PW · PeopleWorks GPT
Now serving A-000 Front desk
EN / ES
AI · Government · Citizens

Now serving: AIWhat it promises public service, and the receipts we should ask for

Shorter lines, less fraud, sharper decisions. AI is arriving at the government service counter with big promises. We take each one, set it next to its risk, and audit the report that lists them, which was itself written by an AI.

PeopleWorks · Nov 7, 2026~15 min readSource: STORM report (Stanford OVAL)
A-01

The report we're auditing

Before you step up to the counter

This article's source is a report called Benefits of Artificial Intelligence in Public Service. It has a table of contents, four major parts and 27 numbered references, and it reads like a consulting brief. It isn't one. It was produced by STORM, a system from Stanford University's Open Virtual Assistant Lab that researches the web and drafts encyclopedia-style articles.

That doesn't disqualify it. STORM is serious academic work, presented at NAACL 2024, and its output is a good map of the territory: it sorts AI's benefits, applications, challenges and future in government. But the cover page carries a warning worth reading slowly:

The generated report can make mistakes. It goes on to ask readers to check important information.

There's a useful irony here. One of the challenges the report describes is documentation: every AI system used in the public sector should explain what it's for, how it was built and where it falls short. An AI-written report about AI in government is a live test of that principle.

This article's rule

Every figure gets a stamp. We checked the key claims against primary sources and gave each one of four marks:

Consistent Shifted ref. Qualified Not traced

"Not traced" doesn't mean "false." It means we couldn't find where it came from, so you shouldn't cite it as fact yet.

A-02

From Dewey to chatbots

Historical context

The report opens far from any server room, in an early-20th-century philosophical argument. John Dewey and Walter Lippmann debated what "the public" actually is. The view they shared reframes government AI: a public isn't a fixed mass of people. It forms through social relations, whenever people find themselves affected by the same problem.

That's why public-sector AI is never only a technical question. Every system that decides who gets a service or who gets investigated creates a new public: the people affected by that decision.

  1. Early 1900s

    The public as a relationship

    Dewey and Lippmann argue that publics emerge from shared problems, not from a census list.

  2. Decades of development

    Machine learning and natural language processing

    AI, the branch of computer science that tries to get machines to reason and decide, picks up two tools that reshape administrative work.

  3. Recent decades

    Chatbots and virtual assistants

    Agencies deploy them to handle inquiries and analyze public feedback, and to lighten administrative load.

  4. Today

    The regulatory question

    The pace of change exposes the lack of clear rules on fairness, accountability and transparency.

A-03

Four benefits, four counters

What the report promises

The report groups AI's benefits into four families. Visit each counter to see the figure the report offers and the stamp it earned once we checked it.

35%potential savings in affected areas over a decade, versus 10–15% from traditional cost-cuttingQualified

Automating repetitive work like data entry, document sorting and case processing cuts errors and speeds up service. The report also claims these tools shave 80 days off average case processing time Not traced.

The 35% is real, but it comes with fine print. We read it at counter A-04.

A-04

Inspecting the 35%

Where the headline number comes from

The report's most eye-catching number is that up-to-35% savings. Tracing it leads to a 2024 Boston Consulting Group study, Spend Smarter and Serve Better. The figure is there, but with three conditions the report drops along the way:

  • BCG gives a range of 25% to 35%, not "up to 35%."
  • It applies to specific functions and agencies, not to a government's whole budget.
  • It needs a long horizon: ten years or more.

For traditional cost optimization, BCG estimates 10% or more across government. The report says 10–15%. Here are both versions side by side:

Estimated cost savings: the report vs. its source

Percent of costs. Hover over or focus each row for details.

Traditional optimizationWith AI (10+ year horizon)
Estimated savings: STORM report vs. BCG 2024 Report: traditional 10–15%, AI up to 35%. BCG: traditional about 10% or more; AI 25–35% in specific functions over 10 or more years. 0%10%20%30%40% The report (STORM) Traditional 10–15% With AI up to 35% The source (BCG, 2024) Traditional ≈10%+ With AI 25–35%
View as table
SourceScenarioSavingsConditions
STORM reportTraditional10–15%Not specified
STORM reportWith AIup to 35%Affected areas, next decade
BCG 2024Traditional≈10% or moreWhole of government, existing tools
BCG 2024With AI25–35%Specific functions, 10+ years, no job cuts assumed

Print your own receipt

Run the number against a real budget. The control that matters is the second one: how much of the budget actually sits in functions AI can touch. That's the share headlines tend to leave out.

Any currency. Example: a mid-sized city.

BCG applies its 25–35% only to specific functions and agencies, not the total.

Receipt for the promise

Affected base—
Traditional (10%)—
AI · low (25%)—
AI · high (35%)—
Gap vs. traditional—

Estimated annual savings once the rollout matures (10+ year horizon). This is a consulting projection, not a guarantee, and it doesn't net out what the AI costs to implement.

A-05

Five scenes of public AI

Applications, and their flip side

The report describes five areas of application. Each card shows the promise. Flip it to see the risk the report itself raises later on, and the question you'd ask as a resident.

