AI for the People:

Use Cases for Government

Mark Fagan
Harvard Kennedy School

2024

M-RCBG Faculty Working Paper Series | 2024-02

Mossavar-Rahmani Center for Business & Government
Weil Hall | Harvard Kennedy School | www.mrcbg.org

The views expressed in the M-RCBG Associate Working Paper Series are those of the author(s) and do
not necessarily reflect those of the Mossavar-Rahmani Center for Business & Government or of
Harvard University. The papers in this series have not undergone formal review and approval; they are
presented to elicit feedback and to encourage debate on important public policy challenges. This paper
is copyrighted by the author(s). It cannot be reproduced or reused without permission. Pursuant to M-
RCBG’s Open Access Policy, this paper is available to the public at hks.harvard.edu/centers/mrcbg free
of charge. Papers may be downloaded for personal use only.

Mossavar-Rahmani Center for Business & Government

Weil Hall | Harvard Kennedy School | www.hks.harvard.edu/mrcbg
 AI for the People: Use Cases for Government

Artificial intelligence (AI) is the news topic of the 2020s. From ChatGPT writing essays
to chatbots answering questions to algorithms identifying cancers, AI is impacting and
often disrupting how we live our lives. AI is proposed or in place in every domain and
touches every function. In government, AI is relevant for departments from education
and transportation to national security and public safety. Within these organizations, AI
can facilitate the hiring of new employees, providing interactive responses with
constituents, planning workloads, identifying anomalies, and forecasting natural
disasters, to name a few applications.

The potential for good is extensive. AI is credited with improving the quality, efficiency
and equity of many government services. Examples include:
• Narita Airport in Japan uses advanced robots to enhance security though
detection of anomalies about people, baggage, and equipment. The airport is
also using facial recognition software to facilitate passenger movement
throughout the airport while maintaining security protocols. Face Express
identifies passengers via a photo image at their first touchpoint at the airport
enabling them to pass through the rest of the airport without interruption.1
• The US Veterans Administration is using AI to synthesize veteran feedback on
the agency’s services to identify performance trends and issues for detailed
analysis.
• The US Patent and Trademark Office is using AI to search for materials that help
determine the current state of the art as it evaluates patent applications.2
• The Rwandan government worked with Babylon Health to create chatbots
capable of assisting the triage process for patients calling the hospital. Upon
hearing the callers’ symptoms, the triage tool would provide recommendations for
accessing care.3
• The US National Oceanic and Atmospheric Administration is using AI to predict
the location of heat islands so proactive responses can be implemented to
reduce the negative health effects of excessive heat. 4
• The Spanish government has teamed up with IBM’s Watson for the use of this
system to address questions about value added taxes. Since the introduction of
Watson, the number of email inquiries to the tax authority has declined by 80
percent as questions were addressed by Watson.5

1
Narita International Airport Corporation, Digital Transformation,” Automation and Labor Efficiency,”
https://www.naa.jp/en/airnarita/automation.html
2
“The Government Is Using AI to Better Serve the Public,” United States, https://ai.gov/ai-use-cases/
3
Arunima Sarkar et al., “Chatbots RESET Framework: Rwanda Artificial Intelligence (AI) Triage,” World Economic
Forum, March 31, 2022, https://www.weforum.org/reports/chatbots-reset-framework-rwanda-artificial-
intelligence-ai-triage-pilot
4
AI Use Case Inventory (link to download spreadsheet), “The Government Is Using AI to Better Serve the Public,”
https://ai.gov/ai-use-cases/
5
Cristina García-Herrera Blanco, “The Use of Artificial Intelligence by Tax Administrations: A Matter of Principles,”
Inter-American Center of Tax Administrations, https://www.ciat.org/the-use-of-artificial-intelligence-by-tax-
administrations-a-matter-of-principles/?lang=en

1
On the other side of the ledger, AI applications have been criticized for biased analysis
often based on limited and nonrepresentative training datasets. One area of concern is
the use of AI in the criminal justice system. Studies raise questions about biased data
used to train the models. The nonprofit news organization ProPublica documented
numerous cases where algorithms appear to favor whites over blacks in their
assessments.6 Autonomous vehicles were found to have a significantly greater ability to
identify adults and light-skinned pedestrian than children and dark-skinned pedestrians
in low-light situations.7

A very public example was Amazon’s employee recruiting algorithm, which was biased
against women as a result of the training data used. “Amazon's computer models were
trained to vet applicants by observing patterns in resumes submitted to the company
over a 10-year period. Most came from men, a reflection of male dominance across the
tech industry. In effect, Amazon's system taught itself that male candidates were
preferable. It penalized resumes that included the word "women's," as in "women's
chess club captain." 8

Beyond concerns about bias, the broad application of AI raises other issues including
cybersecurity, data privacy, deepfakes, and job loss.

Listening to the media, you might conclude that “AI everything” is just around the corner.
That is not the case. The US Census Bureau survey of AI use found that in 2023 only 4
percent of business were using AI to produce goods and services. The most extensive
use was in the information sector, where the number was just under 14 percent.9 There
are several barriers to broad adoption of AI. The first is trust. Until AI demonstrates it is
trustworthy and provides better decisions than humans, its adoption will be limited. The
industry term for fictitious information is “hallucination.” The term means the AI
perceives something is not really present. While this sounds benign, it is not; making up
information, which large language models (LLMs) are prone to do, does not lead to the
confidence necessary for broad use. Progress is being made but made-up AI
information still abounds.

