MGI Artificial Intelligence Discussion Paper PDF

ARTIFICIAL
INTELLIGENCE
THE NEXT DIGITAL
FRONTIER?
DISCUSSION PAPER
JUNE 2017
Jacques Bughin | Brussels

Eric Hazan | Paris
Sree Ramaswamy | Washington, DC
Michael Chui | San Francisco
Tera Allas | London
Peter Dahlstrm | London
Nicolaus Henke | London
Monica Trench | London
Since its founding in 1990, the McKinsey Global Institute (MGI) has sought
to develop a deeper understanding of the evolving global economy. As the
business and economics research arm of McKinsey & Company, MGI aims
to provide leaders in the commercial, public, and social sectors with the facts
and insights on which to base management and policy decisions. The Lauder
Institute at the University of Pennsylvania ranked MGI the worlds number-one
private-sector think tank in its 2016 Global Think Tank Index for the second
consecutive year.
MGI research combines the disciplines of economics and management,

employing the analytical tools of economics with the insights of business
leaders. Our micro-to-macro methodology examines microeconomic
industry trends to better understand the broad macroeconomic forces
affecting business strategy and public policy. MGIs in-depth reports have
covered more than 20 countries and 30 industries. Current research focuses
on six themes: productivity and growth, natural resources, labor markets, the
evolution of global financial markets, the economic impact of technology and
innovation, and urbanization.
Recent reports have assessed the economic benefits of tackling gender

inequality, a new era of global competition, Chinese innovation, and
digital globalization. MGI is led by four McKinsey and Company senior
partners: JacquesBughin, JamesManyika, JonathanWoetzel, and
FrankMattern, MGIs chairman. MichaelChui, SusanLund, AnuMadgavkar,
SreeRamaswamy, and JaanaRemes serve as MGI partners. Project
teams are led by the MGI partners and a group of senior fellows and include
consultants from McKinsey offices around the world. These teams draw on
McKinseys global network of partners and industry and management experts.
Input is provided by the MGI Council, which coleads projects and provides
guidance; members are AndresCadena, SandrineDevillard, RichardDobbs,
KatyGeorge, RajatGupta, EricHazan, EricLabaye, AchaLeke, ScottNyquist,
GaryPinkus, OliverTonby, and EckartWindhagen. In addition, leading
economists, including Nobel laureates, act as research advisers.
McKinsey & Company is a member of the Partnership on AI, a collection of

companies and non-profits that have committed to sharing best practices
and communicating openly about the benefits and risks of artificial
intelligence research. The partners of McKinsey fund MGIs research; it is not
commissioned by any business, government, or other institution.
For further information about MGI and to download reports, please visit
www.mckinsey.com/mgi.
Copyright McKinsey & Company 2017
1 McKinsey Global Institute

CONTENTS
Preface
In brief
1. Artificial intelligence is getting ready for business, but

are businesses ready for AI?
Page 6
2. Artificial intelligence promises to boost profits and

transform industries
Page 20
3. Businesses, developers, and governments need to act

now to realize AIs full potential
Page 31
Appendix A: Five case studies

Page 42
Retail
Page 42
Electric utility
Page 47
Manufacturing
Page 53
Health care
Page 58
Education
Page 65
Appendix B: Technical appendix

Page 70
Bibliography
Page 73
PREFACE
In this independent discussion paper, we examine investment in artificial intelligence (AI),

describe how it is being deployed by companies that have started to use these technologies
across sectors, and aim to explore its potential to become a major business disrupter. To
do this, we looked at AI through several lenses. We analyzed the total investment landscape
bringing together both investment of large corporations and funding from venture capital
and private equity funds. We also reviewed the portfolio plays of major internet companies,
the dynamics in AI ecosystems from Shenzhen to New York, and a wide range of case
studies. As part of our primary research, we surveyed more than 3,000 senior executives
on the use of AI technologies, their companies prospects for further deployment, and AIs
impact on markets, governments, and individuals. This report also leverages the resources
of McKinsey Analytics, a global practice that helps clients achieve better performance
through data. The research was conducted jointly with Digital McKinsey, a global practice
that designs and implements digital transformations.
In addition to identifying a gap between AI investment and commercial application, which

is typical of early technology development curves, we found that the new generation of AI
applications is based on the foundation of digitization. Leading sectors in digital tend to be
leading sectors in AI, and these are predicted to drive growth. We also found that AI has
the potential to accelerate shifts in market share, revenue, and profit poolsall hallmarks
of digitally disrupted sectors. This report leverages two MGI analyses of digitization, Digital
America: A tale of the haves and have-mores, published in December 2015, and Digital
Europe: Pushing the frontier, capturing the benefits, published in June 2016. These reports
introduced the McKinsey Global Institute (MGI) Industry Digitization Index, which combines
dozens of indicators to provide a comprehensive picture of where and how companies are
building digital assets, expanding digital usage, and creating a more digital workforce. This
report also builds on MGIs work on advanced analytics, The age of analytics: Competing
in a data-driven world, published in December 2016, and on automation, A future that
works: Automation, employment, and productivity, published in January 2017, as well as
Artificial intelligence: Implications for China, published in April 2017; and an April 2017 Digital
McKinsey report, Smartening up with artificial intelligence (AI): Whats in it for Germany and
its industrial sector?
This latest research has been led by JacquesBughin, an MGI senior partner based in
Brussels; EricHazan, a member of the MGI Council and a McKinsey senior partner based
in Paris; SreeRamaswamy, an MGI partner based in Washington, DC; MichaelChui,
an MGI partner based in San Francisco; TeraAllas, an MGI visiting fellow in London;
PeterDahlstrm, a McKinsey senior partner in London; NicolausHenke, a McKinsey senior
partner based in London; and MonicaTrench, a McKinsey consultant based in London. The
project team comprised MathildeCastet, FranoisAllaindesBeauvais, LindsayMacdonald,
OlegPynda, DariuszSmolen, and JordanWard. Sincere thanks go to TimothyBeacom,
AprilCheng, Paul-LouisCaylar, and HugoWeber. We would also like to thank MGI senior
editor MarkA.Stein; MattCooke, MGI director of external communications; MGI visual
graphics specialist MarisaCarder, designer MargoShimasaki, and infographic designers
RichardJohnson and JasonLeder; MGI editorial production manager JuliePhilpot; and
DeadraHenderson, MGI manager of personnel and administration.
This report builds on a considerable body of expertise within MGI and McKinsey. We
particularly want to acknowledge TamimSaleh and BrianMcCarthy from McKinsey
Analytics, LouiseHerring and CasperLouw for contributing to the retail sector case
study, ArnoutdePee and MikeMunroe for contributing to the electric utilities case study,
RichardKelly for contributing to the manufacturing case study, MartinMarkus and
SriVelamoor for contributing to the health care case study, and JakeBryant, MikeMunroe,
and JimmySarakatsannis for contributing to the education case study. We would also like
to thank all previous MGI teams that produced reports on digitization, automation, big data
and analytics, the internet of things, and online talent platforms.
Our research was also enriched by insights from EricGoubault and JesseRead from
Ecole Polytechnique.
This report contributes to MGIs mission to help business and policy leaders understand the
forces transforming the global economy, identify strategic imperatives, and prepare for the
next wave of growth. As with all MGI research, this work is independent and has not been
commissioned or sponsored in any way by any business, government, or other institution.
We welcome your comments on the research at MGI@mckinsey.com.
Jacques Bughin
Director, McKinsey Global Institute
Brussels
Eric Hazan
Co-leader, Digital McKinsey
Senior Partner, McKinsey and Company
Member, McKinsey Global Institute Council
Paris
James Manyika
San Francisco
Jonathan Woetzel
Shanghai
June 2017
IN BRIEF
ARTIFICIAL INTELLIGENCE:
THE NEXT DIGITAL FRONTIER?
Artificial intelligence is poised to unleash the next wave vehicles and improve operations, for example, while
of digital disruption, and companies should prepare for financial services firms are more likely to use it in
it now. We already see real-life benefits for a few early- customer experiencerelated functions.
adopting firms, making it more urgent than ever for
Early evidence suggests that AI can deliver real
others to accelerate their digital transformations. Our
value to serious adopters and can be a powerful
findings focus on five AI technology systems: robotics
force for disruption. In our survey, early AI adopters
and autonomous vehicles, computer vision, language,
that combine strong digital capability with proactive
virtual agents, and machine learning, which includes deep
strategies have higher profit margins and expect
learning and underpins many recent advances in the
the performance gap with other firms to widen
other AI technologies.
in the future. Our case studies in retail, electric
AI investment is growing fast, dominated by digital utilities, manufacturing, health care, and education
giants such as Google and Baidu. Globally, we highlight AIs potential to improve forecasting and
estimate tech giants spent $20billion to $30billion on sourcing, optimize and automate operations, develop
AI in 2016, with 90percent of this spent on R&D and targeted marketing and pricing, and enhance the
deployment, and 10percent on AI acquisitions. VC user experience.
and PE financing, grants, and seed investments also
AIs dependence on a digital foundation and the fact
grew rapidly, albeit from a small base, to a combined
that it often must be trained on unique data mean that
total of $6billion to $9billion. Machine learning, as
there are no shortcuts for firms. Companies cannot
an enabling technology, received the largest share of
delay advancing their digital journeys, including AI.
both internal and external investment.
Early adopters are already creating competitive
AI adoption outside of the tech sector is at an early, advantages, and the gap with the laggards looks
often experimental stage. Few firms have deployed set to grow. A successful program requires firms to
it at scale. In our survey of 3,000 AI-aware C-level address many elements of a digital and analytics
executives, across 10 countries and 14 sectors, transformation: identify the business case, set up
only 20percent said they currently use any AI- the right data ecosystem, build or buy appropriate AI
related technology at scale or in a core part of their tools, and adapt workflow processes, capabilities, and
businesses. Many firms say they are uncertain of the culture. In particular, our survey shows that leadership
business case or return on investment. A review of from the top, management and technical capabilities,
more than 160 use cases shows that AI was deployed and seamless data access are key enablers.
commercially in only 12percent of cases.
AI promises benefits, but also poses urgent
Adoption patterns illustrate a growing gap between challenges that cut across firms, developers,
digitized early AI adopters and others. Sectors at the government, and workers. The workforce needs to be
top of MGIs Industry Digitization Index, such as high reskilled to exploit AI rather than compete with it; cities
tech and telecom or financial services, are also leading and countries serious about establishing themselves
adopters of AI. They also have the most aggressive as a global hub for AI development will need to join the
AI investment intentions. Leaders adoption is both global competition to attract AI talent and investment;
broad and deep: using multiple technologies across and progress will need to be made on the ethical, legal
multiple functions, with deployment at the core of their and regulatory challenges that could otherwise hold
business. Automakers use AI to develop self-driving back AI.
Download the full report at www.mckinsey.com/mgi

ARTIFICIAL INTELLIGENCE
The next digital frontier?
The current AI wave is poised to finally break through

Investment in AI is growing at a high rate, but adoption in 2017 remains low
In 2016, companies invested

20% oftheyAI-aware firms say
are adopters
10%
$26B to $39B 3+ technologies

2 technologies
3%
7%
in artificial intelligence 1 technology 40%
Contemplators
TECH GIANTS STARTUPS

$20B to $30B $6B to $9B 41% of firms 31%
Partial adopters
say they are uncertain
about the benefits of AI
10%
3x External investment growth since 2013

Experi-
menters
How companies are adopting AI Six characteristics

AI adoption is greatest in sectors that are already of early AI adopters
strong digital adopters
Assets Larger
High tech / telecom Digitally mature
High AI businesses
adoption
Automotive / assembly
Financial services Us
a
ge
Medium Retail Adopt AI in Adopt multiple

AI Media / entertainment core activities technologies
adoption CPG
Labo
Education
Low AI Health care Focus on growth C-level
r
adoption
Travel / tourism over savings support for AI
Digital maturity
Four areas across the value chain where AI can create value
PRODUCE:
PROJECT: PROMOTE: PROVIDE:
Optimized
Smarter R&D and Targeted sales Enhanced user
production and
forecasting and marketing experience
maintenance
Five elements
of successful AI
transformations Use cases / Data Techniques Workflow Open culture
sources of value ecosystems and tools integration and organization
1. ARTIFICIAL INTELLIGENCE IS
GETTING READY FOR BUSINESS, BUT
ARE BUSINESSES READY FOR AI?
Claims about the promise and peril of artificial intelligence are abundant, and growing.
AI,which enables machines to exhibit human-like cognition, can drive our cars or steal our
privacy, stoke corporate productivity or empower corporate spies. It can relieve workers
of repetitive or dangerous tasks or strip them of their livelihoods. Twice as many articles
mentioned AI in 2016 as in 2015, and nearly four times as many as in 2014.1 Expectations
are high.
AI has been here before. Its history abounds with booms and busts, extravagant promises
and frustrating disappointments. Is it different this time? New analysis suggests yes: AI
is finally starting to deliver real-life business benefits. The ingredients for a breakthrough
are in place. Computer power is growing significantly, algorithms are becoming more
sophisticated, and, perhaps most important of all, the world is generating vast quantities of
the fuel that powers AIdata. Billions of gigabytes of it every day.
Companies at the digital frontieronline firms and digital natives such as Google and
Baiduare betting vast amounts of money on AI. We estimate between $20billion and
$30billion in 2016, including significant M&A activity. Private investors are jumping in, too.
We estimate that venture capitalists invested $4billion to $5billion in AI in 2016, and private
equity firms invested $1billion to $3billion. That is more than three times as much as in
2013. An additional $1billion of investment came from grants and seed funding.
For now, though, most of the news is coming from the suppliers of AI technologies. And
many new uses are only in the experimental phase. Few products are on the market or
are likely to arrive there soon to drive immediate and widespread adoption. As a result,
analysts remain divided as to the potential of AI: some have formed a rosy consensus
about AIs potential while others remain cautious about its true economic benefit. This
lack of agreement is visible in the large variance of current market forecasts, which range
from $644million to $126billion by 2025.2 Given the size of investment being poured into
AI, the low estimate would indicate that we are witnessing another phase in a boom-and-
bust cycle.
Our business experience with AI suggests that this bust scenario is unlikely. In order to
provide a more informed view, we decided to perform our own research into how users
are adopting AI technologies. Our research offers a snapshot of the current state of the
rapidly changing AI industry, looking through the lenses of both suppliers and users to
come up with a more robust view of the economic potential of AI and how it will unfold. To
begin, we examine the investment landscape, including firms internal investment in R&D
and deployment, large corporate M&A, and funding from venture capital (VC) and private
equity (PE) firms. We then look at the demand side, combining use case analyses and our AI
adoption and use survey of C-level executives at more than 3,000 companies to understand
how companies use AI technologies today, what is driving their adoption of AI, the barriers to
further deployment, and the market, financial, and organizational impacts of AI. For further
details on sources of our insights, see Box1, A multi-lens approach to understanding the
AI story.
1
Factiva.
2
Tractica; Transparency Market Research.
6 McKinsey Global Institute Artificial intelligence: The next digital frontier?

AI generally refers to the ability of machines to exhibit human-like intelligencefor example,
solving a problem without the use of hand-coded software containing detailed instructions.
There are several ways to categorize AI technologies, but it is difficult to draft a list that is
mutually exclusive and collectively exhaustive because people often mix and match several
technologies to create solutions for individual problems. These creations sometimes are
treated as independent technologies, sometimes as subgroups of other technologies, and
sometimes as applications. Some frameworks group AI technologies by basic functionality,
such as text, speech, or image recognition, and some group them by business applications
such as commerce or cybersecurity.3
3
Gil Press, Top 10 hot artificial intelligence (AI) technologies, Forbes.com, January 23, 2017; AI100: The
artificial intelligence startups redefining industries, CBinsights.com, January 11, 2017.
Box1. A multi-lens approach to understanding the AI story

For the findings presented in this report, we drew on both Internal investment flows are based on an analysis of the
primary and secondary research. We used six distinct top 35 companies investing in AI globally.
sources of insight to assess the state of AI and its future
Investment hubs. We built a global picture of AI
potential. See Appendix B for further details.
innovation activity, covering 10 countries and over 75
AI adoption and use survey. We surveyed AI-aware cities, and evaluating their local AI ecosystems. This
C-level executives at 3,073 companies about how they included mapping AI investment flows from VC, PE,
are using digital technology and AI today, the drivers and and M&A databases by city, scanning large AI-investing
barriers to further deployment, and the market, financial, companies activities by geography, assessing the
and organizational impacts of AI. Our stratified sample research outputs of and talent flows from universities,
covered 14 sectors of the economy, 10 countries across evaluating the broader business environment for
Europe, North America, and Asia, and companies with entrepreneurs, and interviewing local investors.
workforces ranging from fewer than 10 to more than
Sector case studies. In five sectors, we conducted
10,000. Responses were verified using algorithmic data
industry-expert interviews to understand the specific use
cleansing techniques.
cases today and promising applications for the future.
Use case review. We collated and reviewed over 160 External interviews were complemented by unique
use cases from both public sources and databases insights from across McKinsey sector practices and
assembled for related MGI and Digital McKinsey research. functions, including Digital McKinsey and McKinsey
Use cases were individually assessed to determine the Analytics, as well as advanced-analytics firms that
extent to which they had achieved commercialization, McKinsey has acquired or partnered with.
and were classified according to their primary sector and
Previous MGI and McKinsey research. This report
business function of use.
leverages other recent major research efforts by MGI and
Investment flows. To measure investment in the McKinsey Analytics, including databases of use cases.
development and deployment of new AI technology, In particular, the findings in this report complement and
products, and services, we conducted an extensive build on the following research: A future that works:
review of publicly available information on both external Automation, employment, and productivity (2017), The
funding sources (VC, PE, and M&A), as well as R&D and age of analytics: Competing in a data-driven world (2016),
deployment activities internal to large corporations. Our Digital Europe: Pushing the frontier, capturing the benefits
assessment of external investment is based on global (2016), and Digital America: A tale of the haves and have-
deal databases, scrutinizing deals for AI content, and mores (2015).
classifying them by type of technology and geography.
McKinsey Global Institute Artificial intelligence: The next digital frontier? 7

Trying to pin down the term more precisely is fraught for several reasons: AI covers a broad
range of technologies and applications, some of which are merely extensions of earlier
techniques and others that are wholly new. Also, there is no generally accepted theory
of intelligence, and the definition of machine intelligence changes as people become
accustomed to previous advances.4 Teslers theorem, attributed to the computer scientist
Larry Tesler, asserts that AI is whatever hasnt been done yet.5
The AI technologies we consider in this paper are what is called narrow AI, which performs
one narrow task, as opposed to artificial general intelligence, or AGI, which seeks to be able
to perform any intellectual task that a human can do. We focus on narrow AI because it has
near-term business potential, while AGI has yet to arrive.6
In this report, we focus on the set of AI technology systems that solve business problems.
We have categorized these into five technology systems that are key areas of AI
development: robotics and autonomous vehicles, computer vision, language, virtual agents,
and machine learning, which is based on algorithms that learn from data without relying on
rules-based programming in order to draw conclusions or direct an action. Some are related
to processing information from the external world, such as computer vision and language
(including natural language processing, text analytics, speech recognition, and semantics
technology); some are about learning from information, such as machine learning; and
others are related to acting on information, such as robotics, autonomous vehicles, and
virtual agents, which are computer programs that can converse with humans. Machine
learning and a subfield called deep learning are at the heart of many recent advances
in artificial intelligence applications and have attracted a lot of attention and a significant
share of the financing that has been pouring into the AI universealmost 60percent of all
investment from outside the industry in 2016.
ARTIFICIAL INTELLIGENCES ROLLER-COASTER RIDE TO TODAY

Artificial intelligence, as an idea, first appeared soon after humans developed the
electronic digital computing that makes it possible. And, like digital technology, artificial
intelligence, or AI, has ridden waves of hype and gloomwith one exception: AI has not yet
experienced wide-scale commercial deployment (see Box2, Fits and starts: A history of
artificial intelligence).
That may be changing. Machines powered by AI can today perform many taskssuch
as recognizing complex patterns, synthesizing information, drawing conclusions, and
forecastingthat not long ago were assumed to require human cognition. And as AIs
capabilities have dramatically expanded, so has its utility in a growing number of fields. At
the same time, it is worth remembering that machine learning has limitations. For example,
because the systems are trained on specific data sets, they can be susceptible to bias; to
avoid this, users must be sure to train them with comprehensive data sets. Nevertheless, we
are seeing significant progress.
4
Marvin Minsky, Steps toward artificial intelligence, Proceedings of the IRE, volume 49, number 1, January
1961; Edward A. Feigenbaum, The art of artificial intelligence: Themes and case studies of knowledge
engineering, Stanford University Computer Science Department report number STAN-CS-77621, August
1977; Allen Newell, Intellectual issues in the history of artificial intelligence, in The Study of Information:
Interdisciplinary messages, Fritz Machlup and Una Mansfield, eds., John Wiley and Sons, 1983.
5
Douglas R. Hofstadter, Gdel, Escher, Bach: An eternal golden braid, Basic Books, 1979. Hofstadter writes
that he gave the theorem its name after Tesler expressed the idea to him firsthand. However, Tesler writes in
his online CV that he actually said, Intelligence is whatever machines havent done yet.
6
William Vorhies, Artificial general intelligencethe Holy Grail of AI, DataScienceCentral.com, February 23,
2016.

