Professional Documents
Culture Documents
2. Automation of
knowledge work
The commercialization of computers with this level of intelligence over the coming
decade could have massive implications for how knowledge work is conducted.
Such tools could both extend the powers of human workers and allow them to
offload tedious detail work. But these advanced tools could also ultimately lead to
some jobs being automated entirely.
Automation has already swept through manufacturing and transaction work (tasks
that consist of executing simple exchanges, such as taking deposits or checking
customers out of a grocery store). When it comes to knowledge work, the impact
of automated tools could be less direct. Knowledge work jobs generally consist
of a range of tasks, so automating one activity may not make an entire position
unnecessary (the way welding robots make welders redundant, for example).
In addition, knowledge work has become more complex, in large part due to
information technology, creating demand for workers with new skills who can
perform new kinds of tasks.
39 Erik Brynjolfsson and Andrew McAfee, Race against the machine: How the digital revolution
is accelerating innovation, driving productivity, and irreversibly transforming employment and
the economy, Digital Frontier Press, 2011.
40 Cost of employment numbers include salaries and benefits.
McKinsey Global Institute
Disruptive technologies: Advances that will transform life, business, and the global economy 41
change the nature of work for many people, requiring innovation to fully realize its
potential while managing its risks.
Now it’s time for the tough part of the day: your doctor appointment. You
received a request for an appointment yesterday when your biosensor
alerted your digital physician to a change in your blood pressure. Your vital
signs are scanned remotely, and the system cross-checks this information
with journal cases, your family’s history of hypertension, your diet and
exercise routines, and the vital signs of other men your age. Good news:
“You don’t need drugs, but you do need to stop eating fast food and
skipping the gym,” your computerized doctor says. Relieved, you stop at
the gym on the way home and ask your mobile device to order a salad to be
delivered when you get home.
DEFINITION
We define knowledge work automation as the use of computers to perform
tasks that rely on complex analyses, subtle judgments, and creative problem
solving. Knowledge work automation is made possible by advances in three
areas: computing technology (including processor speeds and memory
capacity), machine learning, and natural user interfaces such as speech
recognition technology.
These capabilities not only extend computing into new realms (for example, the
ability to “learn” and make basic judgments), but also create new relationships
between knowledge workers and machines. It is increasingly possible to
interact with a machine the way one would with a coworker. So, instead of
assigning a team member to pull all the information on the performance of a
certain product in a specific market or waiting for such a request to be turned
into a job for the IT department, a manager or executive could simply ask a
computer to provide the information. This has the potential to provide more timely
access to information and raise the quality and pace of decision making and,
consequently, performance.
42
Today, total global employment costs are $33 trillion a year and, on current trend,
could reach $41 trillion by 2025. We analyzed a subset of knowledge worker
occupations with employment costs that could reach $14 trillion by 2025. These
workers—professionals, managers, engineers, scientists, teachers, analysts, and
administrative support staff—represent some of the most expensive forms of
labor and perform the most valuable work in many organizations. Few of these
workers have benefited from tools that can augment core aspects of their work
involving decision making and judgment.
41 Many current machine learning approaches mimic aspects of the human brain. Neural
networks simulate brain structures via interconnected layers of “artificial neurons,” which
adaptively strengthen or weaken their interconnections based on experience. Deep learning
technology makes use of algorithms that form a learning hierarchy in which higher-level
concepts are defined using layers of lower-level concepts (often using neural networks).
Some machine-learning algorithms don’t require labeling or preclassification of their
training examples and can instead identify their own categories and concepts (e.g., by
cluster analysis).
McKinsey Global Institute
Disruptive technologies: Advances that will transform life, business, and the global economy 43
Exhibit 3
The number of transaction workers in the United States across some major
job types declined more than 50 percent between 1970 and 2010
Decline in transactional jobs between 1970 and 20101
% workforce share decline for select highly automatable jobs
Index: 100 = 1972
120
100
80
General clerks
60
Bookkeeping jobs
40 Secretaries
Typists Jobs
20 almost
Telephone
automated
operators
0 away
1972 1975 1980 1985 1990 1995 2000 2005 2010
1 Job types that can be scripted, routinized, automated (e.g., cashiers, receptionists, stock traders). Data are for the US
private economy. Occupation data normalized in 1983 and 2003 to account for classification differences.
SOURCE: US Bureau of Labor Statistics 1972–2010; McKinsey Global Institute analysis
Exhibit 4
Sized applications of automation of knowledge work could have
direct economic impact of $5.2 trillion to $6.7 trillion per year
in 2025
Potential economic Potential
impact of sized Estimated productivity
Sized knowledge worker occupations in 2025 Estimated scope potential reach or value gains
occupations $ trillion, annually in 2025 in 2025 in 2025
NOTE: Estimates of potential economic impact are for some applications only and are not comprehensive estimates of total
potential impact. Estimates include consumer surplus and cannot be related to potential company revenue, market size,
or GDP impact. We do not size possible surplus shifts among companies and industries, or between companies and
consumers. These estimates are not risk- or probability-adjusted. Numbers may not sum due to rounding.
SOURCE: McKinsey Global Institute analysis
Currently available intelligent personal assistants such as Apple’s Siri and Google
Now illustrate some of the possible uses of these technologies in administrative
support roles. The Google Now service already anticipates user needs, making
recommendations or delivering information based on browser history, calendar
entries, and current location. For example if traffic is bad, Google Now may
suggest that the user leave early, having combined routing and traffic information
with data about the time and location of the user’s next appointment.
The economic impact of such tools in education would come from improving
instructional quality and enabling teachers to provide more one-on-one attention
and coaching. New self-teaching tools could also enable fundamental changes
in scheduling: courses could be tied to subject mastery, rather than semesters or
quarters, allowing students to progress at their own pace.