A second barrier is resources to support and operate AI applications. The development

of robust AI algorithm requires distinctive expertise, including computer and data
science, math and statistics, machine and deep learning, as well as problem-solving

6
Julia Angwin, Jeff Larson, Surya Mattu and Lauren Kirchner, “Machine Bias,” ProPublica, May 23, 2016,
https://www.propublica.org/article/machine-bias-risk-assessments-in-criminal-sentencing
7
“Driverless Cars Worse at Detecting Children and Darker-skinned Pedestrians Say Scientists,” Kings College
London, August 23, 2023, https://www.kcl.ac.uk/news/driverless-cars-worse-at-detecting-children-and-darker-
skinned-pedestrians-say-
scientists#:~:text=The%20researchers%20also%20found%20that,increased%20issues%20for%20nighttime%20drivi
ng.
8
Jeffrey Dastin, “Insight: Amazon Scraps Secret AI Recruiting Tool that Showed Bias Against Women,” Reuters,
October 10, 2018, https://www.reuters.com/article/us-amazon-com-jobs-automation-insight-idUSKCN1MK08G/
9
Cory Breaux and Emin Dinlersoz, “Only 3.8% of Businesses Use AI to Produce Goods and Services, Highest Use in
Information Sector, United States Census Bureau, November 28, 2023,
https://www.census.gov/library/stories/2023/11/businesses-use-ai.html

2
skills. There is a limited number of people with this expertise. This is especially true in
government organizations, which often have difficulty competing with the private sector
for these distinctive skills. A second resource constraint is money. The development of
AI systems is expensive. An AI-powered chatbot that leverages natural language
processing and offers individualized responses to queries can cost in the range of
$50,000 to $100,000 or more.10 Even if funds are available, supply constraints for
hardware can limit AI adoption. One more barrier is the time and energy needed to
integrate the AI program into operations. The requirements included systems
integration, process changes, and developing employees with the necessary skills.
Moreover, AI implementation requires extensive change management internally within
the organization and its customers. Change, especially regarding new technology that is
replacing human decision-making, is never easy.

Given the broad range of AI applications and the need to tap the good and avoid the
bad, the selection and prioritization of the right use cases is essential. This paper
provides a framework for identifying and sequencing AI initiatives. The framework is
then applied to the administration of the judicial system to illustrate how it works in
practice. The overarching philosophy of the approach is to fix existing performance
gaps; longer-term, the focus can shift to making good performance better.

I. AI Basics
AI means different things to different people. The Oxford English Dictionary defines it as
“the theory and development of computer systems able to perform tasks that normally
require human intelligence, such as visual perception, speech recognition, decision-
making, and translating between languages.”11 The operative words are “normally
require human intelligence.”

There are four core characteristics of AI. It can:

• Make or support decisions.
• Combine human attributes.
• Evaluate data and take actions.
• Provide feedback for continuous improvement.

Another way of idenitfying AI systems is that they meet three requirements. The first is
intentionality—they make decisions as humans would. The second is intelligence—AI
uses data to make meaningful decisions. The third is AI systems’ adaptability—they
learn and adjust to new information.

10
“Chatbot Development Cost in 2024,” Antino, March 3, 2024, https://www.antino.com/blog/chatbot-
development-cost

11
Definition of “Artificial Intelligence, Oxford English Dictionary, https://www.oed.com/dictionary/artificial-
intelligence_n

3
Segmentation of AI systems is also helpful context. One differentiator is how intelligence
is imparted into the system. AI models historically have learned through testing different
approaches and refining the approach based on results. Pretrained models (the “P” in
GPT) are trained on massive databases with the intension that they deliver the right
answer based on the data. Another differentiator is whether the models are supervised
or unsupervised. Supervised model results are reviewed by humans and graded to
refine the performance of the AI program. This approach has the benefits of minimizing
hallucinations and finding biases, but it is very time consuming. Unsupervised systems
allow the program to self-learn based on using large amounts of data to calibrate a
model and then take a subset of the data and test the results. Based on the results the
model is refined; human intervention is unneeded. The downside is no human is looking
at the results.

These characteristics and requirements are exemplified by autonomous vehicles (AVs).

AVs use AI to replace a myriad of decisions that humans are particularly good at, such
as when to accelerate, when to brake, setting travel speed, waiting for pedestrians in
the crosswalk, and pulling to the side of the road to allow emergency vehicles to pass.
Making these decisions requires the combination of many human skills, from vision and
hearing to reading and problem-solving. Evaluation of the operating environment and
taking response actions is a continual process when driving. The AV is able to do the
same and it learns as it goes, improving performance just as a human driver does with
more experience. The promise of AVs is that they should significantly reduce the
number of deadly vehicle crashes, since the majority result from human error.

AVs also exemplify the challenges to substituting computer programs for human
decisions. Consider driving and coming to an intersection with stop signs in all
directions. When vehicles arrive at the same time, who goes first? There are rules of the
road to answer this question. In Massachusetts, the answer is the first in is the first to
proceed and if two vehicles arrive at the same time, the vehicle to the right goes first.
Compare the rules with what actually happens. In Boston, the outcome is more a
function of an informal negotiation from afar than based on the rules. The most
aggressive driver goes first. How do you program an autonomous vehicle: with the rules
of the road or what people really do?

