sensible—to automate your internal IT help desk as a step toward the ultimate goal.

3. LAUNCHING PILOTS

Because the gap between current and desired AI capabilities is not always obvious, companies should create pilot projects for cognitive applications before rolling them out across the entire enterprise. Proof-of-concept pilots are particularly suited to initiatives that have high potential business value or allow the organization to test different technologies at the same time. Take special care to avoid “injections” of projects by senior executives who have been influenced by technology vendors. Just because executives and boards of directors may feel pressure to “do something cognitive” doesn’t mean you should bypass the rigorous piloting process. Injected projects often fail, which can significantly set back the organization’s AI program. If your firm plans to launch several pilots, consider creating a cognitive center of excellence or similar structure to manage them. This approach helps build the needed technology skills and capabilities within the organization, while also helping to move small pilots into broader applications that will have a greater impact. Pfizer has more than 60 projects across the company that employ some form of cognitive technology; many are pilots, and some are now in production. At Becton, Dickinson, a “global automation” function within the IT organization oversees a number of cognitive technology pilots that use intelligent digital agents and RPA (some work is done in partnership with the company’s Global Shared Services organization). The global automation group uses end-to-end process maps to guide implementation and identify automation opportunities. The group also uses graphical “heat maps” that indicate the organizational activities most amenable to AI interventions. The company has successfully implemented intelligent agents in IT support processes, but as yet is not ready to support large-scale enterprise processes, like order-tocash. The health insurer Anthem has developed a similar centralized AI function that it calls the Cognitive Capability Office. Business-process redesign. As cognitive technology projects are developed, think through how

©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTS RESERVED. 53 This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

workflows might be redesigned, focusing specifically on the division of labor between humans and the AI. In some cognitive projects, 80% of decisions will be made by machines and 20% will be made by humans; others will have the opposite ratio. Systematic redesign of workflows is necessary to ensure that humans and machines augment each other’s strengths and compensate for weaknesses. The investment firm Vanguard, for example, has a new “Personal Advisor Services” (PAS) offering, which combines automated investment advice with guidance from human advisers. In the new system, cognitive technology is used to perform many of the traditional tasks of investment advising, including constructing a customized portfolio, rebalancing portfolios over time, tax loss harvesting, and tax-­ efficient investment selection. Vanguard’s human advisers serve as “investing coaches,” tasked with answering investor questions, encouraging healthy financial behaviors, and being, in Vanguard’s words, “emotional circuit breakers” to keep investors on plan. Advisers are encouraged to learn about behavioral finance to perform these roles effectively. The PAS approach has quickly gathered more than $80 billion in assets under management, costs are lower than those for purely human-based advising, and customer satisfaction is high. (See the exhibit “One Company’s Division of Labor.”) Vanguard understood the importance of work redesign when implementing PAS, but many companies simply “pave the cow path” by automating existing work processes, particularly when using RPA technology. By automating established workflows, companies can quickly implement projects and achieve ROI—but they forgo the opportunity to take full advantage of AI capabilities and substantively improve the process. Cognitive work redesign efforts often benefit from applying design-thinking principles: understanding customer or end-user needs, involving employees whose work will be restructured, treating designs as experimental “first drafts,” considering multiple alternatives, and explicitly considering cognitive technology capabilities in the design process. Most cognitive projects are also suited to iterative, agile approaches to development.

ONE COMPANY’S DIVISION OF LABOR

Vanguard, the investment services firm, uses cognitive technology to provide customers with investment advice at a lower cost. Its Personal Advisor Services system automates many traditional tasks of investment advising, while human advisers take on highervalue activities. Here’s how Vanguard redesigned its work processes to get the most from the new system.

COGNITIVE TECHNOLOGY

Generates a financial plan Provides goals-based forecasting in real time Rebalances portfolio to target mix Minimizes taxes Tracks aggregated assets in one place Engages clients virtually

ADVISER

Understands investment goals Customizes an implementation plan Provides investment analysis and retirement planning Develops retirement income and Social Security drawdown strategies Serves as a behavioral coach Monitors spending to encourage accountability Offers ongoing wealth and financial-planning support Addresses estate-planning considerations

SOURCE VANGUARD GROUP

4. SCALING UP

Many organizations have successfully launched cognitive pilots, but they haven’t had as much success rolling them out organization-wide. To achieve their goals, companies need detailed plans for scaling up, which requires collaboration between technology experts and owners of the business process being automated. Because cognitive technologies typically support individual tasks rather than entire processes, scale-up almost always requires integration with ©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTS RESERVED. 54 This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

FEATURE ARTIFICIAL INTELLIGENCE FOR THE REAL WORLD

FURTHER READING “Big Idea: The Business of Artificial Intelligence” by Erik Brynjolfsson and Andrew McAfee HBR.org/ai

“Inside Facebook’s AI Workshop” by Scott Berinato HBR.org/ai

“AI Can Be a Troublesome Teammate” by Kurt Gray HBR.org/ai

existing systems and processes. Indeed, in our survey, executives reported that such integration was the greatest challenge they faced in AI initiatives. Companies should begin the scaling-up process by considering whether the required integration is even possible or feasible. If the application depends on special technology that is difficult to source, for example, that will limit scale-up. Make sure your business process owners discuss scaling considerations with the IT organization before or during the pilot phase: An end run around IT is unlikely to be successful, even for relatively simple technologies like RPA. The health insurer Anthem, for example, is taking on the development of cognitive technologies as part of a major modernization of its existing systems. Rather than bolting new cognitive apps onto legacy technology, Anthem is using a holistic approach that maximizes the value being generated by the cognitive applications, reduces the overall cost of development and integration, and creates a halo effect on legacy systems. The company is also redesigning processes at the same time to, as CIO Tom Miller puts it, “use cognitive to move us to the next level.” In scaling up, companies may face substantial change-management challenges. At one U.S. apparel retail chain, for example, the pilot project at a small subset of stores used machine learning for online product recommendations, predictions for optimal inventory and rapid replenishment models, and— most difficult of all—merchandising. Buyers, used to ordering product on the basis of their intuition, felt threatened and made comments like “If you’re going to trust this, what do you need me for?” After the pilot, the buyers went as a group to the chief merchandising officer and requested that the program be killed. The executive pointed out that the results were positive and warranted expanding the project. He assured the buyers that, freed of certain merchandising tasks, they could take on more high-value work that humans can still do better than machines, such as understanding younger customers’ desires and determining apparel manufacturers’ future plans. At the same time, he acknowledged that the merchandisers needed to be educated about a new way of working. If scale-up is to achieve the desired results, firms must also focus on improving productivity. Many, for example, plan to grow their way into productivity— adding customers and transactions without adding staff. Companies that cite head count reduction as the primary justification for the AI investment should ideally plan to realize that goal over time through attrition or from the elimination of outsourcing.

THE FUTURE COGNITIVE COMPANY

Our survey and interviews suggest that managers experienced with cognitive technology are bullish

on its prospects. Although the early successes are relatively modest, we anticipate that these technologies will eventually transform work. We believe that companies that are adopting AI in moderation now— and have aggressive implementation plans for the future—will find themselves as well positioned to reap benefits as those that embraced analytics early on. Through the application of AI, information-intensive domains such as marketing, health care, financial services, education, and professional services could become simultaneously more valuable and less expensive to society. Business drudgery in every industry and function—overseeing routine transactions, repeatedly answering the same questions, and extracting data from endless documents—could become the province of machines, freeing up human workers to be more productive and creative. Cognitive technologies are also a catalyst for making other data-intensive technologies succeed, including autonomous vehicles, the Internet of Things, and mobile and multi­ channel consumer technologies. The great fear about cognitive technologies is that they will put masses of people out of work. Of course, some job loss is likely as smart machines take over certain tasks traditionally done by humans. However, we believe that most workers have little to fear at this point. Cognitive systems perform tasks, not entire jobs. The human job losses we’ve seen were primarily due to attrition of workers who were not replaced or through automation of outsourced work. Most cognitive tasks currently being performed augment human activity, perform a narrow task within a much broader job, or do work that wasn’t done by humans in the first place, such as big-data analytics. Most managers with whom we discuss the issue of job loss are committed to an augmentation strategy—that is, integrating human and machine work, rather than replacing humans entirely. In our survey, only 22% of executives indicated that they considered reducing head count as a primary benefit of AI. We believe that every large company should be exploring cognitive technologies. There will be some bumps in the road, and there is no room for complacency on issues of workforce displacement and the ethics of smart machines. But with the right planning and development, cognitive technology could usher in a golden age of productivity, work satisfaction, and prosperity.  HBR Reprint R1801H THOMAS H. DAVENPORT is the President’s Distinguished Professor of Information Technology and Management at Babson College, a research fellow at the MIT Initiative on the Digital Economy, and a senior adviser at Deloitte Analytics. RAJEEV RONANKI is a principal at Deloitte Consulting, where he leads the cognitive computing and health care innovation practices. Some of the companies mentioned in this article are Deloitte clients.

©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTS RESERVED. 55 This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

Discussion Questions

Artificial Intelligence for the Real World BY THOMAS H. DAVENPORT AND RAJEEV RONANKI

1. Are AI initiatives in your organization “moonshots” or “low-hanging fruit” projects, or is there a mix? Do you think the balance is right? 2. What surprises you about the “Business Benefits of AI” bar chart? How would your business answer this survey? 3. Which factors in the “Challenges of AI” bar chart affect your business most? Can they be changed or improved, or are they simply constraints that you’ll have to work around? 4. Have executives who have been influenced by technology vendors “injected” cognitive projects into your company? How have those projects fared compared to projects that have been rigorously piloted?

Our survey of 250 executives who are familiar with their companies’ use of cognitive technology shows that three-quarters of them believe that AI will substantially transform their companies within three years.

COPYRIGHT © 2018 HARVARD BUSINESS SCHOOL PUBLISHING ALL RIGHTS RESERVED. ©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTSCORPORATION. RESERVED.

11 56

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

REPRINT H046N5 PUBLISHED ON HBR.ORG MARCH 05, 2018

ARTICLE INNOVATION

How AI Is Taking the Scut Work Out of Health Care by Jonathan Bush

©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTS RESERVED.

57

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

INNOVATION

How AI Is Taking the Scut Work Out of Health Care

by Jonathan Bush MARCH 05, 2018

CSA Images/Pattern Collection/Getty Images When we think of breakthroughs in healthcare, we often conjure images of heroic interventions — the first organ transplantation, robotic surgery, and so on. But in fact many of the greatest leaps in human health have come from far more prosaic interventions — the safe disposal of human excrement through sewage and sanitation, for example, or handwashing during births and caesarians.

COPYRIGHT © 2018 HARVARD BUSINESS SCHOOL PUBLISHING ALL RIGHTS RESERVED. ©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTSCORPORATION. RESERVED.

582

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

We have a similar opportunity in medicine now with the application of artificial intelligence and machine learning. Glamorous projects to do everything from curing cancer to helping paralyzed patients walk through AI have generated enormous expectations. But the greatest opportunity for AI in the near term may come not from headline-grabbing moonshots but from putting computers and algorithms to work on the most mundane drudgery possible. Excessive paperwork and red-tape is the sewage of modern medicine. An estimated 14% of wasted health care spending — $91 billion — is the result of inefficient administration. Let’s give AI the decidedly unsexy job of cleaning out the administrative muck that’s clogging up our medical organizations, sucking value out of our economy, and literally making doctors ill with stress. Here’s just one example of the immediate opportunity: Each year, some 120 million faxes still flow into the practices of the more than 100,000 providers on the network of athenahealth, the healthcare technology company where I’m CEO. That’s right: faxes. Remember those? In healthcare, faxes remain the most common method that practitioners use to communicate with each other, and therefore often contain important clinical information: lab results, specialist consult notes, prescriptions and so on. Because most healthcare fax numbers are public, doctors also receive scores of pizza menus, travel specials, and other “junk faxes.” Faxes don’t contain any structured text — so it takes medical practice staff an average of two minutes and 36 seconds to review each document and input relevant data into patient records. Through a combination of machine learning and business-process outsourcing that has automated the categorizing of faxes, we’ve reduced timeper-fax for our practices to one minute and 11 seconds. As a result, last year alone we managed to eliminate over 3 million hours of work from the healthcare system. And that’s just the beginning for our AI team. Next year, we hope to reduce the time it takes to import data from a fax into a patient record to 30 seconds. And we’re developing software that can scan lab results and flag urgent findings for human attention and an algorithm that can help automatically schedule high-risk patients for routine follow-ups. Reading faxes and scheduling appointments don’t exactly quicken the pulse. But here’s why this sort of work is so important. First, we are in the midst of a burnout crisis among U.S. physicians. They’re crushed by administrative overload and feel they are becoming box-tickers rather than clinicians. Patients, too, feel overwhelmed by the cumbersome work required to chase referrals and ensure basic clinical information follows them through the health system. Applying AI to the work that doctors detest presents a path to redemption for the health IT industry. For too many doctors, once-hyped technology such as electronic health records have become part of the problem and added to – rather than mitigated – overload and burnout. We need to rebuild confidence in the promise of technology to free up provider time and enhance care delivery. Will algorithms and AI cause new, unforeseen tensions with physicians in the future? Perhaps. That’s always a risk with new technology. But in my experience, most doctors don’t fear automation, they

COPYRIGHT © 2018 HARVARD BUSINESS SCHOOL PUBLISHING ALL RIGHTS RESERVED. ©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTSCORPORATION. RESERVED.

593

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

fear a loss of autonomy. Using AI to relieve scut work will allow them to focus again on what they love most and where they create the most value: the patient encounter. Finally, focusing artificial intelligence on the goal of eliminating the mundane annoyances of modern medicine might one day lay the groundwork for curing cancer (and other AI moonshots). A recent report by JASON — the elite group of scientists who advise the U.S. government on matters of science and technology — found that poor data management remains a key obstacle to the clinical application of AI. Much heavy lifting remains to be done to improve the data on which the future of AI relies. We need to bust data out of silos so it can be easily accessed, queried, and analyzed. Using AI to correctly identify, categorize and share information will lay the groundwork for future, breakthrough analyses. Most healthcare executives are still unsure of their AI strategy. They sense that AI will be a game changer, but they’re not sure how. I love that healthcare has heroic ambitions for a promising new technology, even after years of high-tech disappointment. But while we shoot for the moon, let’s clean up the muck that’s bogging us down today, unleashing our potential to transform healthcare.

Jonathan Bush is the CEO of athenahealth and the author of Where Does it Hurt? An Entrepreneur’s Guide to Fixing Healthcare.

COPYRIGHT © 2018 HARVARD BUSINESS SCHOOL PUBLISHING ALL RIGHTS RESERVED. ©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTSCORPORATION. RESERVED.

604

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

Discussion Questions

How AI Is Taking the Scut Work Out of Health Care BY JONATHAN BUSH

1. The author begins by noting that prosaic medical innovations—hand washing, waste disposal—have saved far more human lives than technological marvels like organ transplants and robotic surgery. Is there a similar analogy in your industry? Where are there more opportunites for improvements like these? 2. What “administrative muck” is clogging up your organization? Could process automation help alleviate these inefficiencies? 3. If you use AI processes to automate rote back-office tasks in your organization, how can you ensure that the freed-up employee time is dedicated to more creative, fulfilling, and value-adding work?

In my experience, most doctors don’t fear automation, they fear a loss of autonomy. Using AI to relieve scut work will allow them to focus again on what they love most and where they create the most value: the patient encounter.

COPYRIGHT © 2018 HARVARD BUSINESS SCHOOL PUBLISHING ALL RIGHTS RESERVED. ©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTSCORPORATION. RESERVED.

12 61

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

REPRINT H03TDF PUBLISHED ON HBR.ORG AUGUST 23, 2017

ARTICLE TECHNOLOGY

3 Ways Companies Are Building a Business Around AI

by Quentin Hardy

©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTS RESERVED.

62

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

TECHNOLOGY

3 Ways Companies Are Building a Business Around AI

by Quentin Hardy AUGUST 23, 2017

There is no argument about whether artificial intelligence (AI) is coming. It is here, in automobiles, smartphones, aircraft, and much else. Not least in the online search abilities, speech and translation features, and image recognition technology of my employer, Alphabet. The question now moves to how broadly AI will be employed in industry and society, and by what means. Many other companies, including Microsoft and Amazon, also already offer AI tools which,

COPYRIGHT © 2017 HARVARD BUSINESS SCHOOL PUBLISHING ALL RIGHTS RESERVED. ©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTSCORPORATION. RESERVED.

632

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

like Google Cloud, where I work, will be sold online as cloud computing services. There are numerous other AI products available to business, like IBM’s Watson, or software from emerging vendors. Whatever hype businesspeople read around AI — and there is a great deal — the intentions and actions of so many players should alert them to the fundamental importance of this new technology. This is no simple matter, as AI is both familiar and strange. At heart, the algorithms and computation are dedicated to unearthing novel patterns, which is what science, technology, markets, and the humanistic arts have done throughout the story of humankind. The strange part is how today’s AI works, building subroutines of patterns, and loops of patterns about other patterns, training itself through multiple layers that are only possible with very large amounts of computation. For perhaps the first time, we have invented a machine that cannot readily explain itself. In the face of such technical progress, paralysis is rarely a good strategy. The question then becomes: How should a company that isn’t involved in building AI think about using it? Even in these early days, practices of successful early adopters offer several useful lessons: 1. Find and own valuable data no one else has. 2. Take a systemic view of your business, and find data adjacencies. 3. Package AI for the customer experience.

Capture the Scarce Data CAMP3 is a 26-person company, headquartered in Alpharetta, Georgia, that deploys and manages wireless sensor networks for agriculture. The company also sells Google’s G Suite email and collaboration products on a commission basis. Founder and chief executive Craig Ganssle was an early user of Google Glass. Glass failed as a consumer product, but the experience of wearing a camera and collecting images in the field inspired Ganssle to think about ways farmers could use AI to spot plant diseases and pests early on. AI typically works by crunching very large amounts of data to figure out telltale patterns, then testing provisional patterns against similar data it hasn’t yet processed. Once validated, the pattern-finding methodology is strengthened by feeding it more data. CAMP3’s initial challenge was securing enough visual data to train its AI product. Not only were there relatively few pictures of diseased crops and crop pests, but they were scattered across numerous institutions, often without proper identification. “Finding enough images of northern corn leaf blight [NCLB] took 10 months,” said Ganssle. “There were lots of pictures in big agricultural universities, but no one had the information well-tagged. Seed COPYRIGHT © 2017 HARVARD BUSINESS SCHOOL PUBLISHING ALL RIGHTS RESERVED. ©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTSCORPORATION. RESERVED.