Box2. Fits and starts: A history of artificial intelligence
The idea of computer-based artificial intelligence dates to Funding slumped in the 1970s as research backers,
1950, when Alan Turing proposed what has come to be primarily the US government, tired of waiting for practical
called the Turing test: can a computer communicate well AI applications and cut appropriations for further work.5
enough to persuade a human that it, too, is human? A 1
The field was fallow for the better part of a decade.
few months later, Princeton students built the first artificial
University researchers development of expert
neural network, using 300 vacuum tubes and a war-
systemssoftware programs that assess a set of facts
surplus gyropilot.2
using a database of expert knowledge and then offer
The term artificial intelligence was coined in 1955, to solutions to problemsrevived AI in the 1980s.6 Around
describe the first academic conference on the subject, at this time, the first computer-controlled autonomous
Dartmouth College. That same year, researchers at the vehicles began to appear.7 But this burst of interest
Carnegie Institute of Technology (now Carnegie Mellon preceded another AI winter.
University) produced the first AI program, Logic Theorist. 3
Interest in AI boomed again in the 21st century as
Advances followed often through the 1950s: Marvin Lee
advances in fields such as deep learning, underpinned by
Minsky founded the Artificial Intelligence Laboratory
faster computers and more data, convinced investors and
at MIT, while others worked on semantic networks for
researchers that it was practicaland profitableto put
machine translation at Cambridge and self-learning
AI to work.8
software at IBM.4
1
A. M. Turing, Computing machinery and intelligence, Mind, volume 49, number 236, October 1950.
2
Jeremy Bernstein, A.I., The New Yorker, December 14, 1981.
3
Leo Gugerty, Newell and Simons Logic Theorist: Historical background and impact on cognitive modeling, Proceedings of the Human
Factors and Ergonomics Society Annual Meeting, volume 50, issue 9, October 2006.
4
The IBM 700 Series: Computing comes to business, IBM Icons of Progress, March 24, 2011.
5
Michael Negnevitsky, Artificial intelligence: A guide to intelligent systems, Addison-Wesley, 2002.
6
Edward A. Feigenbaum, Expert systems in the 1980s, working paper, 1980.
7
Hans P. Moravec, The Stanford Cart and the CMU Rover, Proceedings of the IEEE, volume 71, issue 7, July 1983; Tom Vanderbilt,
Autonomous cars through the ages, Wired.com, February 6, 2012.
8
Buchanan, Bruce G., A (very) brief history of artificial intelligence, AI Magazine, volume 26, number 4, Winter 2005.
These advances have allowed machine learning to be scaled up since 2000 and used to
drive deep learning algorithms, among other things. The advances have been facilitated
by the availability of large and diverse data sets, improved algorithms that find patterns
in mountains of data, increased R&D financing, and powerful graphics processing units
(GPUs), which have brought new levels of mathematical computing power. GPUs, which are
specialized integrated circuits originally developed for video games, can process images
40 to 80 times faster than the fastest versions available in 2013. Advances in the speed of
GPUs have enabled the training speed of deep learning systems to improve five- or sixfold
in each of the last two years. More datathe world creates about 2.2 exabytes, or 2.2billion
gigabytes, of it every daytranslates into more insights and higher accuracy because it
exposes algorithms to more examples they can use to identify correct and reject incorrect
answers. Machine learning systems enabled by these torrents of data have reduced
computer error rates in some applicationsfor example, in image identificationto about
the same as the rate for humans.
LED BY TECH GIANTS, AI INVESTMENT IS GROWING RAPIDLY,

BUT COMMERCIAL ADOPTION IS LAGGING BEHIND
Tech giants and digital native companies such as Amazon, Apple, Baidu, and Google
are investing billions of dollars in the various technologies known collectively as artificial
intelligence. They see that the inputs needed to enable AI to finally live up to expectations
powerful computer hardware, increasingly sophisticated algorithmic models, and a vast
and fast-growing inventory of dataare in place. Indeed, internal investment by large
corporations dominates: we estimate that this amounted to $18billion to $27billion in

2016; external investment (from VCs, PE firms, M&A, grants, and seed funding) was around
$8billion to $12billion (Exhibit1).7
Exhibit 1
Technology giants dominate investment in AI
Investment in AI, 20161 Investment by tech giants and other corporations

$ billion
Compound annual AI share of

growth rate2 total
% investment
category, 20163
812 201013 201316 %
Other
M&A2
23
M&A 55 85 <1
Internal
corporate 1827
investment1 69 VC, PE, and
other external Private
funding3 equity 15 20 13
Venture
capital 35 40 23
1 Estimate of 2016 spend by corporations to develop and deploy AI-based products. Calculated for top 35 high tech and advanced manufacturing companies
investing in AI. Estimate is based on the ratio of AI spend to total revenue calculated for a subset of the 35 companies.
2 VC value is an estimate of VC investment in companies primarily focused on AI. PE value is an estimate of PE investment in AI-related companies. M&A
value is an estimate of AI deals done by corporations. Other refers to grants and seed fund investments. Includes only disclosed data available in
databases, and assumes that all registered deals were completed within the year of transaction. Compound annual growth rate values rounded.
3 M&A and PE deals expressed by volume; VC deals expressed by value.
SOURCE: Capital IQ; Pitchbook; Dealogic; S&P; McKinsey Global Institute analysis
But for all the recent investment, the scope of AI deployment has been limited so far. That is
partly due to the fact that one beneficiary of that investment, internal R&D, is largely focused
on improving the firms own performance. But it is also true that there is only tepid demand
for artificial intelligence applications for businesses, partly due to the relatively slow pace
of digital and analytics transformation of the economy. Our survey of more than 3,000
businesses around the world found that many business leaders are uncertain about what
exactly AI can do for them, where to obtain AI-powered applications, how to integrate them
into their companies, and how to assess the return on an investment in the technology.
Vivatech Internal investment includes research and development, talent acquisition, cooperation with scientific
7
institutions, and joint ventures with other companies done by corporations. External investment includes
Discussionmergers
paper and acquisitions, private equity funding, venture capital financing, and seed funds and other
early-stage investing. The estimates of external investment are based on data available in the Capital IQ,
0606 mc PitchBook, and Dealogic databases. Provided values are estimates of annual investment in AI, assuming that
all registered deals were completed within the year of transaction. Internal investment is estimated based on
the ratio of AI spend to revenue for the top 35 high tech and advanced manufacturing companies focused on
AI technologies.

Most of the investment in AI has consisted of internal spendingR&D and deployment
by large, cash-rich digital native companies. What is the large corporate investment in
AI focused on? Bigger companies, such as Apple, Baidu, and Google, are working on
suites of technologies internally, but vary in the breadth and focus of their AI investment.
Amazon is working on robotics and speech recognition, Salesforce on virtual agents and
machine learning. BMW, Tesla, and Toyota are among the manufacturers making sizable
commitments in robotics and machine learning for use in driverless cars. Toyota, for
example, set aside $1billion to establish a new research institute devoted to AI for robotics
and driverless vehicles.8 Industrial giants such as ABB, Bosch, GE, and Siemens also are
investing internally, often in machine learning and robotics, seeking to develop specific
technologies related to their core businesses. IBM has pledged to invest $3billion to
make its Watson cognitive computing service a force in the internet of things.9 Baidu has
invested $1.5billion in AI research over the last 2 years. This is in addition to $200million it
committed to a new in-house venture capital fund, Baidu Venture.10
At the same time, big tech companies have been actively buying AI startups, not just to
acquire technology or clients but to secure qualified talent. The pool of true experts in the
field is small, and Alibaba, Amazon, Facebook, Google, and other tech giants have hired
many of them. Companies have adopted M&A as a way to sign up top talent, a practice
known as acqui-hiring, for sums that typically work out to $5million to $10million per
person. The shortage of talent and cost of acquiring it are underlined by a recent report
that companies are seeking to fill 10,000 AI-related jobs and have budgeted more than
$650million for salaries.11
Overall, corporate M&A is the fastest-growing external source of funding for AI companies,
increasing in terms of value at a compound annual growth rate of over 80percent from
2013 to 2016, based on our estimates. Leading high tech companies and advanced
manufacturers have closed more than 100 M&A deals since 2010. Google completed
24 transactions in that time, including eight in computer vision and seven in language
processing. Apple, the second-most-active acquirer, has closed nine, split evenly among
computer vision, machine learning, and language processing.
Companies are also expanding their search for talent abroad. Facebook, for instance,
is opening an AI lab in Paris that will supplement similar facilities in New York and Silicon
Valleyand make it easier for the company to recruit top researchers in Europe.12 Google
recently invested $4.5million in the Montreal Institute for Learning Algorithms, a research
lab at the University of Montreal; Intel donated $1.5million to establish a machine learning
and cybersecurity research center at Georgia Tech; and NVIDIA is working with the National
Taiwan University to establish an AI laboratory in Taipei. 13
8
Craig Trudell and Yuki Hagiwara, Toyota starts $1billion center to develop cars that dont crash, Bloomberg.
com, November 6, 2015.
9
IBM invests to lead global internet of things marketshows accelerated client adoption, IBM press release,
October 3, 2006.
10
Phoenix Kwong, Baidu launches $200m venture capital unit focused on artificial intelligence, South China
Morning Post, September 13, 2016.
11
U.S. companies raising $1billion or more to fuel artificial intelligence (AI) development: Looking to staff
10,000+ openings, cites new Paysa research, Paysa press release, April 18, 2017.
12
Cade Metz, Facebook opens a Paris lab as AI research goes global, Wired.com, June 2, 2015.
13
Cade Metz, Google opens Montreal AI lab to snag scarce global talent, Wired.com, November 12, 2015;
Georgia Tech launches new research on the security of machine-learning systems, Georgia Institute of
Technology press release, October 31, 2016; NVIDIA collaborates with Taipei Tech to establish Embedded
GPU Joint Lab, National Taipei University of Technology press release, September 4, 2014.

The buzz over AI has grown loud enough to encourage venture capital and private equity
firms to step up their investment in AI. Other external investors, such as angel funds and
seed incubators, also are active. We estimate total annual external investment was $8billion
to $12billion in 2016.14
Machine learning attracted almost 60percent of that investment, most likely because it is
an enabler for so many other technologies and applications, such as robotics and speech
recognition (Exhibit2). In addition, investors are drawn to machine learning because, as
has long been the case, it is quicker and easier to install new code than to rebuild a robot or
other machine that runs the software. Corporate M&A in this area is also growing fast, with a
compound annual growth rate of around 80percent from 2013 through 2016.
Exhibit 2
Machine learning received the most investment, although boundaries between technologies are not clear-cut
External investment in AI-focused companies by technology category, 20161

$ billion
Natural language
0.60.9
Autonomous
Computer vision
vehicles
2.53.5
0.30.5 Machine learning
Multiuse and nonspecific applications
5.07.0
Smart
robotics Virtual
0.30.5 agents
0.10.2
1 Estimates consist of annual VC investment in AI-focused companies, PE investment in AI-related companies, and M&A by corporations. Includes only
disclosed data available in databases, and assumes that all registered deals were completed within the year of transaction.
SOURCE: Capital IQ; Pitchbook; Dealogic; McKinsey Global Institute analysis
Estimates of external investment in AI vary widely because measurement standards vary. For example,
14
Venture Scanner puts total funding of AI-related startups in 2016 at $2.5billion, while Goldman Sachs
estimates that the venture capital sector alone made $13.7billion of AI-related investment that year.

Investment in AI is still in the early stages and relatively small compared with the investment
in the digital revolution. Artificial intelligence, for example, attracted 2 to 3percent of all VC
funding by value in 2016, while information technology in general soaked up 60percent.
AI also was a small fraction1 to 3percentof all investment by PE firms in 2016.15 But AI
investment is growing fast.
From 2013 through 2016, external investment in AI technologies had a compound annual
growth rate of almost 40percent. That compares with 30percent from 2010 through 2013.
Not only are deals getting bigger and more numerous, but they require fewer participants
to complete the financing. This suggests that investors are growing more confident in the
sector and may have a better understanding of the technology and its potential.
However, for the most part, investors are still waiting for their investments to pay off. Only
10percent of startup companies that consider machine learning to be a core business
say they generate revenue, according to PitchBook. Of those, only half report more
than $50million in revenue. Moreover, external investment remains highly concentrated
geographically, dominated by a few technology hubs in the United States and China, with
Europe lagging far behind. We explore these issues further in Chapter3.
FIRMS AND INDUSTRIES ALREADY ON THE DIGITAL FRONTIER

ARE ADOPTING AI, BUT OTHERS ARE HESITANT TO ACT
Investors are pouring billions of dollars into AI companies based on the hope that a market of
AI adopters will develop fairly quickly and will be willing to pay for AI infrastructure, platforms,
and services. Clearly, Amazon, Google, and other digital natives are investing for their own
applications, such as optimizing searches and personalizing marketing. But getting a sense
of how much traditional companies in health care, retail, and telecom are spending on AI is
not easy. For this reason, we conducted a survey to understand this situation in more depth.
In general, few companies have incorporated AI into their value chains at scale; a majority
of companies that had some awareness of AI technologies are still in experimental or pilot
phases. In fact, out of the 3,073 respondents, only 20percent said they had adopted one
or more AI-related technology at scale or in a core part of their business.16 Tenpercent
reported adopting more than two technologies, and only 9percent reported adopting
machine learning.17
Even this may overstate the commercial demand for AI at this point. Our review of more
than 160 global use cases across a variety of industries found that only 12percent had
progressed beyond the experimental stage. Commercial considerations can explain why
some companies may be reluctant to act. In our survey, poor or uncertain returns were
the primary reason for not adopting reported by firms, especially smaller firms. Regulatory
concerns, explored further in Chapter3, also have become much more important.
As with every new wave of technology, we expect to see a pattern of early and late adopters
among sectors and firms. We uncover six features of the early pattern of AI adoption, which
is broadly in line with the ways companies have been adopting and using the recent cohort
of digital technologies. Not coincidentally, the same players who were leaders in that earlier
wave of digitization are leading in AIthe next wave.
15
It is worth noting that VC funds were focusing on AI technology when choosing investments, while PE funds
were investing in AI-related companies.
16
Survey results throughout this discussion paper are weighted for firm size; 20percent of firms indicates firms
representing 20percent of the workforce. See Appendix B for an explanation of the weighting methodology.
17
The eight technologies are natural language processing, natural language generation, speech recognition,
machine learning, decision management, virtual agents, robotics process automation, and computer vision.
The five technology systems are robotics and autonomous vehicles, computer vision, language, virtual
agents, and machine learning.

The first feature is that early AI adopters are from sectors already investing at scale in related
technologies, such as cloud services and big data. Those sectors are also at the frontier
of digital assets and usage.18 This is a crucial finding, as it suggests that there is limited
evidence of sectors and firms catching up when it comes to digitization, as each new
generation of tech builds on the previous one.
Second, independently of sectors, large companies tend to invest in AI faster at scale. This
again is typical of digital adoption, in which, for instance, small and midsized businesses
have typically lagged behind in their decision to invest in new technologies.
Third, early adopters are not specializing in one type of technology. They go broader as they
adopt multiple AI tools addressing a number of different use cases at the same time.
Fourth, companies investing at scale do it close to their core business.
Fifth, early adopters that adopt at scale tend to be motivated as much by the upside growth
potential of AI as they are by cutting costs. AI is not only about process automation, but is
also used by companies as part of major product and service innovation. This has been the
case for early adopters of digital technologies and suggests that AI-driven innovation will be
a new source of productivity and may further expand the growing productivity and income
gap between high-performing firms and those left behind.19
Finally, strong executive leadership goes hand in hand with stronger AI adoption.
Respondents from firms that have successfully deployed an AI technology at scale tended
to rate C-suite support nearly twice as high as those from companies that had not adopted
any AI technology.
Early-adopting sectors are closer to the digital frontier

Sector-by-sector adoption of AI is highly uneven right now, reflecting many features of digital
adoption more broadly. Our survey found that larger companies and industries that adopted
digital technologies in the past are more likely to adopt AI. For them, AI is the next wave
of digitization.
This pattern in the adoption of technology is not newwe saw similar behavior in firms
adopting enterprise social technologies.20 But this implies that, at least in the near future, AI
deployment is likely to accelerate at the digital frontier, expanding the gap between adopters
and laggards across companies, industries, and geographic regions.
The leading sectors include some that MGIs Industry Digitization Index found at the digital
frontier, namely high tech and telecom and financial services.21 These are industries with
long histories of digital investment. They have been leaders in developing or adopting digital
tools, both for their core product offerings and for optimizing their operations. However,
even these sectors are far behind in AI adoption when compared with overall digitization
(Exhibit3).
18
Digital Europe: Pushing the frontier, capturing the benefits, McKinsey Global Institute, June 2016; Digital
America: A tale of the haves and have-mores, McKinsey Global Institute, December 2015.
19
Rosina Moreno and Jordi Suriach, Innovation adoption and productivity growth: Evidence for Europe,
working paper, 2014; Jacques Bughin and Nicolas van Zeebroeck, The right response to digital disruption,
MIT Sloan Management Review, April 2017.
20
Jacques Bughin and James Manyika, How businesses are using web 2.0: A McKinsey global survey,
McKinsey Quarterly, December 2007; Jacques Bughin and James Manyika, Bubble or paradigm change?
Assessing the global diffusion of enterprise 2.0, in Alex Koohang, Johannes Britz, and Keith Harman, eds.,
Knowledge management: Research and applications, Informing Science, 2008.
21
Digital America: A tale of the haves and have-mores, McKinsey Global Institute, December 2015.

Exhibit 3
AI adoption is occurring faster in more digitized sectors and across the value chain
AI Index Relatively low Relatively high
Assets Usage Labor
1
MGI Digitization Index
Financial and general

Customer experience
Product development
Supporting digital
Supply chain and
Exposure to AI in
Overall AI index
AI resources per
management
management
technologies
Depth of AI
Operations
distribution
Workforce
workforce
AI spend
worker
assets
High tech and
telecommunications
Automotive and
assembly
Financial services
Resources and
utilities
Media and
entertainment
Consumer
packaged goods
Transportation
and logistics
Retail
Education
Professional
services
Health care
Building materials
and construction
Travel and tourism
1 The MGI Digitization Index is GDP weighted average of Europe and United States. See Appendix B for full list of metrics and explanation of methodology.
SOURCE: McKinsey Global Institute AI adoption and use survey; Digital Europe: Pushing the frontier, capturing the benefits, McKinsey Global Institute, June
2016; Digital America: A tale of the haves and have-mores, McKinsey Global Institute, December 2015; McKinsey Global Institute analysis
Automotive and assembly is also highly ranked. It was one of the first sectors that
implemented advanced robotics at scale for manufacturing, and today is also using AI
technologies to develop self-driving cars.
In the middle are less digitized industries, including resources and utilities, personal and
professional services, and building materials and construction. A combination of factors
may account for this. These sectors have been slow to employ digital tools generally, except
for some parts of the professional services industry and large construction companies. They
are also industries in which innovation and productivity growth has lagged, potentially in part
due to their domestic focus. Some of these sectors have a particularly high number of small
firmsan important predictor for AI adoption, as explored below.

Toward the bottom of the pack for now are traditionally less digital fields such as education
and health care. Despite ample publicity about cutting-edge AI applications in these
industries, the reality is that uptake appears to be low so far. Weaker adoption reflects the
particular challenges faced in these sectors. In health care, for example, practitioners and
administrators acknowledge the potential for AI to reduce costs but quickly add that they
believe that regulatory concerns and customer acceptance will inhibit adoption.
When it comes to adopting AI, the bigger, the bolder

A stylized fact in IT literature is that large firms usually are early adopters of innovative
technology, while smaller firms are more reluctant to be first movers.22 We find the same
digital divide when we look at AI: large firms have much higher rates of adoption and
awareness. Across all sectors, larger firmswhich we define as those with more than 500
employeesare at least 10percent more likely than smaller firms to have adopted at least
one AI technology at scale or in a core part of their business. In sectors with lower rates
of AI uptake, the adoption rate of bigger companies was as much as 300percent that of
smaller companies.
Other digitization indicators reflect this fact, as highlighted in MGIs digitization work.
Larger firms typically have access to more and better-structured data, and are more likely
to have employees with the technical skills needed to understand the business case for
AI investment and to successfully engage suppliers. Bigger firms also have an advantage
because the kind of fixed-cost investment required for AI tends to generate higher returns
when applied to a bigger base of costs and revenue.
Nonetheless, we find success stories among some smaller firms, too. Relative to larger
companies, they can benefit from fewer issues with legacy IT systems and lower levels of
organizational resistance to change. Smaller firms can also benefit from AI tools provided as
a service.
Early AI adopters tend to become serial adopters

We looked at how firms deploy AI across eight different application areas and five
technology systems.23 Our results suggest that early-adopting firms are looking across
multiple AI tools when they begin to adopt, rather than focusing on a particular technology.
This is consistent with adoption patterns in other digital technologies.24
The phenomenon of multitechnology application is persistent at a sector level. Industries

with high rates of adopting one technology have higher rates in adopting others. High tech
and telecom, for example, report the highest rates of adoption across all five technology
groups, while construction is among the lowest among all five.
However, there are anomalies. Education and health care are notable for being slow to
adopt AI technology. In frontier sectorsthose with a relatively high percentage of early
adopterstwo-thirds of firms that had already adopted one of the eight AI technologies
had adopted at least two others as well. In health care, only one-third had, with language
technologies the most likely to be deployed at scale or in a core part of the business.
22
Kevin Zhu, Kenneth L. Kraemer, and Sean Xu, The process of innovation assimilation by firms in different
countries: A technology diffusion perspective on e-business, Management Science, volume 52, number
10, October 2006; Chris Forman, Avi Goldfarb, and Shane Greenstein, The geographic dispersion of
commercial Internet use, in Rethinking rights and regulations: Institutional responses to new communication
technologies, Lorrie Faith Cranor and Steven S. Wildman, eds., MIT Press, 2003.
23
The eight technologies are: natural language processing, natural language generation, speech recognition,
machine learning, decision management, virtual agents, robotics process automation, and computer vision.
The five technology systems are: robotics and autonomous vehicles, computer vision, language, virtual
agents, and machine learning.
24
Sanjeev Dewan, Dale Ganley, and Kenneth L. Kraemer, Complementarities in the diffusion of personal
computers and the internet: Implications for the global digital divide, Information Systems Research, volume
21, number 5, December 2010.

Users are keeping artificial intelligence close to their core
Functionally, AI technologies are finding applications across the value chain, but with some
parts of the value chain getting more attention than others. For example, customer service
functions such as sales and marketing, as well as operations and product development, all
tend to use the most commonly cited AI applications. General and financial management,
by contrast, lag well behind. A similar pattern is found in big data. The literature shows that
the most frequent big data applications originate in sales and marketing functions.25
In general, firms queried in our survey say they tend to adopt AI technologies affecting the
part of their value chain closest to the core. Operations are an important area of adoption in
the automotive and assembly, and consumer packaged goods sectors, as well as utilities
and resources. Operations and customer service are the most important areas for financial
services. This is new. Previously, new digital technology tended to remain on the margins,
away from the core of the business.
However, in line with trends in technology, we also see sectors going deeper and broader
as they increase their degree of AI adoption. Leading sectors are not only more extensively
deploying AI in the core parts of their value chain, they are also deploying it in more parts of
their value chain.
Early adopters see AI increasing revenue while companies

experimenting with AI expect lower costs
As companies become more familiar with AI, their perceptions about its benefits change.
The results of survey analysis show that early AI adopters are driven to employ AI
technologies in order to grow revenue and market share, and the potential for cost reduction
is a secondary idea. Firms that we consider more advanced AI adopters were 27percent
more likely to report using AI to grow their market than companies only experimenting with
or partially adopting AI, and 52percent more likely to report using it to increase their market
share. Experimenters, by contrast, were more focused on costs. They were 23percent
more likely than advanced AI adopters to point to labor cost reductions, and 38percent
more likely to mention non-labor cost reductions.
In other words, the more companies use and become familiar with AI, the more potential
for growth they see in it. Companies with less experience tend to focus more narrowly on
reducing costs. The employment implications are further discussed in Chapter3.
AI is not only about technical adoption, it is about enterprise acceptance

To be successful, AI adoption requires buy-in by the executive suite to generate the
momentum needed to overwhelm organizational inertia.
Successful AI adopters, according to our survey, have strong executive leadership support
for the new technology. Representatives of firms that have successfully deployed an
AI technology at scale tended to rate C-suite support nearly twice as high as those of
companies that had not adopted any AI technology. They added that strong support came
not only from the CEO and IT executivesthat is, chief information officer, chief digital
officer, and chief technology officerbut from all other C-level officers and the board of
directors as well.
Successful adopters also adjusted their firm-wide strategy to become proactive toward AI.
See more details in Chapter2.
Jacques Bughin, Ten big lessons learned from big data analytics, Applied Marketing Analytics, volume 2,
25
number 4, 2017.

