A report from the Deloitte Center for Government Insights

How much time and money

can AI save government?
Cognitive technologies could free up hundreds of millions of
public sector worker hours
 ABOUT THE DELOITTE CENTER FOR
GOVERNMENT INSIGHTS
The Deloitte Center for Government Insights shares inspiring stories of government innovation, looking
at what’s behind the adoption of new technologies and management practices. We produce cutting-
edge research that guides public officials without burying them in jargon and minutiae, crystalizing es-
sential insights in an easy-to-absorb format. Through research, forums, and immersive workshops, our
goal is to provide public officials, policy professionals, and members of the media with fresh insights that
advance an understanding of what is possible in government transformation.

ABOUT THE AUTHORS

PETER VIECHNICKI

Peter Viechnicki is a strategic analysis manager and data scientist with Deloitte Services LP, where he
focuses on developing innovative public sector research using geospatial and natural language process-
ing techniques. Follow him on Twitter @pviechnicki.

WILLIAM D. EGGERS

William D. Eggers is executive director of the Deloitte Center for Government Insights and author of
nine books, including Delivering on Digital: The Innovators and Technologies That Are Transforming Gov-
ernment. His commentary has appeared in dozens of major media outlets including the New York
Times, Wall Street Journal, and Washington Post. He can be reached at [email protected] or on
Twitter @wdeggers.

Deloitte’s “Cognitive Advantage” is a set of offerings designed to help organizations transform de-
cision making, business processes, and interactions through the use of insights, automation, and
engagement capabilities. Cognitive Advantage is tailored to the federal government and powered
by our cognitive platform. Cognitive Advantage encompasses technologies capable of mimicking,
augmenting, and in some cases exceeding human capabilities. With this capability, government
clients can improve operational efficiencies, enhance citizen and end-user experience, and pro-
vide workers with tools to enhance judgment, accuracy, and speed.

COVER IMAGE BY: LIVIA CIVES

CONTENTS

Introduction: AI-based technology brings both

optimism and anxiety | 2

Breaking government work into tasks

to clarify AI’s effects | 4

What do government workers do all day? | 6

Activities most likely to be automated | 9

AI shows enormous potential for

labor time savings | 14

Conclusion: Minimizing disruption and

enabling innovation | 16

Appendix: Data and methods | 17

Endnotes | 20
How much time and money can AI save government?

Introduction
AI-based technology brings both optimism and anxiety

All kinds of institutions today run on data, and that means endless staff hours
spent inputting, processing, and communicating. The work needs to get done,
so someone has to spend that time pecking away at a keyboard, right?

T
HE promise of reducing—or even eliminating— especially since cognitive technologies are increas-
all that drudge work is one reason why many ingly capable of carrying out tasks once reserved for
managers are enthusiastic about new appli- knowledge workers.2
cations based on artificial intelligence (AI). Finally,
Technology, from farm equipment to factory ro-
staff resources could be freed up to do real work,
bots to voice mail, has always displaced low-skilled
with people having time to focus on creative proj-
workers. But only recently has it threatened white-
ects and deal directly with clients and customers.
collar professionals’ positions: Computer scientists
But of course, there’s no guarantee that any new are building machines capable of carrying out al-
labor-saving technology will make everyone’s daily most any task, even those—such as composing mu-
lives more rewarding rather than simply wiping out sic—seemingly at the core of our humanity.3 Knowl-
entire categories of employment.1 And that’s why AI edge workers, whose jobs once seemed secure, are
applications make plenty of people anxious as well, feeling directly threatened for the first time.

So there’s a blend of anticipation and dread within

There’s no guarantee a wide range of organizations and industries—and
public sector agencies are no exception.4 Conversa-
that any new labor- tions with government executives suggest that most

saving technology lack a clear vision of how AI applications might af-

fect their staff and missions, which is understand-
will make everyone’s able, since prior research hardly offers an actionable
forecast. The US Bureau of Labor Statistics optimis-
daily lives more tically predicts that government workforces will see

rewarding rather almost no job losses between now and 2024,5 while
a recent study by Deloitte UK and Oxford University
than simply wiping suggests that up to 18 percent of UK public sector
jobs could be automated by 2030.6
out entire categories
We’ve attempted to bring clarity to the confusion,
of employment. for agency chiefs looking to future workforce needs.

2
 Cognitive technologies could free up hundreds of millions of public sector worker hours

Our view is that the key to planning ahead is un-

derstanding how much time cognitive technologies
could save. And indeed, our research, based on a
new method for studying AI-based technology’s ef-
fects on government workforces, indicates that cog-
nitive technologies could free up large numbers of
labor hours by automating certain tasks and allow-
ing managers to shift employees to tasks requiring
human judgment.

These new applications could save hundreds of mil- free up 30 percent of the government workforce’s
lions of staff hours and billions of dollars annually. time within five to seven years. Lower levels of in-
But the shift’s size and impact will depend on many vestment and support would yield lower savings, of
factors, some political and some financial. With ad- course: Minimal investment in AI would result in
equate investment and support, we believe, AI could savings of just 2 to 4 percent of total labor time.