643

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

companies had pictures too, but no one had pictures of healthy corn, corn with early NCLB, corn with advanced NCLB.” They collected whatever they could from every private, educational, and government source they could, and then took a lot of pictures themselves. Training the data, in this case, may have been easier than getting the data in the first place. That visual training data is a scarce commodity, and a defensible business asset. Initial training for things like NCLB, cucumber downy mildew, or sweet corn worm initially required “tens of thousands” of images, he said. With a system trained, he added, it now requires far fewer images to train for a disease. CAMP3 trains the images on TensorFlow, an AI software framework first developed by Google and then open sourced. For computing, he relied on Amazon Web Services and Google Compute Engine. “Now we can take the machine from kindergarten to PhD-style analysis in a few hours,” Ganssle said. The painful process of acquiring and correctly tagging the data, including time and location information for new pictures the company and customers take, gave CAMP3 what Ganssle considers a key strategic asset. “Capture something other people don’t have, and organize it with a plan for other uses down the road,” he said. “With AI, you never know what problem you will need to tackle next. This could be used for thinking about soils, or changing water needs. When we look at new stuff, or start to do predictive modeling, this will be data that falls off the truck, that we pick up and use.”

Explore Your Data Adjacencies TalkIQ is a company that monitors sales and customer service phone calls, turns the talk into text, and then scans the words in real time for keywords and patterns that predict whether a company is headed for a good outcome — a new sale, a happy customer. The company got its start after Jack Abraham, a former eBay executive and entrepreneur, founded ZenReach, a Phoenix company that connects online and offline commerce, in part through extensive call centers. “I kept thinking that if I could listen to everything our customers were asking for, I would capture the giant brain of the company,” said Abraham. “Why does one rep close 50% of his calls, while the other gets 25%?” The data from those calls could improve performance at ZenReach, he realized, but could also be the training set for a new business that served other companies. TalkIQ, based in San Francisco, took two years to build. Data scientists examined half a million conversations preserved in the company’s computer-based ZenReach phone system. COPYRIGHT © 2017 HARVARD BUSINESS SCHOOL PUBLISHING ALL RIGHTS RESERVED. ©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTSCORPORATION. RESERVED.

654

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

As with CAMP3, part of the challenge was correctly mapping information — in this case, conversations in crowded rooms, sometimes over bad phone connections — and tagging things like product names, features, and competitors. TalkIQ uses automated voice recognition and algorithms that understand natural language, among other tools. Since products and human interactions change even faster than biology, the training corpus for TalkIQ needs to train almost continuously to predict well, said Dan O’Connell, the company’s chief executive. “Every prediction depends on accurate information,” he said. “At the same time, you have to be careful of ‘overfitting,’ or building a model so complex that the noise is contributing to results as much as good data. Built as an adjacency to ZenReach, TalkIQ must also tweak for individual customer and vertical industry needs. The product went into commercial release in January, and according to Abraham now has 27 companies paying for the service. “If we’re right, this is how every company will run in the future.”

Focus on Customer Experience Last March the Denver-based company Blinker launched a mobile app for buying and selling cars in the state of Colorado. Customers are asked to photograph the back of their vehicle, and within moments of uploading the image the car’s year, make and model, and resale value are identified. From there it is a relatively simple matter to offer the car, or seek refinancing and insurance. The AI that identifies the car so readily seems like magic. In fact, the process is done using TensorFlow, along with the Google Vision API, to identify the vehicle. Blinker has agreements with third-party providers of motor vehicle data, and once it identifies the plate number, it can get the other information from the files (where possible, the machine also checks available image data.) Blinker has filed for patents on a number of the things it does, but the company’s founder and chief executive thinks his real edge is his 44 years in the business of car dealerships. “Whatever you do, you are still selling cars,” said Rod Buscher. “People forget that the way it feels, and the pain points of buying a car, are still there.” He noted that Beepi, an earlier peer-to-peer attempt to sell cars online, “raised $150 million, with a great concept and smart guys. They still lost it all. The key to our success is domain knowledge: I have a team of experts from the auto selling business.” That means taking out the intrusive ads and multi-click processes usually associated with selling cars online and giving customers a sense of fast, responsive action. If the car is on sale, the license number is covered with a Blinker logo, offering the seller a sense of privacy (and Blinker some free advertising.)

COPYRIGHT © 2017 HARVARD BUSINESS SCHOOL PUBLISHING ALL RIGHTS RESERVED. ©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTSCORPORATION. RESERVED.

665

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

Blinker, which hopes to go national over the next few years, does have AI specialists, who have trained a system with over 70,000 images of cars. Even these had the human touch — the results were verified on Amazon’s Mechanical Turk, a service where humans perform inexpensive tasks online. While the AI work goes on, Buscher spent over a year bringing in focus groups to see what worked, and then watched how buyers and sellers interacted (frequently, they did their sales away from Blinker, something else the company had to fix). “I’ve never been in tech, but I’m learning that on the go,” he said. “You still have to know what a good and bad customer experience is like.” No single tool, even one as powerful as AI, determines the fate of a business. As much as the world changes, deep truths — around unearthing customer knowledge, capturing scarce goods, and finding profitable adjacencies — will matter greatly. As ever, the technology works to the extent that its owners know what it can do, and know their market.

Quentin Hardy is Head of Editorial at Google Cloud.

COPYRIGHT © 2017 HARVARD BUSINESS SCHOOL PUBLISHING ALL RIGHTS RESERVED. ©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTSCORPORATION. RESERVED.

676

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

Discussion Questions

3 Ways Companies Are Building a Business Around AI BY QUENTIN HARDY

1. Had you previously considered “finding and owning valuable data no one else has” as a business model, the way CAMP3 has? Does your company have access to data that is unlike any other? 2. Could any of your company’s internal AI or data products be valuable to other companies? Could you sell or license them?

How should a company that isn’t involved in building AI think about using it?

COPYRIGHT © 2018 HARVARD BUSINESS SCHOOL PUBLISHING ALL RIGHTS RESERVED. ©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTSCORPORATION. RESERVED.

13 68

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

REPRINT H044R2 PUBLISHED ON HBR.ORG JANUARY 30, 2018

ARTICLE TECHNOLOGY

What Changes When AI Is So Accessible That Everyone Can Use It?

by H. James Wilson and Paul Daugherty

©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTS RESERVED.

69

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

TECHNOLOGY

What Changes When AI Is So Accessible That Everyone Can Use It?

by H. James Wilson and Paul Daugherty JANUARY 30, 2018

Bernard Van Berg/EyeEm/Getty Images Mazin Gilbert has an ambitious goal. As vice president of advanced technologies at AT&T, Gilbert wants to make AI technologies widely available throughout the corporation, especially to those who might not have a computer science background and may not even know how to program. Call it the “democratization of AI.” To accomplish that goal, AT&T is building a user-friendly platform with

COPYRIGHT © 2018 HARVARD BUSINESS SCHOOL PUBLISHING ALL RIGHTS RESERVED. ©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTSCORPORATION. RESERVED.

702

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

point-and-click tools that will enable employees — up to one-quarter of the company’s workforce — to build their own AI applications. AT&T and a host of other companies are trying to address a crucial issue in business: the severe shortage of AI talent. According to some estimates, only about 10,000 programmers in the world have the necessary expertise to develop advanced AI algorithms. But that’s barely a drop in the bucket for what companies will need in their future workforces. Tools like AT&T’s platform will help spread AI technologies well beyond just a limited number of “haves” and reach the “have nots” that may lack the technical knowledge and experience. This democratization of AI will happen in two ways. First, it will enable employees across a large organization like AT&T to develop their own AI applications to make them better at their jobs. But it will also allow smaller firms to deploy some of the same AI capabilities that have heretofore been limited to large corporations. Think of how spreadsheets like Lotus 1-2-3 and Excel helped democratize data analysis, enabling even mom-and-pop shops to perform invaluable “what-if” analyses.

Some Assembly Required AT&T’s in-house platform contains AI “widgets” that can be assembled together to create working applications. A marketer at AT&T might, for example, connect a widget for natural language processing together with other components to create an app for gathering and analyzing unstructured data from social media. In the future, AT&T says that it might begin offering the AI platform as a product to other companies. Somewhat similar tools are already on the market. Consider DataRobot Inc., a Boston-based startup that has developed an automated machine learning platform that enables users to build predictive models that deploy various AI techniques. The firm has more than 100 customers in insurance, banking, and other industries. The product might be deployed, for example, to analyze a huge customer data set to predict which mortgage applicants are most likely to default. Farmers Insurance, for one, is using the DataRobot platform to uncover insights about customer behavior and to improve the design of the company’s different products. Another similar vendor is Petuum, which offers a machine learning platform with a visual interface that enables people to build AI applications quickly without any coding. The company is now working on deploying that general platform to specific industries like manufacturing and health care. And at our company, Accenture, we’ve invested in developing Accenture Insights Platform, which can combine and simplify the tools from the major AI platforms. We’ve seen, firsthand, how democratization increases the capabilities and speed of our professionals using AI in developing business solutions.

AI in the Cloud Meanwhile, high-tech giants Google and Microsoft have been busy adding AI to their cloud services. Initially, the tools were for relatively rudimentary tasks like image classification and voice recognition, but over time, the company will likely increase the technical sophistication of its

COPYRIGHT © 2018 HARVARD BUSINESS SCHOOL PUBLISHING ALL RIGHTS RESERVED. ©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTSCORPORATION. RESERVED.

713

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

offerings. In Google’s AutoML project, the company is building a machine learning system that will be able to develop other machine learning applications. The goal, according to Jeff Dean and Fei-Fei Li, leading engineers at Google, is to open up the use of AI from thousands of companies to millions. For its part, Microsoft has released tools to help people build deep neural networks, which can be difficult to develop and train. “We are eliminating a lot of the heavy lifting,” says Joseph Sirosh, a vice president at Microsoft. Salesforce, a leader in sales automation, has a similar goal. The company offers myEinstein, a suite of tools that enables customers to build their own chatbots and predictive marketing models without having to do any coding. And even companies outside of the traditional high-tech industry are getting into the action. Uber, for one, is now offering Michelangelo, a platform that provides machine learning as a service. Included in the platform are the capabilities to manage data; to train, evaluate, and deploy AI predictive models; and to make and monitor predictions based on those models. According to the company, employees have been using Michelangelo in-house for more than a year now, with dozens of teams building and deploying models on the platform. One early success was Uber Eats, an application that predicts how long a takeout order will take, including the time needed to prepare the food (taking into account how busy a restaurant currently is as well as the complexity of the order) and the time required to deliver the meal (taking into account the route and traffic, among other factors). The company says it wants to make “scaling AI to meet the needs of business as easy as requesting a ride.” Uber’s ambitious goal notwithstanding, it will take considerable advances in the field before AI can be offered to companies as a utility, similar to databases and software testing platforms. But what’s clear is that the democratization of AI is under way, and the competitive advantage could soon be shifting from those companies with advanced in-house AI expertise to those firms with the most innovative worker ideas for utilizing that technology. Rather than displacing workers, AI is actually empowering nontechnical people to use AI to fill today’s growing shortage of technical talent.

H. James Wilson is a managing director of Information Technology and Business Research at Accenture Research. Follow him on Twitter @hjameswilson. Wilson is coauthor with Paul Daugherty of Human + Machine: Reimagining Work in the Age of AI (Harvard Business Review Press, March 2018).

Paul Daugherty is Accenture’s chief technology & innovation officer. Follow him on Twitter @pauldaugh. Daugherty is coauthor with H. James Wilson of Human + Machine: Reimagining Work in the Age of AI (Harvard Business Review Press, March 2018).

COPYRIGHT © 2018 HARVARD BUSINESS SCHOOL PUBLISHING ALL RIGHTS RESERVED. ©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTSCORPORATION. RESERVED.

724

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

Discussion Questions

What Changes When AI Is So Accessible That Everyone Can Use It? BY H. JAMES WILSON AND PAUL DAUGHERTY

1. Does the notion of the democratization of AI make you think differently about developing an AI capability now? How would your investments change if you could be certain that “scaling AI to meet the needs of business as easy as requesting a ride” was just a few years away? 2. What can your company do today to prepare for the coming democratization of AI? Should you be collecting data differently? Should your employees be asking different questions? Should your business model or strategy be different?

The competitive advantage could soon be shifting from those companies with advanced in-house AI expertise to those firms with the most innovative worker ideas for utilizing that technology.

COPYRIGHT © 2018 HARVARD BUSINESS SCHOOL PUBLISHING ALL RIGHTS RESERVED. ©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTSCORPORATION. RESERVED.

14 73

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

REPRINT H049PW PUBLISHED ON HBR.ORG APRIL 09, 2018

ARTICLE BOARDS

Research: Could Machine Learning Help Companies Select Better Board Directors?

by Isil Erel, Léa H. Stern, Chenhao Tan and Michael S. Weisbach

©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTS RESERVED.

74

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

BOARDS

Research: Could Machine Learning Help Companies Select Better Board Directors?

by Isil Erel, Léa H. Stern, Chenhao Tan and Michael S. Weisbach APRIL 09, 2018

Jens Magnusson/Getty Images Ever since Adam Smith published The Wealth of Nations in 1776, observers have bemoaned boards of directors as being ineffective as both monitors and advisors of management. Because a CEO often

COPYRIGHT © 2018 HARVARD BUSINESS SCHOOL PUBLISHING ALL RIGHTS RESERVED. ©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTSCORPORATION. RESERVED.

752

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

effectively controls the director selection process, he will tend to choose directors who are unlikely to oppose him, and who are unlikely to provide the diverse perspectives necessary to maximize firm value. Institutional investors often are critical of CEOs’ influence over boards and have made efforts to help companies improve their governance. Nonetheless, boards remain highly imperfect. Could technology help? Advances in machine learning have led to innovations ranging from facial recognition software to self-driving cars. These techniques are rapidly changing many industries — could they also improve corporate governance? To explore that question, we conducted a study of how machine learning might be used to select board directors, and how the selected directors might differ from those selected by management. Our intent is to demonstrate how a machine-learning model could potentially help investors by helping companies select better directors. The first challenge with such a study is determining what makes a director “better” or “worse.” Most directors’ actions occur in the privacy of the boardroom where they cannot be observed by outsiders. In addition, most of what directors do occurs within the structure of the board, so we cannot isolate their individual contributions. Despite those complications, one clear measure of director performance is publicly available: the fraction of votes a director receives in annual shareholder re-elections. Although the CEO often influences the choice of the person nominated to the board and shareholders have virtually no control over the choice of directors, shareholders vote annually on their re-election. These votes reflect the support the director personally has from the shareholders and should, in theory, incorporate all publicly available information about the director’s performance. Our choice of performance measure is also motivated by the fact that the hiring decision for a corporate director is no different than any other hiring decision: it is fundamentally about predicting the individual’s future performance. Since the mandate of the board is to represent shareholders’ interests, shareholder votes stand out as a natural performance metric. The second challenge we face is that we only have that measure of director performance for directors who are actually selected to join the board. Machine learning is all about prediction, but if we just try to predict how selected directors will fare in shareholder elections we are only looking at half of the problem. Ideally, we also want to predict how would-be directors who were not ultimately nominated would have done if they had the chance to join the board. We address this issue by constructing a pool of potential directors for each board opening from those who, around that time, accept a directorship at a smaller nearby company. We assume that these individuals would have been attracted to a directorship at a larger, neighboring company. For the purposes of our study, we use the fraction of votes these individuals received at the company where they become director as our measure of their potential performance.

COPYRIGHT © 2018 HARVARD BUSINESS SCHOOL PUBLISHING ALL RIGHTS RESERVED. ©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTSCORPORATION. RESERVED.

763

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

We trained a machine learning algorithm to predict directors’ performance, using a dataset of large publicly traded U.S. corporations between 2000 and 2011. We used a machine learning method called gradient boosting, and then evaluated the results using a separate test dataset of directors who joined firms between 2012 to 2014 who the algorithm did not observe during this “training period.” The algorithm was able to identify which directors were likely to be unpopular with shareholders. The directors that were actually hired but that our algorithm predicted would be unpopular with shareholders ended up faring much worse than other available candidates. In contrast, hired directors that our algorithm predicted would do well indeed did better than other available candidates. (Our machine learning model performed substantially better than a standard econometric model such as ordinary least squares.) The differences between the directors suggested by the algorithm and those actually selected by firms allow us to assess the features that are overrated in the director nomination process. We found that firms tend to choose directors who are much more likely to be male, have a large network, have a lot of board experience, currently serve on more boards, and have a finance background. In a sense, the algorithm is telling us exactly what institutional shareholders have been saying for a long time: that directors who are not old friends of management and come from different backgrounds both do a better job in monitoring management and are often overlooked. In light of our findings, it is worth asking: Why do real-world firms appoint directors who they could predict will be unpopular with shareholders? We think there are at least two possible reasons. First, it could be that CEOs do not want effective directors on their boards. Since the publication in 1932 of Adolph Berle and Gardiner Means’ The Modern Corporation and Private Property, economists have argued that managers are able to maintain control over their firms by influencing the director selection process to ensure management-friendly boards. Alternatively, it could be that because of behavioral biases, management is not able to select effective directors as well as an algorithm. In his book Thinking, Fast and Slow, Daniel Kahneman describes a long history of psychological research documenting that, in many circumstances, simple rules can lead to better outcomes than allowing individuals to have discretion over decisions. Machinelearning models, which are much more sophisticated than the rules suggested by psychologists in their experiments, represent a potentially valuable way to operationalize the notion that rules rather than discretion can improve real world decision making. How should our findings be applied in practice? The algorithms we present should be treated as “first pass” approaches; presumably more sophisticated models would predict director performance even better than the ones presented in this paper. In addition, our algorithms rely on publicly available data; if one had more detailed private data on director backgrounds, performance, etc., one could improve the algorithm’s accuracy even more. If algorithms such as these are used in practice in the future, as we suspect they will be, practitioners will undoubtedly have access to much better data

COPYRIGHT © 2018 HARVARD BUSINESS SCHOOL PUBLISHING ALL RIGHTS RESERVED. ©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTSCORPORATION. RESERVED.