AIS NEXT CHALLENGE: GET USERS TO ADAPT AND ADOPT
IT industry analysts concur that the market size for AI technology will experience strong
growth over the next three years. Most of the firms we surveyed expected to increase
spending on AI in the coming three years, a finding echoed in other recent surveys. For
example, 75percent of the 203 executives queried in an Economist Intelligence Unit survey
said AI would be actively implemented in their firms within three years (3percent said it had
already happened).
Expectations of how large this growth will be vary widely. Our survey documented relatively
modest growth projectionsonly one-fifth of firms expected to increase expenditure
by more than 10percent. Industry analysts forecasts of the compound annual growth
rate ranged from just under 20percent to nearly 63percent, including both adoption by
additional companies and increased spending within companies.26 The actual growth rate
may need to be toward the upper end of that range to meet the expectations of investors
piling into the industry.
Growth will hinge on the ability of sectors and firms to overcome technical, commercial, and
regulatory challenges. Our survey respondents and outside forecasters expect financial
services, retail, health care, and advanced manufacturing to be in the AI vanguard. These
are the industries where technical feasibility is relatively high (reflected in the case studies on
the market today) and the business case for AI is most compelling. They are also the sectors
with the highest degree of digital adoption to datea key foundation for AI (Exhibit4).
Technical challenges are an important differentiating factor between industries. While big
tech and academia are pushing advances in the performance of the underlying technology,
engineering solutions need to be worked out for specific use cases, requiring both data and
talent. Industries such as financial services, and high tech and telecom have generated and
stored large volumes of structured data, but others, including construction and travel, lag
far behind.27
Commercial drivers also differ between sectors. Industries most likely to lead the adoption of
AI technologies at scale are those with complex businesses in terms of both operations and
geography, whose performance is driven by forecasting, fast and accurate decision making,
or personalized customer connections. In financial services, there are clear benefits from
improved accuracy and speed in AI-optimized fraud-detection systems, forecast to be a
$3billion market in 2020. In Chapter2 (and the supporting appendix) we explore how these
commercial drivers play out in other industries. For example, in retail, there are compelling
benefits from improved inventory forecasts, automated customer operations, and highly
personalized marketing campaigns. Similarly, in health care, AI-powered diagnosis and
treatment systems can both save costs and deliver better outcomes for patients.
Even where compelling commercial use cases have been engineered and are demanded
by firms, regulatory and social barriers can raise the cost and slow the rate of adoption.
Product liability is one such concern; it is especially troublesome for automakers and other
manufacturers. Privacy considerations restrict access to data and often require it to be
anonymized before it can be used in research. Ethical issues such as trained biases and
algorithmic transparency remain unresolved. (For further discussion, see Box4 in Chapter3,
An overview of ongoing challenges.) Preferences for a human relationship in settings such
as health care and education will need to be navigated. Job security concerns could also
limit market growththere are already serious calls for taxes on robots.
26
The full range of forecasts: BCC Research, 19.7percent; Transparency Market Research, 36.1percent,
Tractica, 57.6percent; IDC, 58percent; and Markets and Markets, 62.9percent.
27
A future that works: Automation, employment, and productivity, McKinsey Global Institute, January 2017.

Exhibit 4
Sectors leading in AI adoption today also intend to grow their investment the most
Future AI demand trajectory1

Average estimated % change in AI spending, next 3 years, weighted by firm size2
13
Leading sectors
12
Financial services
11 High tech and
telecommunications
10
9
8 Transportation and logistics
7 Health care
Travel and tourism Automotive
6 and assembly
Professional services
5 Retail Energy and resources
4 Consumer packaged goods Media and entertainment

Education
3
2
Construction
1
Falling behind
0
0 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30 32
Current AI adoption
% of firms adopting one or more AI technology at scale
or in a core part of their business, weighted by firm size2
1 Based on the midpoint of the range selected by the survey respondent.

2 Results are weighted by firm size. See Appendix B for an explanation of the weighting methodology.
SOURCE: McKinsey Global Institute AI adoption and use survey; McKinsey Global Institute analysis
These forces will help determine the industries that AI is likely to transform the most.
However, if current trends hold, variation of adoption within industries will be even larger than
between industries. We expect that large companies with the most digital experience will be
the first movers because they can leverage their technical skills, digital expertise, and data
resources to develop and smoothly integrate the most appropriate AI solutions.
After decades of false starts, artificial intelligence is on the verge of a breakthrough, with the
latest progress propelled by machine learning. Tech giants and digital natives are investing in
and deploying the technology at scale, but widespread adoption among less digitally mature
sectors and companies is lagging. However, the current mismatch between AI investment
and adoption has not stopped people from imagining a future where AI transforms
businesses and entire industries. In the next chapter, we explore the four major ways in
which AI can create value across the value chain in different sectors.

2. ARTIFICIAL INTELLIGENCE PROMISES TO
BOOST PROFITS AND TRANSFORM INDUSTRIES
AI technologies have advanced significantly in recent years. Adoption, however, remains in

its infancy. This makes it challenging to assess the true potential impact of AI on firms and
sectors. We do know that many non-adopters report that they have trouble making the
business case for AI investment, but what about the firms that have adopted? Looking at
case studies of digital natives and responses from our survey, we find early evidence that
AI implemented at scale delivers attractive returns. Through a review of a large number of
case studies in five sectors, we also show how AI can transform some business activities
and has the potential to fundamentally change others. The cases demonstrate how AI can
shape different functions across the whole value chain and in diverse sectors. The cases
have wide-ranging implications for many stakeholders, including multinational corporations,
startups, governments, and social institutions.
FIRMS THAT COMBINE STRONG DIGITAL CAPABILITY, ROBUST AI ADOPTION,

AND A PROACTIVE AI STRATEGY SEE OUTSIZE FINANCIAL PERFORMANCE
Digital native companies made some of the most significant and earliest investment in
AI, providing test cases for potential return on investments in AI. Amazon has achieved
impressive results from its $775million acquisition of Kiva, a robotics company that
automates picking and packing. Click to ship cycle time, which ranged from 60 to 75
minutes with humans, fell to 15 minutes with Kiva, while inventory capacity increased
by 50percent. Operating costs fell an estimated 20percent, giving a return of close to
40percent on the original investment.28
Netflix has also achieved impressive results from the algorithm it uses to personalize
recommendations to its 100million subscribers worldwide. Helping customers quickly find
desirable content is criticalcustomers tend to give up if it takes longer than 90 seconds
to find a movie or TV show they want to watch. Through better search results, Netflix
estimates that it is avoiding cancelled subscriptions that would reduce its revenue by
$1billion annually.29
Firms adopting AI at scale or in a core part of their business already see the technologys
potential, and those implementing proactive AI strategies are anticipating even greater
benefits. Using our survey results, we compared the current self-reported profit margins of
firms with differing levels of AI adoption, digital maturity (as reflected in their use of big data
and cloud services), and critically, their strategic posture (Exhibit5). We found that serious
adopters have significantly higher projected margins than others.
This suggests AI can deliver significant competitive advantages, but only for firms that
are fully committed to it. Take any ingredient awaya strong digital starting point, serious
adoption of AI, or a proactive strategic postureand profit margins are much less
impressive. This is consistent with our findings in the broader digital space. Technology is a
tool and in itself does not deliver competitiveness improvements. In Chapter3, we explore
what firms need to do to successfully adopt AI, including identifying value-adding use cases,
building capabilities, and embedding a collaborative culture.
28
Eugene Kim, Amazons $775million deal for robotics company Kiva is starting to look really smart, Business
Insider, June 15, 2016.
29
Nathan McAlone, Why Netflix thinks its personalized recommendation engine is worth $1billion per year,
Business Insider, June 14, 2016.

Exhibit 5
AI adopters with a proactive strategy have significantly higher profit margins
AI adopters with proactive strategy1 Partial AI adopters or experimenters Non-adopters
Self-reported current profit margin2

Difference from industry average (unweighted) (percentage points)
-10 -5 0 5 10 15 20
High tech and
telecommunications
Automotive and assembly
Financial services
Utilities and resources
Transportation and logistics
Consumer packaged goods
Retail
Education
Health care
Building materials and

construction
Professional services
Travel and tourism
1 Firms that are big data and cloud services users and report their strategic posture toward AI to be: Disrupting our industry using AI technology is at the core
of our strategy, We have changed our longer-term corporate strategy to address the AI threat or opportunity disruption, or We have developed a
coordinated plan to respond to the AI threat or opportunity but have not changed our longer-term corporate strategy.
2 Operating profit margin for selected sectors as a share of turnover, for continuing operations and before exceptional items.
SOURCE: McKinsey Global Institute AI adoption and use survey; McKinsey Global Institute analysis
The same pattern is apparent when we analyze expectations of future profit margins.
Not only do serious AI adopters with proactive strategies report current profit margins
that are three to 15 percentage points higher than the industry average in most sectors,
but they also expect this advantage to grow in the future, when they could expect their AI
investment to mature and start paying substantial dividends. In the next three years, these
AI leaders expect their margins to increase by up to five percentage points more than the
industry average.
How can AI tangibly provide new sources of value for businesses and institutions? We
have looked at use cases across selected sectors to understand how AI could impact
organizations and their value chains.

INDUSTRY CASE STUDIES DEMONSTRATE THE DISRUPTIVE POTENTIAL OF AI
To develop a vision of AI as it could be applied across business domains, we created five
case studies in different sectors that suggest how AI in its many guises could affect specific
activities (see Appendix A). The sectors we choseretail, electric utilities, manufacturing,
health care, and educationinvolve a mixture of private, public, and social enterprises, as
well as covering the range from highly labor-intensive customer-oriented sectors to more
asset-heavy business-to-business operations.
To fulfil the expectations being heaped upon it, AI will need to deliver economic applications
that significantly reduce costs, increase revenue, and enhance asset utilization. We
categorized the ways in which AI can create value in four areas: enabling companies to
better project and forecast to anticipate demand, optimize R&D, and improve sourcing;
increasing companies ability to produce goods and services at lower cost and higher
quality; helping promote offerings at the right price, with the right message, and to the
right target customers; and allowing them to provide rich, personal, and convenient
user experiences.
These four areas of value creation are based on specific use cases that are being explored
or have been deployed in businesses today. The list, which may not be exhaustive, is based
on our current knowledge of narrow AI technologies. Also, they will have different degrees of
relevance for sectors and industries, with the project and produce levers being particularly
rich with opportunities to leverage AI. In addition, while machine learning can bring highly
valuable benefits to all sectors, some technologies are particularly suited for business
application in specific sectors, such as robotics for retail and manufacturing, computer
vision for health care, and natural language processing and generation for education (see
Exhibit6).
Project: Accurately forecast demand, optimize supply,

and shape future offerings for success
The first area in which AI can create value is projection and forecasting. Organizations need
to constantly anticipate the future to gain competitive advantage. AI allows businesses
to provide better forecasts for their supply chain and design better offerings. Reliably
forecasting demand is a way to use AIs ability to digest disparate data and automatically
adjust to new information. It can discern trends and patterns that can be acted on.
Businesses use this tool in a number of ways, such as forecasting demand to stock only
the specific quantities of specific products they will sell and thus minimize waste, and
anticipating sales trends so they can order more soon-to-be-popular items. Indeed, the
benefits of projecting demand go beyond traditional business sectors: for example, by
using sophisticated algorithms, health systems can increasingly predictand prevent
major epidemics.
The benefits of AI-enabled demand forecasting are impressive in retail, for instance. In some
settings, AI-based approaches to demand forecasting are expected to reduce forecasting
errors by 30 to 50percent from conventional approaches (Exhibit7). Lost sales due to
product unavailability can be reduced by up to 65percent. Costs related to transportation
and warehousing and supply chain administration are expected to decrease by 5 to
10percent and 25 to 40percent, respectively. With AI, overall inventory reductions of 20 to
50percent are feasible.30
Smartening up with artificial intelligence (AI): Whats in it for Germany and its industrial sector? McKinsey &
30
Company, April 2017.

Exhibit 6
Artificial intelligence can create value across the value chain in four ways
Machine learning1 Natural language Applicability
Autonomous vehicles Computer vision High Medium Low
Smart robotics Virtual agents
Project Produce Promote Provide

Enlightened R&D, Operations with Products and Enriched, tailored,
real-time higher services at the and convenient
forecasting, and productivity, lower right price, with the user experience
Applicable smart sourcing cost, and better right message, and
technologies efficiency to the right targets
Retail Anticipate demand Automate Optimize pricing, Personalize tips and
trends, while warehouse and personalize suggestions, offer
optimizing and store operations; promotions, and immediate
automating supplier optimize tailor website assistance with
negotiation and merchandising, displays in real time virtual agents,
contracting product assortment, automate in-store
and microspace checkout, and
complete last-mile
delivery by drones
Electric Enhance demand Optimize preventive Optimize pricing with Automate supplier
utilities and supply maintenance, time-of-day and selection, provide
prediction, assess improve electricity dynamic tariffing; consumption
reliability of production yield, match producers insights, automate
integrated reduce energy and consumers in customer service
generation assets, waste, and prevent real time with virtual agents,
and automate electricity theft and tailor usage to
demand-side consumers
response
preferences
Manufacturing Improve product Improve processes Predict sales of Optimize flight
design yield and by the task, maintenance planning and route
efficiency, automate automate assembly services, optimize and fleet allocation;
supplier lines, reduce errors, pricing, and refine enhance
assessment, and limit product rework, sales-leads maintenance
anticipate parts and reduce material prioritization engineer and pilot
requirements delivery time training
Health care Predict disease, Automate and Predict cost more Adapt therapies and
identify high-risk optimize hospital accurately, focus on drug formulations to
patient groups, and operations; patients risk patients, use virtual
launch prevention automate diagnostic reduction agents to help
therapies tests and make patients navigate
them faster and their hospital journey
more accurate
Education Anticipate job Automate teachers Personalize
market demand, routine tasks, learning, shift from
identify new drivers identify early stop-and-test model
of performance to disengagement to continuous
assess students, signs, and optimize learning cadenced
and help graduates group formation for by virtual coaches
highlight their learning objectives and tutors, and build
strengths student self-
awareness
1 Machine learning for multiuse and nonspecific applications.
SOURCE: McKinsey Global Institute analysis

Exhibit 7
AI can help capture significant gains across the value chain
Examples of AI-related business impact from current use cases
Project Produce Promote Provide

Accurate demand Higher productivity and Products and services Enriched, tailored, and
forecasting, smart minimized maintenance at the right price, with convenient user
sourcing, and and repairs the right message, to experience
enlightened R&D the right targets
Retail 12% EBIT1 30% reduction of 50% improvement of
improvement using stocking time using assortment efficiency
machine learning to autonomous vehicles in 46% sales increase
anticipate fruit and warehouses using geospatial
vegetable sales modeling to improve
20% stock reduction micromarket
using deep learning to attractiveness
predict e-commerce 30% online sales
purchases increase by using
2 million fewer product dynamic pricing and
returns per year personalization
Electric Objective to cut 10% in 20% energy $10$30 savings on
utilities national electricity production increase monthly bills by using
usage by using deep using machine learning machine learning to
learning to predict and smart sensors to automatically switch
power demand and optimize assets yield electricity supply deals
supply 1020% EBIT
improvement by using
machine learning to
enhance predictive
maintenance, automate
fault prediction, and
increase capital
productivity
Manufac- 10% yield 30% increase of 13% EBIT 12% fuel savings for
turing improvement for material delivery time improvement by using manufacturers
integrated-circuit using machine learning machine learning to customers, airlines, by
products using AI to to determine timing of predict sources of using machine learning
improve R&D process goods transfer servicing revenues and to optimize flight routes
39% IT staff reduction 35% production yield optimize sales efforts
by using AI to fully improvement
automate procurement
processes
Health $300 billion possible 3050% productivity 59% health $2 trillion$10 trillion
care savings in the United improvement for expenditure reduction savings globally by
States using machine nurses supported by AI by using machine tailoring drugs and
learning tools for tools learning to tailor treatments
population health Up to 2% GDP treatments and keep 0.21.3 additional
forecasting savings for operational patients engaged years of average life
3.3 billion possible efficiencies in expectancy
savings in the United developed countries
Kingdom using AI to
provide preventive care
and reduce nonelective
hospital admissions
Education Virtual teaching 1% increase in 85% match with
assistants can answer enrollment by using a human grading, using
40% of students' virtual assistant to machine learning and
routine questions follow up with predictive modelling
applicants
1 Earnings before interest and taxes.

When it comes to matching supply and demand, electric utilities are a special case: they
need to do it, literally, in real time. Making short-term load forecasts more accurate in
order to adjust supply to meet anticipated demand can deliver enormous savings, reduce
waste and emissions, and add to system resilience. National Grid in the United Kingdom
is collaborating with DeepMind, an AI startup bought by Google in 2014, to predict
supply and demand variations based on weather-related variables and smart meters as
exogenous inputs. The goal is to cut national energy use by 10percent and maximize the
use of renewable power. AI also is used to briefly switch off air conditioning at participating
businesses as it forecasts the approach of peak consumption, easing the load for all and
postponing or even forgoing the need to fire up peak generating capacity.
Forecasting is not a new idea. A well-functioning supply chain is the backbone of virtually
every industry. Accurate projections to ensure just the right amount of inventory are critical
to achieving a competitive advantage. Factors such as product introductions, distribution
network expansion, weather forecasts, extreme seasonality, and changes in customer
perception or media coverage can severely affect the performance of the supply chain.
Traditional systems for forecasting and replenishment might not take advantage of the
amount of data associated with the internet of things and the sheer number of influencing
factors. Supply-chain leaders are starting to realize the ability of machine learningbased
methods to increase forecasting accuracy and optimize replenishment. The objective is to
reduce swings in inventory levels and increase flexibility. Machine learning approaches not
only incorporate historical sales data and the setup of the supply chains, but also rely on
near-real-time data regarding variables such as advertising campaigns, prices, and local
weather forecasts.
Using AI to forecast demand also allows businesses to optimize their sourcing more broadly,
including fully automating purchases and order processing. The German online retailer
Otto uses an AI application that is 90percent accurate in forecasting what the company
will sell over the next 30 days. The forecasts are so reliable that Otto now builds inventory
in anticipation of the orders AI has forecast, enabling the retailer to speed deliveries to
customers and reduce returns. Otto is confident enough in the technology to let it order
200,000 items a month from vendors with no human intervention.31
But AI is helpful not just for forecasting demand for current products; it can also be used
by R&D departments, partly to help researchers quickly assess whether a prototype
would be likely to succeed or fail in the marketand why. Engineers and researchers
today face difficult challenges, from the sharp growth in demand in emerging countries
to market fragmentation driven by consumers taste for customization. At the same time,
budget constraints require engineering teams to improve their productivity and efficiency,
even as limits on the number of designs that can be tested restrict the predictability
of product performance. These features may be particularly attractive to the health
and pharmaceuticals sectors, which spend $160billion annually on R&D in the United
States alone.
AI-powered technologies can help deliver more efficient designs than were previously
achievable by eliminating waste in the design process. Innovations can be brought to
market faster as AI facilitates faster process cycle times and an increased focus on real-
time negotiations and other interactions. AI-based approaches to increasing R&D project
performance can result in productivity gains of 10 to 15percent. Time to market could be
reduced by 10percent or more.32 Motivo, an artificial intelligence startup, compresses
How Germanys Otto uses artificial intelligence, The Economist, April 12, 2017.
31
Smartening up with artificial intelligence (AI): Whats in it for Germany and its industrial sector? McKinsey &
32
Company, April 2017.

design processes that used to take monthssometimes a year or moreto roughly four
weeks, saving chip makers the cost of iterations and testing.
Produce: Get more out of machines while minimizing maintenance and repairs
The second area where AI can help create value is production, or the transformation of
inputs into outputs, whether products (semiconductors, aircraft engines) or services
(teaching, health care, or power or consumer goods distribution). AI can help businesses
produce by continually optimizing assets and processes, assembling the best teams
of people and robots, improving quality and reliability, and preventing downtime for
maintenanceall of which increase productivity.
The obvious role for AI is to replace humans through automation. However, in some
situations, AI is complementing teams of people. Ocado, the UK online supermarket,
illustrates how this happens when a company embeds AI and robotics at the core of its
operations. In the retailers warehouse, robots steer thousands of product-filled bins over a
maze of conveyor belts and deliver them to human packers just in time to fill shopping bags.
Other robots whisk the bags to delivery vans whose drivers are guided to customers homes
by an AI application that picks the best route based on traffic conditions and weather.
Advances in computer vision are behind many developments in collaborative and

context-aware robots. Enhanced vision is enabled by more powerful computers, new
algorithmic models, and large training data sets. Within the field of computer vision, object
recognition and semantic segmentationthat is, the ability to categorize object type,
such as distinguishing a tool from a componenthave recently advanced significantly in
their performance. They allow robots to behave appropriately for the context in which they
operate, for example by recognizing the properties of the materials and objects they interact
with. They are flexible and autonomous systems that are capable of safely interacting with
the real world and humans.
This is a notable shift since current robotic systems are often limited in their ability to
identify a specific object and react in a flexible manner to changes in their environment,
and therefore are programmed to follow predefined steps. However, new AI-enhanced,
camera-equipped logistics robots can be trained to recognize empty shelf space. This
leads to a dramatic speed advantage over conventional methods in picking objects. Deep
learning can also be used to correctly identify an object and its position. This enables robots
to handle objects without requiring the objects to be in fixed, predefined positions. AI-
enhanced logistics robots are also able to integrate disturbances in their movement routines
via an unsupervised learning engine for dynamics. This capability leads to more precise
makeovers and an overall improved robustness of processes. Rethink Robotics is one
company designing collaborative robots. Its learning algorithms allow for joint human-robot
collaborative work spaces. AI enables the programming of a robot by simply showing it
the desired movements. A human robot instructor can take the machines arm and guide it
through the desired movement, including gripping and releasing objects. Robot movement
is then the result of the robots replication and further improvement of the freshly learned
movement combined with a computer-vision-based assessment of an objects position
in space.
Collaborative robots are particularly relevant with respect to tasks that are not fully
automatable. In such settings they hold the potential to increase productivity by up to
20percent.33 Using AI to improve humans efficiency is a critical enabler of productivity. In
education, AI tools can help teachers accelerate administrative tasks and daily operations.
AI-enabled applications to help teachers quickly grade students work are already available.
They decipher students handwriting, learn how the teacher grades the first few tests, and
33
Ibid.

apply the same standards to subsequent tests. This works well with objectively correct
answers, such as math problems. The true step change will come with the cracking
of natural language understanding, which would allow the automatic grading of more
creative work, such as essays and presentations. While this seems to be feasible, it is still
in development. In a multivendor evaluation at the University of Akron in 2012, grading
software gave marks to 16,000 essays that instructors had already reviewed; computers
matched the teachers grades about 85percent of the time.34 Still unresolved are questions
around creativity and pattern disruption: how can machine learning be taught not to
encourage uniform thinking or the upholding of biases?
AI tools can also improve productivity by better aligning team formation with the teams
objective. Collaboration.ai uses artificial intelligence models to assist teachers in forming
optimized classes. It analyzes data from students education profiles, social media
accounts, and surveys to populate a class or study group with students that have
compatible skills and personalities. The emergence of collaborative robots is further
improving human-machine compatibility.
In addition to speeding up processes, reducing costs, and increasing output, AI has huge
potential to improve quality by reducing errors. Yield losseslosses incurred when products
have to be disposed of or reworked due to defectsplay an important role in complex
manufacturing environments.
The multistep semiconductor chip-production process is a good example because cycle

times from the first processing of the wafer to the final chip are typically several weeks to
months and include various intermediate quality-testing processes. Testing costs and yield
losses in semiconductor production can constitute up to 30percent of the total production
cost. Ample data are usually available in semiconductor fabrication plants, or fabs, due
to their high degree of automation and advanced production equipment. They often have
archives that allow insights into detailed production information dating back months or
years. However, systematic analysis and linkage of data sources across multiple tool groups
is not always performed. Semiconductor manufacturers are starting to use AI engines to
identify root causes of yield losses that can be avoided by changing production processes.
Enhanced applications are designed to monitor and adjust subprocesses in real time. AI
techniques can help determine the optimized product operating conditions or process
conditions to significantly reduce defects in manufacturing.
In asset-heavy businesses, keeping complex systems running with minimal downtime

is another key opportunity for AI. Utility companies can shift from regularly scheduled
maintenance of their extensive electrical grids to condition-based maintenance run by AI.
Using data from sensors, drones, and other hardware, machine learning applications can
help grid operators avoid decommissioning assets before their useful lives have ended,
while simultaneously enabling them to perform more frequent remote inspections and
maintenance to keep assets working well. One European power distribution company was
able to reduce its cash costs by 30percent over five years by changing its maintenance
patterns based on remote analysis of 20 variables to determine the overall health of
power transformers.
AI is enabling the preventive maintenance of people, too, and will do even more in the
future. Applications powered by artificial intelligence will enable health-care providers to
dramatically accelerate the shift toward personalized preventive medicine. Clinicians will
focus on managing patients health remotely via wearable wireless sensors, aiming to keep
them healthy, fit, and out of hospitals. To do this, AI tools will take into account not only
Man and machine: Better writers, better grades, University of Akron press release, April 12, 2012.
34

patients medical histories and genetic makeup but also environmental factors that can
influence health, such as pollution and noise where they live and work.
Promote: Charge the right price and deliver the right message to the right target
The third area where AI can create value is promotion, or marketing offerings at the right
price, with the right message, and to the right target. Armed with enough of the right kind
of data, companies can use artificial intelligence to price goods and services dynamically,
raising prices when demand rises or a consumer appears willing to pay more, and lowering
them when the opposite happens. Yield management programs have been dynamically
pricing airline seats, hotel rooms, and other perishables for years, but AI will allow sellers to
extend dynamic pricing to the rest of the marketplace.
Today, the requirements of intelligent price management are high: customers expect a good
price, and price transparency for brand-name products is close to 100percent. The basic
question to ask for each item is: what price is the customer willing to pay? Hyperconnected
consumers continuously redefine value by comparing prices online, even when browsing
in a brick-and-mortar store. The optimal price for a product depends on many factors: the
day of the week, season, time of day, weather, channel and device, competitors prices, and
much more. The challenge is to set the optimal price in relation to time. The right price at the
right time increases customer satisfaction and leads to more sales and higher profit. AI can
determine the price elasticity for every item and automatically adjust prices according to the
chosen product strategy.
Similarly, energy retailers can use AI to create custom benefits such as low rates or extra
service in order to hold on to their most valuable clients. While price sensitivity is a key
consideration in attracting new customers and reducing churn, machine learning can also
help address another critical component for utilities marketing strategiesdetermining
which customers are the most profitable. They are typically a small proportion (less than
20percent) of a utilitys customer base.
Retailers have accepted that smartphone penetration necessitates an omni-channel

sales strategy, and AI can help optimize, update, and tailor it to each shopper in real time.
Retailers can send mobile couponsusually text messages with a discount codeto
shoppers smartphones as they approach a store, adding others based on how long the
shopper stays. The size of the discount, the merchandise on offer, and other variables can
be determined by an AI program that has looked for clues about what the shopper will like
based on previous purchases, age, home address, web browsing habits, and mounds of
other data. This kind of insights-based selling, including personalized promotions, optimized
assortment, and tailored displays, can increase sales by 1 to 5percent. Online, a focus on
the most valuable customers, combined with dynamic pricing, can lead to a 30percent
growth in sales.
Aerospace companies, too, are using AI technologies to prioritize sales targets and optimize
the price of services. For years, they prioritized maintenance, repair, and overhaul (MRO)
sales leads manually, a cumbersome, resource-heavy, and not always efficient process.
Using AI to improve the accuracy of forecasting MRO work and focusing the firms sales
efforts on the most promising leads can have a significant effect on profitability. One firm
said its profit improved by over $300million by using machine learning to forecast 10years
of repair events for its transportation fleet and by developing a deal-scoring tool to advise on
what good looks like when pricing after-market services.