II. Use Case Considerations

A starting point for identifying high-value use cases is understanding where AI has
comparative advantage. AI is particularly strong in:
• Searching, finding, reading, and summarizing: At the macro level, the most
popular generative Large Language Models (LLMs) draw from the entire internet
to search for information. They are also effective in synthesizing vast amounts of
data and turning it into useful information. At the micro level, AI can also search a
1,000-page legal brief and find every time a specific item such as “endangered”
is mentioned, as well as an array of synonyms including “threatened,” “imperiled,”

4
 “jeopardized,” and “risked.” The user does not even need to specify the
synonyms, the algorithm can do it automatically based on its training.

• Writing prose: Need a birthday message for a friend? Need an op-ed? Need a
training manual? LLMs are your solution. Enter a few prompts and out it comes in
seconds. Need to make it easier to understand? Prompt the program to write for
an eighth-grader and it is produced in seconds. Not only can AI write, but it can
also write in multiple languages.

• Interactively answering questions: We have all become a customer served by

computer response systems. “For a complaint press #2,” for example. These
defined systems are effective if they address your specific concern. If not, the
result is frustration. AI chatbots have the ability to allow you to ask (via voice or
text) a question and get a custom response to your specific concern, in a matter
of seconds. These systems also allow you to refine your ask based on the
response. For example, you might have the following exchange with a school
department chatbot:

o You: “What documents do I need to register my child for school?

o AI: Proof of address such as a utility bill.
o You: “I don’t pay the utilities?”
o AI: You can also use a letter from the city government mailed to your
name and address.
The results are fast and responsive information, 24x7x365 for the constituent and
lower cost for the school department. The interaction can happen in literally
dozens of languages.

• Pattern and anomaly recognition: AI systems are highly effective at finding

patterns in numerical data or text. Humans are also good at these activities, but
AI avoids human biases. We tend to see what we expect to see and what we are
asked to focus on. The video where you count the number of basketball passes
between people but miss the gorilla walking through the group is a powerful
example.12 Beyond biases, AI systems can process much larger volumes of
information. AI anomaly identification capabilities are helping to provide security
at airports, identify traffic bottlenecks, and predict disease outbreaks. AI systems
are also used to identify and predict malfunctions and faults such as weakness in
infrastructure or impending failure of machinery.

• Sensing emotions: Reading the body language of others is a very human trait.
Some do it well, others less well. AI has the ability to sense emotions based on
training data. AI systems have been trained based on facial and voice recognition
as well as text to identify human emotions, from happiness and sadness to anger
and anxiety. At one end of the spectrum, the information can predict mental
12
Daniel Simons, “Selective Attention Test.” March 10, 2010. Video, 1:21.
https://www.google.com/search?client=safari&rls=en&q=gorilla+walking+across+basketball+court&ie=UTF-
8&oe=UTF-8#fpstate=ive&vhid=zephyr:0&vld=cid:46d895f4,vid:vJG698U2Mvo,st:0&vssid=atritem-

5
 health concerns and provide proactive interventions. At the other end, it can be
used to target ads.

• Integrating multiple inputs and projecting outcomes: Accurate forecasts are

the foundation for many services, from predicting storms and managing traffic to
operating the electric grid and predicting health outcomes for individual patients.
Predicting electricity production is illustrative. The AI algorithm integrates
historical production trends, operational efficiency, weather forecasts, and
weather conditions at a very detailed level to estimate the amount of electricity
that will be needed and the optimal portfolio for generation. The better the
prediction, the more reliable the electric grid.13

Note that this list is indicative not exhaustive. New AI capabilities are increasing daily.

AI offers many opportunities, but the public sector manager faces several unique
challenges that impact the use case selection. The first is that AI decisions have a
political overlay. Operational objectives must be combined with political realities in
identifying viable use cases. Using an AI hiring algorithm may create a risk of bias that
is greater than a mayor is willing to accept. A corollary factor is general risk aversion.
Media and public scrutiny may limit risk-taking. Neither the governor nor the cabinet
secretary is willing to lose their job because of an AI implementation gone bad.

A second concern for public organizations is the procurement process. Many

government entities have strict purchasing rules that may dictate the selection of the
low-cost supplier. If the low-cost provider is also the best-qualified supplier, the outcome
is great. However, this is not always the case. Public procurement regulations also can
limit collaboration with potential suppliers. AI is new and ever evolving. Close
collaboration between the government and the provider are needed to modify
specifications as the AI product is being developed and delivered. This is often
challenging in the public sphere.

Multiple customers is another complexity faced as government organization bring in

AI. Who is the customer for a mental health AI product? Is the customer the recipient of
the services? The person’s family? Civil society? Each has their own interests.
Determining how to simultaneously satisfy all needs can be so daunting that the status
quo persists. Finally, there are endless opportunities. The range of AI uses cases in
government is broad and dense. This, combined with limited funds and expertise,
elevates the need to carefully prioritize AI use cases. The next section provides a
framework for identifying and prioritizing AI use cases.