774

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

than we have and should be able to predict director performance more accurately than we do in the paper. Machine learning algorithms are not without their flaws. They are prone to bias, too, depending on the data they are fed and the outcomes they are optimizing for. For the purpose of our study, though, it is clear that algorithms are not prone to agency conflicts and the biases that occur when boards and CEOs meet together to select new directors. Institutional investors are likely to find this attribute particularly appealing and to encourage boards to rely on an algorithm for director selections in the future. How well this approach to selecting directors will be received by management is an open question.

Isil Erel is the Fisher College of Business Distinguished Professor of Finance.

Léa H. Stern is an assistant professor of Finance and Business Economics at the Michael G. Foster School of Business at the University of Washington.

Chenhao Tan is an assistant professor at the Department of Computer Science and the Department of Information Science, by courtesy, at University of Colorado Boulder.

Michael S. Weisbach is the Ralph W. Kurtz Chair in Finance at Ohio State University, as well as a Research Associate of the National Bureau of Economic Research.

COPYRIGHT © 2018 HARVARD BUSINESS SCHOOL PUBLISHING ALL RIGHTS RESERVED. ©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTSCORPORATION. RESERVED.

785

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

Discussion Questions

Research: Could Machine Learning Help Companies Select Better Board Directors? BY ISIL EREL, LÉA H. STERN, CHENHAO TAN, AND MICHAEL S. WEISBACH

1. Are you surprised that machine learning can provide insights on a process as complex as appointing board directors? Are you intrigued, or are you skeptical about the methodology? 2. Having seen the results of this ambitious study, are there complex decisions that you think a similar machine learning approach can facilitate? 3. The authors admit that the machine learning algorithm they used in this study might be prone to bias, based on the training data it is given. How do you think this “unknown” bias compares to the “well-known” biases (such as more directors being male, having a large network, or having lots of board experience) that have an adverse effect on board selection?

In a sense, the algorithm is telling us exactly what institutional shareholders have been saying for a long time: that directors who are not old friends of management and come from different backgrounds both do a better job in monitoring management and are often overlooked.

COPYRIGHT © 2018 HARVARD BUSINESS SCHOOL PUBLISHING ALL RIGHTS RESERVED. ©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTSCORPORATION. RESERVED.

15 79

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

REPRINT H03NFD PUBLISHED ON HBR.ORG MAY 30, 2017

ARTICLE ANALYTICS

How Harley-Davidson Used Artificial Intelligence to Increase New York Sales Leads by 2,930%

by Brad Power

©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTS RESERVED.

80

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

ANALYTICS

How Harley-Davidson Used Artificial Intelligence to Increase New York Sales Leads by 2,930% by Brad Power MAY 30, 2017

It was winter in New York City and Asaf Jacobi’s Harley-Davidson dealership was selling one or two motorcycles a week. It wasn’t enough.

COPYRIGHT © 2017 HARVARD BUSINESS SCHOOL PUBLISHING ALL RIGHTS RESERVED. ©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTSCORPORATION. RESERVED.

812

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

Jacobi went for a long walk in Riverside Park and happened to bump into Or Shani, CEO of an AI firm, Adgorithms. After discussing Jacobi’s sales woes, Shani, suggested he try out Albert, Adgorithm’s AIdriven marketing platform. It works across digital channels, like Facebook and Google, to measure, and then autonomously optimize, the outcomes of marketing campaigns. Jacobi decided he’d give Albert a one-weekend audition. That weekend Jacobi sold 15 motorcycles. It was almost twice his all-time summer weekend sales record of eight. Naturally, Jacobi kept using Albert. His dealership went from getting one qualified lead per day to 40. In the first month, 15% of those new leads were “lookalikes,” meaning that the people calling the dealership to set up a visit resembled previous high-value customers and therefore were more likely to make a purchase. By the third month, the dealership’s leads had increased 2930%, 50% of them lookalikes, leaving Jacobi scrambling to set up a new call center with six new employees to handle all the new business. While Jacobi had estimated that only 2% of New York City’s population were potential buyers, Albert revealed that his target market was larger – much larger – and began finding customers Jacobi didn’t even know existed. How did it do that?

AI at Work Today, Amazon, Facebook, and Google are leading the AI revolution, and that’s given them a huge market advantage over most consumer goods companies and retailers by enabling them to lure customers with highly personalized, targeted advertising, and marketing. However, companies such as Salesforce, IBM, and a host of startups are now beginning to offer AI marketing tools that have become both easier to use (that is, they don’t require hiring expensive data scientists to figure out how to operate the tool and analyze its outputs) and less expensive to acquire, with software-as-aservice (SaaS), pay-as-you-go pricing. And instead of optimizing specific marketing tasks, or working within individual marketing channels, these new tools can handle the entire process across all channels. In the case of Harley-Davidson, the AI tool, Albert, drove in-store traffic by generating leads, defined as customers who express interest in speaking to a salesperson by filling out a form on the dealership’s website. Armed with creative content (headlines and visuals) provided by Harley-Davidson, and key performance targets, Albert began by analyzing existing customer data from Jacobi’s customer relationship management (CRM) system to isolate defining characteristics and behaviors of highvalue past customers: those who either had completed a purchase, added an item to an online cart, viewed website content, or were among the top 25% in terms of time spent on the website. COPYRIGHT © 2017 HARVARD BUSINESS SCHOOL PUBLISHING ALL RIGHTS RESERVED. ©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTSCORPORATION. RESERVED.

823

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

Using this information, Albert identified lookalikes who resembled these past customers and created micro segments – small sample groups with whom Albert could run test campaigns before extending its efforts more widely. It used the data gathered through these tests to predict which possible headlines and visual combinations – and thousands of other campaign variables – would most likely convert different audience segments through various digital channels (social media, search, display, and email or SMS). Once it determined what was working and what wasn’t, Albert scaled the campaigns, autonomously allocating resources from channel to channel, making content recommendations, and so on. For example, when it discovered that ads with the word “call” – such as, “Don’t miss out on a preowned Harley with a great price! Call now!” – performed 447% better than ads containing the word “Buy,” such as, “Buy a pre-owned Harley from our store now!” Albert immediately changed “buy” to “call” in all ads across all relevant channels. The results spoke for themselves.

The AI Advantage For Harley-Davidson, AI evaluated what was working across digital channels and what wasn’t, and used what it learned to create more opportunities for conversion. In other words, the system allocated resources only to what had been proven to work, thereby increasing digital marketing ROI. Eliminating guesswork, gathering and analyzing enormous volumes of data, and optimally leveraging the resulting insights is the AI advantage. Marketers have traditionally used buyer personas – broad behavior-based customer profiles – as guides to find new ones. These personas are created partly out of historic data, and partly by guesswork, gut feel, and the marketers’ experiences. Companies that design their marketing campaigns around personas tend to use similarly blunt tools (such as gross sales) – and more guesswork – to assess what’s worked and what hasn’t. AI systems don’t need to create personas; they find real customers in the wild by determining what actual online behaviors have the highest probability of resulting in conversions, and then finding potential buyers online who exhibit these behaviors. To determine what worked, AI looks only at performance: Did this specific action increase conversions? Did this keyword generate sales? Did this spend increase ROI?

VIDEO A.I. COULD LIBERATE 50% OF MANAGERS' TIME

TO VIEW, PLEASE VISIT THIS ARTICLE AT HBR.ORG

Even if equipped with digital tools and other marketing technologies, humans can only manage a few hundred keywords at a time, and struggle to apply insights across channels with any precision. Conversely, an AI tool can process millions of interactions a minute, manage hundreds of thousands of keywords, and run tests in silica on thousands of messages and creative variations to predict optimal outcomes.

COPYRIGHT © 2017 HARVARD BUSINESS SCHOOL PUBLISHING ALL RIGHTS RESERVED. ©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTSCORPORATION. RESERVED.

834

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

And AI doesn’t need to sleep, so it can do all this around the clock. Consequently, AI can determine exactly how much a business should spend, and where, to produce the best results. Rather than base media buying decisions on past performance and gut instincts, AI acts instantly and autonomously, modifying its buying strategy in real-time based the on everchanging performance parameters of each campaign variable.

Taking the AI Plunge Because AI is new, and because marketers will be wary of relinquishing control and trusting a black box to make the best decisions about what people will or won’t do, it’s wise to adopt AI tools and systems incrementally, as did Harley-Davidson’s Jacobi. The best way to discover AI’s potential is to run some small, quick, reversible experiments, perhaps within a single geographic territory, brand, or channel. Within these experiments, it’s important to define key desired performance results; for example, new customers, leads, or an increased return on advertising spending. When it comes to choosing a tool, know what you want. Some tools focus on a single channel or task, such as optimizing the website content shown to each customer. Others, like IBM’s Watson, offer more general purpose AI tools that need to be customized for specific uses and companies. And still other AI tools produce insights but don’t act on them autonomously. It’s worth taking the plunge, and, in fact, there’s an early adopter advantage. As Harley’s Jacobi told me, “The system is getting better all the time. The algorithms will continue to be refined. Last year, we tripled our business over the previous year.” That’s good news for Jacobi and his employees, and not such good news for his competitors.

Brad Power is a consultant who helps organizations that must make faster changes to their products, services, and systems in order to compete with start-ups and leading software companies.

COPYRIGHT © 2017 HARVARD BUSINESS SCHOOL PUBLISHING ALL RIGHTS RESERVED. ©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTSCORPORATION. RESERVED.

845

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

Discussion Questions

How Harley-Davidson Used Artificial Intelligence to Increase New York Sales Leads by 2,390% BY BRAD POWER

1. What small, quick, reversible experiments could your company perform to test the marketing capability of autonomous AI? Which of your product lines or services would lend themselves to a campaign like this one? 2. One of the reasons Jacobi and Shani were able to measure the success of this campaign was because the “leads generated” is specific and measurable. Which performance metrics do you think would work best for machine learning initiatives for your company and why?

An AI tool can process millions of interactions a minute, manage hundreds of thousands of keywords, and run tests in silica on thousands of messages and creative variations to predict optimal outcomes

COPYRIGHT © 2018 HARVARD BUSINESS SCHOOL PUBLISHING ALL RIGHTS RESERVED. ©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTSCORPORATION. RESERVED.

16 85

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

REPRINT H043SJ PUBLISHED ON HBR.ORG JANUARY 19, 2018

ARTICLE ANALYTICS

Machine Learning Can Help B2B Firms Learn More About Their Customers

by Stephan Kudyba and Thomas H. Davenport

©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTS RESERVED.

86

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

ANALYTICS

Machine Learning Can Help B2B Firms Learn More About Their Customers by Stephan Kudyba and Thomas H. Davenport JANUARY 19, 2018

vincent tsui for hbr Much of the strategic focus in the digital economy thus far has revolved around getting better insights into consumers. B2C firms have been the leaders in customer analytics initiatives. E-

COPYRIGHT © 2018 HARVARD BUSINESS SCHOOL PUBLISHING ALL RIGHTS RESERVED. ©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTSCORPORATION. RESERVED.

872

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

commerce, mobile commerce, and social media platforms have enabled businesses to better sculpt marketing and customer support initiatives and customer services. Extensive data and advanced analytics for B2C have enabled strategists to better understand consumer behavior and corresponding propensities as visitors and purchasers conduct daily activities through online systems. But there is also an emerging capability to gain insights on business customers. B2B, or the process of marketing and selling product and service offerings to business customers, is experiencing an intensified focus with the increased availability of new digital data that describes businesses. Traditional B2B insight activities have involved such limited data as size of companies as measured by revenue, capitalization or employees, and industry type as formally classified by SIC codes. The internet offers a much more detailed level of data, going well beyond standard industry categorization. Web content that provides robust, detailed descriptions of companies provides valuable descriptive information. However, these digital resources yield little value unless individual customers are identified and their detailed backgrounds and interests are analyzed to provide strategic insights for suppliers. And that’s where AI techniques provide can help. Neural networks and “deep learning” algorithms, along with other machine learning methods, enable data scientists to mine the gold in digital formats. These AI-based methods involve advanced search techniques that identify, categorize, and gather user-defined data elements corresponding to search criteria. For example, considerable business description information exists on LinkedIn. But how can organizations analyze each profile on the network? Well-designed AI-based algorithms are the key to extracting information from LinkedIn. These more structured data resources then provide the means for yet another application of AI-based algorithms, where the focus is on identifying patterns in data that ultimately provide the basis for predictive sales and marketing models. These can be used for scoring, forecasting, and classification capabilities. By helping B2B companies gather better data on their customers, AI will help them catch up with their B2C peers. One company focusing on AI-based analytics for B2B applications has adopted a unique way of leveraging the extensive digital footprints that provide descriptive attributes of all types of firms. Its approach to leveraging data assets combines the art and science of producing analytic solutions. EverString Technology considers the diverse sectors of the web that contain descriptive information of businesses (for example, site domains and employee digital footprints) and incorporates input from expert practitioners in the B2B space to help further describe individual businesses. EverString deploys machine learning to identify, extract, and model a categorization scheme of companies so that users in the B2B space can more accurately identify opportunities. B2B companies need to know, for example, how many companies exist in a given market space. How can they identify and access all those firms that fall into the market pertaining to their product or service? And which specific buyers should they target in those firms? By creating a microcategorization scheme and applying guided AI to various sectors of the web, EverString can produce COPYRIGHT © 2018 HARVARD BUSINESS SCHOOL PUBLISHING ALL RIGHTS RESERVED. ©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTSCORPORATION. RESERVED.

883

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

thousands of customer insights in a short period for its B2B customers. The company has created an intelligent system to augment customer data in the B2B space. One B2B company that utilizes EverString’s platform is Autodesk, a multinational software company that provides software for the architectural, engineering, construction, manufacturing, media, and entertainment industries. A major focus in Autodesk’s approaches to B2B sales over the past several years has been on using more data for account selection and understanding. But in large designoriented companies, it is often difficult to understand which individuals might have an interest in computer-aided design software. Prior to working with EverString, Autodesk relied on field experience and customer buying histories. Now it relies increasingly on predictive analytics from EverString to identify likely customers. One key tool is the Enterprise Business Agreement Propensity Model, which suggests which executives in a large customer organization are most likely to engage in an enterprise-level agreement with Autodesk. The company also maintains an overall account potential model that makes use of EverString data and predictions. The primary users of the data and models are, of course, the Autodesk sales force. They are given ranked recommendations and the raw scores created by the EverString models. The Global Sales Strategy organization within Autodesk manages the process and tries to ensure that the data and models check out. It is early days for the use of these capabilities at Autodesk, but thus far both the sales teams and the Global Sales Strategy group feel that the EverString offerings are very helpful to the sales process. As Matthew Stevens, Autodesk’s sales insights manager within Global Sales Strategy, told us: EverString provides key inputs on analytics, which we convert into potential sales opportunities. It’s early to judge the exact payoff, but it’s difficult to imagine making a recommendation without these insights. We are challenged to respond to all the questions about accounts and scores, but at least we have data to support our recommendations now. Stevens also noted that there are many more activities to pursue in the future with this data-driven approach to sales: Finding data on European and Asian companies is challenging due to privacy regulations and language differences. We’re working with EverString to understand these opportunities better. Currently our EverString analytics and data are not connected with Salesforce, our CRM system. But we are at the first stage of a multistage journey to understand analytics and insights in sales. We are definitely moving in the right direction. New tools from organizations like EverString are enabling B2B-oriented firms like Autodesk to develop much-more-data-driven approaches to sales and marketing. The amount and quality of data

COPYRIGHT © 2018 HARVARD BUSINESS SCHOOL PUBLISHING ALL RIGHTS RESERVED. ©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTSCORPORATION. RESERVED.

894

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

on businesses may not yet approach that for consumers, but there is considerable progress being made in achieving parity.

Stephan Kudyba is associate professor of analytics and information systems at the Martin Tuchman School of Business, New Jersey Institute of Technology.

Thomas H. Davenport is the President’s Distinguished Professor in Management and Information Technology at Babson College, a research fellow at the MIT Initiative on the Digital Economy, and a senior adviser at Deloitte Analytics. Author of over a dozen management books, his latest is Only Humans Need Apply: Winners and Losers in the Age of Smart Machines.

COPYRIGHT © 2018 HARVARD BUSINESS SCHOOL PUBLISHING ALL RIGHTS RESERVED. ©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTSCORPORATION. RESERVED.

905

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

Discussion Questions

Machine Learning Can Help B2B Firms Learn More About Their Customers BY STEPHAN KUDYBA AND THOMAS H. DAVENPORT

1. Had you considered how AI-powered customer analytics and research tools have, until now, skewed heavily toward B2C firms rather than B2B firms? 2. Beyond LinkedIn, what other kinds of public-internet-based B2B research does your company engage in? After reading this article, do you think these processes are candidates for machine-learning techniques? What would be the implications of transitioning these processes to machine learning?

B2B companies need to know, for example, how many companies exist in a given market space. How can they identify and access all those firms that fall into the market pertaining to their product or service? And which specific buyers should they target in those firms? By creating a micro-categorization scheme and applying guided AI to various sectors of the web, EverString can produce thousands of customer insights in a short period for its B2B customers.

COPYRIGHT © 2018 HARVARD BUSINESS SCHOOL PUBLISHING ALL RIGHTS RESERVED. ©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTSCORPORATION. RESERVED.

17 91

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

REPRINT H03TAH PUBLISHED ON HBR.ORG JULY 28, 2017

ARTICLE PRODUCT DEVELOPMENT

How We Built a Virtual Scheduling Assistant at Microsoft

by Andrés Monroy-Hernández and Justin Cranshaw

©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTS RESERVED.