Provide: Give customers a rich, personalized, and convenient experience
The fourth area where AI can create value is in enhancing the user experience and creating
new sources of value to make it richer, more tailored, and more convenient. Making your
best customers feel special and welcome is one way to foster loyalty and increase revenue,
but it is difficult and expensive to do and so is often reserved for only the most lucrative
clients. AI technologies like computer vision and machine learning can open a scaled-down
version of the experience to many more people.
If, for example, a regular supermarket shopper puts a bunch of bananas in his cart, cameras
or sensors could relay the information to an AI application that would have a good idea
of what the shopper likes based on previous purchases. The app could then, via a video
screen in the cart, suggest that bananas would be delicious with a chocolate fondue, which
the purchase history suggests the shopper likes, and remind the shopper of where to find
the right ingredients. Or a runner could download an app from an athletic shoe company,
which would monitor her exercise regimen and recommend footwear tailored to her routine
and running paths she may like.
Some of the same AI applications are also being employed in an experimental supermarket
that may give new meaning to the idea of a convenience store. Amazon has built a retail
outlet in Seattle that allows shoppers to take food off the shelves and walk directly out of
the store without stopping at a checkout kiosk to pay. The store, called Amazon Go, relies
on computer vision to track shoppers after they swipe into the store and associate them
with products taken from shelves. When shoppers leave, Amazon debits their accounts
for the cost of the items in their bag and emails them a receipt. In the future, the shopping
experience will be completed by delivery drones for full convenience. Most of todays
effortsby players as big as Amazon and as small as the Reno, Nevada, startup Flirtey
focus on unpiloted aerial drones. Flirtey made its first delivery, a box of snacks from a local
convenience store, to a private residence in July 2016. Delivery drones will significantly
benefit from breakthroughs in deep learning, which will help them categorize and handle
anomalous situations, such as when no one is home to accept a delivery.
Personalizing user experiences has huge advantages in health care and education. In
health care, treatment decisions based on AI analysis of existing science, data from tests,
and patient monitoring with remote diagnostic devices carry the promise of significantly
increased efficacy. Researchers are moving in this direction because standardized
treatments do not work for everyone, given the complexity of each persons history and
genetic makeup. For a cancer patient, the technology models cell biology on the molecular
level and seeks to identify the best drug to use for specific tumors. It can also identify
complex biomarkers and search for combination therapies by performing millions of
simulated experiments each day.
Several companies already use AI technologies to tailor treatments to individuals. Mindmaze

uses AI to optimize rehabilitation for stroke patients. Ginger.io uses it to recommend the best
time to take medication based each patients metabolism and other factors. A startup called
Turbine uses AI to design personalized cancer treatment regimens.
This kind of tailored treatments may reduce health expenditures by 5 to 9percent, add
0.2 to 1.3years to average life expectancy, and increase productivity by $200 per person
per year. Globally, the economic impact of such advances could range from $2trillion to
$10trillion.35
The age of analytics: Competing in a data-driven world, McKinsey Global Institute, December 2016.
35

In education, adaptive learning has been a growing trend, with some 40 companies, such
as Knewton and DreamBox Learning, already marketing adaptive learning systems to
schools in North America, Europe, and Asia.36 Adaptive learning attempts to address the
limitations of conventional classroom teaching by capturing information about what each
student knows and crafting custom lesson plans based on individuals knowledge and
progress. Instead of marching the class through the same lesson (and often covering the
same material repeatedly until most of the students get it), adaptive learning claims to deliver
the right content, at the right time, in the best way for each student. Providing a personalized
experience to students can have significant impact. At Arizona State University, an adaptive
learning program helps students who are struggling with remedial math. Student pass rates
have improved from 66percent to 75percent, and dropout rates are down by 7percent.37
Deep learning algorithms could further unleash the ability of AI techniques to tailor the
teaching experience to students specific needs and progress. Indicators such as facial
expressions, digital interactions, group interactions, and eye tracking can be captured
through computer vision to gauge students engagement and coach and assess students
in real time. A European Union project called iTalk2Learn is currently developing an open-
source intelligent-tutoring platform to help primary school students learn mathematics. A
combination of machine learning, computer vision, user modeling, and natural language
processing enables the platform to interact with and respond to a students speech
throughout a tutoring session.
Common to all sectors is the need for employees to adapt their skills to support and
complement the AI-powered user experience. As teaching becomes more adaptive,
teachers and professors might have to adjust to doing less lecturing and more tutoring and
coaching. Similarly, as sales and call centers become more virtual, human representatives
must hone their emotional intelligence skills to provide a service experience beyond the
capability of machines.
Early adopters and early case studies demonstrate AIs potential to transform business
processes, shake up entire sectors, increase profits, and create new sources of value. AI
applications are starting to reach maturity, and companies with serious, proactive adoption
strategies stand to gain significant competitive advantages. There are many industry- and
sector-specific use cases to inform companies when they define a focused strategy. Also,
while machine learning and deep learning underpin most opportunities, industries will need
to identify the AI technologies that will bring the most benefits to them, and then start to
develop their infrastructure, talent, and knowledge as early as possible to catch up on the
learning and adoption curves. AI is more than the sum of its parts: for truly impressive gains,
companies are building their AI capability across the value chain, integrating it into core
processes, and using it to enable their employees to be more productive.
In this chapter, we presented four areas where AI can create value. Such gains, however, do
not come overnight. In the next chapter, we will look at what is required from businesses, AI
vendors, and governments to seize this opportunity.
36
Learning to adapt: Understanding the adaptive learning supplier landscape, Education Growth Advisors and
Bill and Melinda Gates Foundation, 2013.
37
Arizona State University.

3. BUSINESSES, DEVELOPERS, AND
GOVERNMENTS NEED TO ACT NOW TO
REALIZE AIS FULL POTENTIAL
While AI has the potential to fundamentally reshape society, significant uncertainty remains
about how the technology will develop. For firms, governments and workers, this might
suggest a wait and see approach. However, we think there is a need for urgent but clear-
headed action to respond to the opportunities and risks that are already apparent.
For many firms, this will mean accelerating their digital journey to ensure that they can
effectively deploy AI tools. AI becomes impactful when it has access to large amounts of
high-quality data and is integrated into automated work processes. AI is not a shortcut to
these digital foundations. Rather, it is a powerful extension of them.
Developers have a crucial role to play in helping businesses realize the potential of the
technology. AI products need to address real-world business problems, not just provide
interesting solutions, and they must work at scale.
Finally, governments and workers should prepare for the wide-reaching changes ahead.
Public education systems and workforce training programs will have to be rethought to
ensure that workers have the skills to complement rather than compete with machines.
Cities and countries hoping to establish local AI ecosystems must enter the global
competition for AI talent and investment. And society as a whole will need to navigate the
unresolved legal and ethical issues that could become major barriers to realizing the benefits
of AI.
FIRMS NEED TO FOCUS ON AI USE CASES THAT ADD THE MOST

VALUE AND ENSURE THEY HAVE THE RIGHT DIGITAL ENABLERS
While the overall impact of AI to date may be relatively small in many industries, its potential
for disruption is high. We are already seeing examples of how AI-powered new entrants are
able to take on incumbents and win, as Uber and Lyft have done to the taxi industry. AI can
go beyond changing business processes to changing entire business models with winner-
take-all dynamics, and firms that wait for the AI dust to settle risk being left behind.
This underscores the need for action now. Firms must conduct sober analysis of what the
most valuable AI use cases are. They should also build out the supporting digital assets and
capabilities. Indeed, the core elements of a successful AI transformation are the same as
those for data and analytics generally (Exhibit8). This includes building the data ecosystem,
adopting the right techniques and tools, integrating technology into workplace processes,
and adopting an open, collaborative culture while reskilling the workforce.

Exhibit 8
Successful AI transformations require elements similar to those found in successful digital and
analytics transformations
USE
CASES/ DATA TECHNIQUES WORKFLOW OPEN CULTURE
SOURCES ECOSYSTEMS AND TOOLS INTEGRATION AND ORGANIZATION
OF VALUE
Break down data Identify fit-for- Adopt an open,

collaborative
Scan use-case silos purpose AI tools Integrate AI into
culture
Decide on the Partner or acquire workplace
horizon
level of to plug capability
Build trust in AI
processes
Articulate business
aggregation and gaps Optimize the
insights
needs and create Reskill the
pre-analysis Take an agile, human/machine
business case workforce to
Identify high-value test and learn interface
ensure
data approach
complementarity
SOURCE: The age of analytics: Competing in a data-driven world, McKinsey Global Institute, December 2016; McKinsey Global Institute analysis
Find the true source of value, and build a business case for it
The first step is to establish a solid AI business case and connect it to a firms strategy. This
requires separating the hype and buzz around AI from its actual capabilities in a specific,
real-world context. It includes a realistic view of AIs capabilities and an honest accounting
of its limitations, which requires at least a high-level grasp of how AI works and how it differs
from conventional technological approaches.
A portfolio-based approach, looking at use cases over a one- to five-year horizon, can be
helpful. In the immediate future, focus on use cases where there are proven technology
solutions today that can be adopted at scale, such as robotic process automation and some
applications of machine learning. Further out, identify use cases where a technology is
emerging but not yet proven at scale. Over the longer term, pick one or two high-impact but
unproven use cases and partner with academia or other third parties to innovate, gaining a
potential first-mover advantage in the future. Examples include commercial underwriting in
insurance and geo-modeling of subsurface deposits. Across all horizons, a test and learn
approach can help validate the business case, conducting time-limited experiments to see

what really works and then scaling up successes. The importance of fast, agile approaches
is discussed further later in this chapter.
To ensure a focus on the most valuable use cases, AI initiatives should be assessed and
co-led by both business and technical leaders. Given the significant advancements in AI
technologies in recent years, there is a tendency to compartmentalize accountability for
AI with functional leaders in IT, digital, or innovation. This can result in a hammer in search
of a nail outcome, or technologies being rolled out without compelling use cases. The
orientation should be the opposite: business led and value focused. This business-led
approach follows successful adoption approaches in other digital waves such as mobile,
social, and analytics.38
Build the data ecosystem

Data is at the heart of the disruptions occurring across economies and is recognized
as an increasingly critical corporate asset. Without data, getting the AI engine started is
impossible. Because of this, business leaders should know what data they already have
access to and where they can obtain additional data relevant for their companys future
success. Google and Facebook are well-known examples of companies that obtain most
of their revenue through insights they extract from the enormous quantities of data their
customers generate on a daily basis by using their services.
One important capability will be making data usable that is not available in a relational format
or that cannot be analyzed with traditional methodologies. Much data being produced in
industry today is flat data, without relational structurein manufacturing, an estimated
90percent of data is flat.39 Making this data usable requires new approaches that can
efficiently handle large volumes of different types of datafor example, NoSQL and
Hadoop technologies.
Firms also need to recognize potentially distinctive types of data that can create a
competitive edge in AI-enabled product offerings. Customer sentiment and geo-locational
real-time event data are examples of differentiating data for which competition for exclusive
access is likely to intensify. Certain data may become valuable only if combined with other
data sources in a larger ecosystem.
Given the rapidly increasing data output from sensors, machinery, and social networks,
organizations face challenges in how to handle such massive streams of information. While
some use cases for such data will be very concrete, with clear requirements, other potential
uses will be fuzzy or not yet fully defined. Some use cases will require significant time series
of data, while others may require real-time data. Companies will need to decide which data
to store in their original granularity and which to aggregate or pre-analyze. With increasing
data storage capacities in the cloud as well as more powerful edge computing capabilities
close to sensors, flexibility increases rapidly.
Know whatand whomyou need

To really capture the performance-boosting value of AI, companies need to build internal
capabilities and partner with or acquire additional know-how from AI startups or leading AI
firms. Roles that companies often have to fill are translators and data scientists (quants).
Translators bridge the gap between the techniques available to data scientists and the real-
world problems of management. Quants design, develop, deploy, and train AI technologies.
Raising your Digital Quotient, McKinsey Quarterly, June 2015.

38
Smartening up with artificial intelligence: How AI will transform Germanys industrial sector, Digital McKinsey,
39
April 2017.

Finding the talent to fill these roles is challenging. Out of approximately 150million workers
in the United States, for example, only 235,000 are data scientists.40 Given the scarcity of
talent, it can be useful to partner to get the AI engine up and running while simultaneously
developing and recruiting AI talent. One option for partnership is the build, operate,
transfer model. Regardless of the approach, a key challenge is finding tools that are fit
for the purpose. Off-the-shelf solutions can play an important role, but only if they are
well aligned to company-specific needs and properly account for the costs and time
of integration.
As with other digital technologies, a fast, agile, test and learn approach is important.
Small, fast steps ensure the right focus, for example, through simulation-based pilots that
allow companies to quickly test the impact estimates in the business case. Best-practice
companies set up cross-functional AI task forces that are able to prototype a solution within
weeks (where data is available), test it with the business units, and decide how to proceed.
Two-speed approaches can help, pushing ahead on newer, more flexible IT architecture
while gradually migrating legacy systems.
Internal and external collaboration are also important for agile companies. Internally,
teamwork and collaboration are especially important for digital technologies like AI,
which often cuts across traditionally siloed parts of organizations, from customer service
to fulfillment to supply-chain management to financial reporting. Agile companies also
collaborate beyond the boundaries of the firm, tapping into broader networks of learning
and innovation, and bringing suppliers and other partners along for the ride. This requires
digital leaders to recognize what theyre good at themselves and what others might do
better, and to improve their ability to work collaboratively with people and institutions.
Integrate to capture the benefits

Once companies capabilities are producing AI-powered insights, they must be integrated to
capture the benefits promised in the business case. Typically, this involves process redesign
to incorporate AI insights into the workflow. In some cases, this will involve automation. In
others, it requires getting the right data insights into the hands of the right personnel within
the organization. In either case, optimizing the human-machine interface is critical. Careful
thought needs to be given to determine what tasks the machine automates and how the
hand-off with humans will work (for example, between virtual agents and advisers). In many
cases, the change-management challenges of transforming what many people do within an
organization are greater than many of the technical challenges of implementing AI.
Adopt an open culture and reskill the workforce

To get the most out of AI in the long run, an organizational culture open to the collaboration
of humans and machines is required. Trust is a key enabler here. Due to the interplay of
training and inference in AI, the relationship between a machines inner workings and the
results it produces can become rather obscure. Instead of an algorithms predetermined
steps, it is, in a sense, the data used to train it that leads to a certain outcome. Humans will
need some time to adjust to this paradigm shift. That means the creation of an AI-ready
culture should be a priority early on. It may also require investment to build the capabilities of
workers, especially mid-level managers, to understand how to use data-driven AI insights
and to trust them as the basis for making decisions.
Companies should be forthright in addressing employee concerns if they expect them to

help develop and implement AI tools that will change business processes and potentially
automate some activities. Some workers will have to be retrained to work alongside AI-
directed machines, while others will have to be redeployed within the company or elsewhere
in the economy; businesses have a vital role to play in aiding these transitions. This will
40

require changes in skills, mindsets, and culture as we transition into a world where co-
workers include machines as well as other people.
As with all cultural and organizational changes, leadership is critical, a point reinforced by
our survey findings. Strong leadership support goes hand in hand with stronger AI adoption.
Respondents from firms that have successfully deployed an AI technology at scale tend to
report C-suite support that is nearly twice as high as those from companies that have not
adopted any AI technology.
AI PROVIDER COMPANIES NEED TO MATCH TECHNOLOGICAL

SOLUTIONS TO REAL-WORLD BUSINESS PROBLEMS
AI providers must go beyond offering interesting AI products to offering AI solutions for
real-world business problems. Our survey highlights the scale of this challenge in the
market today. Of the more than 3,000 firms we surveyed, 41percent reported that one of
the biggest barriers preventing them from further adoption of AI was the uncertain return on
investment, while 26percent reported a lack of relevant AI products on the market.
To close that gap, AI providers need to develop sector-specific expertise to understand

where major business opportunities are and what potential sector-specific solutions
would look like. This requires AI developers and engineers to become more familiar with
potential customers value chains and data systems. It also requires more emphasis on
demonstrating the technologies and applications, and the benefits they bring, at scale.
Meeting the market will also require AI developers to address discomfort with human-
machine interactions. In our survey, one-fifth of respondents said their customers were not
interested in AI-enabled products and services, and the figure was significantly higher in
health care and other sectors where the interactions are particularly intimate. Machines can
frustrate people if they cannot understand what the people are saying, deliver inaccurate or
inappropriate answers, or simply do not live up to promises made about them.
There are no easy answers to this evolving problem. AI providers must take the lead in
overcoming this hurdle, as when Amazon reduced the error rate for its intelligent personal
assistant, Alexa, by a factor of two between 2014 and 2016.41 Google took a different
approach by managing customer expectations for what to expect from its voice search
feature, making clear that it was not an intelligent personal assistant, in order to avoid some
of the customer disappointment expressed about Apples Siri.42
GOVERNMENTS SHOULD TAKE THE LEAD IN TACKLING

CROSS-CUTTING AI CHALLENGES
AIs impacts extend well beyond firms. There are serious implications for governments,
the workforce, and broader society. Indeed, the importance is highlighted by the growing
number of governments crafting national AI development plans in the past 18 months,
the United States, the United Kingdom, China, and South Korea have all issued national
strategic plans with significant AI dimensions, in some cases backed up by billions of dollars
of AI-specific funding initiatives. MGI has also looked at country-specific implications of AI
(see Box3, Chinas path forward).
41
Jordan Novet, Amazon has reduced Alexas mistakes in completing tasks by a factor of 2, Venturebeat.com,
July 13, 2016.
42
Farhad Manjoo, Siri is a gimmick and a tease but Google voice search is getting close to fulfilling Apples
broken promise, Slate.com, November 15, 2012.

Box3. Chinas path forward
While many of the challenges of AI are global in nature, the implications for specific
governments vary across countries. In April 2017, McKinsey Global Institute published a
discussion paper assessing the specific priorities for the Chinese government.1
Strategic priority 1: Build a robust data ecosystem. China can move to set and
implement data standards, open public-sector data for private exploration, and
encourage international exchange of data streams.
Strategic priority 2: Broaden adoption of AI within traditional industries. Adoption

of AI by Chinese firms is being held back by a lack of strategic awareness, the costs
of adoption, and a shortage of technical know-how. The government should look to
address these barriers using tax credit and subsidy tools, as well as pioneering adoption
within the government.
Strategic priority 3: Strengthen the pipeline of specialized AI talent. To address

Chinas AI talent gap, the government needs to invest in AI-related education and
research programs, reorient the education system for a greater focus on innovation and
digital skills, and devise an immigration policy to attract the best global talent.
Strategic priority 4: Ensure the education and training systems are prepared
to develop technology skills and retain large segments of the workforce. The
government must proactively identify the jobs that are most likely to be automated and
ensure that retraining programs are made available to the segments of the labor force
whose livelihoods are at risk. These efforts could involve collaborating closely with
vocational training schools and providing educational vouchers to workers.
Strategic priority 5: Establish an ethical and legal consensus among Chinese

citizens and in the global community. The Chinese legislature will need to provide a
framework to deal with the current legal uncertainties around AI. The government should
also look to establish a regulatory body to monitor and regulate AI activities. China
could also take the lead in forming an international governing body to promote peaceful,
inclusive, and sustainable development of AI technology.
1
Artificial intelligence: Implications for China, McKinsey Global Institute, April 2017.
The range of issues confronting governments is vast, ranging from ethical and equity
concerns to the setting of data standards (see Box4, An overview of ongoing challenges).
In this report, we explore two particularly pressing issuesreskilling of the workforce and
supporting local AI industries. These are not issues for government to solve alone. The
Partnership on AI is one effort at helping to ensure progress, bringing together civil society
organizations such as the American Civil Liberties Union, tech leaders such as Amazon,
Apple, Facebook, Google, and Microsoft, and other partners from the private sector,
including McKinsey & Company.

