Governments Are Counting AI Tools. They Should Be Measuring Public Value.
Most public debate about artificial intelligence (AI) in government focuses on two questions: how much is being deployed, and how risky it is. Those questions matter, but they miss the more useful one: Does AI measurably improve people’s experiences of government?
Right now, eligible Americans forgo more than $140 billion in federal benefits each year. This is, in part, because navigating the paperwork, eligibility checks, and renewal processes is simply too hard. AI could help reduce that burden, but it is not the problem government AI is truly being built to solve.
Instead, a significant share of government AI is focused on identifying problems like fraud, noncompliance, and operational risk. That reflects a choice about where AI delivers value—and leaves less attention for improving access, reducing delays, and easing administrative burden. States and cities should learn from that federal pattern before they repeat the same mistakes.
What governments build reveals what they value
The 2025 Federal Agency Artificial Intelligence Use Case Inventory documented 3,611 individual use cases across 56 agencies, up from 1,757 reported the year before. The shape of this growth tells a story about the government’s priorities. Roughly half of recent federal AI use cases are categorized as mission-enabling, covering things like finance, human resources, and facilities management—and many cases involve things like fraud prevention and audit targeting.
Take the Centers for Medicare and Medicaid Services (CMS). CMS’ Center for Program Integrity uses roughly 250 AI models and about 500 staffers to help review 4 to 5 million claims per day. In June 2025, the Department of Justice announced that AI-assisted enforcement had helped stop alleged healthcare fraud totaling more than $14.6 billion.
At the Internal Revenue Services (IRS), the Government Accountability Office (GAO) found 126 active AI use cases, including machine learning models that identify which returns to audit and tools rewriting decades-old code to modernize aging systems.
None of this is inherently wrong. Fraud is real, and taxpayer dollars lost to fraud are dollars not available for anything else. But these systems are oriented toward watching the public and finding exceptions, not helping eligible people navigate services they need.
What “for the public” would actually look like
A case from Wisconsin shows what helping the public might look like.
The Wisconsin Department of Safety and Professional Services used AI to cut the wait time for occupational licenses, or the permits needed by nurses, contractors, or cosmetologists before they can legally start work. In 2023 and 2024, the number of licenses increased by 35% compared to any previous two year period, translating into an estimated $54 million in additional wages for workers starting jobs sooner. That is a story about measuring how many people got through, not how many got caught.
For a nurse waiting on a license, a three-week delay is not an administrative statistic. It’s three weeks of missed wages and a family budget under strain. Removing that delay is the kind of outcome that should define success in government AI, rather than just a cleaner audit trail. It’s also the kind of outcome that improves people’s trust in government.
The Center for Civic Futures’ Public Benefit Innovation Fund points in this direction, with $8.5 million already distributed to projects on unemployment insurance, benefits access, and state AI pilots—and more funds on the way via a second call for proposals.
Eight projects and $8.5 million is real money. It is also a rounding error against the total public sector dollars that could be invested in AI applications with public value at their center. If “AI for the public” is going to be more than a handful of grant-funded pilots, it needs to become a default design question inside agencies at every level—not a side project that outside funders have to underwrite.
The bar that actually matters
If an AI system catches more fraud but also wrongly cuts off eligible people from benefits, that system has failed—even if the fraud numbers look great on a dashboard. The measure of success for government AI should not be how many use cases an agency can list. It should be whether the people on the other end of these systems receive faster, fairer, and more reliable access to the services they are eligible for. That also means we need smart, targeted public investment in these kinds of solutions.
The federal government has spent several years building inventories of AI systems. State and local governments now have a chance to measure what matters: faster approvals, fewer wrongful denials, shorter wait times, and less burden on the people the government is supposed to serve.