Federal AI Intelligence · U.S. Government AI Observatory
Custom inventory report
Filter: AI type = generative; testing = affirmative
Key findings
- 33 use-case records from 7 agencies: 11 deployed (33%), 8 pilots, 13 in development or acquisition, 1 retired, 0 with no stage reported.
- 3 records are designated high-impact; 3 of them (100%) answered at least one minimum-practice question.
- SBA, DOE, SSA report 29 of 33 use cases (88%).
- 100% of reported deployed use cases (11 of 11) are classified as generative or agentic AI.
- Of 2 deployed high-impact use cases, 1 report a completed AI impact assessment and 0 leave every minimum-practice field blank.
- Interpretation: Blank practice fields are missing evidence, not evidence that a safeguard is absent. They mark where follow-up questions to the agency are most useful.
- Sourcing is reported for 100% of use cases. Where reported, 36% involve in-house development.
Lifecycle
| Stage | Records | Share |
|---|---|---|
| Deployed | 11 | 33% |
| Pilot | 8 | 24% |
| Pre-deployment | 13 | 39% |
| Retired | 1 | 3% |
AI technology
| Classification | Records | Share | Deployed |
|---|---|---|---|
| Generative AI | 33 | 100% | 11 |
Agencies
| Agency | Records | Deployed | High-impact | Completeness |
|---|---|---|---|---|
| SBA · Small Business Administration | 14 | 0 | 0 | 100% |
| DOE · Department of Energy | 12 | 7 | 0 | 100% |
| SSA · Social Security Administration | 3 | 2 | 2 | 100% |
| FERC · Federal Energy Regulatory Commission | 1 | 0 | 0 | 100% |
| EAC · Election Assistance Commission | 1 | 0 | 0 | 92% |
| TREAS · Department of the Treasury | 1 | 1 | 0 | 100% |
| STATE · Department of State | 1 | 1 | 1 | 100% |
High-impact designation
| Designation | Records |
|---|---|
| High-impact | 3 |
| Presumed high-impact, determined not | 1 |
| Not high-impact | 29 |
Governance evidence: use cases designated high-impact
| Practice | In place | In progress | Not in place | N/A or precluded | Other | No answer | In scope |
|---|---|---|---|---|---|---|---|
| Pre-deployment testing | 3 | 0 | 0 | 0 | 0 | 0 | 3 |
| AI impact assessment | 2 | 1 | 0 | 0 | 0 | 0 | 3 |
| Independent review | 2 | 1 | 0 | 0 | 0 | 0 | 3 |
| Ongoing monitoring | 2 | 1 | 0 | 0 | 0 | 0 | 3 |
| Operator training | 3 | 0 | 0 | 0 | 0 | 0 | 3 |
| Fail-safe | 3 | 0 | 0 | 0 | 0 | 0 | 3 |
| Appeal process | 0 | 0 | 0 | 3 | 0 | 0 | 3 |
| User and public consultation | 3 | 0 | 0 | 0 | 0 | 0 | 3 |
"In place" means the agency reported the practice as established. "No answer" means the field was blank; it says nothing about whether the practice exists.
Method
All figures are computed directly from the normalised inventory records. Category labels are normalised from the source answer options; original values are kept on each record. Governance answers are classified as in place, in progress, not in place, not applicable, precluded, waived, other or no answer, and blanks are never read as "no". See the Methodology page for full rules.
Limitations
- Figures are what agencies reported in the 2025 inventory. They are not verified, and registration of use cases is not complete for every agency.
- A blank answer is missing evidence, not evidence that a practice or safeguard is absent.
- Counts are use-case records. A larger portfolio does not indicate better performance or greater AI maturity, and agencies differ in how they split systems into use cases.
- The inventory has no budget, cost, accuracy or outcome data. None is shown or inferred here.
- This is a disclosure analysis. It is not an audit, a compliance determination or an official government assessment.
Source records
All 33 records.
Prepared with Federal AI Intelligence by Irakli Petriashvili. Source: https://github.com/ombegov/2025-Federal-Agency-AI-Use-Case-Inventory