92

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

PRODUCT DEVELOPMENT

How We Built a Virtual Scheduling Assistant at Microsoft

by Andrés Monroy-Hernández and Justin Cranshaw JULY 28, 2017

Many people would agree that scheduling meetings is tedious. Perhaps you have experienced an email chain like this: Jenn, a potential client: Hey! What day/time works for a quick call next week? You: (toggling between calendar app and email) I’m wide open Monday. Jenn: (several hours later) Sorry. Traveling that day. How about Wednesday at 10 AM?

COPYRIGHT © 2017 HARVARD BUSINESS SCHOOL PUBLISHING ALL RIGHTS RESERVED. ©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTSCORPORATION. RESERVED.

932

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

You: (checking your calendar app again) That should work. Your office? Jenn: My office is great. Maybe we should see if Emad can join? This back-and-forth can carry on, and it can get even more challenging when people use different calendaring systems or meet across different time zones. Not only are these exchanges timeconsuming, they also obliterate our ability to focus on more demanding tasks. An informal survey confirmed our suspicion that others felt similarly. We asked about 100 information workers, in a wide range of industries and roles, to identify tiresome tasks they regularly do that are not part of their primary work duties. The most onerous task people citied was scheduling meetings. While online calendar sharing tools like Outlook and Google calendar and polling tools like Doodle make scheduling less cumbersome, we still have to stop what we’re doing to use them while we switch from the task at hand to fiddle with our scheduling tool of choice. Things get even more complicated when the people we want to meet with use different tools, since many don’t work well together. With all the progress happening in artificial intelligence (AI), we wondered if we could create a virtual assistant that could handle the conversational back-and-forth required for scheduling meetings, much the same way that executive admins schedule meetings for CEOs. There is a long history of AI research around how to build digital personal assistants, but none of the early work on AI scheduling has taken off. There are several reasons for this. First, in the workplace, business users have very little tolerance for mistakes. If the AI assistant is not a model digital employee, people will quickly lose patience and stop using it. The current state of AI is not yet ready to guarantee such a high-performing assistant. Second, there is the chicken-and-egg problem: Good AI needs a lot of data, and to get a lot of data you need real usage, but that’s hard to achieve without a reliable system. Third, scheduling scenarios can be complex — there are one-on-one versus manyperson meetings, in-person versus remote meetings, meetings that are postponed and need to be rescheduled. People also use their calendars differently: Some use appointments as tasks, and others block out time when they are free. And people also have unique preferences over time that are hard to capture: Some prefer clustered meetings, while others like them spread out. Finally, there are subtle social considerations involved with scheduling, like the relative status between people or the urgency of a meeting. Our virtual assistant solution would need to solve for all three of these issues, as a human assistant would. But where to start? We took a step back and considered long-standing rapid-prototyping approaches in design. These involve building and testing lo-fi prototypes before gradually iterating on higher-fidelity — and more expensive — designs. For example, a designer might initially show paper prototypes to a group of users to rapidly collect user feedback. Then they might build some wire-frame mock-ups to test them with users in a slightly more realistic setting, mirroring the types

COPYRIGHT © 2017 HARVARD BUSINESS SCHOOL PUBLISHING ALL RIGHTS RESERVED. ©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTSCORPORATION. RESERVED.

943

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

of interactions users see in the end product. Finally, the designers move to a “Wizard of Oz” prototype, where users will experience an interface that looks and feels real but behind the curtain is a human researcher pulling the strings and controlling the interface. We decided to go with the Wizard of Oz approach, but took it one step further. Here is how it worked. We invited a handful people from a few companies to sign up for our system, which we later named Calendar.help. Then they simply added the virtual assistant to the Cc line when sending meeting invitation emails. The longer-term goal was that the AI-driven virtual assistant would take it from there, looking at schedules and creating calendar invites for optimal days and times. But at the outset, the virtual assistant was actually us parsing every single invitation email, looking for optimal solutions, and scheduling meetings. Although this was a lot of work, with nearly two years of iterations, it allowed us to get a product into people’s hands early so we could observe their behavior and iterate quickly. It also gave us a deeper understanding of the problem we were trying to solve — and let us start evaluating which portions of the work could eventually be performed through AI. We were delivering accuracy and collecting excellent data that could potentially bootstrap an AI solution down the road. This approach showed us what people would expect from a virtual scheduling assistant. It allowed us to create workflows that could be broken into narrower microtasks, such as extracting the location of a meeting or determining whether the meeting should take place face-to-face or on the phone. With that done, we continued with our humans-in-the-loop model and hired a staff of workers to perform those microtasks. They formed a sort of digital assembly line, with one person surveying people’s calendars to suggest optimal times, another looking at available locations, and another working to reschedule the meeting if needed. The advantage of creating these microtasks was three-fold. First, it focused the worker’s attention on one thing at a time, reducing errors and making the workflow systematic enough that task workers could come in and out of the system. Second, because the microtasks were well-designed, they helped us collect high-quality data that we would later use to automate the process. Third, a system with this level of granularity helped us create a variety of machine learning models to understand natural language, so specific microtasks could be executed automatically. For example, we used people’s responses to time options for the meeting to help us understand how they express their preferences. This enabled us to build and train a machine learning classifier to automatically perform this step in the future. A key takeaway is that the scheduling bot was fully functional from day one but became more efficient through use. A benefit of having humans in the loop early on was that they were able to understand people’s actual needs first, while collecting conversational data on how people interacted with the assistant. Each additional interaction provided data that helped us better understand which scenarios were most important to automate and what data was needed to do so. The team has since continued to improve on Calendar.help, which is now publicly available for people with Office 365 or Google calendars to preview. COPYRIGHT © 2017 HARVARD BUSINESS SCHOOL PUBLISHING ALL RIGHTS RESERVED. ©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTSCORPORATION. RESERVED.

954

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

We learned that keeping humans in the loop when building a virtual assistant does have its limitations and costs. For example, we observed that some invitees did not like the idea of working with a virtual assistant. Some people saw it as producing extra work for them because they still had to respond to the assistant’s emails; others were put off by the perception that the assistant was a bot. The social dynamics of involving bots in existing businesses interactions are evolving, but it’s definitely something to keep an eye on. Over time, if virtual assistants become more common, some of the friction might be reduced. We also learned that it’s important to be transparent about the human-in-the-loop architecture so that users can make product decisions that align with their privacy expectations. Having a virtual assistant with humans in the loop is not always the right option for everyone. For example, medical doctors have heightened responsibilities to their patients’ privacy, and may be selective about including a third party for scheduling. Transparency lets people decide what is right for them. Overall, we believe that creating and using systems like Calendar.help to manage routine tasks is an easy way for companies to leverage AI in their daily business practice. We didn’t want to constrain ourselves to what was currently possible with AI. Instead, we wanted to build something people need and want, and then use that product vision to determine how to make smart investments in automation and language understanding. Furthermore, the approach we took to build Calendar.help can be used to create in-house AI systems that, like ours, make use of off-the-shelf AI technologies. We were not AI experts going into this — and we’re still not. The AI tools are already out there. Our job was to figure out the process that would best take advantage of them.

Andrés Monroy-Hernández was a researcher at Microsoft Research while doing this work. He is now a lead research scientist at Snap Inc. as well as an affiliate faculty at the University of Washington.

Justin Cranshaw is a researcher at Microsoft Research, where he’s been focused on building intelligent systems to make people more productive and effective at work.

COPYRIGHT © 2017 HARVARD BUSINESS SCHOOL PUBLISHING ALL RIGHTS RESERVED. ©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTSCORPORATION. RESERVED.

965

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

Discussion Questions

How We Built a Virtual Scheduling Assistant at Microsoft BY ANDRÉS MONROY-HERNÁNDEZ AND JUSTIN CRANSHAW

1. Microsoft’s approach to creating a virtual scheduling assistant included breaking the process down to three distinct tasks (suggesting times, suggesting locations, rescheduling when necessary) that were trained with separate algorithms. Are the AI processes your company is developing or considering better described as microtasks such as these? Why or why not? How do they compare to the Microsoft example? 2. Does the idea of “keeping humans in the loop” make you think differently about AI training? Had you considered the possibility of human trainers doing complicated tasks like scheduling (rather than rote tasks like labeling pictures)? What kinds of more complex tasks in your business could use human trainers in such a way?

A benefit of having humans in the loop early on was that they were able to understand people’s actual needs first, while collecting conversational data on how people interacted with the assistant. Each additional interaction provided data that helped us better understand which scenarios were most important to automate and what data was needed to do so.

COPYRIGHT © 2018 HARVARD BUSINESS SCHOOL PUBLISHING ALL RIGHTS RESERVED. ©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTSCORPORATION. RESERVED.

18 97

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

REPRINT H03KXF PUBLISHED ON HBR.ORG APRIL 14, 2017

ARTICLE TECHNOLOGY

How Companies Are Already Using AI by Satya Ramaswamy

©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTS RESERVED.

98

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

TECHNOLOGY

How Companies Are Already Using AI

by Satya Ramaswamy APRIL 14, 2017

Every few months it seems another study warns that a big slice of the workforce is about to lose their jobs because of artificial intelligence. Four years ago, an Oxford University study predicted 47% of jobs could be automated by 2033. Even the near-term outlook has been quite negative: A 2016 report by the Organization for Economic Cooperation and Development (OECD) said 9% of jobs in the 21 countries that make up its membership could be automated. And in January 2017, McKinsey’s research arm estimated AI-driven job losses at 5%. My own firm released a survey recently of 835 large companies (with an average revenue of $20 billion) that predicts a net job loss of between 4% and 7% in key business functions by the year 2020 due to AI.

COPYRIGHT © 2017 HARVARD BUSINESS SCHOOL PUBLISHING ALL RIGHTS RESERVED. ©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTSCORPORATION. RESERVED.

992

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

Yet our research also found that, in the shorter term, these fears may be overblown. The companies we surveyed – in 13 manufacturing and service industries in North America, Europe, Asia-Pacific, and Latin America – are using AI much more frequently in computer-to-computer activities and much less often to automate human activities. “Machine-to-machine” transactions are the low-hanging fruit of AI, not people-displacement. For example, our survey, which asked managers of 13 functions, from sales and marketing to procurement and finance, to indicate whether their departments were using AI in 63 core areas, found AI was used most frequently in detecting and fending off computer security intrusions in the IT department. This task was mentioned by 44% of our respondents. Yet even in this case, we doubt AI is automating the jobs of IT security people out of existence. In fact, we find it’s helping such often severely overloaded IT professionals deal with geometrically increasing hacking attempts. AI is making IT security professionals more valuable to their employers, not less. In fact, although we saw examples of companies using AI in computer-to-computer transactions such as in recommendation engines that suggest what a customer should buy next or when conducting online securities trading and media buying, we saw that IT was one of the largest adopters of AI. And it wasn’t just to detect a hacker’s moves in the data center. IT was using AI to resolve employees’ tech support problems, automate the work of putting new systems or enhancements into production, and make sure employees used technology from approved vendors. Between 34% and 44% of global companies surveyed are using AI in in their IT departments in these four ways, monitoring huge volumes of machine-to-machine activities.

COPYRIGHT © 2017 HARVARD BUSINESS SCHOOL PUBLISHING ALL RIGHTS RESERVED. ©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTSCORPORATION. RESERVED.

1003

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

In stark contrast, very few of the companies we surveyed were using AI to eliminate jobs altogether. For example, only 2% are using artificial intelligence to monitor internal legal compliance, and only 3% to detect procurement fraud (e.g., bribes and kickbacks). What about the automation of the production line? Whether assembling automobiles or insurance policies, only 7% of manufacturing and service companies are using AI to automate production activities. Similarly, only 8% are using AI to allocate budgets across the company. Just 6% are using AI in pricing.

Where to Find the Low-Hanging Fruit So where should your company look to find such low-hanging fruit – applications of AI that won’t kill jobs yet could bestow big benefits? From our survey and best-practice research on companies that have already generated significant returns on their AI investments, we identified three patterns that separate the best from the rest when it comes to AI. All three are about using AI first to improve computer-to-computer (or machine-to-machine) activities before using it to eliminate jobs: Put AI to work on activities that have an immediate impact on revenue and cost. When Joseph Sirosh joined Amazon.com in 2004, he began seeing the value of AI to reduce fraud, bad debt, and the number of customers who didn’t get their goods and suppliers who didn’t get their money. By the time he left Amazon in 2013, his group had grown from 35 to more than 1,000 people who used

COPYRIGHT © 2017 HARVARD BUSINESS SCHOOL PUBLISHING ALL RIGHTS RESERVED. ©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTSCORPORATION. RESERVED.

1014

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

machine learning to make Amazon more operationally efficient and effective. Over the same time period, the company saw a 10-fold increase in revenue. After joining Microsoft Corporation in 2013 as corporate vice president of the Data Group, Sirosh led the charge in using AI in the company’s database, big data, and machine learning offerings. AI wasn’t new at Microsoft. For example, the company had brought in a data scientist in 2008 to develop machine learning tools that would improve its search engine, Bing, in a market dominated by Google. Since then, AI has helped Bing more than double its share of the search engine market (to 20%); as of 2015, Bing generated more than a $1 billion in revenue every quarter. (That was the year Bing became a profitable business for Microsoft.) Microsoft’s use of AI now extends far beyond that, including to its Azure cloud computing service, which puts the company’s AI tools in the hands of Azure customers. (Disclosure: Microsoft is a TCS client.)

VIDEO A.I. COULD LIBERATE 50% OF MANAGERS' TIME

TO VIEW, PLEASE VISIT THIS ARTICLE AT HBR.ORG

Look for opportunities in which AI could help you produce more products with the same number of people you have today. The AI experience of the 170-year-old news service Associated Press is a great case in point. AP found in 2013 a literally insatiable demand for quarterly earnings stories, but their staff of 65 business reporters could write only 6% of the earnings stories possible, given America’s 5,300 publicly held companies. The earnings news of many small companies thus went unreported on AP’s wire services (other than the automatically published tabular data). So that year, AP began working with an AI firm to train software to automatically write short earnings news stories. By 2015, AP’s AI system was writing 3,700 quarterly earnings stories – 12 times the number written by its business reporters. This is a machine-to-machine application of AI. The AI software is one machine; the other is the digital data feed that AP gets from a financial information provider (Zacks Investment Research). No AP business journalist lost a job. In fact, AI has freed up the staff to write more indepth stories on business trends. Start in the back office, not the front office. You might think companies will get the greatest returns on AI in business functions that touch customers every day (like marketing, sales, and service) or by embedding it in the products they sell to customers (e.g., the self-driving car, the self-cleaning barbeque grill, the self-replenishing refrigerator, etc.). Our research says otherwise. We asked survey participants to estimate their returns on AI in revenue and cost improvements, and then we compared the survey answers of the companies with the greatest improvements (call them “AI leaders”) to the answers of companies with the smallest improvements (“AI followers”). Some 51% of our AI leaders predicted that by 2020 AI will have its biggest internal impact on their back-office functions of IT and finance/accounting; only 34% of AI followers said the same thing. Conversely, 43% of AI followers said AI’s impact would be greatest in the front-office areas of marketing, sales, and services, yet only 26% of the AI leaders felt it would be there. We believe the leaders have the right idea: Focus your AI initiatives in the back-office, particularly where there are lots of computerto-computer interactions in IT and finance/accounting.

COPYRIGHT © 2017 HARVARD BUSINESS SCHOOL PUBLISHING ALL RIGHTS RESERVED. ©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTSCORPORATION. RESERVED.

1025

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

Computers today are far better at managing other computers and, in general, inanimate objects or digital information than they are at managing human interactions. When companies use AI in this sphere, they don’t have to eliminate jobs. Yet the job-destroying applications of AI are what command the headlines: driverless cars and trucks, robotic restaurant order-takers and food preparers, and more. Make no mistake: Automation and artificial intelligence will eliminate some jobs. Chatbots for customer service have proliferated; robots on the factory floor are real. But we believe companies would be wise to use AI first where their computers already interact. There’s plenty of low-hanging fruit there to keep them busy for years.

Satya Ramaswamy is vice president and global head of Tata Consultancy Services’ Digital Enterprise group.

COPYRIGHT © 2017 HARVARD BUSINESS SCHOOL PUBLISHING ALL RIGHTS RESERVED. ©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTSCORPORATION. RESERVED.

1036

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

Discussion Questions

How Companies Are Already Using AI BY SATYA RAMASWAMY

1. Do the results of the authors’ survey about “How Companies Around the World Are Using Artificial Intelligence” surprise you, given how AI is typically portrayed in the media? Why? What preconceptions did you have? 2. Has your company considered expanding back-office AI or it is only thinking about frontoffice and customer-facing AI? What rationale, if any, went into making that decision? 3. Do you agree with the author’s prescription that companies should start back-office AI initiatives that will keep companies “busy for years”?

The companies we surveyed . . . are using AI much more frequently in computerto-computer activities and much less often to automate human activities. “Machine-to-machine” transactions are the low-hanging fruit of AI, not peopledisplacement.

COPYRIGHT © 2018 HARVARD BUSINESS SCHOOL PUBLISHING ALL RIGHTS RESERVED. ©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTSCORPORATION. RESERVED.

19 104

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

THE NEAR FUTURE: CHANGES TO JOBS, THE ECONOMY, AND SOCIETY

©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTS RESERVED.

105

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

REPRINT R1804J PUBLISHED IN HBR JULY–AUGUST 2018

ARTICLE TECHNOLOGY

Collaborative Intelligence: Humans and AI Are Joining Forces by H. James Wilson and Paul R. Daugherty

This©2018 document is authorized for use only by Jo Teichmann Copying or posting is an infringement of copyright. Please contact [email protected] HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTS([email protected]). RESERVED. 106 or 800-988-0886 for additional copies.

FEATURE COLLABORATIVE INTELLIGENCE: HUMANS AND AI ARE JOINING FORCES

ILLUSTRATION BY KUO CHENG LIAO ©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTS RESERVED. 107 This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

COLLABORATIVE INTELLIGENCE:

Humans and AI Are Joining Forces BY H. JAMES WILSON AND PAUL R. DAUGHERTY

©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTS RESERVED. 108 This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

FEATURE COLLABORATIVE INTELLIGENCE: HUMANS AND AI ARE JOINING FORCES

IN BRIEF THE OUTLOOK

Artificial intelligence is transforming business—and having the most significant impact when it augments human workers instead of replacing them.