3
How much time and money can AI save government?

Breaking government work

into tasks to clarify AI’s effects

Y
OU need to know where you are before you new technology. Over time, technology often results
can decide where you’re going; this truism in a complete rethinking of what organizations pro-
certainly applies to predicting the effects of duce and what the goal of that production is. Recent
AI on government work. Most existing quantitative history shows this pattern has also been true for
models begin by tallying workers by occupation and government work (see sidebar, “How cartography
predicting which jobs will be replaced by technol- went digital”).
ogy. In other words, they rely on occupations as the
Deloitte has developed a new methodology for mea-
unit of analysis.7
suring the amount of time government workers
But we know from a long history with these issues spend on the tasks that fill up their work days. We
that technology typically doesn’t replace jobs whole- believe we’re the first to quantify government work
sale, at least at first.8 Instead, it often substitutes for at the task level. The appendix explains details of
specific tasks, while the workers who previously our method.
performed them shift to jobs complementary to the

HOW CARTOGRAPHY WENT DIGITAL

The US Geological Survey (USGS) began producing topographic maps of the nation in 1879,9 and for most
of its history, it printed its maps on paper. If you were an active hiker or camper in the 1980s, you’ll likely
remember shelves and shelves of USGS topo maps at outdoor stores, but over the following decade,
USGS transformed its mapmaking techniques by embracing digital map production. This transformation,
which relied on a major Reagan-era investment in geospatial information systems technology, was
disruptive and productive. It significantly improved the efficiency of production—and completely changed
the nature of cartographers’ jobs.10

Before the transformation, USGS cartographers worked as skilled craftsmen, performing painstaking
tasks such as drawing elevation contours on acetate sheets. Today, their duties primarily involve
collecting and disseminating digital cartographic data through the National Map program.11

Today, USGS officials recall a bumpy transformation. Veteran cartographer Laurence Moore says, “We
were slow to appreciate how fundamentally GPS and digital map data would change the world, and
tended to think of these technologies as just tools to produce traditional maps faster and cheaper.”

Today, the agency employs only a tenth of the cartographers working there at the peak of the paper-map
production era. But paradoxically, the total number of cartographers and photogrammetrists employed
by federal, state, and local governments has risen by 84 percent since 1999.12 And the Bureau of Labor
Statistics forecasts a 29 percent growth in employment for cartographers and photogrammetrists
through 2024, largely due to “increasing use of maps for government planning.”13

4
 Cognitive technologies could free up hundreds of millions of public sector worker hours

For this article, we’ve applied this method to the because it provides detailed open workforce data
federal civilian workforce and to the workforce of through cutting-edge transparency. We expect pat-
a large, representative Midwestern state (figure 1). terns we find in this state to be broadly applicable to
The state was chosen due to the similarity of its a number of others.
workforce to many other state governments and

Figure 1. Federal civilian and state government workforces at a glance

Representative state
Federal civilian workforce government workforce

2016 employment 2.067 million 58,837

2016 salary
$168 billion $2.4 billion
and wages

1. Miscellaneous administration 1. Corrections officers (7,077)

Top three and program officers (98,405) 2. Administrative staff (2,983)
occupations by 2. Information technology 3. Therapeutic program workers
employment management officers (82,969) (2,608)
3. Nurses (82,875)

Sources: Deloitte analysis of Office of Personnel Management Fedscope March 2016 employment data. Note: Federal
and state data include both full- and part-time employees.
Deloitte University Press | dupress.deloitte.com

5
How much time and money can AI save government?

What do government
workers do all day?

S
O how do government workers spend their
time? We estimate that the two workforces
collectively work 4.3 billion (federal) and 108
million (state government) hours a year. We group
the tasks they perform into “generalized work activ-
ities,” using the US Department of Labor’s (DOL’s)
O*NET activity framework.14

For both federal and state workers, by far the most

time-consuming activity is documenting and re-
cording information, a task capturing 10 percent of
both federal and state government work hours. And
while a few workers undoubtedly love documenta-
tion for its own sake, for most this activity surely
isn’t the most rewarding part of the day.

Few observers will be surprised to find that paper-

For both federal and work can get in the way of government workers’

state workers, by far

more critical functions15—just think of, for instance,
all the times you’ve seen TV police officers groan
the most time- over having to write and file lengthy reports. But
the amount of time devoted to seemingly peripheral
consuming activity activities is sobering.

is documenting and A quick glance at figure 2, unsurprisingly, shows

several tasks that might be highly amenable to auto-
recording information, mation. Now consider figure 3: the five most labor-

a task capturing intensive activities performed by the federal work-

force, and their suitability for automation.
10 percent of AI-based applications can almost certainly improve
both federal and some activities, such as filling out forms or moving
objects. For others, such as caring for patients, cog-
state government nitive technologies aren’t ready to replace people.
(The appendix describes how we rank activities for
work hours. their automation potential.)

6
 Cognitive technologies could free up hundreds of millions of public sector worker hours

Figure 2. A year in the life of the government workforce,

federal vs. a Midwestern state

Documenting/recording information

Handling and moving objects

Assisting and caring for others

Getting information

Communicating with supervisors,

peers, or subordinates
Guiding, directing, and motivating subordinates

Monitoring and controlling resources

Monitoring processes, materials, or surroundings

Performing general physical activities

Inspecting equipment, structures, or material

Providing consultation and advice to others

Performing administrative activities

Evaluating information to determine

compliance with standards

Analyzing data or information

Controlling machines and processes

Thinking creatively

Processing information

Communicating with persons outside organization

Judging the qualities of things, services, or people

Repairing and maintaining mechanical equipment

0 0.02 0.04 0.06 0.08 0.1

Proportion of total annual person-hours

Federal State

Source: Deloitte analysis of Office of Personnel Management Fedscope, state government workforce,
and O*NET data.

Deloitte University Press | dupress.deloitte.com

7
How much time and money can AI save government?

Figure 3. Automation potential of subtasks within the five most labor-intensive

federal activities

Annual hours spent

Documenting/
recording information

Handling and
moving objects

Assisting and caring

for others

Getting information

Communicating with
supervisors, peers,
or subordinates

0 200,000,000 400,000,000

Subtasks with high Subtasks with medium Subtasks with low

automation potential automation potential automation potential

Source: Deloitte analysis of OPM Fedscope and DOL O*NET data.