13
John Villali, “Advanced AI-Powered Energy Forecasting,” IDC Research, Inc., July 2023,
https://www.sas.com/content/dam/SAS/documents/analyst-reports-papers/en/idc-advanced-ai-powered-energy-
forecasting-113587.pdf

6
III. Use Case Identification and Prioritization Framework
The foregoing materials highlight the importance of having a structured process to both
identify and prioritize AI use cases. The goal of the process is to emerge with a list of AI
applications that will enhance the delivery of government services, especially those
where current performance is inadequate. The list should also reflect the ability of the
organization to successfully implement the AI use case at an appropriate level of risk.

The framework consists of the following five steps:

• Step 1: Profile core functions (A specific problem-centered approach is described
in Appendix A.)
• Step 2: Prioritize “reds”
• Step 3: Identify AI applications for top-priority reds
• Step 4: Assess AI solutions
• Step 5: Prioritize AI solutions
Each step is detailed below.

Step 1: Profile Core Functions

The starting point for identifying potential AI use cases is profiling the functions
performed by the agency and highlighting those functions that are both critical and for
which performance is not fully satisfactory. External and internal functions are included.
The external functions are those providing service to constituents/customers. In the
case of the registry of motor vehicles, for example, those functions include conducting
driving permit exams and road tests, providing driver’s licenses, transferring ownership
titles, registering vehicles, certifying insurance coverage, and driver safety awareness,
to name a few. The internal functions range from hiring and developing staff to
designing and implementing IT and operating service centers. Strive for a
comprehensive list.

Next, map the functions based on criticality. A simple test of criticality is: Will managers
be fired if the function is not readily available. This outcome is unlikely if a public safety
campaign is delayed a few weeks, but likely if licenses cannot be issued for a few days.
A more comprehensive approach involves four tasks. The first is to map stakeholder
expectations. Who are the stakeholders and what do they want and expect and how do
these align by function? The second task is to assess the power and interest of the
stakeholders, again by function. Power includes political influence, funding, expertise,
and legitimacy. Interest is how much of a priority this is for the stakeholder. The third
task is to map the criticality of the agency functions on a spectrum from critical to less
important, based on the stakeholder assessment. The final task is to consider the error
of margin associated with the task. Is the error tolerance associated with AI acceptable?
Interestingly, in the case of AVs, human error (i.e. driver caused crashes) is more
acceptable than AI failures.

7
Step 2: Prioritize the “Reds”

Because resources are limited, the focus of adopting AI should be prioritized to the
critical functions where there is significant room for improvement. A green/yellow/red
performance assessment provides an easy to implement and communicate approach.
Where possible, the assessment should be quantitative, such as wait times, processing
times and/or customer satisfaction surveys. Qualitative evidence is also helpful, such as
the number of positive versus negative ratings on Yelp-like sites. Manager observations
and employee input also play a role in the assessment. The green/yellow/red
designation is likely subjective, but having some measurements that define each is
preferable where possible. The AI use case priorities are the most critical “reds.”

For many government services, different constituents/customers have different needs

and expectations. Segmenting users can tease out nuances. For example, in the
registry example, drivers who can renew their license online view wait times differently
from those who need to complete the renewal in person. Geography can be another
important way to segment. Those who live far from the registry are likely to have little
tolerance for failing to complete their transaction in a single trip with minimal waiting
time. Equity is also a consideration. Having multilingual staff at some locations is more
important than at others.

Step 3: Identify AI Applications for Top Priority Reds

AI is unlikely to be the solution to all the reds. There are functions where AI has a
comparative advantage; yet others where non-AI solutions are preferable. Also, there
are applications where AI has a well-established track record and others where it is not
ready for prime time. The risks of adopting AI can be reduced by being a fast follower
rather than a first mover. Regulation is also a factor. In some geographies, higher risk
applications require extra scrutiny.

Another risk mitigation strategy is segmenting the reds based on back-office versus
customer/constituent-facing activities. Back-office activities are often lower risk. Also,
less-critical functions may be a good starting point.14 The counterpoint is that AI is a
major undertaking consuming funds, time, and expertise. Therefore, the use cases
addressed need to generate enough value to ensure there is a strong positive return on
investment.

Back office or constituent facing the applications require data for training and testing. An
additional screen is the availability of that data. Today that data might not be collected
and ready for use. Thinking about the long term, beginning to collect the data now might
enable the effective use of AI in the future as new algorithms are developed.

A good starting point for identifying AI applications is seeing where AI is currently in use
in the public sector. Beginning with existing applications ensures the use case works

14
Insight from Darryl Slabe, HKS Master in Public Admisntration candidate.

8
and provides social proof helpful in selling the idea to internal and external
stakeholders. Moreover, the people running the existing applications can be a source of
insight.

Focusing on existing public sector applications ensures that the AI use is appropriate in
the unique government context. The US federal government maintains an inventory of
AI applications.15 The list includes uses by 27 federal departments and agencies, from
Justice and State to the Environmental Protection Agency and the National Science
Foundation. For example, the Department of Homeland Security’s AI inventory includes
applications for headquarters functions, as well as uses in the subagencies including
the Transportation Security Administration (TSA), Federal Emergency Management
Agency (FEMA), Customs and Border Protection (CBP) and the Cybersecurity and
Infrastructure Security Agency (CISA). Use case examples include:
• Geospatial damage assessments (FEMA)
• AI for Autonomous Situational Awareness (CBP)
• Automated Indicator Sharing – Scoring and Feedback (CISA)
• Touchless Precheck Identity Solution (TSA)
Scanning the hundreds of current AI use cases serves as an effective thought-starter for
addressing the “reds.”