Public safety

Police analyze large datasets to spot crime trends and act earlier. The San Francisco Police Department uses AI to detect emerging threats and link incidents that looked unrelated.

Origin confirmed: Oracle, 2024 Shifted ref.

The report warns that bias can creep in during data collection or model building and carry historical discrimination forward. In policing, that means watching most closely the people who were already watched most.

What historical data trained it, and who checked it for bias?

Healthcare

By analyzing service use and demand patterns, AI helps local governments distribute healthcare resources better, especially in low-resource settings, and surface bias in health data.

Application as described in the report

The same report says health-data uses call for sustained, serious public involvement. And marginalized communities may lack the power to push back on what an algorithm decides.

Were the communities who'll get, or lose, the resource consulted?

Oversight and fraud

Local governments are short on auditors. AI scans financial reports quickly and flags anomalies in invoices. As fraudsters use AI for more sophisticated scams, the defense needs it too.

Origin confirmed: Oracle, 2024

An automated flag isn't proof. The report wants developers held accountable for what they claim their systems do. A false positive in an audit can smear an honest vendor or employee.

Who reviews the flag before it has consequences?

Citizen engagement

Chatbots and self-service portals answer common questions without a wait for a person, cutting response times and administrative load.

Origin confirmed: Oracle, 2024 Qualified

A telling detail: the report says citizens ask the chatbot about payroll and onboarding. In the original source, those questions come from government employees. An internal bot and a public-facing one carry very different risks.

Is there always a human path when the bot can't help?

Crisis management

With real-time data, including aerial drone imagery, AI helps responders assess damage and coordinate relief. In 2023 it was used after Ukraine's Kakhovka Dam failed and during the Maui wildfires.

Origin confirmed: Oracle, 2024

Nobody audits a model mid-emergency. That's why the report insists on documenting each system's purpose, design and limits beforehand, and re-evaluating it when new, unforeseen uses come up.

Was it documented and tested before the crisis hit?

A-06

Auditing the source

Ten claims, ten stamps

The report calls for "accountability inputs": transparent information, independent evaluation and clearly stated limits. We applied them to the report.

We checked ten claims. For each we compared three things: what the report says, the numbered reference it assigns, and what we found when we went looking for the origin.

ConsistentShifted ref.QualifiedNot traced
  • #01Dewey and Lippmann: publics emerge from shared problemsConsistent
    Cited
    [1] "Who are the publics engaging in AI?" (PMC)
    Finding
    The reference title fits the claim. It's the only citation in this set that lines up without a second look.
  • #02Up to 35% savings vs. 10–15% from traditional methodsQualified
    Cited
    [3] "Responsible AI for Public Evaluation"
    Finding
    The figure comes from BCG ("Spend Smarter and Serve Better," 2024). The only BCG entry in the list is [9], not [3]. BCG gives a 25–35% range for specific functions over 10+ years, and about 10% for the traditional route.
  • #0380 days cut from average case processing timeNot traced
    Cited
    [3], the same reference as the savings figure
    Finding
    We couldn't find the origin. There are real cases of large processing-time reductions, but none with this exact number. Better not to cite it.
  • #0430–50% gains in service-delivery efficiencyNot traced
    Cited
    [4] "The three challenges of AI regulation" (Brookings)
    Finding
    Judging by its title, the reference is about regulation, not efficiency measurement. We couldn't locate the figure elsewhere either.
  • #05San Francisco police use AI to link incidentsShifted ref.
    Cited
    [25] "Exploring the Pros and Cons of Regulating AI in Government"
    Finding
    The example appears in Oracle's "Using AI in Local Government: 10 Use Cases" (2024), which the report lists as [16].
  • #06AI at the Kakhovka Dam and the Maui wildfiresShifted ref.
    Cited
    [25]
    Finding
    Also from the Oracle piece: in 2023, responders used AI to analyze drone and helicopter imagery. The content holds up; the citation number doesn't.
  • #07Citizens ask chatbots about payroll and onboardingQualified
    Cited
    [25]
    Finding
    In Oracle's piece, the payroll and onboarding chatbots serve employees. Residents use self-service tools for tasks like getting a building permit. The report swapped the audience.
  • #08Build "accountability inputs"Shifted ref.
    Cited
    [10] "How AI is Being Used to Improve Public Sector…"
    Finding
    The term comes from the US NTIA's AI Accountability Policy Report (March 2024). The list does include an NTIA entry: [20].
  • #09Serious public participation around services and health dataShifted ref.
    Cited
    [9] "Benefits of AI in Government" (BCG)
    Finding
    A benefits piece is an odd home for this idea. By title it fits [19], from the Ada Lovelace Institute, on meaningful public participation. Likely, not confirmed.
  • #10Bias can enter anywhere in the AI lifecycleShifted ref.
    Cited
    [11] "8 Benefits of Artificial Intelligence for Government"
    Finding
    By title it fits [21] better: the US Intelligence Community's AI ethics framework. Likely, not confirmed.

