The AI expansion in government in 2026 comes down to two numbers. US federal agencies reported 3,611 AI use cases in their 2025 inventory, more than double the year before, and GSA's government-wide deals have put AI tools in front of roughly 3.5 million federal employees. The tools got cheap and plentiful. The people who check their output got scarcer.
That gap is where the real story sits. Below are the numbers first, then the places where governance is straining, and a checklist for teams working inside it.
What the AI expansion in government in 2026 looks like
Every civilian federal agency has to publish a list of the ways it uses AI, once a year. The 2025 list, which the Office of Management and Budget (OMB) completed in April 2026, is the best public map of where government AI actually runs.
It shows 3,611 use cases across 56 agencies, up 105% from 1,757 in 2024. Health and Human Services drove the growth in both years. The five largest reporters:
| Agency | 2025 use cases |
|---|---|
| Health and Human Services | 447 (271 in 2024) |
| NASA | 425 |
| Veterans Affairs | 367 |
| Energy | 340 |
| Justice | 314 |
The work itself is unglamorous. Summarising and analysing text tops the list, followed by pattern recognition and chatbots among others. Agencies flagged 445 entries as high-impact, meaning a wrong output could affect someone's rights or safety. Veterans Affairs holds 215 of those, including summaries of clinical data and a voice bot for its call centres.
Each entry also names the software behind it, which turns the inventory into a market map. If you sell to government, it tells you which vendors already sit inside each agency. If you work in government, it tells you which of your systems a journalist can look up.
One caveat changes how you read the total. Use cases at the Pentagon and in the intelligence community are exempt from reporting, and other agencies can leave out national security uses. The inventory is a census that lets the two biggest households skip the form.
AI enterprise governance when every employee gets a chatbot
The cheapest part of government AI is now the model itself. The General Services Administration (GSA) negotiates one deal for the whole government under a program called OneGov, and agencies order against it without running their own procurement.
GSA says OneGov has saved about $1.68 billion since it started, with about $1.4 billion of that from AI agreements. Its newest deal, with OpenAI, takes the next step. From October 1, 2026, GSA's OpenAI agreement runs for 27 months with these terms:
- A 50% discount on token-based usage across ChatGPT models, including FedRAMP-authorised environments.
- Pay-as-you-go terms: no platform fee and no spend commitment.
- Access for every level of government, from federal agencies down to tribal governments.
Token-based pricing means the agency pays per token, a chunk of text the model reads or writes, so the bill follows usage. Fifty percent off with no minimum is the kind of sentence that has launched many unplanned projects.
Here is the governance consequence. Under a seat licence, the contracting officer was the control point: someone had to justify every seat. With no minimum and no commitment, that control moves inside the agency. AI enterprise governance, in this setting, means one set of rules covering every team that touches the same models, with quotas and usage logs among others. Agencies that skip this step will find out what they built when the invoice arrives.
GSA has started building part of that layer. Agencies test models on its USAi platform before buying, and the same platform lets them re-run evaluations later to check for drift. What GSA cannot yet supply is a return figure. Its chief AI officer said in May that working out the return on investment is still the big open question, while adoption keeps climbing.
AI governance leadership with fewer people in the loop
Human in the loop means a person checks the AI's output before it counts. It is the default safeguard in most AI policies, and it costs staff time that many agencies no longer have.
At a summit in October, a USDA IT official called the human-in-the-loop model untenable after staffing and contract cuts. His alternative is agents with minimal privileges, watched through continuous network monitoring and a strong risk framework. A Customs and Border Protection adviser on the same panel argued the opposite and wanted people kept in the loop. Minimal privileges is the agent version of a visitor badge with a chaperone.
Both officials are right about different tasks. A State Department official at the event estimated that StateChat, its internal chatbot, saves staff about two hours a week. Nobody needs a reviewer for a draft email summary. A benefits decision is another matter, as the analysis of AI caseworker policy assistants showed: one wrong rule, applied automatically, becomes thousands of wrong letters.
Other agencies are moving the safeguard earlier. NIH described building a deployment layer where data and spending policies attach to every AI application, after a year in which staff mostly used AI for personal productivity. The State Department now gives new foreign service officers a training module on AI rules in their first days.
So the leadership job is triage. Rank each use case by the cost of a wrong answer and spend scarce reviewers at the top of that list. Where a person can no longer read every output, build the review into the system:
- Sampled review, where a person checks a random slice of outputs, such as 5% a week.
- Logging of every input and output, so a complaint can be traced.
- Alerts when error reports or appeal rates rise.
AI governance failures to watch as agencies scale
Fast adoption tends to fail in predictable ways. Four show up in this year's data.
- Counts that hide the riskiest systems. The public inventory excludes defence and intelligence uses, which carry the highest stakes. Internal registers have to cover what the public one leaves out.
- Vendor concentration. 102 reported use cases run on a version of Microsoft Copilot. When one supplier sits under that many systems, a single policy change hits all of them. Agencies learned this in 2026, when the administration ordered them to stop using Anthropic's Claude, which appeared in 25 reported use cases, after a dispute with the Pentagon.
- Drift with no owner. Model drift means a model's behaviour changes over time, often after a vendor update. If nobody re-tests after a new version ships, a working system can degrade quietly.
- Measuring adoption in place of outcomes. Orders placed and users enabled are easy to report. Hours saved and error rates are harder, and they are the numbers that justify the spend. The same pattern sits behind the AI transformation governance problem in the private sector.
None of these failures is exotic. Each one is a governance gap that gets wider as the use case count grows.
An AI governance and compliance checklist for public teams
If your team works inside this expansion, these steps cover the next quarter. They apply to a state agency as much as a federal one, since the new OneGov terms reach both.
- Keep an internal inventory that includes every use case, published or exempt.
- Tag each use case by the cost of a wrong answer, and assign reviewer coverage from the top down.
- Set per-team budgets and usage alerts on consumption accounts before switching them on.
- Re-run a fixed evaluation set after every model version change.
- Write an exit plan: a second model tested and prompts that can move with you.
- Report outcomes, such as hours saved and error rates, next to adoption numbers.
The first two items take a week of spreadsheet work. The fourth is the one teams postpone, and it is the one that catches a silent model update before the public does.
For the cost and production side of this list, AI Systems in Business covers total cost of ownership and taking a pilot into production. If your use cases involve agents acting on their own, Agentic AI in Business deals with setting boundaries of autonomy. For the regulatory calendar, the monthly AI governance news roundup tracks what changed.
The government's AI census will keep growing. Make sure your count also records who checks the work.