Another source of AI applications in the public sector is the Government AI Coalition

initiated by the city of San Jose, California, in 2023. The mission of the coalition is to
provide a vehicle for state and local government to have a voice in architecting the AI
future.16 The coalition includes representatives from 250 state, county, and local
government organizations. The group’s work is guided by the following goals:
• Use AI for social good,
• Ensure ethical, nondiscriminatory, and responsible AI governance,
• Promote vendor accountability,
• Improve government services, and
• Foster cross-agency collaboration and knowledge-sharing.
They offer a range of resources to support the development of AI tools, including use
case templates and a registry of AI vendors.

Private sector AI applications offer another source of insight, especially where the
application can be implemented in the public sector without extensive modification. AI
use cases in the private sector cover the gamut of industries, from law and insurance to
real estate, finance, and manufacturing. The use cases include:17
• Chatbots for customer service;
• Productivity tools from scheduling to note taking to performance assessment;
• Fraud detection and credit scoring in finance;
• Market analysis and predicting prices in purchasing;

15
“The Government Is Using AI to Better Serve the Public,” United States, https://ai.gov/ai-use-cases/
16
Government AI Coalition, City of San Jose, https://www.sanjoseca.gov/your-government/departments-
offices/information-technology/ai-reviews-algorithm-register/govai-coalition
17
MindTitan, “AI Use Cases,” https://mindtitan.com/resources/industry-use-cases/

9
 • Quality assurance, predictive maintenance, and process automation in
manufacturing;
• Research and records management in law; and
• Call centers in many sectors.

A third source is AI providers. The AI provider community can be segmented into (1)
solution providers; (2) AI-powered product providers; and (3) AI consultants/integration
providers. The solution providers offer AI products that address specific functions and/or
industries. OpenAI’s ChatGPT and IBM’s Watson are examples of products that offer
solutions that cut across functions and industries. For example, Watson can be used to
create chatbots, streamline coding, and find insights in datasets. AI-powered product
providers like Apple (Siri) and Amazon (Alexa) use AI to support use cases including
research, personal assistants, and managing the internet of things at home. A cadre of
AI consulting firms offers services to help design and implement AI applications. Their
comparative advantages are AI expertise and customization to the specific client needs.

Another source is open source projects were developers great AI applications and make
them available to the public. They are free and often have an extensive user base that
continually improves the application. Example span the gamut from productivity to
coding to science and research. These offering might not look pretty but they can be an
effective starting point.

While AI applications are not prevalent in all organizations, it is likely that someone is
using AI to address your reds. Try using AI to find that needle in the haystack, but make
sure you verify.

Step 4: Assess AI Solutions

Just because an AI solution exists does not mean it should be adopted. Like any
initiative, AI applications must have a positive return on investment. The benefits must
outweigh the costs. The benefits include improvements in the quality, efficiency, and
equity of services provided to customers. Internal improvements in effectiveness and
efficiency are also part of the calculus.

A range of costs needs to be considered, including the money to acquire, implement

and maintain the system as well as the staff time to incorporate the AI application into
existing operations. There are also the costs of having staff monitor the application
especially when it is first implemented. Given the newness of the technology the human
intervention might last a long time. New data also adds to costs as more information has
the potential to improve performance but also requires time to retrain and retest.

Change in general, and with AI in particular, introduces risks. These include operational
risk, the costs resulting from the failure to deliver required services, and reputational
risks if the AI implementation fails to meet expectations. As mentioned above, these
risks are greater in the public sector where failures are likely to be very visible. The
media and political critics can easily amplify news of a failure, reinforcing the damage to

10
the organization’s reputation. An additional consideration is the capacity for AI change.
Do you have the expertise to bring in a new application? Also, what other initiatives are
crowded out by the AI adoption?

The costs and benefits should be quantified as much as possible. Many of the costs are
comparatively easy to quantify. The funds paid to vendors are clear. If payments take
place over time, they can be discounted to get a present-value cost. The time needed
for training is also easily estimated, with a drop in productivity during a transition period.
Risk is harder to quantify. One approach is to look at the impact of other failures on an
organization. For example, how was the budget of an organization reduced when a
failure took place? The risk assessment also needs to reflect the strength of the AI
provider—reliability, responsiveness, expertise, bandwidth, and long-term financial
viability.

Benefits are often more difficult to quantify. What is the value of a more responsive
employee who has been hired based on an AI algorithm, for example? The following is
a segmentation and progression for quantifying benefits.
• Segment 1: Direct Financial Gains — If the AI application reduces staffing
levels, the avoided labor costs are in the benefit set. Or if AI reduces raw
material needs and space required, the eliminated expenses are included.
• Segment 2: Improved Accuracy — AI algorithms can often out perform
humans. For example, AI programs often are better at detecting analogies in
medical imaging than humans.
• Segment 3: Customer Time Savings — Returning to the registry of motor
vehicles example, if AI scheduling reduces average wait time by 10 minutes per
customer, the value of the time savings is estimated as a benefit. Time savings
are routinely quantified for transportation infrastructure assessments. The
cumulative time saved by all users is multiplied by an average wage rate to
approximate the benefits.
• Segment 4: Ancillary Benefits — Some gains from AI adoption are indirect. For
example, an AI traffic model that reduces congestion not only saves time but
also reduces carbon emissions. Protocols for quantifying carbon reduction
benefits are now well established.
• Segment 5: Reputational Gains — AI-improved performance can improve an
organization’s reputation, leading to a variety of benefits, including increased
budget, enhanced recruiting of desired candidates, and less stress for
employees.