Deloitte University Press | dupress.deloitte.com

Government employees Figure 4. Government workers spend

20 percent of their time on
spend a day a week on noncore tasks
“supplemental” tasks
Federal employees spend 21 percent of
We estimate that federal and state workers spend at annual hours on noncore work, against
least 20 percent of their time on tasks they consider 19 percent for Midwestern state workers
unimportant (figure 4).16 It’s a low-end estimate
based on the DOL’s restrictive definition of “supple- Federal
employees 79% 21%
mental” tasks. If you asked government workers
directly, they might give you a much higher figure.
State
employees 81% 19%

Core Supplemental

Source: Deloitte analysis of O*NET and federal and state

workforce data.

Deloitte University Press | dupress.deloitte.com

8
 Cognitive technologies could free up hundreds of millions of public sector worker hours

Activities most likely to be

automated

S
“ UPPLEMENTAL tasks” is a very broad descrip- the maximum benefit from technology’s cost-
tion, of course, and can mean different things effectiveness and reliability. The opposite is often
in different contexts. As agency executives true, however—automation usually begins with
consider incorporating AI-based technology into unimportant tasks or, at least, those perceived as
their work, where should they begin? unimportant.

Just because a task can be automated doesn’t mean The same is true for work activities. We studied 13
it will or should be anytime soon. Several factors years of changes to the length of time spent on indi-
tend to influence which tasks are both most condu- vidual tasks, using data from the DOL O*NET data-
cive to automation and most likely to be automated. base. Over the study period, tasks considered less
We’ve identified these from our research on 13-year important consumed less and less time, implying
trends in work activities as well as the widely some degree of technological substitution (figure 5).
accepted findings of labor market economists.
In our data set, tasks with above-average impor-
The factors are task importance, skill requirements, tance gained labor inputs by 4.6 percent, while tasks
work volume, and technological barriers. We with below-average importance lost labor inputs by
examine each below. 1.3 percent. A task’s importance correlated posi-
tively and significantly (rho = .09, p < .0001) with a
change in the amount of time spent on it.
1. Peripheral tasks
Thus, we can comfortably expect that agencies will
It would be logical to assume that industries would look to begin integration of AI-based technology
automate their most important tasks first, to gain with tasks considered less important.

Figure 5. Peripheral tasks and declining labor inputs, 2003–16

Change in labor allocation, time 1 to time 2

Above-average
importance 4.6%

Below-average
-1.3%
importance

-2% -1% 0% 1% 2% 3% 4% 5%

Source: Deloitte analysis of O*NET database, 13,356 pairs of tasks observed at two different
years between 2003 and 2016.
Deloitte University Press | dupress.deloitte.com

9
How much time and money can AI save government?

2. Middle-skilled tasks technological replacement because enough people

perform them (providing enough “volume”) and the
A task’s skill requirements also affect its likelihood wages paid are high enough to justify investing in
of automation. In employment settings, “middle- the technology.
level” skills generally refer to positions requiring
American labor market economists usually high-
education beyond high school but less than a four-
light skills-biased technological change by showing
year college degree. More broadly, one author has
that employment for high-skilled workers has risen
defined middle-level tasks as “cognitive or manual
rapidly over time, while the middle-skilled work-
in nature and requir[ing] one to follow precise
force has shed jobs.20 And as middle-skilled workers
procedures.”17 In government, various clerking
lose jobs, they’re forced to compete for lower-skilled
positions provide good examples.
jobs, driving down wages.
In the future tasks requiring middle-skill levels will
Employment trends in government jobs follow the
likely be automated sooner, on average, than both
pattern you’d expect for skills-biased technological
high- and low-skill tasks. Many low-skilled tasks
change. In the past decade, middle-skill govern-
have already been replaced by previous waves of
ment employment fell while high-skilled employ-
automation, and those yet to be automated may
ment rose (figure 6).
pose some barrier to automation (such as requiring
a worker to navigate an unpredictable physical envi- Figure 6 shows 10 years of federal jobs data broken
ronment), or wages may be so low as not to justify into five skill levels, using the DOL’s formula for
investing in automation technology.18 “job zones.” The share of federal workers in higher-
skilled jobs (job zones 4 and 5) rose in every year
It may seem counterintuitive, but this tendency
of the study period, while middle-skill employ-
to hollow out the middle of the labor market first
ment (zones 2 and 3) shrank. Many of the jobs lost
is a well-known characteristic of technological
in government were positions such as clerks or
change. Multiple studies have demonstrated how
administrative professionals. Though considered
well it explains historical trends in employment
white-collar work, the tasks involved were routine
and wages.19 These tasks are the easiest targets for
enough to allow them to be automated by what Tom

Figure 6. Higher-skilled workers gain share in the federal workforce

0.45
Share of federal workforce

0.40
0.35
0.30
0.25
0.20
0.15
0.10
0.05
0.00
1 2 3 4 5

Job zone (1–5: Education and experience level)

2006 2008 2010 2012 2014 2016

Source: Deloitte analysis of O*NET and OPM Fedscope data.

Deloitte University Press | dupress.deloitte.com

10
 Cognitive technologies could free up hundreds of millions of public sector worker hours

AUTOMATING MIDDLE-SKILL TASKS TO SPEED UP CIRCUIT TESTING

In 2016, the US Army Research Labs (ARL) automated testing of electronic silicon wafers used in military
radios and cellphones (figure 7). Testing circuits is critical to making sure that soldiers’ communication
equipment functions properly, but the testing process was time-consuming and dull, requiring mid-level
skills (such as those possessed by engineering graduate students) and painstaking attention to detail.
Testing was viewed as a bottleneck in the production process, and delays encouraged ARL to automate
the testing tasks.23

ARL developed an automated probe that can test the circuits imprinted on the wafers, freeing up the
engineers to focus on core responsibilities. “Those core responsibilities such as forming hypotheses and
designing experiments to test them, or designing systems using input from the data analysis, are much
more difficult to perform and require high skill and creative intelligence,” in the words of ARL scientist
Ryan Rudy. Automation has sped up testing time by a multiple of 60. Previously, an ARL intern might
test 10 percent of one silicon wafer in three months; after automation, an entire wafer can be tested in
two weeks.