THE DETAILS

Companies see the biggest performance gains when humans and smart machines collaborate. People are needed to train machines, explain their outputs, and ensure their responsible use. AI, in turn, can enhance humans’ cognitive skills and creativity, free workers from low-level tasks, and extend their physical capabilities.

THE PRESCRIPTION

Companies should reimagine their business processes, focusing on using AI to achieve more operational flexibility or speed, greater scale, better decision making, or increased personalization of products and services.

ARTIFICIAL INTELLIGENCE IS BECOMING GOOD AT MANY “HUMAN” JOBS—DIAGNOSING DISEASE, TRANSLATING LANGUAGES, PROVIDING CUSTOMER SERVICE—AND IT’S IMPROVING FAST. THIS IS RAISING REASONABLE FEARS THAT AI WILL ULTIMATELY REPLACE HUMAN WORKERS THROUGHOUT THE ECONOMY. BUT THAT’S NOT THE INEVITABLE, OR EVEN MOST LIKELY, OUTCOME. NEVER BEFORE HAVE DIGITAL TOOLS BEEN SO RESPONSIVE TO US, NOR WE TO OUR TOOLS. WHILE AI WILL RADICALLY ALTER HOW WORK GETS DONE AND WHO DOES IT, THE TECHNOLOGY’S LARGER IMPACT WILL BE IN COMPLEMENTING AND AUGMENTING HUMAN CAPABILITIES, NOT REPLACING THEM.

©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTS RESERVED. 109 This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

THE VALUE OF COLLABORATION Companies benefit from optimizing collaboration between

humans and artificial intelligence. Five principles can help them do so: Reimagine business processes; embrace experimentation/employee involvement; actively direct AI strategy; responsibly collect data; and redesign work to incorporate AI and cultivate related employee skills. A survey of 1,075 companies in 12 industries found that the more of these principles companies adopted, the better their AI initiatives performed in terms of speed, cost savings, revenues, or other operational measures.  7x PERFORMANCE IMPROVEMENT

Certainly, many companies have used AI to automate processes, but those that deploy it mainly to displace employees will see only short-term productivity gains. In our research involving 1,500 companies, we found that firms achieve the most significant performance improvements when humans and machines work together (see the exhibit “The Value of Collaboration”). Through such collaborative intelligence, humans and AI actively enhance each other’s complementary strengths: the leadership, teamwork, creativity, and social skills of the former, and the speed, scalability, and quantitative capabilities of the latter. What comes naturally to people (making a joke, for example) can be tricky for machines, and what’s straightforward for machines (analyzing gigabytes of data) remains virtually impossible for humans. Business requires both kinds of capabilities. To take full advantage of this collaboration, companies must understand how humans can most effectively augment machines, how machines can enhance what humans do best, and how to redesign business processes to support the partnership. Through our research and work in the field, we have developed guidelines to help companies achieve this and put the power of collaborative intelligence to work.

6x 5x 4x 3x 2x 1x 0x 0

1

2

3

4

5

NUMBER OF HUMAN-MACHINE COLLABORATION PRINCIPLES ADOPTED (0 INDICATES THE ADOPTION OF ONLY BASIC, NONCOLLABORATIVE AI)

HUMANS ASSISTING MACHINES

Humans need to perform three crucial roles. They must train machines to perform certain tasks; explain the outcomes of those tasks, especially when the results are counterintuitive or controversial; and sustain the responsible use of machines (by, for example, preventing robots from harming humans). Training. Machine-learning algorithms must be taught how to perform the work they’re designed to do. In that effort, huge training data sets are amassed to teach machine-translation apps to handle idiomatic expressions, medical apps to detect disease, and recommendation engines to support financial decision making. In addition, AI systems must be trained how best to interact with humans. While organizations across sectors are now in the early stages of filling trainer roles, leading tech companies and research groups already have mature training staffs and expertise. Consider Microsoft’s AI assistant, Cortana. The bot required extensive training to develop just the right personality: confident, caring, and helpful but not bossy. Instilling those qualities took countless hours of attention by a team that included a poet, a novelist, and a playwright. Similarly, human trainers were needed to develop the personalities of Apple’s Siri and Amazon’s Alexa to ensure that they accurately reflected their companies’ brands. Siri, for example, has just a touch of sassiness, as consumers might expect from Apple.

AI assistants are now being trained to display even more complex and subtle human traits, such as sympathy. The start-up Koko, an offshoot of the MIT Media Lab, has developed technology that can help AI assistants seem to commiserate. For instance, if a user is having a bad day, the Koko system doesn’t reply with a canned response such as “I’m sorry to hear that.” Instead it may ask for more information and then offer advice to help the person see his issues in a different light. If he were feeling stressed, for instance, Koko might recommend thinking of that tension as a positive emotion that could be channeled into action. Explaining. As AIs increasingly reach conclusions through processes that are opaque (the so-called black-box problem), they require human experts in the field to explain their behavior to nonexpert users. These “explainers” are particularly important in evidence-­based industries, such as law and medicine, where a practitioner needs to understand how an AI weighed inputs into, say, a sentencing or medical recommendation. Explainers are similarly important in helping insurers and law enforcement understand why an autonomous car took actions that led to an accident—or failed to avoid one. And explainers are becoming integral in regulated industries—indeed, in any consumer-facing industry where a machine’s

©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTS RESERVED. 110 This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

FEATURE COLLABORATIVE INTELLIGENCE: HUMANS AND AI ARE JOINING FORCES

output could be challenged as unfair, illegal, or just plain wrong. For instance, the European Union’s new General Data Protection Regulation (GDPR) gives consumers the right to receive an explanation for any algorithm-based decision, such as the rate offer on a credit card or mortgage. This is one area where AI will contribute to increased employment: Experts estimate that companies will have to create about 75,000 new jobs to administer the GDPR requirements. Sustaining. In addition to having people who can explain AI outcomes, companies need “sustainers”—employees who continually work to ensure that AI systems are functioning properly, safely, and responsibly. For example, an array of experts sometimes referred to as safety engineers focus on anticipating and trying to prevent harm by AIs. The developers of industrial robots that work alongside people have paid careful attention to ensuring that they recognize humans nearby and don’t endanger them. These experts

Human-machine collaboration enables companies to interact with employees and customers in novel, more effective ways. may also review analysis from explainers when AIs do cause harm, as when a self-driving car is involved in a fatal accident. Other groups of sustainers make sure that AI systems uphold ethical norms. If an AI system for credit approval, for example, is found to be discriminating against people in certain groups (as has happened), these ethics managers are responsible for investigating and addressing the problem. Playing a similar role, data compliance officers try to ensure that the data that is feeding AI systems complies with the GDPR and other consumer-protection regulations. A related data-­u se role involves ensuring that AIs manage information responsibly. Like many tech companies, Apple uses AI to collect personal details about from users as they engage with the company’s devices and software. The aim is to improve the user experience, but unconstrained data gathering can compromise privacy, anger customers, and run

afoul of the law. The company’s “differential privacy team” works to make sure that while the AI seeks to learn as much as possible about a group of users in a statistical sense, it is protecting the privacy of individual users.

MACHINES ASSISTING HUMANS

Smart machines are helping humans expand their abilities in three ways. They can amplify our cognitive strengths; interact with customers and employees to free us for higher-level tasks; and embody human skills to extend our physical capabilities. Amplifying. Artificial intelligence can boost our analytic and decision-making abilities by providing the right information at the right time. But it can also heighten creativity. Consider how Autodesk’s Dreamcatcher AI enhances the imagination of even exceptional designers. A designer provides Dreamcatcher with criteria about the desired product—for example, a chair able to support up to 300 pounds, with a seat 18 inches off the ground, made of materials costing less than $75, and so on. She can also supply information about other chairs that she finds attractive. Dreamcatcher then churns out thousands of designs that match those criteria, often sparking ideas that the designer might not have initially considered. She can then guide the software, telling it which chairs she likes or doesn’t, leading to a new round of designs. Throughout the iterative process, Dreamcatcher performs the myriad calculations needed to ensure that each proposed design meets the specified criteria. This frees the designer to concentrate on deploying uniquely human strengths: professional judgment and aesthetic sensibilities. Interacting. Human-machine collaboration enables companies to interact with employees and customers in novel, more effective ways. AI agents like Cortana, for example, can facilitate communications between people or on behalf of people, such as by transcribing a meeting and distributing a voice-­ searchable version to those who couldn’t attend. Such applications are inherently scalable—a single chatbot, for instance, can provide routine customer service to large numbers of people simultaneously, wherever they may be. SEB, a major Swedish bank, now uses a virtual assistant called Aida to interact with millions of customers. Able to handle natural-language conversations, Aida has access to vast stores of data and can answer many frequently asked questions, such as how to open an account or make cross-border payments. She can also ask callers follow-up questions to solve their problems, and she’s able to analyze a caller’s tone of voice (frustrated versus appreciative, for instance) and use that information to provide better service later. Whenever the system can’t resolve an

©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTS RESERVED. 111 This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

issue—which happens in about 30% of cases—it turns the caller over to a human customer-service representative and then monitors that interaction to learn how to resolve similar problems in the future. With Aida handling basic requests, human reps can concentrate on addressing more-complex issues, especially those from unhappy callers who might require extra hand-holding. Embodying. Many AIs, like Aida and Cortana, exist principally as digital entities, but in other applications the intelligence is embodied in a robot that augments a human worker. With their sophisticated sensors, motors, and actuators, AI-enabled machines can now recognize people and objects and

work safely alongside humans in factories, warehouses, and laboratories. In manufacturing, for example, robots are evolving from potentially dangerous and “dumb” industrial machines into smart, context-aware “cobots.” A cobot arm might, for example, handle repetitive actions that require heavy lifting, while a person performs complementary tasks that require dexterity and human judgment, such as assembling a gear motor. Hyundai is extending the cobot concept with exoskeletons. These wearable robotic devices, which adapt to the user and location in real time, will enable industrial workers to perform their jobs with superhuman endurance and strength.

©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTS RESERVED. 112 This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

FEATURE COLLABORATIVE INTELLIGENCE: HUMANS AND AI ARE JOINING FORCES

REIMAGINING YOUR BUSINESS

In order to get the most value from AI, operations need to be redesigned. To do this, companies must first discover and describe an operational area that can be improved. It might be a balky internal process (such as HR’s slowness to fill staff positions), or it could be a previously intractable problem that can now be addressed using AI (such as quickly identifying adverse drug reactions across patient populations). Moreover, a number of new AI and advanced analytic techniques can help surface previously invisible problems that are amenable to AI solutions (see the sidebar “Revealing Invisible Problems”). Next, companies must develop a solution through co-creation—having stakeholders envision how they might collaborate with AI systems to improve a process. Consider the case of a large agricultural company that wanted to deploy AI technology to help farmers.

At Mercedes-Benz, cobot arms guided by human workers pick up and place heavy parts, becoming an extension of the worker’s body. An enormous amount of data was available about soil properties, weather patterns, historical harvests, and so forth, and the initial plan was to build an AI application that would more accurately predict future crop yields. But in discussions with farmers, the company learned of a more pressing need. What farmers really wanted was a system that could provide real-time recommendations on how to increase productivity— which crops to plant, where to grow them, how much nitrogen to use in the soil, and so on. The company developed an AI system to provide such advice, and the initial outcomes were promising; farmers were happy about the crop yields obtained with the AI’s guidance. Results from that initial test were then fed back into the system to refine the algorithms used. As with the discovery step, new AI and analytic techniques can assist in co-creation by suggesting novel approaches to improving processes. The third step for companies is to scale and then sustain the proposed solution. SEB, for example, originally deployed a version of Aida internally to assist ©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTS RESERVED.

15,000 bank employees but thereafter rolled out the chatbot to its one million customers. Through our work with hundreds of companies, we have identified five characteristics of business processes that companies typically want to improve: flexibility, speed, scale, decision making, and personalization. When reimagining a business process, determine which of these characteristics is central to the desired transformation, how intelligent collaboration could be harnessed to address it, and what alignments and trade-offs with other process characteristics will be necessary. Flexibility. For Mercedes-Benz executives, inflexible processes presented a growing challenge. Increasingly, the company’s most profitable customers had been demanding individualized S-class sedans, but the automaker’s assembly systems couldn’t deliver the customization people wanted. Traditionally, car manufacturing has been a rigid process with automated steps executed by “dumb” robots. To improve flexibility, Mercedes replaced some of those robots with AI-enabled cobots and redesigned its processes around human-machine collaborations. At the company’s plant near Stuttgart, Germany, cobot arms guided by human workers pick up and place heavy parts, becoming an extension of the worker’s body. This system puts the worker in control of the build of each car, doing less manual labor and more of a “piloting” job with the robot. The company’s human-machine teams can adapt on the fly. In the plant, the cobots can be reprogrammed easily with a tablet, allowing them to handle different tasks depending on changes in the workflow. Such agility has enabled the manufacturer to achieve unprecedented levels of customization. Mercedes can individualize vehicle production according to the real-time choices consumers make at dealerships, changing everything from a vehicle’s dashboard components to the seat leather to the tire valve caps. As a result, no two cars rolling off the assembly line at the Stuttgart plant are the same. Speed. For some business activities, the premium is on speed. One such operation is the detection of credit-card fraud. Companies have just seconds to determine whether they should approve a given transaction. If it’s fraudulent, they will most likely have to eat that loss. But if they deny a legitimate transaction, they lose the fee from that purchase and anger the customer. Like most major banks, HSBC has developed an AI-based solution that improves the speed and accuracy of fraud detection. The AI monitors and scores millions of transactions daily, using data on purchase location and customer behavior, IP addresses, and other information to identify subtle patterns that signal possible fraud. HSBC first implemented the system in the United States, significantly reducing the rate of undetected fraud and false positives, 113

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

ENHANCING PERFORMANCE

At organizations in all kinds of industries, humans and AI are collaborating to improve five elements of business processes.

FLEXIBILITY

SPEED

SCALE

DECISION MAKING

PERSONALIZATION

Auto manufacturing

Mercedes-Benz

Assembly robots work safely alongside humans to customize cars in real time.

Product design

Autodesk

Software suggests new product design concepts as a designer changes parameters such as materials, cost, and performance requirements.

Software development

Gigster

AI helps analyze any type of software project, no matter the size or complexity, enabling humans to quickly estimate the work required, organize experts, and adapt workflows in real time.

Fraud detection

HSBC

AI screens credit- and debit-card transactions to instantly approve legitimate ones while flagging questionable ones for humans to evaluate.

Cancer treatment

Roche

AI aggregates patient data from disparate IT systems, speeding collaboration among specialists.

Public safety

Singapore government

Video analytics during public events predicts crowd behavior, helping responders address security incidents rapidly.

Recruiting

Unilever

Automated applicant screening dramatically expands the pool of qualified candidates for hiring managers to evaluate.

Customer service

Virgin Trains

Bot responds to basic customer requests, doubling the volume handled and freeing humans to address more-complex issues.

Casino management

GGH Morowitz

Computer-vision system helps humans continuously monitor every gaming table in a casino.

Equipment maintenance

General Electric

“Digital twins” and Predix diagnostic application provide techs with tailored recommendations for machine maintenance.

Financial services

Morgan Stanley

Robo-advisers offer clients a range of investment options based on real-time market information.

Disease prediction

Icahn School of Medicine at Mount Sinai

Deep Patient system helps doctors predict patients’ risk of specific disease, allowing preventive intervention.

Guest experience

Carnival Corporation

Wearable AI device streamlines the logistics of cruise-ship activities and anticipates guest preferences, facilitating tailored staff support.

Health care

Pfizer

Wearable sensors for Parkinson’s patients track symptoms 24/7, allowing customized treatment.

Retail fashion

Stitch Fix

AI analyzes customer data to advise human stylists, who give customers individualized clothing and styling recommendations.