The focus of the cost and benefit analysis should be on quantitative data, but qualitative
information is also valuable. Descriptive narratives, especially of benefits, can often
bring the benefits of AI to life. A customer who says they were able to get immediate and
accurate help from a tax authority chatbot rather than wait on the phone for 20 minutes
is powerful evidence of the value of AI.

There may be multiple AI options to address a specific red opportunity. In those cases,
the benefit-cost analysis should be completed for each option. Choose the option with

11
the best return on investment. In making this determination, consider the working
relationship you will establish with the AI provider. You will be partners for a long time,
so make sure you have the right “chemistry.”

A final assessment is a make-buy decision. Should you develop and implement the AI
application internally or partner with a third party? The assessment criteria include:
• Expertise
• Cost
• Time-to-use
• Customization
• Maintenance
• Data protection
• IP control

Expertise, cost and time are typically the threshold considerations. Where possible,
quantify the assessment, but in most cases the final decision will be a qualitative.

Step 5: Prioritize AI Solutions

The prior step is completed for each of the top priority reds. Before launching initiatives
that have a strong positive benefit-cost analysis, develop a road map to the overall AI
plan. AI applications need to fit into the overall strategy and system of the organization.
An AI application that, for example, provides more accurate forecasts of demand is
necessary but not sufficient to improve customer satisfaction unless it is accompanied
by a delivery system that is agile enough to use the improved forecasts.

Create a road map of how to sequence the AI applications that address immediate
needs, the reds. Divide the map into a near-term plan (2–24 months) and a long-term
view of 2–4 years. The sequencing should balance the comparative importance of the
applications to performance with the risk-adjusted net benefits. Review and refine a
draft mapping reflecting the organization’s overall strategy.

This task is competed by identifying one or two initiatives to adopt as a pilot, taking AI
for a test drive in a low-stakes environment to learn what works and what does not.
Those applications selected to pilot should have a high probability of success in light of
the capabilities of the technology, the provider, and the implementing team.

IV. Framework Applied to the Judicial Court System

This section shows how the framework could be applied to the administration of the
judicial system, specifically the courts. The content below is only illustrative; it is not
based on a specific judicial system. The materials are intended to provide a general
template and thought-starters for how to apply the framework in your own specific
context. Citations are provided when factual information is used.

12
The core process stages of the court system are fourfold: pleadings, discovery, trial, and
judgement. The completion of this process involves administrative tasks and
substantive decisions. The five-step AI use case framework is applied below.

Step 1: Profile and Assess Core Functions

The starting point for identifying and prioritizing AI use cases in the court system is
detailing the core functions performed. From the administrative perspective, managing
the workforce, court facilities and physical evidence, documents, and digital content are
core. The key functions are detailed in Exhibit 1.

The substantive tasks often involve legal research and analysis. Decision-making is the
primary substantive responsibility of the courts and covers a broad spectrum of
activities, the most important of which are setting bail, determining fault, and
determining sanctions; sentencing in criminal cases; and compensation/restitution in
civil matters. (Exhibit 2)

13
The assessment criteria for the core functions are quality, efficiency, and equity. From
an administrative viewpoint, quality is a clear, predictable and reliable process.
Efficiency is achieved when all resources, people, facilities, equipment/technology are
used with minimal waste. Equity refers to equal treatment of all participants in the
process regardless of race, socioeconomic status, location, gender, language, etc.
Using red, yellow, and green as an assessment grade for “poor,” “all right,” and “good,”
the following is the evaluation of the administrative core functions in our hypothetical
court system:
• Managing the workforce: Red
o Quality: Unpredictable schedules, poor retention (20% annual turnover
versus 12% in other government agencies).
o Efficiency: Recruitment process takes 3 weeks longer than benchmark.
o Equity: Failing to meet diversity objectives.
• Managing the court facilities: Red
o Quality: Facilities need $25 million to be in a state of good repair, frequent
HVAC problems, spotty internet.
o Efficiency: Maintenance costs 15% above industry standard; utilization of
courtrooms in only 55%, 15% below industry standards.
o Equity: Unreliable translation services cause delays and associated costs;
courts in lower socioeconomic locations in a worse physical condition.
• Managing documents and digital content: Red
o Quality: Document files well organized and easily accessed, but limited
use of digital archival technology.
o Efficiency: Accelerated use of digital storage offers $1.5 million in
document retention budget.

14
 o Equity: All courts have access to the same document management
process and technology.

The bottom line is that there is a significant opportunity for improvement in the
administration of this judicial system.

Turning to the substantive functions, quality is defined as making decisions based on

comprehensive, unbiased, and apolitical legal scholarship. It also reflects the accuracy
of judicial decision-making. For example, was Driver A or Driver B at fault in a traffic
crash? Efficiency addresses the time it takes to complete legal proceedings. Indicators
are the total elapsed time for the process and the time judges can work on legal
decisions rather than on administration. Equity centers on decisions that are consistent
regardless of race, gender, socioeconomic status, or age, and judicial biases are
minimized.