Davenport and Julia Kirby call the second era of Figure 7. Testing silicon wafer circuits
automation—when computers take over the “dull for military radios at Army Research Labs
jobs.”21

Since overall government employment trends

follow skills-biased trends, we expect similar trends
at the task level, determining which government
tasks will be replaced sooner than others.22 (See
sidebar “Automating middle-skill tasks to speed
up circuit testing” for a discussion of how Army
Research Labs is automating middle-skill tasks to
free up scientists for higher-level work.)

3. High-volume tasks
A third factor determining where AI investments
may be most effective is volume of business. Decades
of economic research support the idea that indus-
tries with more business volume are better able to
invest in expensive labor-saving technologies.24

The volume concept can help guide government

executives in targeting AI investments. Since we can
break government work into activities and estimate
how many hours are spent on each, we can iden-
tify time-consuming tasks with high potential for
automation—a useful tool for government agencies Pictures used with permission of ARL.
directing precious investment funds. Deloitte University Press | dupress.deloitte.com

11
How much time and money can AI save government?

Figure 8. Midwestern state government occupations ranked according to

automation potential of their activities

Mail clerks and mail machine operators

Bookkeeping, accounting, and auditing clerks
10
Data entry keyers

8 Highway maintenance workers

2012 employment

0
50 100 150 200 250 300
Automation potential rank

Note: Green dots show selected occupations that have high relative employment and perform many activities with high
automation potential—e.g., mail clerks and mail machine operators, data entry keyers, highway maintenance workers,
and bookkeeping, accounting, and auditing clerks.

Source: Deloitte analysis of O*NET data, state government job classification catalog, and state transparency portal
salaries database.

Deloitte University Press | dupress.deloitte.com

Figure 8 ranks state government occupations: on three types of “intelligence” as current challenges
the horizontal axis by the automation potential of to AI: social intelligence, creative intelligence, and
their associated tasks (low ranks being easier to perception and manipulation. In their analysis,
automate) and on the vertical axis by employment. social intelligence tasks comprise those requiring
traditionally human traits: “negotiation, persuasion
Activities performed by the occupations that figure
and care.” Creative intelligence involves the basic
8 shows in green (such as data entry workers) could
human ability to generate ideas and things that are
be good starting points for AI investment.
novel and interesting, whether a theory or a recipe.
Perception and manipulation tasks use our ability to
4. Special skill requirements comprehend and interact with the chaotic patterns
of real life—the irregular, object-filled worlds of
prevent some tasks from airports, supermarkets, and our own homes.25
automation—for now These are the tasks that will be more difficult—
The fourth factor is the type of skill required to though not necessarily impossible—to hand over
complete the task in question. Oxford economists to AI technology. For now, “cognitive collaboration”
Carl Frey and Michael Osborne have identified between humans and machines will likely be the
most efficient way of carrying out such tasks.26

12
 Cognitive technologies could free up hundreds of millions of public sector worker hours

We employ the Oxford study’s occupational criteria Figure 9 shows the relation of social intelligence,
to identify jobs requiring social intelligence, creative creative intelligence, and perception/manipula-
intelligence, or perception and manipulation. Jobs tion with automation at the occupational level. We
requiring any of these show a lower degree of auto- expect the same characteristics to constrain AI
mation in a sample of 964 occupations from the development at the task level as well.28
O*NET database (figure 9).27

Figure 9. Social intelligence, creative intelligence, and perception and manipulation

correlate with lower average automation index

Three barriers to automation

31
30
29
Degree of automation

28
27
26
25
24
23
Job requires Job requires Job requires perception/
social intelligence creative intelligence manipulation

Yes No

Source: Deloitte analysis of O*NET database. Deloitte University Press | dupress.deloitte.com

13
How much time and money can AI save government?

AI shows enormous potential

for labor time savings

D
ECISIONS concerning how to invest in cogni- change in labor inputs to each government task and
tive technology, and how much, could have adjust it according to intrinsic task characteristics.
major implications for government efficien- We then simulate changes to task labor inputs by
cy and effectiveness. Our research quantifies the sampling from the normal distribution using the
likely upper and lower bounds of these effects over adjusted mean, with standard deviation chosen us-
the next five to seven years. We don’t use predictive ing O*NET values (figure 10).
analytics to model these scenarios because cognitive
Given low, medium, and high levels of government
technology is changing so fast that extrapolations
resourcing and investment in AI, our simulations
are likely to fail. Only 12 years ago, for example, MIT
generate the scenarios shown in figure 11.
researchers confidently predicted that AI would
never replace human drivers on America’s roads.29 Figure 11 shows that even low levels of effort behind
AI adoption could save government workforces be-
Instead, we use Monte Carlo simulation—a method
tween 2 to 4 percent of all their labor hours. With
for modeling the probability of different outcomes—
middling investment levels, much bigger savings
to describe three different scenarios for the likely
become possible. The midrange scenario, which we
near-term effects of automation on government
consider realistic based on our experience with pub-
work.30 For each, we select the base mean of the
lic and private sector automation projects, indicates

Figure 10. Simulation parameters: low, medium, and high levels of effort

Level of investment Base mean for simulation How value was chosen

Task labor inputs decline on average Low-end threshold of time savings

Low
by 20% for process automation

100% approximates average percent

Task labor inputs decline on average time saved on back-office functions
Medium
by 100% through robotic process automation
projects

200% approximates the savings in

testing time for silicon wafer circuits
Task labor inputs decline at Army Research Labs (see page
High
on average by 200% sidebar “Automating middle-skill
tasks to speed up testing”); reflects
the higher end of time savings

Deloitte University Press | dupress.deloitte.com

14
 Cognitive technologies could free up hundreds of millions of public sector worker hours

Figure 11. Time and money savings from AI under three levels of investment

Level of investment Savings category Federal State government

Annual person-hours 96.7 million 4.3 million

Hours as percentage
Low 2.23% 3.94%
of total

Salary $3.3 billion $119 million

Annual person-hours 634 million 15.3 million

Hours as percentage
Medium 14.63% 13.93%
of total

Salary $21.6 billion $420 million

Annual person-hours 1.2 billion 33.8 million

Hours as percentage
High 27.86% 30.84%
of total

Salary $41.1 billion $931 million

Source: Deloitte simulation of likely changes to labor inputs to government tasks.

