The Federal AI Use Case Inventory
Forty-one US federal agencies are required to publish what AI they run. They report 3,611 systems. 445 of those are designated high-impact — the designation that triggers an obligation to test the system, assess its effects on people's rights, review it independently, monitor it, and give anyone it decides against a way to appeal. On every one of those safeguards, roughly seven in ten agencies have said nothing at all.
- Reported AI systems
- 3,611
- Designated high-impact
- 445
- Agencies reporting
- 41
- With an appeal process
- 16
Every value here is an agency describing its own systems. This register records what agencies did and did not claim — it is not an audit of what they actually did. A blank safeguard field is published as not stated and never as a “no”: an agency that stayed silent has not denied doing the work, and the distinction is the whole point of publishing the column. Withheld entries (24) are systems an agency declined to detail, mostly citing disclosure risk or law. To dispute or correct an entry, contact us. Built under our data standards.
What the safeguards look like
These eight obligations attach to the 445 high-impact systems. The source is not a yes/no field — agencies answer in their own vocabulary (CAIO Review, Appeal Precluded by Law, Yes – Monitoring Established), so each answer is normalised to a comparable state with the agency's own wording kept alongside in the JSON.
| Safeguard | Established | In progress | Not applicable | Not stated | Silent |
|---|---|---|---|---|---|
| Pre-deployment testing | 45 | 82 | — | 318 | 71.5% |
| Impact assessment | 37 | 91 | — | 317 | 71.2% |
| Independent review | 37 | 90 | — | 318 | 71.5% |
| Ongoing monitoring | 40 | 87 | — | 318 | 71.5% |
| Operator training | 45 | 82 | — | 318 | 71.5% |
| Human failsafe | 37 | 81 | 10 | 317 | 71.2% |
| Appeal process | 16 | 78 | 34 | 317 | 71.2% |
| Public consultation | 47 | 81 | — | 317 | 71.2% |
Read the appeal row first. Of 445 systems designated as materially affecting people's rights, 16 report an established way to appeal the decision.
Which agencies
Agencies with at least one high-impact system, by high-impact count. The Department of Veterans Affairs reports the most, followed by Justice and Homeland Security. 130 high-impact systems sit in the law enforcement topic area.
| Agency | High-impact | All reported | Withheld |
|---|---|---|---|
| Department of Veterans Affairs | 215 | 367 | — |
| Department of Justice | 114 | 314 | — |
| Department of Homeland Security | 55 | 238 | — |
| Department of Energy | 29 | 340 | — |
| Social Security Administration | 9 | 33 | — |
| National Credit Union Administration | 5 | 6 | — |
| Department of Agriculture | 4 | 162 | — |
| Department of the Treasury | 4 | 129 | — |
| Department of State | 3 | 60 | — |
| Department of Health and Human Services | 2 | 447 | — |
| Environmental Protection Agency | 2 | 29 | 3 |
| National Aeronautics and Space Administration | 1 | 425 | 15 |
Where it is deployed
| Topic area | Systems |
|---|---|
| Science | 758 |
| Admin Functions | 424 |
| IT | 394 |
| Law Enforcement | 285 |
| Health and Medical | 272 |
| Other | 173 |
| Service Delivery | 150 |
| Energy and Environment | 131 |
| Procurement and Finance Mgmt | 90 |
| Transportation | 79 |
| Benefits Processing | 56 |
| Cybersecurity | 53 |
Development stage across all 3,611 reported systems: Pre-deployment 1,479 · Deployed 1,040 · Pilot 440 · Retired 314.
Source and method
Published by the US Office of Management and Budget, Office of the Federal CIO, consolidating the inventories each agency must post under EO 13960 Sec. 5; Advancing American AI Act; OMB M-25-21. The register here is rebuilt from OMB's own machine-readable file; the repository is on GitHub.
Licence. US federal government work, public domain under 17 U.S.C. 105 (by authorship — the publisher ships no explicit LICENSE file). Stated plainly because it matters: OMB ships no LICENSE file and the GitHub API reports no licence, so the public-domain basis is federal authorship under 17 U.S.C. 105 rather than an explicit dedication by the publisher.
Personal data. None. The source carried one contaminated column — agency contact addresses, mostly named federal scientists — and it is dropped in full rather than filtered, because it also held a bare personal name with no address at all and a personal consumer address, neither of which a pattern-match would catch. Every remaining field is swept at build time and the build fails closed if anything person-shaped appears. Note that the has_pii field describes whether a system processes personal data; it is a fact about an institution, not about a person.
One counting note. The source designation field has four values, one of which is Presumed High-Impact, but Not High-impact (110 systems). Those are excluded from the 445 count. Any analysis matching on the substring “high-impact” will overcount by exactly that number.
Edition: 2025 consolidated inventory. Rebuilt 2026-08-20. See also the Data Protection and Genetic Privacy registers for what the law permits agencies to do with the data these systems process.