The assessment reveals the research and analysis activities are yellow and decision-
making is green. Details and supporting rationale are below.
• Research and analysis: Yellow
o Quality: Several decisions were overturned based on weak
research/analysis scholarship.
o Efficiency: Information is readily and easily accessible to all parties at the
court and remotely.
o Equity: Few decisions are overturned based on bias/discrimination.
• Decision-making: Green
o Quality: Few decisions are appealed, fewer are overturned.
o Efficiency: Court backlog and total elapsed time are less than
benchmarks; judges spend less than 15% of their time on administration
compared to the 20% benchmark.
o Equity: Academic study of court decisions reveals minimal instances of
bias in decision-making.

Exhibit 3 provides a summary of the assessments. These assessments are the inputs to
segmentation and prioritization, the next step.

15
Step 2: Segment and Prioritize the “Reds”

The potential applications of AI in most organization are expansive yet the capacity for
adoption is limited. Therefore, prioritization is key. The functions profiled in Step 1
provided the reds, those activities where there is significant opportunity for
improvement. In this illustration, these include managing the workforce, court facilities
and evidence/document/digital content. These broad categories should be unpacked to
specific activities, as AI applications are generally task specific. For example, the need
for better management of the workforce is unpacked into hiring, training and retention,
and scheduling.

The AI priorities emerge by considering the criticality of each specific task: critical to
performance and critical to customers/constituents. Exhibit 4 shows the criticality
mapping. Managing several aspects of the workforce, translation services and
scheduling courtrooms are top priorities.

16
Step 3: Identify AI Applications for Top-priority Reds

The critical reds are the most pressing areas for improvement in the court system.
Adopting AI is predicated on the availability of AI applications. As mentioned above,
being a fast follower (using existing AI applications) is preferable to being on the leading
(or often “bleeding”) edge.

In this illustrative example, the one priority is providing translation services. The court
system guarantees litigants, attorneys, and witnesses real-time translation to their
native language for both oral and text communications. Routinely, translators are not
available, leading to delays in proceedings which are costly for the court and its
“customers.” The root causes of this failure are (1) a lack of translators, especially in
minority areas; (2) lack of translators in less widely used languages; (3) low
compensation levels and no payment for travel time and cost; and (4) lack of sufficient
lead time to arrange for services.

AI is routinely being used for translation in business and government. For example, US
Customs and Border Protection developed CBP Translate to enabler its agents to
communicate effectively with travelers who are not proficient in English when
conventional translation services are not available. Protections are in place to obtain
consent for the traveler to use the program; travel documents are not photographed;

17
biographical information is not captured for US citizens or Lawful Permanent Residents;
and the conversation history is deleted after 90 days.18

Another example is the use of an AI translation chatbot by the Minnesota Department of

Public Safety’s Driver and Vehicle Services. The AI program provides translation in
multiple languages including Spanish and Somali. The system provides interactive
information such as updating insurance and updates in title and license status. The
project was initiated as a result of customer input that indicated the language barrier
was a major challenge for many in the state. The chatbot is powered by Google AI.19

In legal settings, however, the use of AI translation is still under consideration. Given the
importance of addressing the court translation problem and that few options that exist,
assessing AI translation is justified. Of course, the return on investment needs to be
high to overcome the leading-edge risk.

This same process is completed for each of the red functions.

Step 4: Assess AI Solutions

Knowing that AI translation is used in government settings, the next task is determining
if AI is a viable solution in your context and, indeed, the best solution. AI may be the
best solution if the following conditions are met:
• AI providers have the capacity to design, implement and support your solution.
• Your organization has the expertise to use the AI application and management is
prepared to support the change management process.

If these conditions are met, the next question is: Does using AI generate not only a
positive return on investment but also a superior return compared to the non-AI
alternatives? The return on investment is estimated based on the costs and benefits of
the AI application. Exhibit 5 provides an illustrative example of the key factors for
consideration.

18
US Department of Homeland Security, “Privacy Impact Assessment for the CBP Translate Application,” March 16,
2021, https://www.dhs.gov/sites/default/files/publications/privacy-pia-cbp069-cbptranslateapplication-
march2021.pdf
19
Julia Edinger, “Translation AI Helps Bridge Language Barrier for Minnesota DVS,” Government Technology,
September 28, 2023, https://www.govtech.com/artificial-intelligence/translation-ai-helps-bridge-language-barrier-
for-minnesota-dvs

18
The costs include:
• Initial development or use fee
• Ongoing operating/computing and maintenance
• Implementation, change management and training
• Quality control
• Privacy and security functions
Some of these are cash outlays for hardware, software, and expertise; others are using
employee time and other organization resources. The costs are readily expressed in
dollars. Of note, the cost of computing can be substantial. There is a tradeoff between
more comprehensive models but their higher computing costs. Narrower models are
likely to be more feasible for courts where budgets are likely constrained.