©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTS RESERVED. 114 This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

FEATURE COLLABORATIVE INTELLIGENCE: HUMANS AND AI ARE JOINING FORCES

REVEALING INVISIBLE PROBLEMS

Former U.S. Defense Secretary Donald Rumsfeld once famously distinguished among “known knowns,” “known unknowns,” and “unknown unknowns”— things you’re not even aware you don’t know. Some companies are now using AI to uncover unknown unknowns in their businesses. Case in point: GNS Healthcare applies machine-learning software to find overlooked relationships among data in patients’ health records and elsewhere. After identifying a relationship, the software churns out numerous hypotheses to explain it and then suggests which of those are the most likely. This approach enabled GNS to uncover a new drug interaction hidden in unstructured patient notes. CEO Colin Hill points out that this is not garden-variety data mining to find associations. “Our machine-learning platform is not just about seeing patterns and correlations in data,” he says. “It’s about actually discovering causal links.” and then rolled it out in the UK and Asia. A different AI system used by Danske Bank improved its fraud-­ detection rate by 50% and decreased false positives by 60%. The reduction in the number of false positives frees investigators to concentrate their efforts on equivocal transactions the AI has flagged, where human judgment is needed. The fight against financial fraud is like an arms race: Better detection leads to more-devious criminals, which leads to better detection, which continues the cycle. Thus the algorithms and scoring models for combating fraud have a very short shelf life and require continual updating. In addition, different countries and regions use different models. For these reasons, legions of data analysts, IT professionals, and experts in financial fraud are needed at the interface between humans and machines to keep the software a step ahead of the criminals. Scale. For many business processes, poor scalability is the primary obstacle to improvement. That’s particularly true of processes that depend on intensive human labor with minimal machine assistance. Consider, for instance, the employee recruitment process at Unilever. The consumer goods giant was looking for a way to diversify its 170,000-person

workforce. HR determined that it needed to focus on entry-level hires and then fast-track the best into management. But the company’s existing processes weren’t able to evaluate potential recruits in sufficient numbers—while giving each applicant individual attention—to ensure a diverse population of exceptional talent. Here’s how Unilever combined human and AI capabilities to scale individualized hiring: In the first round of the application process, candidates are asked to play online games that help assess traits such as risk aversion. These games have no right or wrong answers, but they help Unilever’s AI figure out which individuals might be best suited for a particular position. In the next round, applicants are asked to submit a video in which they answer questions designed for the specific position they’re interested in. Their responses are analyzed by an AI system that considers not just what they say but also their body language and tone. The best candidates from that round, as judged by the AI, are then invited to Unilever for in-person interviews, after which humans make the final hiring decisions. It’s too early to tell whether the new recruiting process has resulted in better employees. The company has been closely tracking the success of those hires, but more data is still needed. It is clear, however, that the new system has greatly broadened the scale of Unilever’s recruiting. In part because job seekers can easily access the system by smartphone, the number of applicants doubled to 30,000 within a year, the number of universities represented surged from 840 to 2,600, and the socioeconomic diversity of new hires increased. Furthermore, the average time from application to hiring decision has dropped from four months to just four weeks, while the time that recruiters spend reviewing applications has fallen by 75%. Decision making. By providing employees with tailored information and guidance, AI can help them reach better decisions. This can be especially valuable for workers in the trenches, where making the right call can have a huge impact on the bottom line. Consider the way in which equipment maintenance is being improved with the use of “digital twins”—virtual models of physical equipment. General Electric builds such software models of its turbines and other industrial products and continually updates them with operating data streaming from the equipment. By collecting readings from large numbers of machines in the field, GE has amassed a wealth of information on normal and aberrant performance. Its Predix application, which uses machine-learning algorithms, can now predict when a specific part in an individual machine might fail. This technology has fundamentally changed the decision-intensive process of maintaining industrial equipment. Predix might, for example, identify some unexpected rotor wear and tear in a turbine, check the

©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTS RESERVED. 115 This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

turbine’s operational history, report that the damage has increased fourfold over the past few months, and warn that if nothing is done, the rotor will lose an estimated 70% of its useful life. The system can then suggest appropriate actions, taking into account the machine’s current condition, the operating environment, and aggregated data about similar damage and repairs to other machines. Along with its recommendations, Predix can generate information about their costs and financial benefits and provide a confidence level (say, 95%) for the assumptions used in its analysis. Without Predix, workers would be lucky to catch the rotor damage on a routine maintenance check. It’s possible that it would go undetected until the rotor failed, resulting in a costly shutdown. With Predix, maintenance workers are alerted to potential problems before they become serious, and they have the needed information at their fingertips to make good decisions—ones that can sometimes save GE millions of dollars. Personalization. Providing customers with individually tailored brand experiences is the holy grail of marketing. With AI, such personalization can now be achieved with previously unimaginable precision and at vast scale. Think of the way the music streaming service Pandora uses AI algorithms to generate personalized playlists for each of its millions of users according to their preferences in songs, artists, and genres. Or consider Starbucks, which, with customers’ permission, uses AI to recognize their mobile devices and call up their ordering history to help baristas make serving recommendations. The AI technology does what it does best, sifting through and processing copious amounts of data to recommend certain offerings or actions, and humans do what they do best, exercising their intuition and judgment to make a recommendation or select the best fit from a set of choices. The Carnival Corporation is applying AI to personalize the cruise experience for millions of vacationers through a wearable device called the Ocean Medallion and a network that allows smart devices to connect. Machine learning dynamically processes the data flowing from the medallion and from sensors and systems throughout the ship to help guests get the most out of their vacations. The medallion streamlines the boarding and debarking processes, tracks the guests’ activities, simplifies purchasing by connecting their credit cards to the device, and acts as a room key. It also connects to a system that anticipates guests’ preferences, helping crew members deliver personalized service to each guest by suggesting tailored itineraries of activities and dining experiences.

THE NEED FOR NEW ROLES AND TALENT

Reimagining a business process involves more than the implementation of AI technology; it also requires

a significant commitment to developing employees with what we call “fusion skills”—those that enable them to work effectively at the human-machine interface. To start, people must learn to delegate tasks to the new technology, as when physicians trust computers to help read X-rays and MRIs. Employees should also know how to combine their distinctive human skills with those of a smart machine to get a better outcome than either could achieve alone, as in robot-assisted surgery. Workers must be able to teach intelligent agents new skills and undergo training to work well within AI-enhanced processes. For example, they must know how best to put questions to an AI agent to get the information they need. And there must be employees, like those on Apple’s differential privacy team, who ensure that their companies’ AI systems are used responsibly and not for illegal or unethical purposes. We expect that in the future, company roles will be redesigned around the desired outcomes of reimagined processes, and corporations will increasingly be organized around different types of skills rather than around rigid job titles. AT&T has already begun that transition as it shifts from landline telephone services to mobile networks and starts to retrain 100,000 employees for new positions. As part of that effort, the company has completely overhauled its organizational chart: Approximately 2,000 job titles have been streamlined into a much smaller number of broad categories encompassing similar skills. Some of those skills are what one might expect (for example, proficiency in data science and data wrangling), while others are less obvious (for instance, the ability to use simple machine-­ learning tools to cross-sell services). MOST ACTIVITIES AT the human-machine interface require people to do new and different things (such as train a chatbot) and to do things differently (use that chatbot to provide better customer service). So far, however, only a small number of the companies we’ve surveyed have begun to reimagine their business processes to optimize collaborative intelligence. But the lesson is clear: Organizations that use machines merely to displace workers through automation will miss the full potential of AI. Such a strategy is misguided from the get-go. Tomorrow’s leaders will instead be those that embrace collaborative intelligence, transforming their operations, their markets, their industries, and—no less important— their workforces.  HBR Reprint R1804J H. JAMES WILSON is a managing director of information technology and business research at Accenture Research. PAUL R. DAUGHERTY is Accenture’s chief technology and innovation officer. They are the coauthors of Human + Machine: Reimagining Work in the Age of AI (Harvard Business Review Press, 2018).

©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTS RESERVED. 116 This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

Discussion Questions

Collaborative Intelligence: Humans and AI Are Joining Forces BY H. JAMES WILSON AND PAUL DAUGHERTY

1. In what ways have jobs or job functions at your company been redesigned, created, or eliminated since you’ve introduced AI initiatives? 2. The authors describe six types of human-machine collaboration: training, explaining, sustaining, amplifying, interacting, and embodying. Do any of these functions exist at your company now? Which have the largest potential for growth? Which jobs do you think will be the hardest to fill? 3. How do you feel differently about the future of work after reading this article?

To take full advantage of this collaboration, companies must understand how humans can most effectively augment machines, how machines can enhance what humans do best, and how to redesign business processes to support the partnership.

©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTS RESERVED.

117

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

REPRINT H0433D PUBLISHED ON HBR.ORG JANUARY 12, 2018

ARTICLE DATA

Robo-Advisers Are Coming to Consulting and Corporate Strategy by Thomas H. Davenport, Barry Libert and Megan Beck

©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTS RESERVED.

118

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

DATA

Robo-Advisers Are Coming to Consulting and Corporate Strategy by Thomas H. Davenport, Barry Libert and Megan Beck JANUARY 12, 2018

CSA Images/Printstock Collection/Getty Images Does a robot manage your money? For many of us, the answer is yes. Online and algorithmic investment and financial advice is easy to come by these days, usually under the moniker of “roboadvisor.” Startups such as Wealthfront, Personal Capital, and Betterment launched robo-advisors as industry disruptors, and incumbents, such as Schwab’s (Intelligent Advisor), Vanguard (Personal

©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTS RESERVED.

119

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

Advisor Services), Morgan Stanley and BlackRock have joined the fray with their own hybrid machine/advisor solutions. It’s clear that robo-advisors and AI play an important and growing role in the financial services industry, but a question remains. Will robo-advisors disrupt corporate capital allocation the same way they have personal capital allocation? And, will they shake up the trilliondollar corporate consulting and advisory industry? Robo-advisors, which were introduced in 2008, are steadily eating up market share from their human counterparts much the way that Amazon and Netflix have taken share from Walmart and Regal Cinemas. A study by Deloitte estimated that “assets under automated management” (including hybrid offerings) in the U.S. will grow to U.S. $5 trillion to U.S. $7 trillion by the year 2025 from about U.S. $300 billion today. This would represent between 10% and 15% of total retail financial assets under management. At the end of 2016, Fitch Ratings estimated that all robo-advisors managed under U.S. $100B in assets, and predicts double-digit growth in assets under management over the next several years. Finally, A.T. Kearney predicts that assets under “robo-management” will total $2.2 trillion by 2021. If AI-guided investing can work for a person, can it also work for a company? Corporations buy and employ human advice from many wise advisors—consultants, lawyers, investment bankers—in the same fashion that investors did in the past. Corporate strategy is complex, and the advice is expensive. However, the approaches advisors take are usually data-driven and guided by previous experiences. This is just the sort of problem that can benefit from machine intelligence. This makes corporate strategy an enormous and untapped prize for “robos” and “AI-enabled” expert advice across the entire enterprise; this market is ripe for disruption much the way the financial investing industry was in 2008. Marketing and sales, manufacturing, recruiting (including people assessment), customer service, and support are all fields that can benefit from artificial intelligence according to McKinsey’s recent research. The reasons for this potential disruption now are many: • There is an explosion in the amount of corporate data. In fact, it is doubling every 14 months and it will reach 10.5 ZB by 2020. This data is both financial (revenues, profits, growth) and non-financial (customer sentiment, employee engagement, marketing effectiveness, product feedback, and partner ecosystems). The availability of this data creates fertile ground for robos to provide algorithmic insights and recommendations that deliver highly predictive, error-proof, and low-cost advising. • Companies are both operators and investors.Research by McKinsey shows that US companies allocate about $650B a year across all their activities—be it financial, physical, human, intellectual, or customer capital. However, they don’t have the tools or practices to best allocate capital, and as a result, 92% of companies allocate their capital the same way year over year. Just like individual investors, most corporations could probably use some help in making wise investment decisions. ©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTS RESERVED.

120

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

• AI is growing exponentially in enterprises. By almost all accounts, companies at the digital frontier such as Google, Facebook, and Microsoft are investing vast amounts in AI—somewhere between $20 billion and $30 billion alone in 2016. Many established firms—a 2017 Deloitte survey suggested about 20% in the U.S.—are making substantial investments in AI as well. Further, venture capitalists are jumping in with both feet. $4 to $5 billion was invested by VCs in AI in 2016. Lastly, private equity firms invested another $1 billion to $3 billion. These numbers represent more than three times as much as was invested in 2013. • The costs of AI-enabled tools are falling, and availability is rising. Both proprietary tools, like IBM’s Watson, and open-source tools from firms like Google, Microsoft, and Amazon, are widely available. Cloud-based hardware is also increasingly available to any business at low cost. • Companies in every industry can benefit from making more data and algorithm-based decisions in areas of internal operations and finance. Analytics are growing in every business function and industry. “Robo-advice” is a straightforward extension of these analytical tools. Each one of us is becoming increasingly more comfortable being advised by robots for everything from what movie to watch to where to put our retirement. Given the groundwork that has been laid for artificial intelligence in companies, it’s only a matter of time before the $60 billion consulting industry in the U.S. is going to be disrupted by robotic advisors. For those who want to stay ahead of the curve, there are three strategies you can take: Build a pure-play solution: Several robo-advice companies started their offerings with machine-only advice. Their goal was to hit the lowest possible price point, and to appeal to “digital native” customers. However, as the companies providing hybrid advice have grown rapidly, most startups now also offer some level of human advice—typically for a higher fee. Only Wealthfront remains a machine-only robo-advisor. This suggests that corporate robo-advice providers should think carefully before abandoning the human component completely. At Vanguard, the Personal Advisor Services offering features advisors as “investing coaches” who are able to answer investor questions, encourage healthy financial behaviors, and be, in Vanguard’s words, “emotional circuit breakers” to keep investors on their plans. There are likely to be corporate equivalents of these functions. Create your own internal robo-advisory service: Companies could develop their own robotic or semirobotic advice for key decision domains. This is in fact what cancer hospitals, for example, are attempting to do with IBM Watson in cancer care, and what customers of semi-automated machine learning platforms are doing for highly quantitative decisions (DataRobot is one example; Google’s new AutoML is another). However, developing a robo-advisor only for one’s own internal issues may be more difficult and expensive than many companies are willing to venture into. Further, it is decidedly outside the wheelhouse for most established firms, which brings us to the third option. Partner with or acquire an existing provider: In financial robo-advice, firms that were not first to market are now moving quickly to either partner with a startup or acquire one. Examples include ©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTS RESERVED.

121

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

BlackRock, which recently acquired FutureAdvisor for a reported $150-200 million; JPMorgan’s recent partnership with Motif Investing, and UBS’ equity investment in SigFig. There are likely to eventually be a number of vendors of corporate robo-advice, though they are not widely available at this point. Regardless of which strategy you pursue, it seems likely that corporate robo-advisors are coming to many parts of the organization, just as software has spread through the value chain over the last two decades. Robo-advisors have the potential to deliver a broader array of advice and there may be a range of specialized tools in particular decision domains. These robo-advisors may be used to automate certain aspects of risk management and provide decisions that are ethical and compliant with regulation. In data-intensive fields like marketing and supply chain management, the results and decisions that robotic algorithms provide is likely to be more accurate than those made by human intuition. Finally, it is becoming clear that serious AI adopters with proactive business strategies based on it benefit from higher profit margins. In fact, a McKinsey survey suggests that these front runners report current profit margins that are 3 to 15 percentage points higher than the industry average in most sectors, and they also expect this advantage to grow in the future. In the next three years, these AI leaders expect their margins to increase by up to 7 percentage points more than the industry average. Of course, traditional consultants and other providers of corporate advice are unlikely to disappear. Like the human advisors that still complement robo-advisors in the investment world, they can provide a number of key functions. Here are several ways existing corporate advisors can complement their future robot partners: • Integrate different online advice sources, and help clients and investment firms to understand what systems to use for what purposes. Human advisors could also, like hedge fund managers, analyze the results from machine-advised decisions and advise clients on whether changes are necessary in the algorithms and logic employed by the machines. • Shift to providing advice on business models, not just strategy and operations. We suggested in a recent article that pure advice from even the most elite consultants would be put at risk by machine learning. However, our research as well as others’ suggest that consultants can focus on their clients’ business models rather than just strategy, operations, and best practices to insure their future growth, relevance and success.

©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTS RESERVED.

122

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

• Deliver behavioral coaching. As corporate strategy advice is increasingly disrupted by algorithms and artificial intelligence, corporate advisors could coach leaders on the best approach to success using their EQ skills. As with behavioral coaches in individual investing, corporate coaches could, for example, dissuade leaders and boards from buying companies at the top of the market or selling when the markets crash. They can help them with change management as smart machines provide new insights at increasing speeds. While the details of adoption of automated advice from robo advisors in all industries are unclear, it is likely that the future will include automated advisors in many fields. They already exist in personal investing, driving navigation (Google Maps, Waze), matchmaking (EHarmony, Match.com), and healthcare (WebMD Symptom Checker). It seems only logical that they would extend into corporate strategy and finance. Financial services firms, financial advisors, and their clients were the first to witness substantial disruption, but they won’t be the last. The days of only face-to-face discussions between client and consultant may not vanish altogether, but they shift from crunching the numbers to changing behaviors and nurturing relationships with clients. As Ian Dodd, Director of legal analytics firm Premonition, said to the BBC, “The knowledge jobs will go. The wisdom jobs will stay.”

Thomas H. Davenport is the President’s Distinguished Professor in Management and Information Technology at Babson College, a research fellow at the MIT Initiative on the Digital Economy, and a senior adviser at Deloitte Analytics. Author of over a dozen management books, his latest is Only Humans Need Apply: Winners and Losers in the Age of Smart Machines.

Barry Libert is a board member and CEO adviser focused on platforms and networks. He is chairman of Open Matters, a machine learning company. He is also the coauthor of The Network Imperative: How to Survive and Grow in the Age of Digital Business Models.

Megan Beck is Chief Product and Insights Officer at OpenMatters, a machine learning startup, and a digital researcher at the SEI Center at Wharton. She is the coauthor of The Network Imperative: How to Survive and Grow in the Age of Digital Business Models.

©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTS RESERVED.

123

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

Discussion Questions

Robo-Advisers Are Coming to Consulting and Corporate Strategy BY THOMAS H. DAVENPORT, BARRY LIBERT, AND MEGAN BECK

1. The authors propose that the success algorithms have had in managing personal wealth suggests that robo-advisers will soon be managing corporate capital allocation as well. Does this seem like a logical conclusion? Why does this seem likely or unlikely to you? What are the risks and benefits? 2. What consulting and advice functions does your company provide that it might soon be possible to deliver with AI? 3. Could your company engage robo-advisers for strategy and operations? How would these AI-based capabilities have to demonstrate their merit before you use them?

Given the groundwork that has been laid for artificial intelligence in companies, it’s only a matter of time before the $60 billion consulting industry in the U.S. is going to be disrupted by robotic advisers.

©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTS RESERVED.

124

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

REPRINT H04BM0 PUBLISHED ON HBR.ORG MAY 10, 2018

ARTICLE TECHNOLOGY

10 Promising AI Applications in Health Care by Brian Kalis, Matt Collier and Richard Fu

©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTS RESERVED.

125

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

TECHNOLOGY

10 Promising AI Applications in Health Care by Brian Kalis, Matt Collier and Richard Fu

MAY 10, 2018

Peter Dazeley/Getty Images There’s a lot of excitement right now about how artificial intelligence (AI) is going to change health care. And many AI technologies are cropping up to help people streamline administrative and clinical health care processes. According to venture capital firm Rock Health, 121 health AI and machine learning companies raised $2.7 billion in 206 deals between 2011 and 2017.

©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTS RESERVED.

126

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

The field of health AI is seemingly wide—covering wellness to diagnostics to operational technologies —but it is also narrow in that health AI applications typically perform just a single task. We investigated the value of 10 promising AI applications and found that they could create up to $150 billion in annual savings for U.S. health care by 2026. We identified these specific AI applications based on how likely adoption was and what potential exists for annual savings. We found AI currently creates the most value in helping frontline clinicians be more productive and in making back-end processes more efficient—but not yet in making clinical decisions or improving clinical outcomes. Clinical applications are still rare.

