AI in Government: How US Agencies Are Adopting New Tools

For most of the last decade, artificial intelligence in the US government meant isolated projects: a fraud model at one bureau, an image classifier at another, a chatbot answering a narrow set of questions on a benefits website. That has changed. Agencies now buy general purpose generative AI tools across whole workforces, the White House has rewritten the rules to encourage adoption, and the General Services Administration has turned AI into a shared procurement item rather than a bespoke engineering effort.

This report sets out what the federal rules now require, how agencies are actually buying and using the tools, where the evidence of value is strong or weak, and the risks that remain unresolved. It focuses on federal civilian and defence agencies, with a short note on states.

The policy shift: from guardrails first to adoption first

The framework changed in two steps. In October 2023, Executive Order 14110 set out a broad safety agenda, and in March 2024 the Office of Management and Budget issued memorandum M-24-10, which required agencies to name Chief AI Officers, publish inventories of AI use cases and apply minimum risk practices to systems that affect rights or safety.

In January 2025, the new administration revoked that order and replaced it with Executive Order 14179, titled Removing Barriers to American Leadership in Artificial Intelligence. In April 2025 OMB issued two replacement memoranda: M-25-21 on accelerating federal use of AI, and M-25-22 on acquisition. In July 2025 the White House published America's AI Action Plan.

The tone moved decisively toward speed. M-25-21 tells agencies to remove barriers to adoption and to treat AI as a way to deliver services more efficiently. But the structure built in 2024 largely survived. Agencies still need Chief AI Officers, still publish use case inventories, and still apply minimum practices to what the new memo calls high-impact AI: systems whose outputs serve as a principal basis for decisions with legal, material or significant effects on people. The label changed. The core obligation to test, monitor and offer human review for consequential systems did not disappear.

How agencies are buying: the OneGov model

The most visible change is in procurement. In 2025 GSA launched its OneGov strategy, negotiating government-wide terms directly with large software providers instead of leaving each agency to run its own contract. The AI deals drew attention for their price. GSA announced agreements under which OpenAI offered ChatGPT Enterprise to federal agencies for one dollar per agency for a year, followed by similar nominal-price offers from Anthropic for Claude and from Google for Gemini for Government.

GSA also added major model providers to its Multiple Award Schedule, so agencies can buy through an existing contract vehicle, and introduced USAi, a shared platform that lets federal staff try models from several vendors in a government-managed environment. The logic is sensible: evaluating tools once, centrally, is cheaper than having hundreds of offices repeat the same security review.

The pricing deserves scepticism rather than celebration. A one dollar first year is a market entry strategy. The meaningful costs arrive later: renewal pricing, usage-based charges for heavier workloads, integration with agency systems, training and the staff time needed to govern the tools. Agencies that measure value only against a nominal licence fee will understate what they are committing to. The useful question for any programme office is what the second and third years will cost, and how hard it will be to switch.

Defence is moving on a separate track

The Department of Defense runs its own adoption effort through the Chief Digital and Artificial Intelligence Office. In July 2025 the CDAO awarded agreements with a ceiling of up to 200 million dollars each to Anthropic, Google, OpenAI and xAI to develop agentic and frontier AI capabilities for national security work. Ceiling values are maximums, not money spent, and the work involved is largely prototyping. Still, the awards signal that the military intends to work directly with frontier labs rather than wait for capabilities to filter through traditional defence contractors.

What agencies are actually using AI for

A useful public record is the federal 2024 AI use case inventory. Its published summary reports 2,133 AI use cases from 41 agency submissions as of January 23, 2025, and the repository's revised v2 CSV contains 2,133 entries. These are reported use cases, not 2,133 deployed generative AI tools: the dataset includes planned, developing, operational and retired systems, and covers AI more broadly than chatbots. The reporting scope excludes Department of Defense use cases, Intelligence Community systems, National Security Systems and certain other categories. This is a dated inventory snapshot, not a current total for all US government AI activity or a measure of successful adoption.

Internal agency support is a prominent category. Examples include summarising documents, drafting correspondence, searching policy manuals, classifying incoming mail and forms, translating text and supporting software developers. But the inventory also covers health and medicine, law and justice, science and other mission work. Its summary lists 271 use cases at the Department of Health and Human Services and 229 at the Department of Veterans Affairs; those totals should not be read as counts of paperwork tools or generative AI deployments.

A smaller set touches the public directly: fraud detection on payments, triage of claims, prioritisation of inspections and customer service tools. These are the cases where the high-impact rules matter, because an error can delay a benefit, flag an innocent person or skew enforcement toward particular communities.

Where the evidence of value is strong, and where it is thin

The strongest evidence for generative AI productivity comes from controlled studies outside government. Experiments on writing tasks and customer support, such as the study of support agents by Brynjolfsson, Li and Raymond, found meaningful gains concentrated among less experienced workers. Those results transfer reasonably well to government drafting and correspondence work, which looks much like the tasks tested.