Deloitte University Press | dupress.deloitte.com

savings of 13 to 15 percent in time requirements within five to seven years. Since IT costs continue to
within five to seven years. Finally, with strong sup- plummet and cognitive technologies are developing
port for AI adoption, we can simulate a ceiling of rapidly, even the high-end scenario may be within
potential benefits: 27 to 30 percent time savings reach.

15
How much time and money can AI save government?

Conclusion
Minimizing disruption and enabling innovation

E
XPERIENCE teaches us that AI, like other Finally, after the IT department installs AI applica-
forms of technology, will likely cause disrup- tions, the technology doesn’t run itself; often, main-
tion among government workers whose jobs taining it requires a surprising amount of human
it changes. But agency heads can take steps in ad- labor. Asking software vendors to design acces-
vance to minimize the effects.31 sible training, tuning, and maintenance interfaces
for their AI products would help ensure that the
First, agencies should provide maximum advance
employees asked to incorporate AI technology into
notice of plans to replace or augment certain tasks
their work can participate in its use.
with AI-based applications. Good communication
lowers employee stress levels as they undergo tech- We’ve seen that cognitive technologies can poten-
nological transformation.32 tially free up millions of labor hours for government
workers, with the magnitude of those savings de-
Second, agency technology leaders should coordi-
pendent on policy decisions. But what will govern-
nate with human-capital planners to synchronize
ment workers do with those liberated hours?
their upgrades with workforce trends. For example,
if an agency anticipates a high rate of retirement Senior policymakers will have a choice—one that
within a given occupation, it might prioritize AI in- mirrors our perennial national debate about big ver-
vestments in that area.33 sus small government. Some may see AI-based tech-
nology as a lever to shrink government workforces,
Third, HR executives can cushion the effects of
aiming to deliver the same services with fewer em-
disruption by encouraging employees to develop
ployees. Other jurisdictions may choose to use the
new skills. Government might create program of-
applications as tools for their workers, encouraging
fices to oversee curricula and learning incentives
them to find new ways to use liberated work hours
relevant to cognitive technologies. Such programs
to improve the services they provide to citizens. The
could boost the skills of wide swathes of govern-
most forward-leaning jurisdictions will see cogni-
ment employees. Just as foreign-language program
tive technologies as an opportunity to reimagine the
offices boosted government skills in mission-critical
nature of government work itself, to make the most
languages such as Farsi and Arabic in the last two
of complementary human and machine skills.
decades, AI training offices could promote targeted
curricula and incentives for data analytics, machine AI will support all these approaches. It will be up to
learning, and designing human-to-machines in- government leaders to decide which will best serve
terfaces. More broadly, government organizations their constituents.
can improve their training for human skills that are
most likely to complement AI-based technology in
the long run: problem solving, social intelligence,
and creativity.34

16
 Cognitive technologies could free up hundreds of millions of public sector worker hours

Appendix
Data and methods

D
ATA used in this research originates from We use 1,043 as the equivalent for “hourly” on the
two main sources: information on numbers assumption that even tasks performed around the
of workers, their demographic characteris- clock take up no more than half of a worker’s time,
tics, and their salaries collected by the federal Of- with the other half used for non-occupation-specific
fice of Personnel Management (OPM) and our large activities. Multiplying by the proportion of respon-
Midwestern state’s Department of Administrative dents, choosing each value, and summing over the
Services; and data on tasks performed by 1,110 oc- task, we calculate the average annual hours for the
cupations collected by the US Department of Labor activity. This provides annual task-hours.
as part of its O*NET OnLine database. The first
We then tally the annual task-hours performed by
source provides information on who is in the work-
each occupation, multiply by the workforce-specific
force; the second tells us what they do.
employment in that occupation, and apply a scale
Analyzing the data requires linking both sources via factor (0.45 for the federal workforce and 0.25 for
a crosswalk, and OPM helpfully publishes one at the state workforce) to estimate total task-hours
www.eeoc.gov/federal/directives/00-09opmcode. performed by all members of the workforce. This
cfm. The Midwestern state does not provide such a provides the labor inputs to a task.
crosswalk, so we created one using state employee
The 19,125 O*NET tasks are further linked to more
salary data and the state’s online job classification
than 2,000 “detailed work activities,” 331 “interme-
handbook.
diate work activities,” and 37 “general work activi-
ties,” allowing us to analyze annual task-hours and
Establishing the labor inputs for work tasks at any desired level of
specificity.
current baseline
O*NET contains the results of worker surveys ask-
ing respondents to estimate the time spent on each
Understanding changes
of their work activities for 19,125 detailed, occu- in task labor inputs
pation-specific tasks. We convert those frequency
The O*NET program surveys workers in each oc-
scale ratings to annual task-hours, assuming 2,080
cupation repeatedly, but at irregular intervals. For
total person-hours per full-time equivalent, using
13,356 of the 19,125 detailed ONET tasks (70 per-
these equivalences:
cent), ONET reports two or more observations
of task frequency at different time points. In this
Less than yearly 0.5 hours/year
sample, the earliest observation of a task took place
Yearly 1 hours/year
in 2003; the latest was in 2016. The length of time
Monthly 12 hours/year between observations averaged 7.03 years, with a
Weekly 52 hours/year minimum of two years and a maximum of 13.
Daily 260 hours/year
Given two observations of the labor inputs to a task
More than daily 520 hours/year at time 1 (t1) and time 2 (t2), we calculate the percent
Hourly 1,043 hours/year change in annual task-hours for that task. We use