Let’s take a look at a few examples of AI on the frontline of care. AI has demonstrated its aptitude for improving the efficiency of image analysis by quickly and accurately flagging specific anomalies for a radiologist’s review. In 2011, researchers from NYU Langone Health found that this type of automated analysis could find and match specific lung nodules (on chest CT images) between 62% to 97% faster than a panel of radiologists. Our findings suggest such AI-generated efficiencies in image analysis could create $3 billion in annual savings by giving radiologists more time to focus on reviews that require greater interpretation or judgement. Another area is AI-assisted robotic surgery. In orthopedic surgery, a form of AI-assisted robotics can analyze data from pre-op medical records to physically guide the surgeon’s instrument in real-time

©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTS RESERVED.

127

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

during a procedure. It can also use data from actual surgical experiences to inform new surgical techniques. A study of 379 orthopedic patients across nine surgical sites found that an AI-assisted robotic technique created by Mazor Robotics resulted in a five-fold reduction in surgical complications compared to when surgeons operated alone. When applied properly to orthopedic surgery, our analysis found AI-assisted robotic surgery could also generate a 21% reduction in patients’ length of stay in the hospital following surgery, as a result of fewer complications and errors, and create $40 billion in annual savings. AI techniques are also being applied to the costly problem of dosage errors—where our findings suggest AI could generate $16 billion in savings. In 2016, a ground breaking trial in California found that a mathematical formula developed with the help of AI had correctly determined the correct dose of immunosuppressant drugs to administer to organ patients. Determining the dose has traditionally depended on a combination of guidelines and educated guesswork—and dosing errors make up 37% of all preventable medical errors. While this type of AI technique is nascent, the example is powerful considering that the correct dose is critical to making sure a graft is not rejected after an organ transplant. Using AI to aid clinical judgement or diagnosis still remains in its infancy, but some results are emerging to illustrate the possibility. In 2017, a group at Stanford University tested an AI algorithm against 21 dermatologists on its ability to identify skin cancers. The clinical findings, as reported by Nature last year, “achieve performance on par with all tested experts …demonstrating an artificial intelligence capable of classifying skin cancer with a level of competence comparable to dermatologists.” Our findings suggest AI could yield $5 billion in annual savings by doing a preliminary diagnosis before a patient enters the emergency department. We’re also starting to see potential of AI-powered virtual nurse assistants in helping patients. For example, Sensely’s “Molly” is an AI-powered nurse avatar being used by UCSF and the UK’s NHS to interact with patients, ask them questions about their health, assess their symptoms, and direct them to the most effective care setting. Our findings estimate AI-powered nurse assistants could save $20 billion annually by saving 20% of the time nurses spend on patient maintenance tasks. AI also holds promise for helping the health care industry manage costly back-office problems and inefficiencies. Activities that have nothing to do with patient care consume over half (51%) of a nurse’s workload and nearly a fifth (16%) of physician activities. AI-based technologies, such as voice-to-text transcription, can improve administrative workflows and eliminate time-consuming non-patient-care activities, such as writing chart notes, filling prescriptions, and ordering tests. We estimate that these applications could save the industry $18 billion annually. For example, while Beth Israel Deaconess Medical Center garnered attention for an AI-enabled cancer screen, its first foray into AI was more prosaic: using it to reduce hospital readmission rates and identify possible no-shows. Using machine learning, technologists at Beth Israel Medical Center

©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTS RESERVED.

128

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

developed an application to predict which patients are likely to be no shows or lapse on treatment so they can intervene ahead of time. Errors and fraud are a similarly expensive problem for health care organizations and also for insurers. Fraud detection has traditionally relied on a combination of computerized (rules-based) and manual reviews of medical claims. It’s a time-consuming process that hinges on being able to quickly spot anomalies after the incident occurs in order to intervene. Health insurers are experimenting with AIsupported data mining, coupled with AI-based neural networks (which mimic the processes of the human brain, but much more quickly) to search Medicare claims for patterns associated with medical reimbursement fraud. We estimated that AI could create $17 billion in annual savings by improving the speed and accuracy of fraud detection in Medicare claims. Beyond fraudulent activity, the litany of data breaches, such as WannaCry or Petya, over the past few years has made cybersecurity a major concern for health care organizations. Health care breaches are estimated to cost organizations $380 per patient record. Using AI to monitor and detect abnormal interactions with proprietary data could create $2 billion in annual savings by reducing health record breaches. As AI technologies become more prevalent, health care organizations will have to invest in those that deliver the most value. Uses of AI for clinical judgement still remains in its infancy and will need time to fully take root in a meaningful way. But the AI applications that can deliver the most value today (AI-assisted surgery, virtual nurse, administrative workflow) should be prioritized and invested in, so health care providers and insurers are free to focus on better care.

Brian Kalis leads digital health for Accenture.

Matt Collier leads health strategy consulting globally for Accenture Strategy.

Richard Fu is a coauthor of Accenture’s research report on artificial intelligence.

©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTS RESERVED.

129

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

Discussion Questions

10 Promising AI Applications in Health Care BY BRIAN KALIS, MATT COLLIER, AND RICHARD FU

1. This article presents 10 AI applications in health care in a table with “potential annual value” and “key drivers for adoption.” Could you make a similar table for possible AI initiatives in your company? 2. Amid all of the pessimism around cost controls in the U.S. health care system, does this article make you feel more optimistic about the future of health care? How so, or why not? 3. How would you feel about using a virtual nursing assistant or having an AI application determine dosage for a medication you are taking? Some people will perceive applications such as these as creating value for the medical providers at the expense of individual patients. Would your customers view any of your company’s AI-enabled offerings in a similar light? Could this lead to backlash and how should your company address it?

We found AI currently creates the most value in helping frontline clinicians be more productive and in making back-end processes more efficient—but not yet in making clinical decisions or improving clinical outcomes. Clinical applications are still rare.

©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTS RESERVED.

130

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

REPRINT H03LEM PUBLISHED ON HBR.ORG APRIL 18, 2017

ARTICLE EXPERIMENTATION

The First Wave of Corporate AI Is Doomed to Fail

by Kartik Hosanagar and Apoorv Saxena

©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTS RESERVED.

131

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

EXPERIMENTATION

The First Wave of Corporate AI Is Doomed to Fail

by Kartik Hosanagar and Apoorv Saxena APRIL 18, 2017

Artificial intelligence is a hot topic right now. Driven by a fear of losing out, companies in many industries have announced AI-focused initiatives. Unfortunately, most of these efforts will fail. They will fail not because AI is all hype, but because companies are approaching AI-driven innovation incorrectly. And this isn’t the first time companies have made this kind of mistake.

COPYRIGHT © 2017 HARVARD BUSINESS SCHOOL PUBLISHING ALL RIGHTS RESERVED. ©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTSCORPORATION. RESERVED.

1322

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

Back in the late 1990s, the internet was the big trend. Most companies started online divisions. But there were very few early wins. Once the dot-com bust happened, these companies shut down or significantly downscaled their online efforts. A few years later they were caught napping when online upstarts disrupted industries such as music, travel, news, and video, while transforming scores of others. In the mid-2000s, the buzz was about cloud computing. Once again, several companies decided to test the waters. There were several early issues, ranging from regulatory compliance to security. Many organizations backed off from moving their data and applications to the cloud. The ones that persisted are incredibly well-positioned today, having transformed their business processes and enabled a level of agility that competitors cannot easily mimic. The vast majority are still playing catch-up. We believe that a similar story of early failures leading to irrational retreats will occur with AI. Already, evidence suggests that early AI pilots are unlikely to produce the dramatic results that technology enthusiasts predict. For example, early efforts of companies developing chatbots for Facebook’s Messenger platform saw 70% failure rates in handling user requests. Yet a reversal on these initiatives among large companies would be a mistake. The potential of AI to transform industries truly is enormous. Recent research from McKinsey Global Institute found that 45% of work activities could potentially be automated by today’s technologies, and 80% of that is enabled by machine learning. The report also highlighted that companies across many sectors, such as manufacturing and health care, have captured less than 30% of the potential from their data and analytics investments. Early failures are often used to slow or completely end these investments. AI is a paradigm shift for organizations that have yet to fully embrace and see results from even basic analytics. So creating organizational learning in the new platform is far more important than seeing a big impact in the short run. But how does a manager justify continuing to invest in AI if the first few initiatives don’t produce results?

VIDEO A.I. COULD LIBERATE 50% OF MANAGERS' TIME

TO VIEW, PLEASE VISIT THIS ARTICLE AT HBR.ORG

We suggest taking a portfolio approach to AI projects: a mix of projects that might generate quick wins and long-term projects focused on transforming end-to-end workflow. For quick wins, one might focus on changing internal employee touchpoints, using recent advances in speech, vision, and language understanding. Examples of these projects might be a voice interface to help pharmacists look up substitute drugs, or a tool to schedule internal meetings. These are areas in which recently available, off-the-shelf AI tools, such as Google’s Cloud Speech API and Nuance’s speech recognition API, can be used, and they don’t require massive investment in training and hiring. (Disclosure: One of us is an executive at Alphabet Inc., the parent company of Google.) They will not be transformational, but they will help build consensus on the potential of AI. Such projects

COPYRIGHT © 2017 HARVARD BUSINESS SCHOOL PUBLISHING ALL RIGHTS RESERVED. ©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTSCORPORATION. RESERVED.

1333

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

also help organizations gain experience with large-scale data gathering, processing, and labeling, skills that companies must have before embarking on more-ambitious AI projects. For long-term projects, one might go beyond point optimization, to rethinking end-to-end processes, which is the area in which companies are likely to see the greatest impact. For example, an insurer could take a business process such as claims processing and automate it entirely, using speech and vision understanding. Allstate car insurance already allows users to take photos of auto damage and settle their claims on a mobile app. Technology that’s been trained on photos from past claims can accurately estimate the extent of the damage and automate the whole process. As companies such as Google have learned, building such high-value workflow automation requires not just off-the-shelf technology but also organizational skills in training machine learning algorithms. As Google pursued its goal of transitioning into an AI-first company, it followed a similar portfoliobased approach. The initial focus was on incorporating machine learning into a few subcomponents of a system (e.g., spam detection in Gmail), but now the company is using machine learning to replace entire sets of systems. Further, to increase organizational learning, the company is dispersing machine learning experts across product groups and training thousands of software engineers, across all Google products, in basic machine learning. This all leads to the question of how best to recruit the resources for these efforts. The good news is that emerging marketplaces for AI algorithms and datasets, such as Algorithmia and the Googleowned Kaggle, coupled with scalable, cloud-based infrastructure that is custom-built for artificial intelligence, are lowering barriers. Algorithms, data, and IT infrastructure for large-scale machine learning are becoming accessible to even small and medium-size businesses. Further, the cost of artificial intelligence talent is coming down as the supply of trained professionals increases. Just as the cost of building a mobile app went from $200,000–$300,000 in 2010 to less than $10,000 today with better development tools, standardization around few platforms (Android and iOS), and increased supply of mobile developers, similar price deflation in the cost of building AIpowered systems is coming. The implication is that there is no need for firms to frontload their hiring. Hiring slowly, yet consistently, over time and making use of marketplaces for machine learning software and infrastructure can help keep costs manageable. There is little doubt that an AI frenzy is starting to bubble up. We believe AI will indeed transform industries. But the companies that will succeed with AI are the ones that focus on creating organizational learning and changing organizational DNA. And the ones that embrace a portfolio approach rather than concentrating their efforts on that one big win will be best positioned to harness the transformative power of artificial learning.

Kartik Hosanagar is a Professor of Technology and Digital Business at The Wharton School of the University of Pennsylvania. He was previously a cofounder of Yodle Inc. Follow him on Twitter @khosanagar.

COPYRIGHT © 2017 HARVARD BUSINESS SCHOOL PUBLISHING ALL RIGHTS RESERVED. ©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTSCORPORATION. RESERVED.

1344

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

Apoorv Saxena is a product manager at Google leading AI products and also Co-Chair of AI Frontier conference.

COPYRIGHT © 2017 HARVARD BUSINESS SCHOOL PUBLISHING ALL RIGHTS RESERVED. ©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTSCORPORATION. RESERVED.

1355

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

Discussion Questions

The First Wave of Corporate AI Is Doomed to Fail BY KARTIK HOSANAGAR AND APOORV SAXENA

1. The authors suggest that in the near future, many companies will retreat from AI initiatives after their first attempts to use the technology are unsuccessful. In what ways is your company investing in AI using a portfolio approach, balancing goals of short-term results, and building capabilities for the long term? 2. How can your company ensure that it retains and builds upon its learnings from its earlystage AI initiatives, regardless of whether they are successful? 3. Is your company an incumbent that went through the dot-com bust and the rise of cloud computing? How does your company’s history with these technology adoption cycles affect the way it thinks about AI now? What might be done to move past this?

Unfortunately, most of these efforts will fail. They will fail not because AI is all hype, but because companies are approaching AI-driven innovation incorrectly. And this isn’t the first time companies have made this kind of mistake.

COPYRIGHT © 2018 HARVARD BUSINESS SCHOOL PUBLISHING ALL RIGHTS RESERVED. ©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTSCORPORATION. RESERVED.

24 136

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

REPRINT H03NES PUBLISHED ON HBR.ORG MAY 08, 2017

ARTICLE SECURITY & PRIVACY

AI Is the Future of Cybersecurity, for Better and for Worse by Roman V. Yampolskiy

©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTS RESERVED.

137

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

SECURITY & PRIVACY

AI Is the Future of Cybersecurity, for Better and for Worse

by Roman V. Yampolskiy MAY 08, 2017

In the near future, as artificial intelligence (AI) systems become more capable, we will begin to see more automated and increasingly sophisticated social engineering attacks. The rise of AI-enabled cyberattacks is expected to cause an explosion of network penetrations, personal data thefts, and an epidemic-level spread of intelligent computer viruses. Ironically, our best hope to defend against AIenabled hacking is by using AI. But this is very likely to lead to an AI arms race, the consequences of

COPYRIGHT © 2017 HARVARD BUSINESS SCHOOL PUBLISHING ALL RIGHTS RESERVED. ©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTSCORPORATION. RESERVED.

1382

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

which may be very troubling in the long term, especially as big government actors join the cyber wars. My research is at the intersection of AI and cybersecurity. In particular, I am researching how we can protect AI systems from bad actors, as well as how we can protect people from failed or malevolent AI. This work falls into a larger framework of AI safety, attempts to create AI that is exceedingly capable but also safe and beneficial. A lot has been written about problems that might arise with the arrival of “true AI,” either as a direct impact of such inventions or because of a programmer’s error. However, intentional malice in design and AI hacking have not been addressed to a sufficient degree in the scientific literature. It’s fair to say that when it comes to dangers from a purposefully unethical intelligence, anything is possible. According to Bostrom’s orthogonality thesis, an AI system can potentially have any combination of intelligence and goals. Such goals can be introduced either through the initial design or through hacking, or introduced later, in case of an off-the-shelf software — “just add your own goals.” Consequently, depending on whose bidding the system is doing (governments, corporations, sociopaths, dictators, military industrial complexes, terrorists, etc.), it may attempt to inflict damage that’s unprecedented in the history of humankind — or that’s perhaps inspired by previous events. Even today, AI can be used to defend and to attack cyber infrastructure, as well as to increase the attack surface that hackers can target, that is, the number of ways for hackers to get into a system. In the future, as AIs increase in capability, I anticipate that they will first reach and then overtake humans in all domains of performance, as we have already seen with games like chess and Go and are now seeing with important human tasks such as investing and driving. It’s important for business leaders to understand how that future situation will differ from our current concerns and what to do about it. If one of today’s cybersecurity systems fails, the damage can be unpleasant, but is tolerable in most cases: Someone loses money or privacy. But for human-level AI (or above), the consequences could be catastrophic. A single failure of a superintelligent AI (SAI) system could cause an existential risk event — an event that has the potential to damage human well-being on a global scale. The risks are real, as evidenced by the fact that some of the world’s greatest minds in technology and physics, including Stephen Hawking, Bill Gates, and Elon Musk, have expressed concerns about the potential for AI to evolve to a point where humans could no longer control it.

VIDEO BOARDS NEGLECT CYBERSECURITY AT THEIR COMPANIES’ PERIL

TO VIEW, PLEASE VISIT THIS ARTICLE AT HBR.ORG

When one of today’s cybersecurity systems fails, you typically get another chance to get it right, or at least to do better next time. But with an SAI safety system, failure or success is a binary situation: Either you have a safe, controlled SAI or you don’t. The goal of cybersecurity in general is to reduce the number of successful attacks on a system; the goal of SAI safety, in contrast, is to make sure

COPYRIGHT © 2017 HARVARD BUSINESS SCHOOL PUBLISHING ALL RIGHTS RESERVED. ©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTSCORPORATION. RESERVED.

1393

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

no attacks succeed in bypassing the safety mechanisms in place. The rise of brain-computer interfaces, in particular, will create a dream target for human and AI-enabled hackers. And braincomputer interfaces are not so futuristic — they’re already being used in medical devices and gaming, for example. If successful, attacks on brain-computer interfaces would compromise not only critical information such as social security numbers or bank account numbers but also our deepest dreams, preferences, and secrets. There is the potential to create unprecedented new dangers for personal privacy, free speech, equal opportunity, and any number of human rights. Business leaders are advised to familiarize themselves with the cutting edge of AI safety and security research, which at the moment is sadly similar to the state of cybersecurity in the 1990s, and our current situation with the lack of security for the internet of things. Armed with more knowledge, leaders can rationally consider how the addition of AI to their product or service will enhance user experiences, while weighing the costs of potentially subjecting users to additional data breaches and possible dangers. Hiring a dedicated AI safety expert may be an important next step, as most cybersecurity experts are not trained in anticipating or preventing attacks against intelligent systems. I am hopeful that ongoing research will bring additional solutions for safely incorporating AI into the marketplace.

Roman V. Yampolskiy is a tenured associate professor in the department of computer engineering and computer science at the Speed School of Engineering, University of Louisville. He is the founding and current director of the university’s cybersecurity lab and an author of many books, including Artificial Superintelligence: a Futuristic Approach. Follow him on Twitter @romanyam.

COPYRIGHT © 2017 HARVARD BUSINESS SCHOOL PUBLISHING ALL RIGHTS RESERVED. ©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTSCORPORATION. RESERVED.