Box4. An overview of ongoing challenges
AI presents a wide range of issues for government and society. We outline some of these issues
below, along with some ideas about how to address them. Progress on these issues will be
critical both to realizing the potential benefits of AI and to safeguarding people from the risks.
Encourage broader uptake of AI. Adoption of AI tends to be concentrated in relatively

digitized industries and, within those industries, in the firms that are already at the digital
frontier. Broader adoption of AI and supporting digital technologies, especially in smaller
firms, could be important to supporting productivity growth across the economy and to
healthy, competitive markets. Broader adoption could also support more equal wage
growth. AI can drive improvements in labor productivity, which in turn drives higher wages.
Broadening adoption will ensure that the productivity benefits of AI are spread across more
firms and workers, rather than pushing up the wages of only those in frontier firms, who are
already located near the top of the income distribution.
Address employment and income-distribution concerns. AI-powered automation

could have a profound impact on jobs and wages. In our survey, the majority of firms did
not expect AI to significantly reduce the size of their workforce. However, there might still
be people whose skills and capabilities are mismatched to the work that needs doing.
Governments may have to rethink models of social support. Various ideas are under
consideration, including work sharing, negative income taxes, and universal basic income.1
Resolve ethical, legal, and regulatory issues. AI presents a range of ethical, legal, and
regulatory issues. Real-world biases risk being embedded into training data. Since the real
world is racist, sexist, and biased in many other ways, real-world data that feeds algorithms
will also have these featuresand when AI algorithms learn from biased training data,
they internalize the biases, exacerbating those problems.2 There are also concerns about
the algorithms themselveswhose ethical guidelines will be encoded into them, what
rights should people have to understand the decision-making process, and who will be
responsible for their conclusions? This has led to calls for algorithmic transparency and
accountability.3 Privacy is likewise a concernwho should have ownership of data, and
what safeguards are needed to protect highly sensitive data, such as health-care data,
without destroying its usefulness? Organizations leading efforts to tackle these questions
include the Partnership on AI, OpenAI, the Foundation for Responsible Robotics, and the
Ethics and Governance of Artificial Intelligence Fund.
Ensure the availability of training data. An abundance of data is critical for AI training
systems. Opening up public-sector data can spur private-sector innovation. Setting
common data standards can also help. In the United States, for example, the Securities
and Exchange Commission mandated in 2009 that all public companies must disclose
their financial statements in XBRL (extensible business reporting language) format, thereby
ensuring that public data is machine readable.
Deploy AI within government. AI has tremendous potential benefits for the public sector
as well as the private sector. Improved planning, targeting, and personalization could
deliver a much-needed step change in both the efficiency and effectiveness of government
services.4 In our sector deep dives (see Appendix A), we explore what an AI-powered future
could look like for two particular areas of governmenthealth care and education.
1
2
3
See, for example, Statement on algorithmic transparency and accountability, Association for Computing
Machinery US Public Policy Council, January 2017.
4
For further discussion of digital technology and analytics on government productivity, see Government
productivity: Unlocking the $3.5trillion opportunity, McKinsey Center for Government, April 2017.

Prepare the workforce for a lifetime of reskilling
Artificial intelligence tools have the promise to change our lives as fundamentally as personal
computers did a generation ago. But, as we saw then, gains will be accompanied by losses.
In some cases, full or partial automation from AI will displace labor. In A future that works, we
show that almost half of all work activities have the potential to be automated by adapting
currently proven technology. We estimate that 60percent of all occupations have at least
30percent technically automatable activities (Exhibit9). However, automation will change far
more occupationsby, for example, partially automating themthan it will replace.43
Exhibit 9
While few occupations are fully automatable, 60 percent of all occupations have at least 30 percent technically
automatable activities
Automation potential based on demonstrated technology of occupation titles in the United States (cumulative)1
Share of roles (%)

Example occupations 100% = 820 roles
100 1 <5% of occupations consist of

Sewing machine operators, activities that are 100% automatable
graders and sorters of
agricultural products >90 8
Technical automation potential (%)
>80 18
Stock clerks, travel agents,

>70 26
watch repairers
>60 34
In about 60% of
Chemical technicians, >50 42 occupations, at least 30%
nursing assistants, of activities are automatable
Web developers >40 51
>30 62
Fashion designers, chief

>20 73
executives, statisticians
>10 91
Psychiatrists, legislators
>0 100
1 We define automation potential according to the work activities that can be automated by adapting currently demonstrated technology.
SOURCE: US Bureau of Labor Statistics; A future that works: Automation, employment and productivity, McKinsey Global Institute, January 2017; McKinsey
Global Institute analysis
Not all AI innovations will displace labor. Many AI applications target non-labor cost savings,
as when an AI algorithm reduces a factorys energy use or when AI is used in predictive
maintenance. Indeed, less than a fifth of AI adopters in our survey said their primary driver
for adoption was to reduce labor costs. Improving capital efficiency was cited more often, as
were revenue-focused drivers such as enhancing their product offering.
Although one of the near-term benefits of automation is labor cost reduction, our survey
shows 24percent of firms that have adopted AI at scale expect to increase the size of their
workforce in response to AI, anticipating growth in their business activities. Moreover,
82percent of AI-aware firms do not expect to significantly reduce the size of their workforce,
43

owing in part to increased demand for new skills, such as those from data scientists and
data translators.
The implication of these changes is clear: companies need to update the skills in their
workforces, and individuals need to acquire skills that work with, not compete against,
machines.44 For people already in the workforce, reskilling will be essential. Much of this
reskilling can occur on the job through stronger professional development programs. For
people transitioning between jobs, vocational and adult education programs must be
strengthened. These work best when they are short, affordable, and closely linked to the job
market. Nanodegree programs are one recent innovation in this space. Governments and
training institutions also need to do better to ensure alignment of training programs with the
jobs of the future. The World Economic Forum argues that one bet governments should look
to make is on the care economy. AI-powered predictive analysis could also be put to use to
anticipate future pockets of skill shortages and oversupply.
For students, a future-focused curriculum is a necessity. The World Economic Forum

identified 16 skills that are needed in the 21st centuryincluding creativity, collaboration,
initiative, and adaptabilitybut are not included in standard curricula.45 People are also
questioning the learn then earn model, wondering if lengthy degree programs still
make sense in a world of fast-changing jobs. Instead, there are calls for a new deal on
lifelong learning.
Building local industries

Governments serious about establishing their cities or countries as hubs for AI development
will need to join the global competition to attract AI talent and investment.
Our global review finds that AI investment is highly concentrated geographically: in 2016,
the United States absorbed around 66percent of external investment (VC, PE, and M&A
activity). China was a distant second, at 17percent, but it is growing fast. Unsurprisingly, the
San Francisco Bay Area and Silicon Valley is the biggest and most important ecosystem,
attracting around 40percent of global external investment in 2016 (Exhibit10). Other US
cities, such as Boston and New York, have also drawn AI investment, and are followed by
two strong China AI tech ecosystems: Beijing and Shenzhen. In Europe the only clear and
strong ecosystem is currently London. Other vibrant European AI ecosystems may appear
within the next few years, given exceptional startup activity and VC investment in 2016,
especially in Germany, France, and the Nordic region.46
Many of the most important hubs for external investment are also key locations for tech
giants AI research and development work. However, we also see a pattern of big tech
investing in new areas to scout and attract local talent, including Google at the University
of Montreal, Intel at Georgia Tech, NVIDIA at the National Taiwan University, and Facebook
in Paris.
These hubs benefit not only from the creation of highly skilled, highly paid jobs, but also,
critically, knowledge and innovation spillovers. Employees become AI entrepreneurs, AI-
savvy workers switch between companies, and innovative AI products can be developed
for and deployed in the local market. Indeed, our survey suggests that countries that lead
the way in AI investment and innovationthe United States and Chinaalso lead the way in
AI adoption.
44
45
Jenny Soffel, What are the 21st-century skills every student needs? World Economic Forum, March 10,
2016.
46
Galina Degtyareva, European AI startups landscape, Medium.com, March 21, 2017.

Exhibit 10
The United States and China dominate the AI landscape, with Europe falling behind
The most vibrant AI hubs
Silicon Valley New York Beijing

Top global hub for startups Leading hub for financial and Leading in volume of academic
12,70015,600 active media industries research output in AI coming from
startups AI talent pipeline from universities Tsinghua, Beihang, and Peking
2 million tech workers such as Cornell universities
Global leader for VC Strong funding ecosystem Extensive involvement of tech leaders,
investment second in the world after Silicon especially Baidu
Headquarters of many top Valley for the absolute number of AI identified as a strategically important
high-tech companies early-stage investments technology by the Chinese government
Boston London Shenzhen

Long history of cooperation Global finance center, supporting both Hub for electronics
between science and industry investment and fin-tech applications manufacturing firms such as
World-class universities such European leader of VC startup investment Huawei and ZTE
as MIT developing advanced Presence of top high-tech companies Strong expertise in hardware
technologies and providing a Talent pipeline and research expertise from AI identified as a strategically
talent pipeline universities such as University of important technology by the
Cambridge, Imperial College, and Oxford Chinese government
and the external investment behind their growth1

$ billion (estimate)
North America Asia Europe
2010 1.02.0 ~0.1 ~0.1
2013 2.54.0 ~0.2 ~0.1
2016 5.08.0 1.52.5 1.11.7
1 Estimates consist of annual VC investment in AI-focused companies, PE investment in AI-related companies, and M&A done by corporations. Includes only
disclosed data available in databases, and assumes that all registered deals were completed within the year the transactions were announced.
SOURCE: Capital IQ; Pitchbook; Dealogic; S&P; McKinsey Global Institute analysis

There are opportunities for governments to act on multiple fronts. Thoughtful incentives to
attract both investment and talent are helpfultax breaks for AI entrepreneur immigrants
and special tech visa quotas are two approaches governments have adopted. The United
Kingdom, for example, has established the Tech Nation Visa Scheme, which awards up to
200 visas annually, without work sponsorship requirements, for applicants with exceptional
talent or promise in the digital space.47
Funding for leading-edge science programs is also important, be it through grants to

universities, creation of government laboratories, or joint research initiatives with the
private sector. US federal government investment into unclassified AI R&D topped
$1billion in 2015.48 The South Korean government has announced it is investing 1trillion
won ($900million) to build a public-private AI research center jointly with leading Korean
conglomerates.49 Meanwhile, Chinas National Development and Research Commission
has launched a national engineering laboratory for deep learning, led by Baidu.50
Governments also have other important levers to foster and consolidate AI hubs. They
can act as lead customers, providing an important boost to startups by giving them a
credible reference case. They can ensure that local regulations are AI-friendly, providing
space for new technologies to be safely piloted and for greater clarity around legal liability
and data-ownership issues. As mentioned above, they can also improve the availability
of training data, both by opening up their own data banks and by setting data standards.
These actions can help secure an environment that is uniquely attractive to AI investors and
entrepreneurs alike.
After decades of hopes and disappointment, AI is back and could be set to drive profound
changes in the global economy. Investment has been growing fast since 2013, with
tech giants making huge plays on AI technology development and deploying it across
their businesses. We are already seeing examples of real-life business benefits for early-
adopting firms.
However, adoption of AI technologies remains largely at an experimental stage. Indeed, the

gap between early adopters and the rest is set to grow. While many companies have yet
to be convinced of AIs benefits, frontier firms are charging ahead. Early AI adopters tend
to be larger firms with mature digital strategies that display both deep and broad adoption
patterns. Their focus on AI is driven by the desire to increase revenue and reduce cost, and
they have momentum and support for AI from the top of the company.
Significant gains are there for the taking. For many companies, this means accelerating the
digital transformation journey. They will have to put the right digital assets and skills in place
to be able to effectively deploy AI.
For governments, there is an urgent need to support both firms and citizens to ensure that
they benefit from the AI-powered digital economy. This means preparing workers for rapidly
changing job requirements, facilitating AI investment and adoption, and solving for broad-
based, productivity-driven economic growth.
47
Tech Nation Visa Scheme: High demand as scheme enters fourth year, Tech City UK blog, April 13, 2017.
48
National Science and Technology Council, Preparing for the future of artificial intelligence, Executive Office of
the President, October 2016.
49
Mark Zastrow, South Korea trumpets $860-million AI fund after AlphaGo shock, Nature, March 18, 2016.
50
Weining Hu, How China is becoming a world leader in artificial intelligence, China Briefing blog, March 14,
2017.

APPENDIX A: FIVE CASE STUDIES
RETAIL
AI CAN HELP PREDICT DEMAND, AUTOMATE OPERATIONS,
AND DELIVER A BETTER SHOPPING EXPERIENCE
Retailers are already beginning to apply AI, machine learning, and robotics in major
parts of the value chain. Most importantly, AI technologies could eliminate many levels of
manual activities in areas such as promotions, assortments, and supply chain. The three
areas of greatest opportunity in the short to medium term are promotions, assortment,
and replenishment. Major retailers are experimenting with AI in all of these areas.
Digital native e-commerce companies are leading the way, using AI to predict trends,
optimize warehousing and logistics, set prices, and personalize promotions. Some even
aim at the full anticipation of customers orders, shipping goods without waiting for a
purchase confirmation.
The use of AI in retail can generate several benefits. First, it helps people make smarter
decisions, with more accurate and real-time forecasting. Good forecasts help improve
supply management, define impactful thematic promotions, and optimize assortment and
pricing. Second, AI can make operations more efficient, thanks to a combination of robotics
and process optimization that enhances productivity and reduces manual labor costs. AI
will enable retailers to increase both the number of customers and the average amount they
spend by creating personal and convenient shopping experiences.
Can artificial intelligence help traditional, non-digital retailers catch up, or will it further widen
the divide between the agile and data-driven internet pure players and historical brands
that are lagging behind? Success will hinge on the capacity of retailers to get on board and
secure access to strategic data while reinventing the shopping experience. But before we
delve into the conditions of full achievement, lets explore what the future could look like
in 2030.
Retailers can know more about what shoppers want

sometimes before shoppers themselves
In the future, artificial intelligence could help forecast and automate retailers decision
making in real time. By identifying and learning from patterns in large volumes of data,
spanning many disparate sourcesprevious transactions, weather forecasts, social media
trends, shopping patterns, online viewing history, facial expression analysis, seasonal
shopping patternsAI can help companies adjust to and master an increasingly dynamic
market environment. By improving forecasting accuracy, machine learning and computer
vision can help better anticipate consumer expectations while optimizing and automating
supplier negotiations.
The impact of AI-enabled forecasting is already being demonstrated. For instance, a

European retailer was able to improve its earnings before interest and taxes (EBIT) by 1 to
2percent by using a machine learning algorithm to anticipate fruit and vegetable sales. The
company automatically orders more produce based on this forecast to maximize turnover
and minimize waste. Similarly, German e-commerce merchant Otto has cut surplus stock by
20percent and reduced product returns by more than two million items a year, using deep
learning to analyze billions of transactions and predict what customers will buy before they
place an order. The system is 90percent accurate in forecasting what the firm will sell over

the next 30 days, so Otto allows it to order 200,000 items a month from vendors with no
human intervention.
AI technologies can also help retailers predict future store performance when expanding
their physical footprints. As more sales migrate online, non-digital store sales per square
meter are declining. In the United Kingdom, retailers would need to shave space by more
than 20percent to return to 2010 sales densities in real terms if all other factors were
held equal. It has become crucial for retailers to optimize store space and location. One
Japanese retailer applied machine learning to understand profitability drivers when picking
the location of a new concept store.
Operations are ripe for automation

Warehousing and store operations present a rich set of optimization opportunities for AI
application. For some non-digital retailers, notably supermarkets, automation of operations
would actually make an existential difference. Many supermarkets offer online sales and
home delivery to match online grocers but still carry the full cost of physical stores, so the
costs of online servicein the United Kingdom, roughly 5 to pick items off the shelves for
an 80 order and 8 to deliver themwipe out the industrys 2percent profit margin many
times over.
Autonomous robots can work alongside people to increase productivity and reduce injuries.
Swisslog has reduced stocking time by 30percent since it began using autonomous guided
vehicles in its warehouses. DHL unleashed a pair of fully automated trolleys last year that
follow pickers through the warehouse and relieve them of physical work.
In store, machine learning can help optimize merchandising, with opportunities to improve
assortment efficiency by 50percent. A retailer was able to generate a sales uplift of 4 to
6percent by using geo-spatial modeling to determine micromarket attractiveness and
leveraging statistical modeling to predict and minimize running out of stock. With machine
learning, these efficiencies would be realized in real time and would gain in accuracy as they
learn from new data.
Ocado, a UK online supermarket, is one company that has embedded AI at the core of
its operations. In the retailers warehouse, machine learning algorithms steer thousands
of products over a maze of conveyor belts and deliver them to humans just in time to fill
shopping bags. Other robots whisk bags to delivery vans whose drivers are guided by an AI
application that picks the best route based on weather and traffic conditions.
Retailers are getting personal

Empowered by the ease, economy, and immediacy of online shopping, many consumers
already expect personalized, immediate, pitch-perfect help. In the future, AI will be
invaluable to marketers trying to reach hyperconnected consumers who continuously
redefine value by comparing prices onlineeven, and particularly, when browsing in a
non-digital store. Smartphone penetration necessitates an omni-channel strategy, and AI
can help optimize, update, and tailor it to each shopper in real time. Insights-based selling,
including personalized promotions, optimized assortment, and tailored displays, could
increase sales by 1 to 5percent. Online, this kind of personalization, combined with dynamic
pricing, can lead to a 30percent growth in sales.
The pure internet players are several steps ahead in targeted marketing, thanks to data
gathered online. Traditional retailers need to start gaining access to data assets to
compete. Carrefour, the global retailer based in France, and Target in the United States
have both deployed electronic beacons in stores to collect data about customer behaviors
and purchasing patterns, and they use machine learning algorithms to determine which

personalized promotions to send customers as they shop. Carrefour reported a 600percent
increase in app users after it deployed beacons in just 28 stores.51
As natural language processing develops, AI-enabled personalization could go far

beyond the realm of targeted promotions. In store, virtual assistants could identify repeat
customers using facial recognition, analyze their shopping history to make suggestions, and
communicate in a conversational way using natural-language processing and generation. In
the meantime, online retailers are trying to give more of a human touch to the Web and make
personal recommendations to shoppers. Stitch Fix, an online personal-shopper service, has
an algorithm that reviews the images clients display on Pinterest to better understand their
styles, even if they have a difficult time articulating it online. Online retailers also use smart
agents to understand shoppers needs. One example is floral retailer 1800-Flowers digital
gift concierge, powered by machine learning and language recognition, which proposes a
selection of products based on a chat with the shopper.
Bringing it all back home

Enhanced user experience is the area that offers probably the most futuristic perspectives
for AI in retail. Deep learning and computer vision technologies also will help store owners
compete with the one-click convenience of online retailers by eliminating checkout
altogether. Amazon Go, an experimental grocery in Seattle, allows shoppers to take
goods off shelves and leave without seeing a cashier or stopping at a self-checkout kiosk.
Computer vision identifies them as they enter the store, then links them with products taken
from shelves. When shoppers leave, the system deducts the cost of the items in their bag
from their Amazon accounts and sends an email receipt.
At home, virtual assistants further push the convenience boundaries. In the future, they
could alert users that theyre about to run out of a product and suggest buying more.
Googles smart speaker service, Google Home, allows shoppers to complete orders with
50 participating Google Express retailers, such as Costco, Whole Foods, and PetSmart,
while Amazons Alexa has partnerships with more than 100 third-party services. Recent
developments in smart home assistants pave the way for a significant shopping disruption,
where computer vision helps identify desired goods by taking a picture or the assistant
identifies preference patterns from images and videos liked online by consumers. Amazons
new Echo Look device, for instance, which was unveiled in April 2017, incorporates a
camera into Alexas virtual assistant function and recommends styles based on the users
wardrobe and body shape, combining machine learning and computer vision.
In the future, artificial intelligence technologies could also be used at scale to deliver goods
minutes after purchase. Most of todays effortsby players as big as Amazon and as small
as the Reno, Nevada, startup Flirteyfocus on unpiloted aerial drones. Flirtey made its
first delivery to a private residence, a box of snacks from a local convenience store, in July
2016. In Europe, an Estonian startup, Starship Technologies, has taken a different path: six-
wheeled delivery robots that putter along city pavements at 4 miles per hour.52 The ramp-
up of drones and robots hinges on the application of deep learning technology to allow
innovative problem solving and decision making, and on airspace regulation, as civil aviation
authorities question the operating of drones over populated areas and close to the flight
paths of piloted aircraft.
51
Shubhi Mittal, 25 retailers nailing it with their proximity marketing campaigns, Beaconstac.com, February 11,
2016.
52
Starship Technologies launches testing program for self-driving delivery robots with major industry partners,
Starship Technologies press release, July 6, 2016.

Retailers can know more about what shoppers
wantsometimes before shoppers themselves
Case study: Retail
Machine learning Autonomous

personalizes drones using deep
promotions to match learning technology
shoppers' profiles; complete last-mile
in-store beacons send delivery, and are
offers to their smart- able to handle
phones as they browse obstacles or absent
through the store recipients
Facial recognition Computer vision with
software, machine deep learning
learning, and natural identifies articles
language enable bagged by shoppers;
virtual agents to greet adding data from
you personally, sensors, AI allows
anticipate orders, and non-stop checkout
provide directions and autmatic payment
Interactive screens and Stores update and

tabletops enabled with optimize prices in real
computer vision and time, with machine
deep learning can identify learning leveraging
articles and recommend data on competitors
complementary products prices, weather, and
and uses that fit inventory levels to
shoppers' lifestyle profile maximize revenues
An autonomous AI-enhanced robots
shopping cart follows continuously track
you in the store, and inventory, recognize
can find its way to empty shelves, and
your vehicle or to a replenish them; other
robot or drone for robots fill bags in the
home delivery warehouse

Moving forward
Keeping up with the competition in this new realm of retailing will be as hard as it is
important. Retailers would be wise to invest in data gathering up and down their supply
chain to seek a competitive advantage. This implies shifting to a collaborative mindset to
allow insights to flow through the entire value chain.
On the one hand, partnerships between retailers and their suppliers will become
important to improve supply chains and marketing, optimize pricing, and achieve efficient
merchandising. Major retailers are already experimenting with data integration. For instance,
Walmart has started sharing real-time data with major consumer packaged goods players
using a data lake.
On the other hand, cross-industry partnerships between retailers and other players will
evolve to create better customer insights. Ecosystems will emerge with third-party legal
entities that bring together shareholders such as retailers, loyalty card providers, and banks.
The trend is already under way in Brazil, Turkey, Thailand, Indonesia, the United Kingdom,
and other countries.
Computer vision technology in physical retail outlets can give tremendous insight through
customer footfall data, customer reactions, and promotions (for example, how long
shoppers stand in front of a promotion and whether they take the product), and can enable
automatic checkout. Natural language processing and deep learning will facilitate the rise of
virtual assistants by making dialogue more intuitive and improving it faster.
However, the rise of virtual assistants that take orders directly from shoppers may someday
lead to the disintermediation of retailers. With automation, retailers also will need to rethink
their sales teams skills, as human interaction will be increasingly reserved for higher
engagement types, involving emotional intelligence, excellent product knowledge, and
brand championing. The retailing revolution has room to run.