17
How much time and money can AI save government?

the formula (t1-t2)/average(t1, t2) to calculate per- combined index values are easier to automate than
cent change for this and other time trends in this activities with higher index values.
paper.
We rank the 331 IWAs according to automation po-
A decrease in labor inputs to a task over time can tential and combine that ranking with employment
have many explanations, including structural to rank occupations. We do so by linking each occu-
changes to the occupation and changes in customer pation to the IWAs it performs. We use the average
demand; dry cleaners don’t do much sewing any- combined automation index for all the IWAs linked
more. One explanation, however, is that technology to an occupation, weighted by number of task-hours
has substituted for part of the labor of the task. spent on each IWA, to represent the automation po-
tential of the activities of the occupation. We rank
We calculate the correlation between percentage
occupations according to combined IWA automa-
change in task labor input and task importance us-
tion index and employment.
ing Pearson product-moment correlation to dem-
onstrate that, on average, peripheral tasks are auto- When applied to the 669 federal occupation series
mated before core tasks. established by the Office of Personnel Management,
this method yields the following 20 jobs with both
We measure the standard deviation of the changes
the highest automation potential and highest em-
to task labor input and use that value to constrain
ployment (figure 12).
the Monte Carlo simulation of levels of AI invest-
ment described in the following section.
Monte Carlo simulation
Ranking activities and of AI technology
occupations according to adoption scenarios
automation potential We begin with the data set of 19,125 detailed O*NET
task descriptions, representing each using intrinsic
The realization that tasks requiring social intelli-
task characteristics discussed above: task impor-
gence, creative intelligence, or perception and ma-
tance and the binary variables for whether the oc-
nipulation are less easy to adapt to AI technology
cupation requires social intelligence, creative intel-
allows us to rank O*NET’s 331 intermediate work
ligence, or perception and manipulation.
activities (IWAs) according to their automation po-
tential. For the three levels of effort in the scenarios, we
choose a base mean for the normal distribution as
For each of the 19,125 O*NET detailed tasks, we
shown in figure 10 and set the standard deviation to
code it with three binary variables according to
0.63 based on the percentage changes to 13,356 task
whether the associated occupation requires social
labor inputs described above.
intelligence, creative intelligence, or perception and
manipulation. We use the same O*NET indicators We run the simulation as follows. For each task, if
that Carl Frey and Michael Osborne used35 to assign the task requires social intelligence, creative intel-
those binary values. Each IWA is linked through ligence, or perception/manipulation, we set the dis-
O*NET’s database structure to one or more tasks. tribution mean to zero. Otherwise, we set the distri-
For each, we average the binary values for social in- bution mean to the base mean times the reciprocal
telligence and call this the IWA’s social index, which of task importance, on a scale of one to six. We then
measures how many of the tasks included in the sample percentage change to the annual task-hours
IWA are performed by occupations requiring social from that distribution and store the results. We re-
intelligence. We do the same to build a creative in- port scenario results by running the simulation 10
dex and a perception/manipulation index. We then times and averaging the results.
sum the indices for each IWA, ranking them accord-
ing to the sum of the three indices. IWAs with lower

18
 Cognitive technologies could free up hundreds of millions of public sector worker hours

Figure 12. Federal jobs with high employment and automation potential

OPM occupation Average IWA

OPM occupation title Employment
series automation rank
AGRICULTURAL COMMODITY
1980 1572 29.57
GRADING

1981 AGRICULTURAL COMMODITY AID 1194 29.57

0305 MAIL & FILE 5547 52.59

FINANCIAL CLERICAL &

0503 8104 58.53
TECHNICIAN

0525 ACCOUNTING TECHNICIAN 6251 58.53

MISCELLANEOUS GENERAL
4701 MAINTENANCE & OPERATIONS 2423 55.97
WORK

UTILITY SYSTEMS REPAIRING-

4742 1854 55.97
OPERATING

6907 MATERIALS HANDLER 6872 66.14

4754 CEMETERY CARETAKING 693 49.68

0540 VOUCHER EXAMINING 1585 58.53

7408 FOOD SERVICE WORKING 7637 77.54

0561 BUDGET CLERICAL & ASSISTANCE 838 58.53

5003 GARDENING 480 54.25

7407 MEATCUTTING 1380 63.94

3566 CUSTODIAL WORKING 12908 86.07

MISCELLANEOUS CLERK &

0303 56589 100.54
ASSISTANT

CLAIMS ASSISTANCE &

0998 3180 74.36
EXAMINING

0356 DATA TRANSCRIBER 2142 72

5703 MOTOR VEHICLE OPERATING 5363 81.21

4102 PAINTING 3979 78.52

Source: Deloitte analysis of OPM Fedscope and O*NET data.

Deloitte University Press | dupress.deloitte.com

19
How much time and money can AI save government?

ENDNOTES

1. Claire Cain Miller, “Evidence that robots are winning the race for American jobs,” New York Times, March 28, 2017,
https://nyti.ms/2ouv37V.

2. Neuroscientist Sam Harris’s June 2016 TED talk “Can we build AI without losing control over it?” is a great exposition
of the pessimistic view of AI. See www.ted.com/talks/sam_harris_can_we_build_ai_without_losing_control_over_it.