1404

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

Discussion Questions

AI Is the Future of Cybersecurity, for Better and for Worse BY ROMAN V. YAMPOLSKIY

1. Yampolskiy states that leaders can “rationally consider how the addition of AI to their product or service will enhance user experiences, while weighing the costs of potentially subjecting users to additional data breaches and possible dangers.” What measures does your company have in place to attain this knowledge and make decisions surrounding machine intelligence in this way? How might it further develop this ability? 2. Who has the responsibility to ensure that machine intelligence stays safe? Is considering and assessing the potentially catastrophic long-term consequences of AI beyond the scope of business? What actions would you like your business to take to ensure that AI doesn’t harm society?

The rise of AI-enabled cyberattacks is expected to cause an explosion of network penetrations, personal data thefts, and an epidemic-level spread of intelligent computer viruses. Ironically, our best hope to defend against AI-enabled hacking is by using AI.

COPYRIGHT © 2018 HARVARD BUSINESS SCHOOL PUBLISHING ALL RIGHTS RESERVED. ©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTSCORPORATION. RESERVED.

25 141

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

REPRINT H044HL PUBLISHED ON HBR.ORG JANUARY 26, 2018

ARTICLE TECHNOLOGY

How AI Could Help the Public Sector by Emma Martinho-Truswell

©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTS RESERVED.

142

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

TECHNOLOGY

How AI Could Help the Public Sector

by Emma Martinho-Truswell JANUARY 26, 2018

UPDATED JANUARY 29, 2018

omair khan/unsplash Last Thanksgiving, I watched my father-in-law evaluate over one hundred exams for the high school class he teaches on the U.S. government. They were mostly short answer questions: matching different provisions of the U.S. Constitution, and explaining the contents of the Bill of Rights. The grading was tedious and time consuming, and took him hour after hour during what should have been a holiday. I started to wonder whether there could be a faster way.

COPYRIGHT © 2018 HARVARD BUSINESS SCHOOL PUBLISHING ALL RIGHTS RESERVED. ©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTSCORPORATION. RESERVED.

1432

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

Automatic computer grading could do exactly that, learning from previous answers and getting better as it goes — and it is already being used in some universities and for large online courses (MOOCs). It could grade bundles of student papers quickly, perhaps flagging those with unusual elements that need a bit of human oversight. Teachers would get time back to plan new lessons, give extra tutorials to students who are struggling, do extra reading, or simply get their holiday time back. A public school teacher grading papers faster is a small example of the wide-ranging benefits that artificial intelligence could bring to the public sector. A.I could be used to make government agencies more efficient, to improve the job satisfaction of public servants, and to increase the quality of services offered. Talent and motivation are wasted doing routine tasks when they could be doing more creative ones. Applications of artificial intelligence to the public sector are broad and growing, with early experiments taking place around the world. In addition to education, public servants are using AI to help them make welfare payments and immigration decisions, detect fraud, plan new infrastructure projects, answer citizen queries, adjudicate bail hearings, triage health care cases, and establish drone paths. The decisions we are making now will shape the impact of artificial intelligence on these and other government functions. Which tasks will be handed over to machines? And how should governments spend the labor time saved by artificial intelligence? So far, the most promising applications of artificial intelligence use machine learning, in which a computer program learns and improves its own answers to a question by creating and iterating algorithms from a collection of data. This data is often in enormous quantities and from many sources, and a machine learning algorithm can find new connections among data that humans might not have expected. IBM’s Watson, for example, is a treatment recommendation-bot, sometimes finding treatments that human doctors might not have considered or known about. Machine learning program may be better, cheaper, faster, or more accurate than humans at tasks that involve lots of data, complicated calculations, or repetitive tasks with clear rules. Those in public service, and in many other big organizations, may recognize part of their job in that description. The very fact that government workers are often following a set of rules — a policy or set of procedures — already presents many opportunities for automation. To be useful, a machine learning program does not need to be better than a human in every case. In my work, we expect that much of the “low hanging fruit” of government use of machine learning will be as a first line of analysis or decision-making. Human judgment will then be critical to interpret results, manage harder cases, or hear appeals. When the work of public servants can be done in less time, a government might reduce its staff numbers, and return money saved to taxpayers — and I am sure that some governments will pursue that option. But it’s not necessarily the one I would recommend. Governments could instead choose to invest in the quality of its services. They can re-employ workers’ time towards more rewarding COPYRIGHT © 2018 HARVARD BUSINESS SCHOOL PUBLISHING ALL RIGHTS RESERVED. ©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTSCORPORATION. RESERVED.

1443

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

work that requires lateral thinking, empathy, and creativity — all things at which humans continue to outperform even the most sophisticated AI program. Deciding who qualifies for unemployment benefits, for example, is an important task with major consequences. Machine learning applications might speed up decisions, either giving a clear answer or indicating which cases need a human to take over. Sometimes, a citizen’s most valuable response from her government is a fast “yes” or “no.” At other times, the question might be more complicated. Perhaps someone has been unemployed for several months, and wants a longer conversation that includes some coaching, advice, and encouragement. A human will do this far better than a computer, and it might also be the best part of a public servant’s job: he gets to think about a new problem, and to truly help someone. On the other hand, asking a human to act like a computer, processing simple claims and hiding empathy or creativity, creates a tedious job for the government worker and a depressing experience for the citizen interacting with government. Writing as a former government worker — and now a full-time consultant for governments — I am very familiar with the high proportion of government work that is mundane. Complicated processes that leave little room for new ideas turn enthusiastic new public servants into cynics (and encourage them to leave government work). This is bad for public servants, but more importantly, it is bad for government. Regular surveys of trust in government, including by the OECD and Edelman, show that trust in government is low, and falling. Increasing the space for government workers to use their more human skills — empathy, creativity, and lateral thinking — may help. Humans are much better at this kind of thinking (and feeling) than machines are, and it is often the meaningful connection, the good sense, and the understanding that citizens are seeking when they deal with their government. If they are used well, artificial intelligence programs can make our government services faster and more tailored. The critical decision to be made by governments is how the time won by the best technology can be given back to citizens. At a time when many industries and jobs will change quickly, citizens may find that opportunities to have longer conversations with more engaged public servants may be much more important than a cheaper government. With thanks to Richard Stirling and Antone Martinho-Truswell. Editor’s note: this article has been updated to clarify the role of IBM Watson in making treatment recommendations.

Emma Martinho-Truswell is the co-founder and Chief Operating Officer of Oxford Insights, which advises organizations on the strategic, cultural, and leadership opportunities from digital transformation and artificial intelligence.

COPYRIGHT © 2018 HARVARD BUSINESS SCHOOL PUBLISHING ALL RIGHTS RESERVED. ©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTSCORPORATION. RESERVED.

1454

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

Discussion Questions

How AI Could Help the Public Sector BY EMMA MARTINHO-TRUSWELL

1. The author suggests that in the near future, AI programs, when applied in the public sector, will either free employees to “use their more human skills—empathy, creativity, and lateral thinking”—for the benefit of citizens or save money with a less expensive government. Which path do you think your own national and local governments will follow, and why? 2. Do you foresee similar questions about “how to apply the human time freed up by automation” arising in your company? Do you think the private sector and the public sector should address these questions in different ways?

Deciding who qualifies for unemployment benefits, for example, is an important task with major consequences. Machine learning applications might speed up decisions, either giving a clear answer or indicating which cases need a human to take over. Sometimes, a citizen’s most valuable response from her government is a fast “yes” or “no.”

COPYRIGHT © 2018 HARVARD BUSINESS SCHOOL PUBLISHING ALL RIGHTS RESERVED. ©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTSCORPORATION. RESERVED.

26 146

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

REPRINT H0449F PUBLISHED ON HBR.ORG JANUARY 24, 2018

ARTICLE TECHNOLOGY

How Will AI Change Work? Here Are 5 Schools of Thought by Mark Knickrehm

©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTS RESERVED.

147

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

TECHNOLOGY

How Will AI Change Work? Here Are 5 Schools of Thought

by Mark Knickrehm JANUARY 24, 2018

UPDATED JANUARY 25, 2018

Yaroslav Kushta/Getty Images The future of the workforce is one of the biggest issues facing CEOs today. It’s abundantly clear to all that artificial intelligence, big data analytics, and advanced robotics make it possible for machines to take on tasks that once required a person to do them. How should companies prepare, strategically, to thrive in this world?

COPYRIGHT © 2018 HARVARD BUSINESS SCHOOL PUBLISHING ALL RIGHTS RESERVED. ©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTSCORPORATION. RESERVED.

1482

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

Views on what to expect vary dramatically. By some accounts, almost half of all jobs in the U.S. economy could be made obsolete. Others have described how intelligent machines will actually create jobs — including entirely new categories of jobs. Some people even talk about a world of superabundance where work will be about pursuing your passion, on your own terms. It’s critical for companies to understand the range of opinions on this issue, because implicitly or explicitly, they will influence the way business leaders create the workforce of the future. And while a lot will shake out in years to come, this issue is already front and center. Companies are making decisions today that will matter hugely to their ability to compete tomorrow and throughout the 2020s. Most companies are already moving rapidly to acquire new capabilities. In a new Accenture survey (“Reworking the Revolution,” which published on January 23rd) of 1,200 C-level executives worldwide, 75% say that they are currently accelerating investments in AI and other intelligent technologies. And 72% say they are responding to a competitive imperative — they recognize the need for new tools to keep up with rivals, both by improving productivity and by finding new sources of growth. Some companies are transforming themselves into “intelligent enterprises,” in which all processes are digitized, decisions are data-driven, and machines do the heavy lifting — both physical and cognitive. So, there’s a great deal at stake in the debate over productivity and jobs. Leaders must understand the debate and be prepared to address tough questions: What kind of new skills do we need? How should we be organized? How do we define jobs? How can we bring our people along with us, in a way that benefits everyone? Through research, we’ve identified five schools of thought in this debate. The Dystopians Position: Man and machine will wage a Darwinian struggle that machines will win. AI systems will take on tasks at the heart of middle- and high-skill jobs, while robots will perform menial work that requires low-skill labor. The result will be massive unemployment, falling wages, and wrenching economic dislocation. Falling incomes will have grave consequences in places like the United States and Europe, where consumption accounts for 56% or 69% of GDP, respectively, requiring new social supports, such as a universal basic income. The Utopians Position: Intelligent machines will take on even more work, but the result will be unprecedented wealth, not economic decline. AI and computing power will advance in the next two decades to achieve “the singularity” — when machines will be able to emulate the workings of the human brain in its entirety. Human brains will be “scanned” and “downloaded” to computers and billions of replicated human brains will do most of the cognitive work, while robots will do all the heavy lifting. Economic output could double every three months. The singularity may even lead to a world where COPYRIGHT © 2018 HARVARD BUSINESS SCHOOL PUBLISHING ALL RIGHTS RESERVED. ©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTSCORPORATION. RESERVED.

1493

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

little human labor is required, a universal income program covers basic needs, and people apply their talents to meaningful pursuits. The Technology Optimists Position: A burst of productivity has already begun but is not captured in official data because companies are still learning how intelligent technologies can change how they operate. When companies do take full advantage of intelligent technologies, a leap in productivity will produce a digital bounty — creating both economic growth and improvements in living standards not counted in GDP, such as consumer surplus (from better, cheaper products) and the value of free apps and information. However, based on current trends, the bounty won’t be distributed evenly, and many jobs will be displaced. To avoid negative income and employment effects, there will be a need to invest in education and training alongside investments in technology. The Productivity Skeptics Position: Despite the power of intelligent technologies, any gains in national productivity levels will be low. Combine that with headwinds from aging populations, income inequality, and the costs of dealing with climate change, and the United States will have near-zero GDP growth. In the end, there isn’t much to do except brace for stagnant growth in advanced economies. The Optimistic Realists Position: Digitization and intelligent machines can spur productivity gains that match previous technology waves. Productivity will advance rapidly in certain sectors and for high-performing companies. New jobs will be created, but intelligent technologies may exacerbate the trends of the recent past, in which demand rose for both high- and low-skill workers whose jobs could be easily automated, while demand for middle-skill workers fell. With no simple solutions, more research is needed into the true relationship between productivity, employment, and wages to uncover effective responses.

Three Actions for Shaping the Future Our crystal ball for what things might look like in 10 years is cloudy. What we do know is that business leaders must take steps now to shape their workforces for the emerging intelligent enterprise. Our research and experience point to three critical imperatives: Use technology to augment human skills and reinvent operating models. Companies that think beyond labor substitution and cost savings will see a much greater payoff. For example, a new class of adaptive robots can function safely alongside workers and can take on difficult and tedious work. Consider this example: At BMW’s Spartanburg, S.C. plant, robots are installing door-sealing gaskets, an awkward and tiring job for workers. This speed up the line, improves quality, and gives workers more time to do higher-value work. Researchers estimate that using adaptive robots this way could cut time wasted on non-value-added work by 25%. Employee surveys show that workers have more positive views of the new robots, which they regard as useful helpers. Away from the factory, companies are using AI to offload routine work from employees and to give them new analytical tools COPYRIGHT © 2018 HARVARD BUSINESS SCHOOL PUBLISHING ALL RIGHTS RESERVED. ©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTSCORPORATION. RESERVED.

1504

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

to improve customer experience and discover new possibilities for products, services, and business models that drive growth. Take the opportunity to redefine jobs and rethink organizational design. Companies cannot optimize their investments if they have the same old job descriptions and organizational structures. Executives should assess the tasks that need to be done, anticipate which ones will be transferred to machines, then reconfigure jobs by adding new tasks or creating entirely different roles that are needed for managing intelligent technologies. A factory worker, for example, can be trained to run robots. AI systems also need human help to train and correct algorithms and override fallible machine judgment. For example, at Stitch Fix, an online clothing subscription service, 3,400 human stylists work with an AI recommendation engine to make personalized suggestions for customers. The machines give stylists the speed they need to be productive, and the stylists provide the additional judgment needed for accurate recommendations (and fewer returns). To function effectively, an intelligent enterprise should have a non-hierarchical organization, in which employees collaborate across functional and operational silos. This enables the intelligent enterprise to act quickly on the insights from data-crunching machines and deploy human talent to swarm on problems, experiment, iterate, and get solutions into the market. Make employees your partners in building the intelligent enterprise. To strike the right balance between investing in intelligent technologies and maintaining existing businesses, companies need help from their employees. In our (above-referenced) research, we have found that employees are far more willing — even eager — to master new technologies than employers appreciate. They want to learn new skills, not least because they know they will need them to remain employed. Investments in both technology and training will help companies make a smooth transition to the intelligent enterprise. Companies that do this stand to outperform competitors because they will unleash the human talents that machines still can’t match and that are essential to growth — creativity, empathy, communications, adaptability, and problem-solving. “As basic automation and machine learning move toward becoming commodities,” says Devin Fidler, research director at the Institute for the Future, “uniquely human skills will become more valuable.” The debate over technology and jobs will rage on. Business leaders must follow this debate — and participate in it, too. And much more research is needed to fully understand the implications of intelligent technologies on work. In the meantime, companies that actively seize control of what can be done to prepare will position themselves to thrive in this exciting new era. The author thanks his colleagues in Accenture Research, Svenja Falk, David Light, and Geoffrey Lewis, for their contributions to this article. Editors’ note: We’ve updated this article with the correct number of human stylists at Stitch Fix.

Mark Knickrehm is group chief executive for Accenture Strategy.

COPYRIGHT © 2018 HARVARD BUSINESS SCHOOL PUBLISHING ALL RIGHTS RESERVED. ©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTSCORPORATION. RESERVED.

1515

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

Discussion Questions

How Will AI Change Work? Here Are 5 Schools of Thought BY MARK KNICKREHM

1. Which of the five “schools of thought” seems most likely to you? Would you have answered differently before reading this collection of articles? 2. Do your company’s senior executives subscribe to any one of these five beliefs? What assumptions are they making about the future of AI when they set company strategy? 3. Has your company begun to reshape its workforce for an AI-driven future? What would be the benefits and risks of redefining jobs and rethinking organizational design now? What pros and cons are there to taking a wait-and-see approach? 4. Knickrehm says that “the debate over technology and jobs will rage on” and that business leaders must follow and participate in the debate. Where does your company see itself in this debate? Is it looking to lead or follow other companies?

The debate over technology and jobs will rage on. Business leaders must follow this debate—and participate in it, too. And much more research is needed to fully understand the implications of intelligent technologies on work. In the meantime, companies that actively seize control of what can be done to prepare will position themselves to thrive in this exciting new era.

COPYRIGHT © 2018 HARVARD BUSINESS SCHOOL PUBLISHING ALL RIGHTS RESERVED. ©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTSCORPORATION. RESERVED.

27 152

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

Related Resouces

HBR’s 10 Must Reads on AI, Analytics, and the New Machine Age The Big Idea: The Business of Artificial Intelligence The Big Idea: Internet Insecurity Prediction Machines: The Simple Economics of Artificial Intelligence by Ajay Agrawal, Joshua Gans, and Avi Goldfarb Human + Machine: Reimagining Work in the Age of AI by Paul R. Daugherty and H. James Wilson Competing on Analytics, Updated, with a New Introduction: The New Science of Winning by Thomas H. Davenport and Jeanne G. Harris Reinventing Jobs: A 4-Step Approach for Applying Automation to Work by Ravin Jesuthasan and John W. Boudreau Smart Business: What Alibaba’s Success Reveals About the Future of Strategy by Ming Zeng

©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTS RESERVED.

28 153

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.

Product #ARTINT Questions? Contact [email protected] or tel. 800-988-0886. Feedback? We want to ensure this product meets your needs. Please share any comments about this collection or topics you’d like to see us cover in the future at [email protected]. Requests for permission should be directed to [email protected]. Harvard Business Review Press products are available at significant quantity discounts when purchased in bulk for client gifts, sales promotions, and premiums. Special editions, including versions with corporate logos, customized cover pages, and letters from the company or CEO included, can also be created in large quantities for special needs. For details and discount information, contact [email protected], tel. 800-988-0886, or www.hbr.org/bulksales. www.hbr.org ©2018 HARVARD BUSINESS SCHOOL PUBLISHING. ALL RIGHTS RESERVED.

29 154

This document is authorized for use only by Jo Teichmann ([email protected]). Copying or posting is an infringement of copyright. Please contact [email protected] or 800-988-0886 for additional copies.