ELECTRIC UTILITY
AI CAN MAKE THE SMART GRID SMARTER AND REDUCE
THE NEED FOR UTILITIES TO ADD POWER PLANTS
The electric utilities sector has great potential to embrace artificial intelligence in the
coming years. At every step of the value chain, from power generation to end consumers,
opportunities for machine learning, robotics, and decision-making automation exist that
could help electric utilities better predict supply and demand, balance the grid in real time,
reduce downtime, maximize yield, and improve end-users experience.
The increasing deployment of renewable energy sources has introduced significant volatility
in energy supply, with variation of up to 60percent. Demand also fluctuates dramatically
and frequently by time and by region, with weather and events such as the Super Bowl
creating demand peaks sometimes for period of less than an hour. Following an investment
wave during the 1970s and 1980s, transmission and distribution companies, under financial
constraints and regulatory requirements to hold down costs and rates, put far less money
into improving their networks over the following decades. As a result, current energy grids
are ill-equipped to smooth out these spikes, and excess power is regularly lost at top dollar.
The increasingly complex network of stakeholders and assets, combined with aging
critical capital assets, highly uncertain demand and supply, non-linear power loads, cost
pressures, and price deregulation, is engineering real momentum for artificial intelligence
and robotics.
Ultra-accurate forecasts can make possible the integration of both

additional renewable energy and excess power into the grid
One of the most acute cases for AI in electricity is around demand and supply prediction. An
inaccurate load forecast in the power grid can affect many stakeholders: power generation
determines which power sources should be allocated for the next 24 hours based on one-
to two-day forecasts; transmission grids assign resources based on power transmission
requirements; and electricity retailers calculate energy prices based on estimated demand.
Electric utilities are starting to explore artificial intelligence technologies to produce more
accurate short-term load forecasts. DeepMind, the AI startup bought by Google in 2014,
is currently working with National Grid to predict supply and demand peaks in the United
Kingdom by using weather-related variables and smart meters as exogenous inputs, hoping
to cut national energy usage by 10percent and maximize the use of renewable power
despite its intermittence.
In the future, machine and deep learning technologies could forecast demand and supply
in real time and optimize load dispatch, thereby saving energy and cost. For a network
that experiences demand ranges between 10 and 18 gigawatts, savings could reach 100
megawatts over periods of one to four hours per day.53 More reliable forecasts would allow
utilities to delay or even avoid ramping up a fossil-fuel-powered station. It would also offer
cost-effective alternatives to operators, who currently consider building new plants to
absorb seemingly impossible variability.
Grid modernization and deployment of smart meters are already under way in most
countries, aiming at creating a more dynamic matching of supply and demand; the use
of artificial intelligence enables suppliers to better predict and optimize load dispatch.
Smart grid initiatives allow small, private energy producers (even individual homeowners)
to sell excess capacity back to the grid. The technology is developing quickly: the United
Jaime Buitrago and Shihab Asfour, Short-term forecasting of electric loads using nonlinear autoregressive
53
artificial neural networks with exogenous vector inputs, Energies, volume 10, number 1, 2017.

States alone has committed more than $9billion in public and private funds since 2010.54
In Europe, Sweden and Italy have replaced nearly all meters with smart meters; other
European countries are on track to finish the conversion within 10years. In 2011, Chinas
State Grid Corporation announced plans to invest $45billion in smart grid technologies and
another $45billion between 2016 and 2020.55
AI could also help utilities assess the reliability of new small supply players, such as
households, by predicting the lifetime of their storage units and their suitability for integration
in a power storage scheme. Demand-side response aggregators currently focus only on
larger firms due to the cost of scanning hundreds of small participants. In the future, the
grid could become a marketplace where grid operators bid on power offered for sale by a
large number of small players with electricity from sources such as electric-car batteries and
rooftop solar cells. Machine learning could help automatically assess this large volume of
minisuppliers, who, collectively, could be instrumental in addressing demand peaks.
Yield optimization, predictive outage, and preventive

maintenance can help better plan the grid
The other lever where AI and robotics could help reduce costs is operations, from power
generation to transmission and distribution.
With AI, power providers could maximize their generation efficiency with real-time
adjustments across assets. For instance, machine learning can help optimize wind turbines
yield based on their own past performance, real-time communication with other wind
farms, the grid status, and changes in wind speed and direction. GE Renewables recently
introduced a digital wind farms concept, which optimizes yields with machine learning
applied to turbine sensors data, and modular turbines that can be customized to conditions
at each installation site.56 GE says the technology could boost a wind farms energy
production by as much as 20percent and create $100million in extra value over the lifetime
of a 100-megawatt farm.
Power generation yield can also be bolstered by reducing downtime and improving
preventive maintenance. To date, preventive maintenance efforts have had a limited impact
because firms can be overwhelmed by the sheer volume of sensor data and inaccurate
alerts. This is an opportunity for AI technologies, which thrive on mountains of information.
Advanced analytics already demonstrate the benefit of intelligent maintenance. Some coal
power plants, for instance, were able predict the timing of failures within one week six to
nine months in advance, with 74percent accuracy. Overall, we estimate that optimizing
preventive maintenance, automating fault prediction, and increasing capital productivity
through AI applications could increase power generation earnings before interest, taxes,
depreciation, and amortization (EBITDA) by 10 to 20percent.
Machine learning can be instrumental in reducing energy losses in transmission and

distribution. Looking across multiple assets on a customer site, AI tools can make intelligent,
real-time decisions to balance the grid, such as rescheduling the consumption of power
assets within a building in synchrony and leveraging the inherent flexibility of air conditioning
and refrigeration systems to maximize the benefit of solar panels and batteries installed on
site. Upside Energy, a startup, received a UK government grant in 2017 to use AI to manage
a portfolio of batteries and other storage assets and provide real-time energy reserves to
the grid. The company estimates that machine learning could be used to help unlock up to
54
2014 smart grid system report, US Department of Energy, August 2014.
55
Melanie Hart, China pours money into smart grid technology, Center for American Progress, October 24,
2011.
56
GE launches the next evolution of wind energy making renewables more efficient, economic: The digital wind
farm, General Electric press release, May 19, 2015.

six gigawatts of demand-side flexibility that can be shifted during the evening peak without
affecting end users.57
By shaving off demand peaks during the day, utilities could postpone or even forgo the need
to add generating capacity that would be requiredand would generate revenueonly for
short periods.
Transmission and distribution firms also could shift from time-based maintenance to
condition-based maintenance. Leveraging data from sensors, communications devices,
and other hardware that tracks and controls objects remotely, machine learning applications
could liberate grid operators from decommissioning assets before their useful lives have
ended, while enabling them to perform more frequent inspections and maintenance to keep
assets working well. A European power distribution company was able to reduce its cash
costs by 30percent over five years by analyzing 20 variables to determine the overall health
of power transformers and diagnose the condition of individual components. The company
could predict parts failures more accurately and prioritize repairs based on which faults
cause the most disruption; it can even identify faults that do not yet have enough data to be
spotted by todays systems.
In the future, one can imagine that operational trade-offs among several power stations
or within the distribution networksuch as outage planning refinement, replacement
vs. run-to-failure decisions, and cutoff for spare deliverywill be automatically made by
advanced analytics and machine learning algorithms. When necessary, inspections could
be automatically scheduled, with machine learning algorithms making a judgment call on
whether the deployment of drones and smart robots could be sufficient at first, or whether
human intervention is needed right away. Drone-based surveillance can replace the
time-intensive and risky manual inspection of turbines, for example. This reduces turbine
downtime because inspections by drones or robots can happen while turbines remain
running, helping keep costs down.
At last, AI tools can help tackle non-technical energy losses, such as electricity theft, which
is a significant problem in some developing countries. In Brazil, electricity theft accounts for
up to 40percent of electricity distributed.58 In Hungary, Eon, after realizing that consumption
in one town was three times the value of the electricity invoiced, found that 50percent of
businesses inspected and more than 80percent of households in the town were stealing
electricity.59 The utility started to use machine learning to narrow down the list of suspicious
users for its private detectives to investigate, and it was able to reduce theft by 30percent.
Machine learning can analyze customer data, including usage patterns and payment history,
and compare it to known irregular behavior.
By our reckoning, network operators that adopt the full suite of AI appsincluding
inspection automation, preventive maintenance, demand management, and theft
detectioncould raise their EBITDA by 20 to 30percent.
57
Michael Bironneau, How AI is shaping the future of energy, Clean Energy News, February 21, 2017.
58
Patrick Glauner, Large-scale detection of non-technical losses in imbalanced data sets, PhD dissertation,
University of Luxembourg, March 2016, and expert interviews.
59
Kester Eddy, Hungarys power thieves, Financial Times, March 25, 2011.

AI can unleash market forces on retail electricity markets
Machine learning applications have the power to tailor electricity prices based on the large
volume of data currently being provided by the growing number of smart meters and other
devices and sensors. In the future, if regulators allow dynamic tariffs, utilities could adopt
machine learningbased dynamic pricing that would permit them to protect their margins
and reduce customer churn while maximizing their assets. For instance, they could use
time-of-day pricing to encourage customers to shift nonessential consumption to earlier in
the morning or later in the evening, when demand is lower.
Energy retailers also would be able to use AI to create custom benefits such as low rates
or extra service in order to hold on to their most valuable, high-volume clients. While price
sensitivity is a key consideration in attracting new customers and reducing churn, machine
learning can also help address another critical component for utilities marketing strategies,
that is, identifying which customers are the most profitable.
The emergence of smart grids around the world also creates an opportunity for AI to
support energy trading, not only for utilities but also for prosumers, consumers who will
be able to sell excess power back to grid operators. Data and analytics are transforming the
way markets connect sellers and buyers for many products and services, and this holds
particularly true for grid operators. Hyperscale digital platforms can have a notable impact
as electricity demand and supply fluctuate frequently. AI can help produce better and faster
matches. AI-powered hyperscale platforms could transform energy markets by enabling
smart grids to deliver distributed energy from many small producers. In the Netherlands,
some startups are using the peer-to-peer model to match individual households directly
with small providers (such as farmers) who produce excess energy. Vandebron, for instance,
charges a fixed subscription fee to connect consumers with renewable energy providers;
in 2016, this service provided electricity to about 80,000 Dutch households. Utilities could
also inform their purchases and sales for trading activities on liquid, unpredictable over-the-
counter markets or through more reliable power-purchasing agreements.
AI can transform the user experience with consumption

tailoring and automation for more convenience
Energy consumers also can benefit from AI. Since the liberalization of energy retailing, new
entrants have piled into the market. In Europe, customers can choose from more than 20
suppliers, many competing on price alone. AI can help understand consumption patterns,
tailor the value proposition as well as consumption to the users preferences, and limit the
hurdles for switchers.
Machine learning can help consumers deal with the complex task of selecting their
electricity supplier based on users preferences in terms of pricing and energy generation
type, as well as metering measurements. Lumator has developed software with Carnegie
Mellon University in Pittsburgh, Pennsylvania, that scans the market for the most suitable
electricity supply deal. Lumator claims it can save people between $10 and $30 a month
on their bills. In the future, AI could automatically switch energy plans, without consulting
consumers or interrupting service, as the best deals become available for that specific
users profile.
Consumers could also benefit from detailed real-time insights on their energy consumption.
Machine learning applied to metering data could extract energy profiles of the homes
largest appliances and see how much each device contributes to the electricity bill.
Startup Bidgeley mines home-meter data and disaggregation as well as machine learning
algorithms to isolate the consumption of energy-hungry appliances such as ovens and
clothes dryers. The company claims that its customers who use this information save
between 4 and 12percent on their electricity bills.

AI can make the smart grid smarter and reduce
the need forCaseutilities
study: Utilities
to add power plants
Machine learning Drones and
enabled forecasting insect-size robots
anticipates supply identify defects,
and demand peaks, predict failures, and
and maximizes the inspect assets
use of intermittent without interrupting
renewable power production
Sensors and machine Smart wires combine

learning allow for with machine learning to
by-the-minute adjustments enable real-time power
to maximize generation dispatching, and
efficiency by adjusting to optimize it to current grid
changes in wind load and to buildings
conditions, for example asset portfolios
Few technicians remain, Virtual agents automate

but they spend more time call centers, and
on problem solving; in automatically segment
place of logging consumers based on
inspection status by service history; machine
hand, documents are learning offers early
automatically logged and warning of bad debts
routed
Field workforce Smart-meter data
receives real-time and machine
updates to decrease learning enable
response times and utilities to offer
reduce the impact services based on
of outages usage, weather and
other factors

Users could also benefit from tailored consumption management. AI-enabled devices
can identify heating usage patterns and learn about inhabitants preferences to tailor
heating regulation. For instance, Google-owned Nests Wi-Fi-enabled thermostat creates a
personalized heating schedule by monitoring users habits with a motion sensor to detect
when the home is empty and doesnt need to be heated. The hope is that eventually users
will never have to change the temperature by hand.
Another area of user experience enhancement that AI can help with is customer service.
Virtual agents offer the opportunity to reduce customer service costs. Some retail utilities
are already testing virtual agents to answer consumers queries and provide instant help,
using natural language to understand consumers comments and machine learning to help
customers deploy and use their power. The development of natural language technologies
will eventually unlock the capacity to fully automate customer service.
Toward an entirely independently run grid?

Employing AI opens a vista of possibilities for the electricity sector. It may lead to a world
where power generation, distribution, and transmission operations are automatically
optimized, where the grid is balanced independently of any human interventions, where
trading and arbitrage decisions are made in nanoseconds at a scale that only machines
could tackle, and where end-users never have to worry about searching for a better supplier
or changing the temperature manually.
As the price of solar cells and battery storage falls and their popularity rises, it is not
inconceivable to imagine a day when distributed generationpower generation at the
point of consumptionbecomes the primary source of electricity and century-old grids,
with their tens of thousands of generating stations and hundreds of thousands of miles of
transmission and distribution lines, become the backups.
A study by the Lawrence Berkeley National Laboratory, for example, suggests that a
2.5percent penetration rate of home-installed generation sources would reduce the
earnings of a utility by 4percent.60 Utilities will have to work closely with policy makers
to balance their own interests with the lowest economic cost and highest efficiency of
resource use. To address this issue, a number of interventions could be explored. These
include redefining rate structures for distributed power, creating a more flexible rate
structure such as time-of-use pricing to drive efficiency and demand management, and
expanding storage to balance the load and avoid massive spikes in demand. Not all of
these changes will be easy, of course. Nor will they be without some discomfort for some
industries and individuals. Regulators will be cautious in reviewing time-of-day pricing and
other fundamental changes, considering the fraud, mis-selling, and consumer ignorance
encountered when the United States deregulated energy markets in the 1980s.
60
Andrew Satchwell et al., Financial impacts of net-metered PV on utilities and ratepayers: A scoping study of
two prototypical U.S. utilities, Lawrence Berkeley National Laboratory, September 2014.

MANUFACTURING
SMARTER, MORE NIMBLE, AND LESS PRONE TO ERROR
Manufacturing is on the verge of a revolution in which artificial intelligence applications, from
virtual assistants to advanced robotics, will disrupt end-to-end value chains amid radical
shifts in demand. The scope of change will compel many manufacturers to adopt new plant
designs, reshape their manufacturing footprints, and devise new supply chain models.
Advances in AI technologies will enable the industry to leverage rapid growth in the volume
of data to optimize processes in real time. They can shorten development cycles, improve
engineering efficiency, prevent faults, increase safety by automating risky activities, reduce
inventory costs with better supply and demand planning, and increase revenue with better
sales lead identification and price optimization.
However, many companies are unprepared to face the future: they lack visibility on the
nature of changes such as new plant models, the implications for the way they do business,
and what they require to manage the transition toward more a collaborative working model.
Intelligent manufacturing is a smart approach to production where machines linked

through the internet assemble parts and adapt to new processes with minimal guidance
from human operators. It has become a priority for governments and corporations around
the world as they prepare for a restructuring of global industrial value chains driven by
artificial intelligence.
A glimpse of the future of intelligent manufacturing may be had at Siemens Electronic Works
Amberg. People manage and control the production of programmable logic circuits through
a virtual factory that replicates the factory floor. Via bar codes, products communicate
with the machines that make them, and the machines communicate among themselves to
replenish parts and identify problems. Nearly 75percent of the production process is fully
automated, and 99.99988percent of the logic circuits are defect-free.61
In this section, we detail how AI could transform manufacturing in two industries, aerospace
and semiconductors.
For engineering and R&D, artificial intelligence tools can

mean quicker turnaround and fewer iterations
Engineers and researchers today face difficult challenges, from the sharp growth in
demand in emerging countries to market fragmentation driven by consumers taste for
customization. At the same time, budget constraints require engineering teams to improve
their productivity and efficiency, even as limits on the number of designs that can be
considered and optimized against process capabilities restrict the predictability of product
performance. AI-powered technologies can help deliver more efficient designs than
previously achievable by eliminating waste in the design process. Innovation can be brought
to the market faster as AI facilitates lower process cycle times and an increased focus on
real-time negotiations and other interactions.
Siemens, Welcome to Electronics Works Amberg (EWA), presentation to analysts, September 29, 2015;
61
Massimo Barbato, Inside Amberg: Industry 4.0 in action, Chartered Management Institute, November 4,
2015.

Intel deployed a team of data scientists in its R&D department to speed up data integration
and the use of advanced analytics. Afterward, the company achieved 10percent higher
yield for integrated-circuit products compared with other players pursuing similar designs at
a similar, pre-production development stage.
Predictive analytics using machine learning is indeed a powerful tool to reduce the time
required to solve design problems for semiconductor manufacturers. Motivo, an artificial
intelligence startup, managed to compress semiconductor design processes from years to
a few weeks, saving chip makers the cost of iterations and testing.
Without machine learning, an aerospace manufacturer used advanced analytics to develop

productivity tools for engineering teams, such as team travel norms, team composition, and
supplier communication. The firm reduced its development costs by nearly 200million.
Machine learning could multiply this performance, unleashing the speed, accuracy, and
relevance of advanced analytics.
We expect that manufacturers will apply machine learning to consolidated comprehensive

data, enhancing the design for manufacturability process from beginning to end. AI also
will permit more explicit enforcement of value-based decisions when making trade-offs, as
when balancing safety and cost weighting. AI allows manufacturers to integrate production
and client feedback in real time to refine the product design. With suppliers, AI-based tools
could provide better accountability throughout the supply chain, which helps aerospace
manufacturers, for instance, conform to safety regulations. Meanwhile, deep learning
and network theory will help engineering development teams to better optimize their
composition, ways of working, and key performance indicators, all in real time.
Untangling the procurement process to get a better

grip on costs and supplies around the world
Keeping manufacturers stocked with parts is a complex challenge. Thousands of different
parts must be sourced from tens of thousands of suppliers all over the world. When
manufacturers are digitally linked with their suppliers systems, AI technologies can provide
transparency on supplier machine availability, performance, and downtime. They can also
help balance the supply chain and optimize inventories in real time.
One aerospace manufacturer attacked the problem by applying advanced algorithms

to spending data to search for discrepancies between what it paid suppliers and what
it actually owed them. The company also sought data that would help it understand
differences across suppliers to develop effective procurement levers and reduce
administrative costs.
In the future, supplier research and analysis will be automated and optimized with machine
learning algorithms; e-auctions will be supplemented with virtual agents and, when needed,
will automatically program in-person interactions. We estimate that full automation could
reduce IT staff numbers by 39percent.

Program managers can use advanced analytics to
improve the effectiveness of review processes
Program reviews sometimes fail to detect emerging problems or prioritize critical decisions
and tasks, resulting in delays that cost millions of dollars. Faulty communications between
the marketing and manufacturing functions, which are still mostly manual, hinder
the process.
One aerospace manufacturer sought to improve the effectiveness of its program reviews
by using advanced analytics to predict key performance indicators and identify traffic
lights, such as heavy email traffic, that could be a bellwether of problems that would appear
later on. Doing this, the manufacturer generated a pre-tax run-rate cash flow impact of
approximately 40million.
In the future, manufacturers will be able to use deep learning technology to optimize the
key performance indicators of program reviews in real time. Deep learning networks are
already applied to make predictions in real time but are trained with batches of historic
data (that is, not updated with real-time data streams). The tailoring of a model in real time
implies performing deep learning on real-time information. This will help to better predict,
identify, and prevent material and staffing bottlenecks and optimize energy consumption.
Virtual agents could alert program leadership and team members when problems arise
and recommend solutions. The step change will come with the development of natural
language, because it will allow those virtual agents to engage with team members and help
them solve problems.
AI can rethink manufacturing processes and assembly line practices

to cut costs, reduce waste, and speed time to market
Manufacturing and assembly inefficiencies cost manufacturers billions of dollars every year.
Existing fault detection and classification tools can be wildly inaccurate and cause expensive
and unnecessary interruptions on the assembly line. An aerospace manufacturer applied AI
technologies to these problems and reaped 350million in savings. Nearly 60percent came
from using advanced analytics to review data from every step in the assembly process and
then rewriting standard operating procedures based on the results. The rest came from
using machine learning algorithms, collaborative robots, and self-driving vehicles to improve
warehouse costs and reduce inventory levels.
A semiconductor maker reduced its material-delivery time by 30percent by using machine

learning to propose the best time to leave the office or warehouse. It also improved its
production yield by 3 to 5percent.
Factory managers can apply deep learning to real-time information flows in order to update
and increase the accuracy of standard-operating-procedure predictions, especially during
ramp-up, with visibility on component availability and risk management. Asset reliability can
also be enhanced with AI tools, notably thanks to machine learning improving the predictive
accuracy of defaults or production interruptions.
Virtual agents will deliver instructions and information on tablets or other interactive
personal-communications devices to reduce assembly errors and flatten the learning curve
for new operators.
Assembly lines will be significantly automated and optimized in real time, with control and
scheduling directly connected with real-time dispatching systems.

Manufacturers can deliver after-sales service just when it is needed
Aircraft maintenance and service is a large part of the value chain for aerospace firms. In the
late 2000s, GE broke new ground in this space with its power by the hour service concept,
which calls for operators to pay for aeroengines only when planes are flying. This business
model is becoming increasingly dominant in aerospace manufacturing, notably thanks to
the internet of things. The network of sensors and actuators enables companies to monitor
the actual use of their products, provides customized pay-as-you-go services, and better
prevents service disruption and downtime. AI technologies are ideally suited to leverage
the vast amount of data collected from operating engines, while the manual prioritization
of maintenance, repair, and overhaul sales leads is cumbersome, resource-heavy, and not
always optimized. This results in unnecessarily keeping aircraft groundednot generating
revenueand inefficiently allocating expensive engineering labor. With real-time feedback
between plane and ground support facilities, machine learning algorithms will be able to
make judgment calls on when to deploy drones and smart microrobots to conduct aircraft
inspections and quality checks on intermediate and final products. Manufacturers with
the best algorithms and data will be able to offer service contracts that promise better
performance at a lower cost.
Improving the accuracy of forecasting MRO work and focusing sales efforts on the most
promising leads can have a significant effect on manufacturers EBIT. One firm reported a
profit improvement of approximately 300million from using machine learning to forecast
10years of repair events for a fleet of over 17,000 commercial aircraft, and to develop a deal-
scoring tool to advise on what good looks like when pricing MRO work.
Advanced analytics and AI tools are also used to optimize processes around unplanned
maintenance events, allow the manufacturer to respond more effectively to disruptions,
and increase uptime. For instance, GE turned to Kaggle, a platform for predictive modeling
and analytics competitions, and invited data scientists to design new routing and machine
learning algorithms for flight planning that optimized fuel consumption by looking at
variables such as weather patterns, wind, and airspace restraints. The winning routing
algorithm showed a 12percent improvement in efficiency over actual flight data.
In the future, AI tools will shift predictive analytics to cognitive assessments. The algorithms
themselves will discover new rules, automatically optimizing sales and servicing for
manufacturers. Preventive maintenance will be conducted with real-time feedback
between plane and ground support facilities, using machine learning and virtual assistants
to identify issues and using drones for inspections and quality checks on intermediate
and final products. Robots the size of insects will be able to inspect airframes without
removing panels and identify common defects with computer vision and machine learning.
Smart flight systems will use predictive models based on deep learning technology to
provide real-time feedback to flight crews, helping them to optimize energy consumption
throughout flight.
We may see manufacturers shift from the traditional business model of providing spare
parts and MRO services to an end-to-end approach amid an increase in joint ventures or
industry consolidation.
In any case, virtual agents and deep learning technology will be able to improve the training
delivered to maintenance operators and pilots, and machine learning could optimize global
supply networks for spare parts to proactively stock critical and non-critical parts, thereby
limiting aircraft downtime and reducing inventory costs.