3. Anna Gaca, “‘The world’s first songs composed by artificial intelligence’ are neither first nor entirely artificial,”
Spin, September 22, 2016, www.spin.com/2016/09/first-song-written-by-ai-really-isnt/.

4. See, for instance, Executive Office of the President, “Artificial intelligence, automation, and the economy,”
December 20, 2016, https://obamawhitehouse.archives.gov/sites/whitehouse.gov/files/documents/Artificial-
Intelligence-Automation-Economy.pdf.

5. The Bureau of Labor Statistics predicts federal employment will shrink by 1.5 percent between now and 2024,
while state and local employment will grow by 0.4 percent. See BLS, “Employment by major industry sector,”
www.bls.gov/emp/ep_table_201.htm, accessed April 12, 2017.

6. Deloitte, “Deloitte: Automation set to transform public services,” October 25, 2016, https://www2.deloitte.com/
uk/en/pages/press-releases/articles/automation-set-to-transform-public-services.html.

7. Three well-known examples are Marten Goos and Alan Manning’s 2007 Review of Economics and Statistics article
“Lousy and lovely jobs: The rising polarization of work in Britain,” Alan Blinder’s 2009 World Economics study
“How many US jobs might be offshorable?”, and David Autor and David Dorn’s 2013 American Economic Review
study, “The growth of low-skill service jobs and the polarization of the US labor market.”

8. See, for example, Bureau of Labor Statistics, “Technological change and employment,” Monthly Labor Review,
April 1987, p. 28, https://stats.bls.gov/opub/mlr/1987/04/art3full.pdf: “The content of jobs is being modified by
technological change. Although job titles frequently remain the same while innovation is taking place, over time,
employers have less demand for manual dexterity, physical strength for materials handling, and for traditional
craftsmanship.”

9. US Geological Survey, “topoView,” https://ngmdb.usgs.gov/maps/TopoView/, accessed April 12, 2017.

10. Deloitte interview with US Geological Survey officials, November 2016.

11. USGS National Map program, https://nationalmap.gov/, accessed April 12, 2017.

12. Taken from Bureau of Labor Statistics, “Occupational Employment Survey, National Industry Series.” In 1999,
2,140 cartographers and photogrammetrists were employed in government jobs; in 2015, the number had
climbed to 5,240.

13. Bureau of Labor Statistics, “Occupational Outlook Handbook: Cartographers and photogrammetrists,” www.bls.
gov/OOH/architecture-and-engineering/cartographers-and-photogrammetrists.htm, accessed April 12, 2017.

14. US Department of Labor, “O*NET OnLine,” www.onetonline.org/, accessed April 12, 2017.

15. The obstacle posed by paperwork is well known in the human services professions. Surveys have shown it’s
considered one of the top three most challenging aspects of human services work, along with caseloads and
issues confronting families: See, for instance, National Association of Social Workers, “If you’re right for the job,
it’s the best job in the world,” June 2004, www.socialworkers.org/practice/children/NASWChildWelfareRpt062004.
pdf.

20
 Cognitive technologies could free up hundreds of millions of public sector worker hours

16. We use the Department of Labor’s definition of “core” and “supplemental” tasks, whereby core tasks have
importance ratings greater than or equal to 3 and relevance ratings greater than or equal to 67 percent. See
“O*NET OnLine Help,” www.onetonline.org/help/online/scales, accessed April 12, 2017.

17. Bureau of Labor Statistics, “Middle-skill jobs decline as U.S. labor market becomes more polarized,” by Demetrio
Scopelliti, Monthly Labor Review, October 2014, www.bls.gov/opub/mlr/2014/beyond-bls/middle-skill-jobs-
decline-as-us-labor-market-becomes-more-polarized.htm.

18. The second author has a vivid memory of visiting India in 2013 and seeing armies of gardeners down on their
knees trimming lawns with sickles. (Sickles!) He was told this was possible simply because wages were so low that
it was cost-competitive with mechanical lawn mowers.

19. A widely cited example is David Autor, Frank Levy, and Richard Murnane, “The skill content of recent technological
change: An empirical exploration,” Quarterly Journal of Economics, 2003, pp. 1,279–33, https://economics.mit.edu/
files/11574.

20. See, for example, Daron Acemoglu, “The impact of IT on the labor market,” Massachusetts Institute of Technology,
September 2016, http://economics.mit.edu/files/12118.

21. Tom Davenport and Julia Kirby, “Beyond automation,” Harvard Business Review, June 2015, https://hbr.
org/2015/06/beyond-automation.

22. We have not yet been able to fit quantitative models of the relationship of skill levels with task labor inputs due
to the limitations of O*NET task data, but we expect to demonstrate this as our data improves.

23. Deloitte interview with Ryan Rudy and Jonathan Puder, US Army Research Labs, December 2016.

24. See, for instance, Jerome A. Marks, “Technological change in employment: Some results from BLS research,”
Monthly Labor Review, April 1987, https://stats.bls.gov/opub/mlr/1987/04/art3full.pdf.

25. Carl Frey and Michael Osborne, “The future of employment: How susceptible are jobs to computerisation?”,
University of Oxford, Martin School, September 2013, pp. 27–30, www.oxfordmartin.ox.ac.uk/publications/
view/1314.

26. For a complete discussion of humans and AI working as partners, see Jim Guszcza, Harvey Lewis, and Peter
Evans-Greenwood, “Cognitive collaboration,” Deloitte Review 20, January 23, 2017, https://dupress.deloitte.com/
dup-us-en/deloitte-review/issue-20/augmented-intelligence-human-computer-collaboration.html.

27. Pearson correlation coefficients from a sample of 961 occupations with social intelligence (rho=-.14, p < .0001);
creative intelligence (rho=-.09, p < .007); and perception and manipulation (rho=-.07, p < .03).