The benefits are internal efficiencies and external customer satisfaction. Internal
efficiencies could include reduction in translation costs as AI substitutes for labor. The
availability of AI translation at scale could also speed court proceedings, improving
utilization of juridical resources. Moreover, AI is likely to enhance consistency through
uniform translations. The labor reduction benefits can be directly converted into dollar
savings. Resource efficiencies can also be estimated. For example, if proceeding
delays waiting for translators are reduced by 10 minutes per day and there are 6 staff
members impacted, there is an hour per day of labor savings. The value of this is the
average labor cost per hour multiplied by the days per year the court operates.

The external customer benefits have both tangible and intangible dimensions. Tangible
benefits center on faster proceedings. Less time waiting for translators is valued at an
average wage rate. Likely of greater value is a better understanding of the proceedings

19
for the foreign language speaker and the other participants in the proceeding, which
should lead to a more accurate and equitable outcome.

Valuing the customer benefits in general and the right outcome in particular is very
difficult. One way to complete the cost-benefits analysis despite this challenge is
conducting a breakeven analysis. Compare the costs with the directly quantified
benefits. The difference represents the value that the indirect benefits must account for.
If the number is small, it is likely the return on investment is positive.

Implementation of any new technology, including AI, has risks. The risks are on the cost
side—higher than planned expenses and longer than anticipated implementation. There
are also risks that the anticipated benefits will not materialize. Therefore, use
conservative estimates and proceed only if the return is strongly positive.

A strong positive return on AI is necessary but not sufficient. The final requirement is
that the return is higher than the return on non-AI alternatives. In the forgoing scenario,
there were no other alternatives.

The bottom line in this example: If AI translation offers a strong positive return on
investment, it should be on the list of AI applications for prioritization.

Step 5: Prioritize AI Solutions

The final step in the AI use case prioritization process is developing an overall AI road
map. Two time frames are recommended: near-term (1–2 years) and long-term (3–5
years.) The near-term list of AI initiatives should focus on the current reds, where
existing AI applications are ideally in use in the courts, but if not, at least in government.
Use the “crawl-walk-run” approach to sequencing the projects. Do the easiest first to
build internal capacity and minimize risk. Here the goal is to understand how best to
bring AI into your organization. This also enables your organization to establish a track
record of success and value creation, which will fuel subsequent initiatives. View the
first few efforts as pilots, where learning is as important as generating the intended
return on investment.

The long-term mapping reflects how the initiatives support the organization’s
overarching strategy. Rather than addressing individual red functions, the long-term
plan is directly aligned with mission and vision. A sample map for using AI in the courts
is provided in Exhibit 6. Given the pace of innovation in the AI space, the long-term plan
informs the indicative direction of the organization’s adoption of AI but should be
revisited on a semi-annual basis to incorporate internal progress on AI use and new
external applications being developed.

20
VI. Conclusion
AI is here to stay, and its impacts will continue to grow. Governments have an
opportunity to leverage AI to delivery greater value to its constituents. The framework
detailed above provides a structured approach to identifying and prioritizing AI use
cases that maximize value and minimize risk.

Appendix A: Specific Problem-Centered Approach

Steps 1 and 2 of the framework are designed to examine the entirety of the organization
for AI use cases. However, some organizations may have a specific problem they need
to tackle and seek to determine if AI offers a solution. In those cases, an alternative
approach that narrows the focus of the investigation to addressing the problem is more
efficient. This is accomplished by substituting the following tasks for Steps 1 and 2.

Task 1: Define the Problem and Assess Its Root Cause

Solving a problem begins with a clear articulation of the concern. The Kepner-Tregoe
framework is effective in defining the problem. The framework asks the typical questions
of what (identify the issue), where (locate the issue), when (timing of the concern and
when did it first appear) and the magnitude of the issue. Importantly, it also asks the
questions: When it is not a problem and where does it not appear? The answers offer
important insights, enabling more specific problem definition.

21
A clear and crisp problem definition is necessary but not sufficient. The root cause of the
issue must be determined. There are several approaches to getting to the root cause of
the problem. An effective and intuitive approach is the Ishikawa or fishbone diagram.
The head of the fish is the effect (i.e., the problem) and the bones are the causes. To
ensure a comprehensive identification of the causes, ask: What is the role of polices,
process, people, and plant/technology? These four dimensions form the core elements
of the fish’s skeletal structure. The smaller bones are the specific causes. When
drawing the diagram, the more important the issue, the thicker the bone. An example of
a fishbone root cause analysis is shown in Exhibit A-1..

Task 2: Identify Solution Options

There are often several options for solving a problem. Some might be policy or process
related, others involve people. Others require a technology solution including AI.
Several tools support the ideation process. One is SCAMPER. This acronym stands for
substitute, combine, adapt, modify, put to another use, eliminate and reverse. These
words provide a checklist to prompt your thinking on different ways the problem could
be solved. Another tool is the ideation loop. You start by imagining a solution. Next you
sketch what that solution looks like in reality. Then you create a model or test to see if it
works. The idea of looking is that the process is iterative, continually refining and
improving the solution.

Once you have a comprehensive list of potential solutions, you can evaluate them to
find the most promising approach. Screening through the options is facilitated with a set
of evaluation criteria. Efficacy, the direct improvement in the situation, is one key
criterion. Others include operational, political, and financial feasibility. The criteria can be
weighted to reflect their comparative importance.

22
The top option or two will become the focus for the AI assessment that begins in Step 3
of the AI use case framework presented in the main article.

23