People and factories must collaborate and communicate
better to realize the full potential of AI
The key feature of the future manufacturing paradigm will be collaborative agility, which is
the ability to adapt almost instantly to changes in demand and the evolution of regulation,
input prices, technologies, and other parts of the industry landscape. From a human capital
perspective, manufacturing will become more collaborative, based on increasingly complex
activities and interdependent workers. From a technological perspective, manufacturing
plants around the globe, supply chains, and value chains will be more interconnected and
collaborative via a global digital backbone. This will combine highly automated plants that
extensively use smart robotics for mass production of limited product ranges as well as a
network of customer-centric plants close to higher-end market segments, and e-plants in a
boxsmall-scale, low-capital-expenditure, mobile plants that can produce a limited range
of products at a competitive cost. These e-plants will work well in niche markets where
demand is temporarily strong and in remote markets where production must be local and
low cost. Machine learning applications will help these integrated manufacturing systems
self-adjust in real time.
The risk-conscious aerospace industry adopts new technology only as fast as it can ensure
long-term safety. The US Federal Aviation Administrator and other regulators have change-
management systemsfor example, the parts manufacturer approval procedures in the
United Statesto monitor engineering alterations as they are made and then subject the
changes to two to three years of testing before approving their use.
Manufacturers will need to significantly invest in training as technicians roles become more
focused on knowledge and exception handling. Machine learning, advanced robotics,
virtual agents, and other technologies may be new and difficult to master for mechanics,
96percent of whom are more than 30years old.
Aircraft maintenance automation can create approximately $5billion to $15billion of value

for the industry annually, about 35percent from performance gains and the rest from labor
substitution. Given the high level of investment required in both technology and knowledge
for full AI scale-up and automation, airlines may rely on partial measure for many years.
Real-time automation of the supplier selection process will blur where legal liability will
lie if an accident occurs, because machine learning does not yet allow humans to fully
understand how the algorithm got to the assessment.
Procurement agents roles will focus more on joint value creation than on negotiation
prowess, requiring more cross-functional skills. Certain AI integration changes will require
significant capital investment. However, after the technology demonstrates its value,
automakers will be quick to invest because shortening time to market offers a significant
competitive advantage.

HEALTH CARE
AI WILL MAKE QUICKER DIAGNOSES, CREATE BETTER TREATMENT
PLANS, AND ENABLE NEW APPROACHES TO INSURANCE
Health care is a promising market for AI. There is enormous potential in its ability to draw
inferences and recognize patterns in large volumes of patient histories, medical images,
epidemiological statistics, and other data. AI has the potential to help doctors improve their
diagnoses, forecast the spread of diseases, and customize treatments. Artificial intelligence
combined with health care digitization can allow providers to monitor or diagnose patients
remotely as well as transform the way we treat the chronic diseases that account for a large
share of health-care budgets.
AIs impact on the health-care industry could be significant. In our survey, we asked
industry executives to review case studies about AI adoption. Executives from health-
care companies that were early adopters of AI said they expect the technologies will raise
operating profit margins by five percentage points within the next three years.
Several AI technologies appear to be suitable for use in medical practices. Machine learning
is starting to be applied in payments and claims management, but its further application
in health care may arrive at scale soon. Machine learning is suited to analyzing the data in
millions of medical histories to forecast health risks at the population level. This could be an
early win for AI because it brings the potential for large savings and would not require the
regulatory scrutiny to be expected when trying to anticipate individual health risks.
Using AI to diagnose illnesses may not happen so fast. While machine learning is able to use
data to make a diagnosis, completely automated diagnosis is not likely to happen quickly,
partly because of questions about whether patients will accept it, and partly because of
the technical difficulty of integrating data from multiple sources and complying with strong
regulatory requirements.
Hospitals also could improve their capacity utilization by employing AI solutions to optimize
many ordinary business tasks. Virtual agents could automate routine patient interactions.
Speech recognition software has been used in client services, where it has reduced the
expense of processing patients by handling routine tasks such as scheduling appointments
and registering people when they enter a hospital. Natural language processing can
analyze journal articles and other documents and digest their contents for quick access by
doctors. These kinds of applications can have a significant impact without needing to pass a
regulatory review.
Before the medical profession can realize this potential, however, health care providers must
adopt significant changes in the way they do business, commit to a substantial investment
in computing power and technical expertise, and work to increase the availability of the fuel
that will power progress: data, including medical records. (Specialized data brokers, such as
Explorys, which IBM bought in 2015, already offer to aggregate health care data and sell it to
potential AI solutions providers and users.).

In any case, the success of AI-based tools in medicine will hinge on whether public
officials sign on and pitch in with financing, research support, and legislation that protects
patients privacy and gives medical professionals access to anonymized data on illnesses,
treatments, and outcomes to teach computers how to identify and treat a wide range
of maladies.
Making these changes will not be easy. But there are considerable rewards for success:
AI is capable of improving care while reducing costsno small matter when health-care
spending globally reached 9.9percent of GDP in 2014 (it was 11.5percent in France and
17.1percent in the United States), according to the World Health Organization.62
Despite AIs potential, health care trails other industries in adopting the technology,
according our survey. AI use is concentrated in operations and customer service; the
technologies adopted most often are speech recognition and computer vision, by 9 and
7percent, respectively, of health care companies in our survey sample, which included
organizations that already were aware of AI. In most hospitals, operations management
functions such as appointment scheduling are still done manually.
We have found that if a sector was slow to adopt digital technologies, it tends to trail the
pack in putting AI to use, too. Our report Digital America found that almost one-quarter
of the nations hospitals and more than 40percent of its office-based physicians have not
yet adopted electronic health record systems.63 Even those that do have electronic record
systems may not be sharing data seamlessly with the patient or with other providers; tests
are repeated needlessly and patients are required to recount their medical histories over and
over because these systems are not interoperable. Another MGI report, The age of analytics,
found that the US health-care sector has realized only 10 to 20percent of its opportunities to
use advanced analytics and machine learning.64
This slow progress does not stem from a lack of interest among medical professionals
and executives. There is interest, but medicine faces some uniquely high hurdles to
adoption. The sensitive nature of medical records and strict regulations to keep them
private has stymied the collection of the high-quality aggregated data required by deep
learning applications and other AI tools. Also slowing adoption are the complexity of both
that data and the industry itself, the fragmentation of the health-care industry, and other
regulatory barriers.
AI can identify public-health threats and the most at-risk patients

AI technology adoption rates are low. The most advanced application area currently is
payment and claims management supported by machine learning algorithms. Some
clinicians are using AI to forecast the spread of certain diseases and try to anticipate which
patients would be most likely to succumb. Armed with this information, they offer preventive
care. They also use the forecasts to help hospital administrators schedule staff members,
negotiate reimbursement rates with insurers, set budgets, and optimize inventory levels.
62
World Health Organization, Total expenditure on health as a percentage of gross domestic product (US$),
Global Health Observatory, February 13, 2017.
63
64

This idea of leveraging medical and social data to better manage costs has made
forecasting one of the few areas of active AI applications in health care, attracting top tech,
pharma, and medical players as well as small startups. Johnson & Johnson, in partnership
with SAP, has used machine learning to anticipate customer demand, inventory levels, and
product mix. Careskore, a predictive analytics platform, uses machine learning to determine
the likelihood of a patients being be readmitted to a hospital.
Indeed, in the future, AI tools will enable health care to dramatically accelerate its shift
toward preventive medicine. Medical professionals will focus on managing patients health
remotely and keeping them out of hospitals. To do this, AI tools will analyze not only patients
medical histories but also environmental factors that can influence health, such as pollution
and noise where they live and work. This can identify risk groups and inform local authorities
decisions about where to implement preventive-care programs.
Primary care providers will have information to engage patients about preventive actions
that involve both medical services and lifestyle and environmental factors such as nutrition,
exercise, and pollution avoidance. Hospital administrators will be better equipped to
forecast spikes in admissions in ways not available today. Tracking the incidence of
communicable diseases, combined with personal medical records, weather data, and other
information will help an AI tool estimate how many people will need hospitalization.
In routine times, population health management analytics can identify service gaps. An AI
application, for example, could use medical and demographic data to anticipate a rise in
births and alert health care administrators if obstetric clinics require additional staff.
We estimate the full potential health care service cost savings of AI-enabled initiatives would
be $300billion a year in the United States, or about 0.7percent of GDP.65 In the United
Kingdom, using AI to target preventive care and reduce non-elective hospital admissions
can save 3.3billion annually.66
AI can help medical professionals diagnose disease and improve operations

Machine learning has enormous potential to enhance diagnostic accuracy. The Sloan
Kettering Institute estimates that doctors use only 20percent of the available trial-based
knowledge when diagnosing cancer patients and prescribing treatment. AI applications
can sift through millions of pages of medical evidence to provide a diagnosis and treatment
options in seconds.
AI-based image recognition and machine learning can see far more detail in MRI and X-ray
images than human eyes can register.67 For example, different types of glioblastomas have
distinct genetic abnormalities, and doctors treat each one based on those abnormalities.
But radiologists cannot identify genetic abnormalities of these brain cancers from images
alone. The Mayo Clinic has a machine learning program that can quickly and reliably identify
the abnormalities.68
65
Ibid.
66
Camdens CCGs analysis of the scale of opportunity for reducing non-elective admissions, in Better Care
Fund, Appendix BDefining our Focus and Ambitions, Camden Clinical Commissioning Group, September
2014; and Alicia OCathain et al., A system-wide approach to explaining variation in potentially avoidable
emergency admissions: National ecological study, BMJ Quality and Safety, volume 23, issue 1, January
2014.
67
Bertalan Mesk, Top artificial intelligence companies in healthcare to keep an eye on, MedicalFuturist.com,
February 2, 2017.
68
Marianne Matthews, Machine learning can bring more intelligence to radiology, Healthdatamanagement.
com, May 2, 2017.

Other institutions also have recognized the medical potential of machine learning and
moved to harness it. Moorfields Eye Hospital in London has teamed with Googles
DeepMind, and the Cleveland Clinic Lerner College of Medicine is collaborating with IBM
Watson.69 Software developers do not need to partner with a medical center to innovate.
GE Healthcare and Alterys have built ViosWorks, which uses an AI algorithm to improve and
accelerate MRI scanning. ViosWorks can use a series of MRI scans to create a 3-D moving
picture of a beating heart while also showing in real time how much blood is pumped with
each contraction.70
Innovation is not limited to imaging. Entrepreneurs are working to change each step in the
patient care process. A startup called Enlitic is working on a deep learning app that could
improve the accuracy of disease diagnostics. Oncora Medical has developed an AI tool that
helps oncologists draft personalized radiation treatment plans for cancer patients.
AI-powered automation has the potential to increase health care productivity by relieving
doctors and nurses of routine activities. Someday, chatbots equipped with deep learning
algorithms could relieve emergency room personnel of tending to large numbers of walk-in
patients with non-emergencies like sore throats and urinary tract infections.
Together, AI-enabled operational efficiencies could represent sizable savings in developed

countries. Estimates for the United States range from 1 to 2percent of GDP. Across other
high-income countries, estimated savings would be 0.5 to 1percent of GDP.71 Full AI
adoption could raise the productivity of registered nurses by 40 to 50percent.72 McKinsey
research has found that this could allow hospitals to cut staffing costs in half while still
significantly reducing patient waiting time.
Insurers can devise new ways to encourage

preventive care and incentivize providers
The ability of machine learning technologies to predict patient behavior and calculate
disease probabilities better than current methods will lift the profitability of life- and health-
insurance providers.
New business models can use AI combined with behavioral health interventions to focus on
prevention, disease management, and wellnessaddressing unhealthy behaviors before
people become patients. A South African insurer, Discovery Health, tracks the diet and
fitness activity of people it insures and offers incentives for healthy behaviors.73
AI also will encourage new partnerships among payers, providers, and pharma companies
and will facilitate pay-for-performance models that will accelerate the shift toward preventive
care. Payers may become more involved in care management or encourage their providers
to do so by introducing contract models based on risks uncovered by machine learning or
the potential for AI-based risk-management modeling.
69
Bertalan Mesk, Top artificial intelligence companies in healthcare to keep an eye on, MedicalFuturist.com,
February 2, 2017.
70
Drew Field, See the heart in 7 dimensions: This team of researchers attacks worlds biggest killer with
software, GE Reports, December 3, 2015.
71
72
The age of analytics: Competing in a data-driven world, McKinsey Global Institute, December 2016; A future
that works: Automation, employment, and productivity, McKinsey Global Institute, January 2017.
73
Discovery Holdings.

Episode-based payment plans, which reimburse doctors and hospitals based on the
average cost of treatment across all providers in the group, will be significantly extended
when more insurers use machine learning to analyze historical inpatient data. Based
on McKinseys client experience, we believe that this approach can have a clear impact
on costs, reducing orthopedic surgeons fees by 8 to 12percent and the fees paid to
diagnosing physicians by 4 to 5percent.
Doctors will be able to tailor treatmentseven drugsto individual patients

Patients also can benefit directly from the rise of AI in health care. Standardized treatments
do not work for every patient, given the complexity of each persons history and genetic
makeup, so researchers are using advanced analytics to personalize regimens. Decisions
can be based on data analysis and patient monitoring with use of remote diagnostic
devices. A startup called Turbine uses AI to design personalized cancer-treatment
regimens. The technology models cell biology on the molecular level and seeks to identify
the best drug to use for specific tumors. It can also identify complex biomarkers and search
for combination therapies by performing millions of simulated experiments each day.
AIs ability to use a universe of data to solve a narrow problem resonates with advocates of
customized medical treatments. Knowing the health outcomes of millions of other people
with similar symptoms, prognoses, and ages is invaluable to providers that promise one-
of-a-kind pharmaceuticals, physiotherapies, and other actions designed to deliver the
greatest benefit with the fewest side effects. Several companies are already using machine
learning or other AI technologies to tailor their treatments to individual patients. Mindmaze
uses machine learning to optimize rehabilitation activities for stroke patients. Ginger.io uses
machine learning to recommend the best time to take medication based each patients
metabolism and other factors.
Tailored treatments may reduce health expenditures by 5 to 9percent, add 0.2 to 1.3years
to average life expectancy, and increase productivity by $200 per person annually. Globally,
the economic impact could range from $2trillion to $10trillion.74
Virtual agents can serve as primary touchpoints for patients

Medical practices have taken small steps toward incorporating AI into patient management,
introducing speech recognition and other language AI technologies to automate steps
in the process. In the future, virtual assistants equipped with speech recognition, image
recognition, and machine learning tools will be able to conduct consultations, make
diagnoses, and even prescribe drugs. If these systems lack enough information to reach a
conclusion, a virtual agent could order additional tests and schedule them with the patient.
In rural areas, virtual agents will be able to conduct remote consultations. However, this
scenario would require patients, providers, and regulators to become comfortable with fully
automated diagnosis and prescriptions.
Less controversially, in hospitals, virtual agents will be able to register patients and refer
them to the appropriate doctor to address their issues. Virtual agents would be able to help
patients navigate hospital bureaucracy, prepare them for tests, and make sure they are on
time for appointments.
74

AI in health care: quicker diagnoses, better
treatment plans, and improved health insurance
Case study: Health care
Autonomous AI algorithms
diagnostic devices optimize hospital
using machine learning operations, staffing
and other AI techno- schedules, and
logies can conduct inventory by using
simple medical tests medical and
without human environmental
assistance, relieving factors to forecast
doctors and nurses of patient behavior and
Machine learning AI-powered
routine activities disease probabilities
program analyzes diagnostic tools
patients health identify diseases
remotely via mobile faster and with
device, compares it to greater accuracy,
medical records, and using historical
recommends a fitness medical data and
routine or warns of patient records
possible disease
AI tools analyze Personalized

patients medical treatment plans
histories and designed by
environmental machine learning
factors to identify tools improve
people at risk of therapy efficiency
an illness and by tailoring
steer them to treatment to
preventive care specific patients
programs Virtual agents in needs and medical AI insights from
the form of population health
interactive kiosks analyses give payers
register patients an opportunity to
and refer them to reduce
appropriate hospitalization and
doctors, improving treatment costs by
their experience encouraging care
and reducing providers to manage
waiting time 1 Weeks 15
patients wellness

Several hurdles stand in the way, starting with data availability
One of the biggest potential bottlenecks that could inhibit or derail AI development and
adoption in health care is the availability of sufficient quantities of high-quality data in
standardized formats. As noted earlier, information today is highly fragmented and spread
across the industry, residing in diverse, mostly uncoordinated repositories like electronic
medical records, laboratory and imaging systems, physician notes, and health-insurance
claims. Merging this information into large, integrated databases, which is required to
empower AI to develop the deep understanding of diseases and their cures, is difficult.
Cultural barriers, for example, stand in the way of partnerships among health-care data
ownershospitals, insurers, drug makers, and diagnostic companies. And, of course, there
is the highly sensitive nature of the data itself. While people routinely allow access to certain
kinds of personal data when they buy online or join loyalty programs, they may resist any
attempt to give wider access to their more intimate medical histories, particularly if they do
not see it as a necessity and the potential benefit is abstract. They may also be concerned
that centralized collections of intimate health details would be a natural target for hackers
and data thieves. Regulators will need to proactively develop clear rules defining who would
be able to use the data, what they could be used for, how they must be stored, and how they
would be anonymized.
Technological limitations are another hurdle. To work, AI technologies must know a great
deal about the patient and other humans, but humans know little about how AI technologies
actually make a diagnosis or choose a treatment plan. How much patients would trust
AI tools and be willing to believe an AI diagnosis or follow an AI treatment plan remains
unresolved. Regulators would not be eager to risk an incorrect computer decision harming a
patient when no one would be able to explain how the computer made its decisionor how
to prevent a repeat of the situation. This is a particular issue for the most powerful AI tools,
such as deep neural networks, and could remain the case for some time, even though AI
tools are, in theory, less likely to make mistakes than individual human clinicians.
AI companies will have to address market fragmentation. Hundreds of vendors offer

thousands of different machine learning programs, each one designed for a specific
clinical situation. But in daily practice, physicians need platforms that can address many
different situations.
Health-care providers also have work to do if they want to take advantage of AI capabilities.
To start, they must hire or develop people with the education and skills needed to deploy,
maintain, and operate AI systems. In addition to data analysts and technical staff, this
includes people with project management, team development, and problem-solving skills.
At the same time, legacy staffdoctors, nurses, and other medical professionalswill have
to become accustomed to working with the support of machines and AI tools. This may
require them to overcome significant skepticism, although it will give them the opportunity
to focus more on clinical cases and leave administration and low-risk work to AI and
digital solutions.

EDUCATION
VIRTUAL TUTORS POWERED BY AI CAN PERSONALIZE
LEARNING AND OPTIMIZE TEACHING
For decades people have discussed how to revolutionize education with technology,
whether gamifying instructional materials or expanding access to knowledge via massive
open online courses. EdTechXGlobal and Ibis capital estimated that schools spent nearly
$160billion on education technology, or ed tech, in 2016, and forecast spending to grow
17percent annually through 2020.75 Private investment in educational technology, broadly
defined as the use of computers or other technology to enhance teaching, grew 32percent
annually from 2011 through 2015, rising to $4.5billion globally.
AIs share of these flows has not been tabulated, but it is likely to increase because artificial
intelligence technologies are well suited to achieving crucial education objectives, such as
enhancing teaching efficiency and effectiveness, providing education for all, and developing
the skills that will be essential in the 21st century. So where will education be in 2030
in terms of artificial intelligence? Most probably, it will play an important part. However,
success hinges not only on technical issues but on ethical issues, starting with who owns
data on students, who can see it, who can use it, and for what purposes.
Bridging the skills gap

Many countries suffer from significant skills mismatches, brought about both by the inability
of the education system to accurately reflect the demands of employers and by the frictions
in the labor market that prevent optimal matching of individuals to jobs.76 In a survey across
10 developed and developing countries, only half of students believed their post-secondary
studies improved their employability, and more than a third of employers thought skills
shortages are a leading reason for entry-level vacancies.77 Not only does the resulting skills
gap lead to economic underperformance, it also means that many individuals are not given
the opportunity to reach their full potential.
Artificial intelligence will also play a key role in better connecting education systems and
labor markets. Digital technologies are already making a difference by connecting talent
with opportunities in the job market. A recent MGI study estimated that by 2025 online
talent platforms could enable as many as 60million people find work that more closely suits
their skills or preferences and reduce the cost of human resources management, including
recruitment, by as much as 7percent.78 With an increasing emphasis on lifelong learning,
the opportunities for artificial intelligence in employment-to-education settings have already
started to attract new players. In 2015, the employment-oriented social network LinkedIn
acquired the educational website Lynda.com, hoping to leverage AI to offer a personalized
online class selection for members considering a new job or career.
75
2016 Global EdTech industry report: A map for the future of education, EdTechXGlobal and Ibis Capital.
76
Mge Adalet McGowan and Dan Andrews, Labor market mismatch and labor productivity: Evidence from
PIAAC data, OECD, April 28, 2015.
77
Education to employment: Designing a system that works, McKinsey & Company, 2013.
78
A labor market that works: Connecting talent with opportunity in the digital age, McKinsey Global Institute,
June 2015.

In the future, improved pattern recognition enabled by machine learning, along with detailed
data on potential employees, could improve recruitment results further. It could enable
hiring companies to pinpoint the precise skill sets and personal traits that would enable
someone to be successful in a job, and surface insights that have previously not been used
in talent management. Artificial intelligence may also help detect promising candidates with
less conventional credentials and free recruiters from using school reputation as a proxy to
assess candidates potential.
More fundamentally, artificial intelligence will enable education systems to better meet the
needs of future employers. AI can be used by governments to forecast detailed job-market
demand more accurately and steer educational institutions to adapt their curricula and
approaches accordingly, making sure students have the skills required to fill those jobs. An
example of this is Saudi Arabias current exploration of machine learning as a tool to reduce
unemployment. The administration hopes to leverage large amounts of past and forecasted
economic and social information about the country to possibly guide students toward an
education best matched to their abilities.
There is a risk, however: in an Orwellian scenario, AI could also be used to optimize labor
markets without regard to nuanced social preferences, make education decisions on
citizens behalf, and sell valuable data on peoples skills to private companies or political
parties. Protecting individuals data privacy is therefore critical for enabling AI to help bridge
the skills gap.
Attracting students, and keeping them

Alongside governments, educators themselves will be able to use personal, academic,
and professional data to ensure that students benefit from the courses they choose. The
value stems not only from the ability of students to flourish academically but also from the
institutions ability to help them to find meaningful jobs. Machine learning could also be
used to identify people who appear unsuitable based on traditional measures of academic
success but have high potential based on other abilities and traits. Better targeting will
benefit students and enable institutions attract the most appropriate mix of individuals,
improve learning outcomes, and help schools and universities continually improve
their offerings.
Universities are already exploring AI applications that can improve student retention. Some
schools and colleges are testing advanced analytics and machine learning to identify
students in trouble and offer them support before they drop out. Civitas Learning and
Salesforce collaborate on a service for universities that identifies and engages with students
at risk of quitting. The Salesforce tools use machine learning to recommend engagement
strategies to optimize retention and graduation rates.
In the future, computer vision could identify signs of students disengagement by monitoring
them as they work, tracking their eye movements and observing their expressions to check
whether they are engaged, confused, or bored. In the United Kingdom, some institutions
are experimenting with computer vision, natural language, and deep learning algorithms
to better understand students learning difficulties and learning preferences, incorporating
novel types of data, such as students activities on social media.