28. We have not yet been able to model this constraint quantitatively at the task level, again because of limitations
of O*NET data. Specifically, O*NET does not code tasks according to what skills each requires—it codes only
occupations in this way.

29. Frank Levy and Richard J. Murnane, The New Division of Labor (Princeton, N.J.: Princeton University Press, 2005),
cited in Erik Brynjolfsson and Andrew McAfee, The Second Machine Age (New York: W.W. Norton, 2016), pg. 18.

30. We constrain the scenarios using findings from our research on task characteristics that affect automation
potential: task importance and requirements for social, creative, or perceptual intelligence. We do not yet use
task skill level to constrain our model, as we do not yet have good quantitative models of the effect of task skill
level on automation. The three scenarios reflect three levels of organizational effort behind the push for AI
technology, both in funding and in political will for process change.

31. Analysts have long studied how to manage the disruptive effects of technology. Again, see Bureau of Labor
Statistics, “Technological change and employment.”

21
How much time and money can AI save government?

32. For a more complete discussion of how organizations can smooth the transition to cognitive technology, see Jeff
Schwartz, Laurence Collins, Heather Stockton, Darryl Wanger, and Brett Walsh, The future of work: The augmented
workforce, Deloitte University Press, February 28, 2017, https://dupress.deloitte.com/dup-us-en/focus/human-
capital-trends/2017/future-workforce-changing-nature-of-work.html.

33. Ibid. Schwartz and colleagues’ article lays out some ways in which organizations can redesign multiyear strategic
and annual operational workforce planning in light of artificial intelligence.

34. See Angus Knowles-Cutler and Harvey Lewis’s treatment of which skills will be most valuable in the era of
cognitive technology: Talent for survival: Essential skills for humans working in the machine age, Deloitte UK, 2016
https://www2.deloitte.com/content/dam/Deloitte/uk/Documents/Growth/deloitte-uk-talent-for-survival-report.
pdf.

35. Frey and Osborne, “The future of employment.”

22
 Cognitive technologies could free up hundreds of millions of public sector worker hours

ACKNOWLEDGEMENTS

Center colleagues Purva Singh and Pankaj Kishnani were the lead researchers for the study. Mitali
Chatterjee and Sushumna Agarwal assisted in the data analysis.

A number of colleagues generously contributed their time and insights to this report, including Daniel
Bachman, Patricia Buckley, Akrur Barua, and Mahesh Kelkar from Deloitte Services LP and Harvey
Lewis, Laura Webb, James Munk, and Jemma Venables from Deloitte UK. In addition, the authors
would like to thank Vikrant Jain, Ziaul Khan, Abbas Naqvi, Saravanan Ramamurthy, Tarakeshwar
Allamsetty, and Vipul Sangoi from Deloitte Services LP and Jakob Deel of Deloitte Consulting for help-
ing build out the state government data analysis.

We would like to thank Ryan Rudy, Jonathan Puder, and Joyce Martin of Army Research Labs for tell-
ing us about the advances they’re making in automating circuit testing. We would also like to thank Larry
Moore, Kristin Fishburn, Helmut Lestinski, and Bob Davis of the US Geologic Survey for sharing their
experiences of the transition to digital cartography. Finally, we would like to offer a special thanks to
Kwame Porter Robinson of Kwamata Inc. and the University of Chicago, for sharing his time and exper-
tise in helping us make sense of the O*NET data.

23
How much time and money can AI save government?

CONTACTS

William D. Eggers Christopher Rose

Executive director, Deloitte Center for Govern- Partner, Deloitte Consulting LLP
ment Insights [email protected]
Deloitte Services LP +1 703 980 1359
[email protected]
+1 571 882 6585 Thomas Beyer
Principal, Deloitte Consulting LLP
Marc Mancher [email protected]
Principal, Deloitte Consulting LLP +1 619 237 6659
[email protected]
+1 860 488 5071

Mark White
Global Consulting Technology CTO
Partner, Deloitte Consulting LLP
[email protected]
+1 571 277 0771

24
 Follow @DU_Press
Sign up for Deloitte University Press updates at www.dupress.deloitte.com.

About Deloitte University Press

Deloitte University Press publishes original articles, reports and periodicals that provide insights for businesses, the public
sector and NGOs. Our goal is to draw upon research and experience from throughout our professional services organization,
and that of coauthors in academia and business, to advance the conversation on a broad spectrum of topics of interest to
executives and government leaders.
Deloitte University Press is an imprint of Deloitte Development LLC.

About this publication

This publication contains general information only, and none of Deloitte Touche Tohmatsu Limited, its member firms, or its
and their affiliates are, by means of this publication, rendering accounting, business, financial, investment, legal, tax, or other
professional advice or services. This publication is not a substitute for such professional advice or services, nor should it be
used as a basis for any decision or action that may affect your finances or your business. Before making any decision or taking
any action that may affect your finances or your business, you should consult a qualified professional adviser.
None of Deloitte Touche Tohmatsu Limited, its member firms, or its and their respective affiliates shall be responsible for any
loss whatsoever sustained by any person who relies on this publication.

About Deloitte
Deloitte refers to one or more of Deloitte Touche Tohmatsu Limited, a UK private company limited by guarantee (“DTTL”), its
network of member firms, and their related entities. DTTL and each of its member firms are legally separate and independent
entities. DTTL (also referred to as “Deloitte Global”) does not provide services to clients. In the United States, Deloitte refers to
one or more of the US member firms of DTTL, their related entities that operate using the “Deloitte” name in the United States
and their respective affiliates. Certain services may not be available to attest clients under the rules and regulations of public
accounting. Please see www.deloitte.com/about to learn more about our global network of member firms.
Copyright © 2017 Deloitte Development LLC. All rights reserved.
Member of Deloitte Touche Tohmatsu Limited