A curious pattern

In three cases, the best-matching reference sits exactly ten positions after the one cited:

[9] → [19] participation
[10] → [20] NTIA
[11] → [21] ethics and bias

We can't know why. But it fits what Wikipedia editors found when they evaluated STORM: the model sometimes draws connections between facts that are actually independent. The content is usually good; the trail isn't always.

Audit verdict: the report's map is useful, and most of its concrete cases are real. What breaks is the paper trail, which is exactly what the report demands of public AI systems. If you cite a figure, cite it from its original source.

A-07

The five brakes

Challenges and considerations

The report spends as much space on risks as on benefits. It sorts them into five brakes worth knowing before anyone floors the accelerator.

B-1Public engagement

When AI touches civil rights, as with access to government services or the use of health data, public involvement has to be sustained and serious, not a box to tick. That includes consumers and marginalized groups. Without them there's no trust, and no system that actually meets the community's needs.

B-2Accountability and transparency

To earn trust, each system needs transparent information, independent evaluation and plain communication of its limits and risks. And governments should hold developers to the claims they make about their products.

B-3Ethics and bias

Bias can slip in during data collection or model development. It has to be found, documented and fixed so historical discrimination doesn't repeat. The report goes further: AI shouldn't reduce people to mere "data subjects." And it warns against exploiting communities that lack the power to contest automated decisions.

B-4Documentation and communication

Every system should document its purpose, design, limitations and the rules for its use. That record helps manage it well and helps evaluate it when new uses appear that nobody planned for. It's the brake this article stress-tested at counter A-06.

B-5Ethical applications

Assistive AI can improve decisions and services in healthcare and public safety. But telling ethical uses from unethical ones takes continuous reflection, not one-time approval.

Resident's checklist

Think of an AI system at your city, your college or a public agency you deal with. Can it answer "yes" to these eight questions? Every one comes from the five brakes.

0/8
A-08

What comes next

Future prospects

The report ends on conditional optimism. AI will help agencies allocate resources, anticipate demand and personalize services. But its most human promise is a different one: taking routine work off staff so they can spend more time with the people they serve.

Opportunity

Analyzing large datasets for actionable insight. In healthcare, that means decisions informed by data and aware of each community's needs.

Regulation

Clear rules on privacy, bias and accountability, existing regulation updated to keep pace, and ongoing dialogue among policymakers, technologists and community representatives.

Collaboration

India is building inclusive AI strategies tailored to its social and cultural context, with grassroots participation. That experience can guide other low- and middle-income countries.

The broader lesson from India: imported AI rarely fits. A public system designed for another country, language or data infrastructure can fail precisely the people who need it most. The report calls for industry, academia and government to work out best practices together.

A-09

Counterpoints

Editorial addition by PeopleWorks

AI in government is politically charged territory. It runs into spending efficiency, surveillance and public employment. Here are five reasonable objections, and the reply the report's framework would most likely give.

Objection

The savings figures come from consultancies

Ten-year projections from firms that sell digital transformation aren't independent evidence.

Likely reply

That's exactly why the report wants independent evaluation as an accountability input. The figure points the way; local measurement settles it.

Objection

Predictive policing amplifies bias

If the data reflects where officers patrolled, the model sends more patrols to the same places, in a feedback loop.

Likely reply

The report acknowledges the risk and asks for bias to be identified, documented and fixed across the lifecycle. It isn't an unconditional endorsement of police use.

Objection

"Efficiency" can be a euphemism for cuts

Automation can become cover for shrinking staff and degrading human service.

Likely reply

The report argues the opposite: freeing time for closer contact with residents. And BCG sizes its range without assuming job cuts. Either way, the choice is political, not technical.

Objection

Chatbots shut out the people least able to use them

Older residents, people without connectivity or with low digital literacy can be left behind.

Likely reply

The report's call to involve marginalized groups is how you catch this early. Self-service should add channels, not replace them.

Objection

Demanding perfection means nothing ships

If every system must clear every test, government will never innovate.

Likely reply

The fair comparison is with today's performance. If human auditors catch about 5% of fraud, a well-supervised AI doesn't need to be perfect. It needs to be better, and answerable.

A-10

Takeaways

Five tickets for Monday
  1. The promise is real; the exact number rarely is. Savings are possible, but 35% is a ceiling for specific functions over a decade, not a guaranteed result.
  2. Assisting isn't deciding. AI adds the most and risks the least when it helps a person read, summarize and anticipate, rather than deciding alone.
  3. Every deployment creates a public. As Dewey argued, the people a system affects deserve a say in it before launch, not after.
  4. Ask for the receipt. Purpose, limits, independent evaluation and a way to appeal. If those don't exist, the system isn't ready.
  5. Audit your research AI, too. This article started from an AI-generated report: useful as a map, unreliable as a paper trail. Verify before you publish.
If an AI decided your place in line at a government office tomorrow, would you know who to hold accountable?That question, more than any percentage, measures whether a government is ready for AI.