Unleashing personalized learning
Attracting and retaining students is key, but the real step change for education will likely
emerge from a radically different approach to learning itself, whether in the classroom or
outside of it. For the last few decades, many efforts have been made to tailor learning to
each student and to shift away from a standardized approach. Adaptive learning solutions
aim to address the limitations of conventional classroom teaching by personalizing lesson
plans to the students existing knowledge, particular learning preferences, and individual
progress. Instead of delivering one lesson to the entire class, which can leave behind
struggling students or disengage fast learners, adaptive learning claims to deliver the right
content, at the right time, in the best way to each student. As of 2015, some 80 companies,
such as Knewton and DreamBox Learning, were marketing adaptive-learning systems to
schools in North America, Europe, and Asia.79
Artificial intelligence could improve adaptive learning and personalized teaching by

identifying factors or indicators of successful learning for each student that were previously
not possible to capture. In addition to monitoring such variables as the number of times
a student pauses during a lesson, the amount of time needed to answer a question, and
the number of times a question was attempted before getting it right, computer vision and
deep learning could call in new information such as mouse movements, eye tracking, and
sentiment analysis, delivering a deeper insights on the students performance, confidence,
mindset, and cognitive ability.
Implemented at scale, AI-enabled adaptive learning could restructure education. First,

it could put an end to traditional testing systems and measure academic abilities and
achievement in a more nuanced way. Class formats would give more room for students to
learn according to their own preferences, with teachers focusing less on lecturing and more
on coaching, aided by prescriptive analytics to choose the most effective methods.
Finally, artificial intelligence could empower students by providing them with control over
how fast they learn, awareness of how they learn best, and the lifelong feedback of ones
own cognitive and behavioral preferences. Ultimately, deep learning algorithms could not
only predict outcomes and prescribe accurate solutions, but explain how the algorithm
reached its conclusion and help retro-engineer drivers of educational success. This will
allow students to reflect on their cognitive abilities and understand their own optimal learning
setting, even after they leave school. Empowered by AI, students could build their own
virtual robot teachers to help them navigate lifelong learning experiences.
The promise of AI is to use digital tools as a cognitive window into students minds, and help
tailor learning to maximize individuals potential. However, the data it would leverage is by
nature private and sensitive. Education institutions and education technology providers will
need to put in place solid guarantees of data protection for students and families to embrace
the use of AI-powered tools.
Learning to Adapt 2.0: The evolution of adaptive learning in higher education, Tyton Partners, 2016
79

Releasing teachers true value add
In the future, the role of teachers could be stripped of time-consuming administrative tasks,
such as supervising and answering routine questions. Teachers would have more time to
mentor and coach studentsvalue-adding tasks that are (at least currently) uniquely suited
to human beings.
Natural language, computer vision, and deep learning could help replace teachers in
answering students routine questions or acting as tutorial supervisors. In 2014, a Georgia
Tech professor and his team created a robot teaching assistant, which provided responses
to students online questions for five months without the students noticing.80 The professor
estimates that in 2017, the robot will be able to answer 40percent of the students questions.
A virtual supervisor could harness AI to track students work and behavior and to support
teachers by supplying statistically based insights on students progress and constructive
feedback. Courseras online classes rely on machine learning to alert teachers when a large
number of students make similar errors on an assignment, suggesting possible gaps in the
teachers lectures or course materials. In the future, AI solutions may also leverage voice and
facial recognition to supervise an entire classroom and call out students individually.
Finally, AI could assist teachers in forming the most effective groups or classes by applying
machine learning algorithms to data from students education profile, social media, and
surveys. Startups such as Collaboration.ai use artificial intelligence to process data on
each students experience, knowledge, and capabilities; to create instantaneous maps of
connections and networks; to highlight each students specific potential; to break down
preferences and bias; and to recommend group formations best suited for the learning
objective. Machine learning can identify complementary skills that will maximize critical
thinking and test students capacity to adapt and collaborate.
Toward virtual teachers

UNESCO estimates that the world will need to recruit and train 24.4million primary school
teachers in order to achieve universal primary education by 2030 and another 44.4million
teachers to fill openings at secondary schools. Many of these new hiresmore than
85percent of them, in the case of primary schoolswill be required just to replace teachers
who leave education. Artificial intelligence could be one part of a solution. With a much wider
reach, AI-assisted teaching could have a significant impact in third world countries and
remote locations by supporting two key enablers of teaching: coaching and assessing.
Coaching and assessing require specific skills, such as emotional intelligence, creativity, and
communication, that are beyond machines current capabilities. Yet with new indicators,
such as facial expressions, digital interactions, group interactions, and attention tracking,
deep learning algorithms could recognize patterns, attitude toward the learning situation,
and affective states, and could support students in real time. A European Union project
called iTalk2Learn is currently developing an open-source intelligent tutoring platform to
help primary school students learn mathematics. Using a combination of machine learning,
user modeling, and natural language processing, the tool is already able to interact with and
respond to a students speech throughout a tutoring session.
Jason Maderer, Artificial intelligence course creates AI teaching assistant, Georgia Tech press release, May
80
9, 2016.

AI-powered machines also are making headway in student assessment. Companies like
GradeScope already use computer vision and machine learning to grade students work
quicker than a teacher, starting by deciphering handwriting and remembering the teachers
initial decisions on marks to automatically grade subsequent students. The technology
today can assess only work with objectively correct answers, such as math problems, and
rule-based learning, such as orthography, languages, and history events.
In the future, advances in natural language could expand AIs usefulness in automatically
grading more creative work, such as essays and presentations. In a multivendor evaluation
conducted by the University of Akron in 2012 and hosted by predictive modeling and
analytics competitions platform Kaggle, some 16,000 essays were assessed by both
grading software and teachers. Computers matched the grades from human teachers as
much as 85percent of the time.81 However, automatic grading immediately raises questions
around creativity and pattern disruption: how can machine learning be tempered to prevent
uniform thinking or the upholding of biases?
An important role for AI

Artificial intelligence clearly presents significant opportunities to raise the quality of
education to a level that our current standardized-curriculum-and-testing systems have
not been able to achieve and allow a shift of teachers focus on higher-value creative and
interpersonal tasks. Yet implementation at scale hinges on overcoming a structurally
fragmented sector and will require support from key stakeholdersstudents, parents,
teachers, administrators, and policy makers. Investment in technology and new capabilities
in the classroom and in educational administration will also be crucial.
Most importantly, data privacy issues must be addressed. The use of AI in education
raises legitimate concerns about how educational data, like other intimate personal data,
are gathered and used. What are the risks that a students longitudinal performance data,
intended for teachers to improve instruction, become public, or that poorly performing
students are denied educational and employment opportunities? To reap the benefits of
AI in education, it will be essential that students and families take on roles as advocates
and advisers in the design of AI solutions, that teachers and administrators familiarize
themselves with AI technologies, and that policy makers and regulators create a safe and
protected environment for AI-enabled education.
Man and machine: Better writers, better grades, University of Akron press release, April 12, 2012.
81

APPENDIX B: TECHNICAL APPENDIX
ESTIMATE OF INVESTMENT IN AI
Investment flows into AI technologies were divided into two categories: external and internal
investment. The external investment category captures annual investment in artificial
intelligence companies by VC and PE funds, as well as M&A activities by corporations.
The values provided are estimates of annual investment in AI based on data available
in the Capital IQ, PitchBook, and Dealogic databases. The estimate assumes that all
registered deals were completed within the year of transaction. For VC and M&A deals,
only AI companies whose core technology is AI were included. For PE investment, target
companies needed to be strongly related to AI. Deals announced but not completed by the
end of 2016 were excluded.
The internal investment category includes expenditure on all corporate activities both for
developing AI-based products or services and for building and leveraging the companys
AI capabilities. It excludes M&A. The estimate of internal investment is based on a sector-
specific AI spend-to-revenue ratio for 35 major companies (18 from North America, ten
from Asia, and seven from Europe) making strong plays in AI. Companies were divided
into the following sectors: advanced manufacturing, automotive, hardware/electronics,
internet services, retail, and systems/software. For each sector, companies that disclosed
information on internal AI expenditure were used to estimate a sector-wide average ratio.
The estimate of total internal investment was arrived at by applying these calculated ratios to
total revenues by sector.
SURVEY OF AI ADOPTION AND USE

As part of this research, we executed a survey designed to understand trends in AI
investment, adoption, use, and business implications by sector, geography, type of
company, technology, and application. The survey was conducted in March 2017. The final
survey sample after quality checks consisted of C-level executives from 3,073 companies.
The survey targeted respondents who were aware of at least one AI technology or
application and its use in business from the following list: natural language processing,
natural language generation, speech recognition, image recognition and video processing,
machine learning and deep learning, virtual agents or artificial conversational entities,
robotics, robotics process automation, and decision management (Exhibit11).
The survey sample covered 14 sectors of the economy, ten countries (from Europe, North
America, and Asia), and companies with workforces ranging from fewer than ten to more
than 10,000 employees.
The survey consisted of four groups of questions, in addition to basic information about
the company and the respondent. The first group of questions asked respondents about
their awareness and adoption rates of AI technologies or applications. The second group
asked about the current and future impact of AI in the respondents sector, including the
most important AI technology. It also asked about the parts of the business in which AI was
being deployed, and drivers and barriers for AI deployment in the respondents business.
The third group of questions investigated the financial impact of AI, asking respondents
to report current and future operating profit margins as well as spend on AI as a share of
digital investment. The last group addressed the organizational impact of AI technologies,
specifically on levels of employment and skills requirements.

Results were weighted by the share of employment represented by each firm size category,
such that the overall results reflect the relative economic importance of different sizes of
firms. Firm size weights were calculated separately for each sector, comparing the share
of respondents in a particular firm size category with the share of the workforce in that
category, using an average of the 10 countries in the survey.
Exhibit 11
MGI AI adoption and use survey sample overview
% of respondents (n = 3,073)
Geography Company size Sector
Sweden 5 >10,000 7
Other 12
South Korea 9 5,00010,000 6
Energy and resources 3
Travel and tourism 4
China 10 1,0005,000 15 Automotive and assembly 4
Transportation and logistics 4
Germany 10 Telecommunications 5
5001,000 10
Consumer packaged goods 5
Japan 10 Education 5
250500 10
Media and entertainment 5
Italy 11 Financial services 5

50250 11
Health-care systems and services 7
Canada 11
Construction 8
1050 14
France 11 Retail 8
United High tech 10

11
States
<10 27
United Professional services 14

12
Kingdom
NOTE: Numbers may not sum due to rounding.
AI INDEX
The AI index measures the extent of AI adoption and usage in 13 sectors across the 10
countries in the survey. (The telecommunications and high tech sectors were merged
to align with the MGI Industry Digitization Index.) The index is based on 16 input metrics,
divided into three categories: AI assets (three metrics), AI usage (11 metrics), and AI-enabled
labor (two metrics). Using principal component analysis, the input metrics were combined
into an overall AI adoption score. The data for these metrics were primarily obtained from the
AI adoption and use survey, proprietary databases, and the MGI Industry Digitization Index
(Exhibit12).

Exhibit 12
Metrics included in the industry AI index
Metric Description
Assets Depth of AI technologies Average number of AI technologies adopted at scale or in core part of
business per company
AI spend Average AI spend as share of total annual investment1
Supporting digital assets Percentage of firms using cloud and big data
Usage Product An entirely new Percentage of firms in sector using AI for entirely new product or
development product or service service
Research and Percentage of firms in sector using AI in R&D

development
Operations Percentage of firms in sector using AI in operations
Supply chain Supply chain Percentage of firms in sector using AI in supply chain management
and management
distribution
Distribution Percentage of firms in sector using AI in distribution
Customer Customer services Percentage of firms in sector using AI in customer services

experience
Sales and marketing Percentage of firms in sector using AI in sales and marketing
Financial and Executive Percentage of firms in sector using AI in executive management

general management
management
Financial and risk Percentage of firms in sector using AI in financial and risk
management management
Workforce Management of Percentage of firms in sector using AI in operational staff management

management operational staff
HR Percentage of firms in sector using AI in HR
Labor Exposure to AI in workforce Percentage of workforce in firms adopting AI at scale or in core part of
business
AI resources per worker Average AI spend per employee ( thousand)
1 Results unweighted.
In order to calculate the overall AI score for an industry, a weight was assigned to each
variable. Principal component analysis was used to determine the weights. The analysis is
a mathematical transformation that converts a set of potentially correlated input variables
into principal components, or new sets of values that explain the variance in the input
variables. In this case, the resulting principal components aim to explain the variance across
the 16 input variables in use. A principal component analysis yields multiple components,
so the component that explains the most variance of the original 16 variables was used.
Each component has corresponding variable loadings or weights, which were applied as
the weights for each value in the index calculation. Since the 16 input variables are not in
the same units, each value was standardized by subtracting from it the mean value for the
variable and dividing the result by the standard deviation for the variable. The standardized
scores for each variable were added to arrive at the overall AI index value for each sector.
The analysis was performed within each of the three categories of variables: assets, usage,
and labor.

BIBLIOGRAPHY
Buchanan, Bruce G., A (very) brief history of artificial intelligence, AI

Magazine, volume 26, number 4, winter 2005.
Bughin, Jacques, and James Manyika, Bubble or paradigm change?

Assessing the global diffusion of enterprise 2.0, in Alex Koohang, Johannes
Britz, and Keith Harman, eds., Knowledge management: Research and
applications, Informing Science, 2008.
Bughin, Jacques, and James Manyika, How businesses are using web 2.0: A
McKinsey global survey, McKinsey Quarterly, December 2007.
Bughin, Jacques, and Nicolas van Zeebroeck, The right response to digital
disruption, MIT Sloan Management Review, April 2017.
Bughin, Jacques, Ten big lessons learned from big data analytics, Applied
Marketing Analytics, volume 2, number 4, 2017.
Buitrago, Jaime, and Shihab Asfour, Short-term forecasting of electric loads

using nonlinear autoregressive artificial neural networks with exogenous
vector inputs, Energies, volume 10, number 1, 2017.
Catlin, Tanguy, Jay Scanlan, and Paul Willmott, Raising your Digital Quotient,
McKinsey Quarterly, June 2015.
Dewan, Sanjeev, Dale Ganley, and Kenneth L. Kraemer, Complementarities

in the diffusion of personal computers and the internet: Implications for the
global digital divide, Information Systems Research, volume 21, number 5,
December 2010.
Feigenbaum, Edward A., Expert systems in the 1980s, working paper, 1980.
Feigenbaum, Edward A., The art of artificial intelligence: Themes and case
studies of knowledge engineering, Stanford University Computer Science
Department report number STAN-CS-77621, August 1977.
Forman, Chris, Avi Goldfarb, and Shane Greenstein, The geographic

dispersion of commercial Internet use, in Rethinking rights and regulations:
Institutional responses to new communication technologies, Lorrie Faith
Cranor and Steven S. Wildman, eds., MIT Press, 2003.
Glauner, Patrick, Large-scale detection of non-technical losses in imbalanced

data sets, PhD dissertation, University of Luxembourg, March 2016.
Gugerty, Leo, Newell and Simons Logic Theorist: Historical background

and impact on cognitive modeling, Proceedings of the Human Factors and
Ergonomics Society Annual Meeting, volume 50, issue 9, October 2006.
Hofstadter, Douglas R., Gdel, Escher, Bach: An eternal golden braid, Basic
Books, 1979.
McGowan, Mge Adalet, and Dan Andrews, Labor market mismatch and
labor productivity: Evidence from PIAAC data, OECD, April 28, 2015.
McKinsey & Company, Education to employment: Designing a system that

works, 2013.

McKinsey & Company, Smartening up with artificial intelligence (AI): Whats in
it for Germany and its industrial sector? April 2017.
McKinsey Center for Government, Government productivity: Unlocking the

$3.5trillion opportunity, April 2017.
McKinsey Global Institute, The age of analytics: Competing in a data-driven

world, December 2016.
McKinsey Global Institute, Artificial intelligence: Implications for China,

April 2017.
McKinsey Global Institute, Digital America: A tale of the haves and have-
mores, December 2015.
McKinsey Global Institute, Digital Europe: Pushing the frontier, capturing the
benefits, June 2016.
McKinsey Global Institute, A future that works: Automation, employment, and

productivity, January 2017.
McKinsey Global Institute, Harnessing automation for a future that works,

2017.
McKinsey Global Institute, A labor market that works: Connecting talent with
opportunity in the digital age, June 2015.
McKinsey Global Institute, Technology, jobs, and the future of work, 2017.
Minsky, Marvin, Steps toward artificial intelligence, Proceedings of the IRE,

volume 49, number 1, January 1961.
MIT Technology Review, Asias AI agenda: How Asia is speeding up global

artificial adoption, November 2016.
Moravec, Hans P., The Stanford cart and the CMU Rover, Proceedings of the
IEEE, volume 71, issue 7, July 1983.
Moreno, Rosina, and Jordi Suriach, Innovation adoption and productivity

growth: Evidence for Europe, working paper, 2014.
Narrative Science, Outlook on artificial intelligence in the enterprise, 2016.
National Science and Technology Council, Preparing for the future of artificial
intelligence, Executive Office of the President, October 2016.
Negnevitsky, Michael, Artificial Intelligence: A guide to intelligent systems,

Addison-Wesley, 2002.
Newell, Allen, Intellectual issues in the history of artificial intelligence, in

The study of information: Interdisciplinary messages, Fritz Machlup and Una
Mansfield, eds., John Wiley and Sons, 1983.
OCathain, Alicia et al., A system-wide approach to explaining variation in

potentially avoidable emergency admissions: National ecological study, BMJ
Quality and Safety, volume 23, issue 1, January 2014.

Satchwell, Andrew et al., Financial impacts of net-metered PV on utilities
and ratepayers: A scoping study of two prototypical U.S. utilities, Lawrence
Berkeley National Laboratory, September 2014.
Turing, A. M., Computing machinery and intelligence, Mind, volume 49,

number 236, October 1950.
Zastrow, Mark, South Korea trumpets $860-million AI fund after AlphaGo

shock, Nature, March 18, 2016.
Zhu, Kevin, Kenneth L. Kraemer, and Sean Xu, The process of innovation
assimilation by firms in different countries: A technology diffusion perspective
on e-business, Management Science, volume 52, number 10, October 2006.

RELATED MGI
AND MCKINSEY RESEARCH
A future that works: Automation, employment, and productivity

(January2017)
Improvements in robotics, artificial intelligence, and machine learning have put
us on the cusp of a new automation age. This will change work for everyone,
from miners to bankers.
The age of analytics: Competing in a data-driven world

(December2016)
Data and analytics are advancing rapidly. Organizations that harness them will
create significant value and differentiate themselves; those that dont will be at
a disadvantage.
The rise of the machines: How Chinese executives think about

developments in artificial intelligence (December2016)
Software developers have approached the dream of teaching computers to
think, aided by advances in computing power, the development of machine
learning, and an explosion in data.
Digital globalization: The new era of global flows (March2016)

Conventional wisdom says globalization has stalled because the global goods
trade has flattened and cross-border capital flows have declined. But global
flows of data and information are soaring.
The Internet of Things: Mapping the value beyond the hype (June2015)
If policy makers and businesses get it right, linking the physical and digital
worlds could generate up to $11.1trillion a year in economic value by 2025.
No ordinary disruption: The four global forces breaking all the trends
(April2015)
The global economys operating system is being rewritten. Four great forces
are colliding and transforming the global economy, and the worlds leaders
need to adjust to a new reality.
Smartening up with artificial intelligence: How AI will transform

Germanys industrial sector (April2017)
Artificial intelligence is finally bringing a multitude of capabilities to machines
that were long thought to belong to humans. We explain how and where AI
could affect the German industrial sector.
Artificial intelligence: Implications for China (April2017)

Artificial intelligence has gone from speculation to reality. This paper
explores AIs potential to fuel Chinas productivity growthand to disrupt the
nations workforce.
Open data: Unlocking innovation and performance with liquid

information (October2013)
Open datafrom both public and private sourcesare adding a new
dimension to analytics and giving rise to novel, data-driven innovations.
McKinsey Global Institute
June 2017
Copyright McKinsey & Company
www.mckinsey.com/mgi
@McKinsey_MGI
McKinseyGlobalInstitute

MGI Artificial Intelligence Discussion Paper PDF

Uploaded by

Copyright:

Available Formats

MGI Artificial Intelligence Discussion Paper PDF

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

MGI Artificial Intelligence Discussion Paper PDF

Uploaded by

Copyright:

Available Formats

ARTIFICIAL

Jacques Bughin | Brussels

MGI research combines the disciplines of economics and management,

Recent reports have assessed the economic benefits of tackling gender

McKinsey & Company is a member of the Partnership on AI, a collection of

Copyright McKinsey & Company 2017

1 McKinsey Global Institute

1. Artificial intelligence is getting ready for business, but

2. Artificial intelligence promises to boost profits and

3. Businesses, developers, and governments need to act

Appendix A: Five case studies

Appendix B: Technical appendix

In this independent discussion paper, we examine investment in artificial intelligence (AI),

In addition to identifying a gap between AI investment and commercial application, which

Download the full report at www.mckinsey.com/mgi

The current AI wave is poised to finally break through

In 2016, companies invested

$26B to $39B 3+ technologies

TECH GIANTS STARTUPS

3x External investment growth since 2013

How companies are adopting AI Six characteristics

Medium Retail Adopt AI in Adopt multiple

6 McKinsey Global Institute Artificial intelligence: The next digital frontier?

Box1. A multi-lens approach to understanding the AI story

McKinsey Global Institute Artificial intelligence: The next digital frontier? 7

ARTIFICIAL INTELLIGENCES ROLLER-COASTER RIDE TO TODAY

8 McKinsey Global Institute Artificial intelligence: The next digital frontier?

LED BY TECH GIANTS, AI INVESTMENT IS GROWING RAPIDLY,

McKinsey Global Institute Artificial intelligence: The next digital frontier? 9

Technology giants dominate investment in AI

Investment in AI, 20161 Investment by tech giants and other corporations

Compound annual AI share of

10 McKinsey Global Institute Artificial intelligence: The next digital frontier?

McKinsey Global Institute Artificial intelligence: The next digital frontier? 11

External investment in AI-focused companies by technology category, 20161

SOURCE: Capital IQ; Pitchbook; Dealogic; McKinsey Global Institute analysis

12 McKinsey Global Institute Artificial intelligence: The next digital frontier?

FIRMS AND INDUSTRIES ALREADY ON THE DIGITAL FRONTIER

McKinsey Global Institute Artificial intelligence: The next digital frontier? 13

Fourth, companies investing at scale do it close to their core business.

Early-adopting sectors are closer to the digital frontier

14 McKinsey Global Institute Artificial intelligence: The next digital frontier?

AI Index Relatively low Relatively high

Assets Usage Labor

Financial and general

Supply chain and

Travel and tourism

McKinsey Global Institute Artificial intelligence: The next digital frontier? 15

When it comes to adopting AI, the bigger, the bolder

Early AI adopters tend to become serial adopters

The phenomenon of multitechnology application is persistent at a sector level. Industries

16 McKinsey Global Institute Artificial intelligence: The next digital frontier?

Early adopters see AI increasing revenue while companies

AI is not only about technical adoption, it is about enterprise acceptance

McKinsey Global Institute Artificial intelligence: The next digital frontier? 17

18 McKinsey Global Institute Artificial intelligence: The next digital frontier?

Future AI demand trajectory1

4 Consumer packaged goods Media and entertainment

1 Based on the midpoint of the range selected by the survey respondent.