Pennsylvania provides a state-government example, not evidence of federal results. Its ChatGPT Enterprise pilot report describes a yearlong programme from March 2024 to March 2025: over 175 employees accessed the tool, and 136 provided direct feedback. Participants estimated an average of 95 minutes saved per day using ChatGPT. The report identifies that figure as an estimate reported in the exit survey, not an independently timed productivity gain; the 136 feedback contributors are not identified as the denominator for that particular estimate. The Office of Administration's announcement associates the savings with writing, research, summarization and IT support. Participants volunteered, and the report explicitly cautions that they were not a representative sample of Commonwealth employees. The result describes perceived savings among pilot participants, not a demonstrated 95-minute gain for every state or federal employee.

The gap matters. Without baselines and measured outcomes, an agency cannot tell whether a tool saves time, shifts work to reviewers, or simply produces more text. Our earlier report on how to use AI to save time at work explains why gains depend heavily on task selection and organisational design.

The oversight record

The Government Accountability Office has reviewed federal AI governance repeatedly. Its December 2023 report on agency implementation of AI requirements found that many agencies had not fully met earlier obligations, including complete inventories and planning, and made dozens of recommendations. GAO has also warned about the federal AI workforce, noting that agencies struggle to hire and keep people with the technical skills to evaluate the systems they buy.

That last point is the quiet constraint on the whole effort. Buying a tool is easy. Knowing whether a model is accurate enough for a benefits decision, whether a vendor's security claims hold, or whether an inventory entry is honest requires skilled staff inside government. Workforce reductions across several agencies in 2025 make that harder, not easier.

Risks that have not been solved

**Accountability for decisions.** When an AI system contributes to a decision about a person, the agency remains legally responsible. Administrative law expects reasons that can be explained and reviewed. A model that produces a recommendation without a traceable rationale fits poorly with that expectation, which is why the high-impact rules require human review and appeal routes.

**Federal records.** Prompts and outputs created in the course of government business can be federal records. The National Archives has issued guidance that records obligations apply to new technologies. Agencies need retention settings, export capabilities and clear rules on what staff may enter, or they risk losing records or exposing them in ways the Freedom of Information Act and the Privacy Act do not anticipate.

**Security and data handling.** Commercial tools must clear FedRAMP authorisation for the data levels involved, and generative systems add new risks such as prompt injection and data leakage through connected tools. The NIST AI Risk Management Framework remains the common reference, but it is voluntary guidance, not a test that a system passes once.

**Vendor concentration.** Government-wide deals with a handful of model providers make adoption fast and also make agencies dependent on those providers. Portability of data, prompts and workflows should be a contract requirement, not an afterthought.

**Bias and error in public-facing systems.** Fraud and eligibility models can produce error rates that fall unevenly across groups. Past automated systems in state unemployment and benefits programmes generated wrongful flags that took years to unwind. The lesson is to test on representative data, publish error rates where possible and keep a fast human appeal route.

A note on states

States are moving in parallel, often with more published detail. Governors in several states have issued executive orders on generative AI, and legislatures have passed laws on government use, deepfakes and automated decisions. The result is uneven: a resident's experience of public sector AI increasingly depends on where they live. For vendors, it means a patchwork of procurement rules and disclosure requirements.

What a well-run agency programme looks like

The agencies most likely to get value share a few habits. They start with internal, low-risk tasks where errors are cheap to catch. They measure time and quality before and after deployment instead of relying on surveys. They keep an honest inventory, including systems that were tried and dropped. They set records and data rules before rollout, not after an incident. They budget for years two and three. And they keep enough technical staff in house to challenge a vendor's claims.

What to watch

Three things will show whether federal adoption is working. First, renewal terms when the nominal-price introductory deals expire, which will reveal the true cost of government-wide AI. Second, the next public use case inventories, and whether they include measured outcomes rather than descriptions. Third, how agencies handle the first serious failures of high-impact systems, because the response to an error will say more about governance than any memo.

Limitations

This report is based on public policy documents, procurement announcements, GAO reports, published inventories and peer-reviewed research. We did not interview agency officials or test the tools described. Use case inventories are agency-reported, include multiple development stages and exclude specified categories; their totals do not measure deployed systems or proven benefits. Contract ceilings are maximum values, not spending. The Pennsylvania time savings are exit-survey estimates from a nonrepresentative volunteer pilot, not independently measured or workforce-wide results. Federal AI policy has changed frequently since 2023 and may change again; readers should check the current text of OMB guidance before relying on any specific requirement.

Correction, October 23, 2026: We checked the linked inventory and Pennsylvania pilot sources. The federal inventory paragraph now cites the repository's January 23, 2025 summary of 2,133 reported use cases and removes the unsupported claim that the total was roughly double the previous year. The Pennsylvania paragraph now gives the March 2024 to March 2025 pilot period, identifies the 95-minute figure as an exit-survey estimate and states the sample limitations. A broken Pennsylvania source link has been replaced with the official announcement and report.