43 Roadside assistance customers benefit from smart support during peak and after hours,
SmartAction case study of Canadian Automobile Association Saskatchewan, June 2012.
44 Measurement Incorporated, “MI’s automated essay scoring system takes first prize in national
competition,” press release, October 4, 2012.
46
plan with the highest probability of success.45 This could be the first of many
applications of knowledge work automation in medical diagnostics, given the
costs of misdiagnoses.
Technical professions
The ability to use deep learning techniques to discover new relationships in huge
amounts of data and to determine which relationships are the most important
amounts to an enormous shortcut in many kinds of technical work—from
software design to drug discovery. For example, by applying deep learning to
drug development data, researchers can quickly narrow the field of possible
formulations from thousands to dozens, drastically speeding up the discovery
process and saving thousands of hours of labor. A team of researchers in a
contest sponsored by Merck recently proved that a deep-learning computer
could examine an unfamiliar data set of chemical structures and develop its
own rules to narrow down the thousands of unique molecules to those with
the greatest potential to be effective.46 Software engineers are using machine
learning to speed up software development through automated testing and
algorithm performance optimization, as well as project management tasks such
as managing code libraries, tracking version control, and dividing tasks between
developers.47
Management
Machine learning excels at the complex analytics that managers use to monitor
activities under their responsibility, understand the root causes of issues as
they arise, and accurately forecast future trends on the horizon. For example,
managers currently use machine-learning technology to monitor, control, and
diagnose faults in manufacturing plants.
45 Jonathan Cohn, “The robot will see you now,” The Atlantic, February 20, 2013.
46 John Markoff, “Scientists see promise in deep-learning programs,” The New York Times,
November 23, 2012.
47 Jitesh Dundas, “Machine learning helps software development,” Software Magazine,
June 2012.
McKinsey Global Institute
Disruptive technologies: Advances that will transform life, business, and the global economy 47
Professional services
Fields such as law and financial services are already beginning to see the benefits
of knowledge worker automation. Law firms, for example, are using computers
that can scan thousands of legal briefs and precedents to assist in pretrial
research—work that would once have taken hundreds or thousands of hours of
paralegal labor. Symantec’s Clearwell system uses language analysis to identify
general concepts in documents and present the results graphically. In one case,
this software was able to analyze and sort more than 570,000 documents in two
days.48
Artificial intelligence (AI) has played a role in financial transactions for some time.
AI algorithms are able to parse myriad news stories, financial announcements,
and press releases, make decisions regarding their trading relevance, and then
act in slivers of a second—faster and with greater information recall than any
human trader.49 Banks can also use machine learning to detect fraud, finding
charges or claims outside a person’s normal buying behavior. Even services like
Future Advisor use AI to offer personalized financial advice inexpensively and
at scale.
Cultural and organizational hurdles also exist. Risk-averse firms may delay
adoption until the benefits of these technologies have been clearly proven. And,
in some cases, there could be resistance. Many attorneys were initially hesitant
to use computerized research systems because they did not trust the machines
to catch every document. Some business leaders might have concerns about
legal liability regarding situations in which these technologies make mistakes (for
example, with a patient diagnosis).
48 John Markoff, “Armies of expensive lawyers replaced by cheaper software,” The New York
Times, March 4, 2011.
49 For one interesting example of the use of AI in stock trading, see Christopher Mims, “AI that
picks stocks better than the pros,” MIT Technology Review, June 10, 2010.
48
IMPLICATIONS
The automation of knowledge work has the potential to become pervasive,
transforming the economics of many industries, but also posing challenges and
opportunities for technology providers, virtually all business leaders, individuals,
and policy makers.
Technology providers (both software and hardware) will play a critical role in this
nascent field by designing powerful, easy-to-use knowledge work applications
and supporting adoption within organizations. There could be many opportunities
across a range of possible capabilities and approaches. Some technology
providers might focus on high-end, advanced systems such as the Watson
supercomputer, perhaps configured as enterprise solutions or programmed
for specific verticals such as medicine. Others might focus on next-generation
assistants similar to Apple’s Siri for both businesses and consumers. And many
might focus on special-purpose tools for analytics, search functions, or a host
of other potential applications. These knowledge work automation tools could
be delivered in many ways, including via enterprise solutions, apps, or Web
services. They could be delivered via the cloud (see Chapter 4) and on mobile
Internet devices (see Chapter 1). They could also integrate with Internet of Things
devices, both to analyze additional data and to directly control processes and
environments (see Chapter 3).
The first task for business leaders is to understand how knowledge work is (and
will be) carried out in their organizations, including where the most time and
money are spent, which functions contribute the most value, and which contribute
the least, and where productivity is lowest. Answering such questions will help set
priorities regarding areas in which the adoption of tools to automate knowledge
work might be both feasible and able to create consistently higher performance.
Given the potential power of these tools, the biggest benefits may come from
applying knowledge work automation to boost the productivity of employees in
high-value-added functions, rather than focusing on simple tasks that might be
turned over entirely to machines.
McKinsey Global Institute
Disruptive technologies: Advances that will transform life, business, and the global economy 49
Automation of knowledge work could drive the creation of many new types of
jobs if businesses and governments can innovate effectively and adjust education
and training to focus on new skills. As with advanced robotics, these technologies
could also create jobs for experts who can create and maintain the technology
itself. However, increased productivity without innovation and retraining could
ultimately exert downward pressure on wages and increase income disparities.
50 See The world at work: Jobs, pay, and skills for 3.5 billion people, McKinsey Global Institute,
June 2012.
50