Before I Begin: The views represented here are my own and have never been views of my past employer(s). Last Updated: June 2026

The 2025 Federal AI Use Cases Inventory: What Changed

OMB dropped the 2025 Federal Agency AI Use Case Inventory recently, and the headline is 3,611 use cases across 56 agencies - up from 2,133 across 41 agencies in 2024. That’s 15 more agencies and roughly 1,500 more reported use cases year over year.

I dug into the 2024 inventory when it came out. Curious to see what actually moved.

The Big Picture

Some of the jump is real growth. Some of it is better reporting - more agencies in the pool and more consistent definitions of what gets counted. I’d be careful reading the 69% increase as 69% more AI running in government. It’s both of those things mixed together and there’s no clean way to separate them.

Who’s Leading

Department of Veterans Affairs is the standout - and more sharply than last year. The VA reports more than 200 high-impact AI use cases out of 445 total. Nearly half the high-impact inventory is at one agency.

High-impact means the system affects rights, safety, or consequential decisions about individuals. The VA running more of those than everyone else by a wide margin makes sense given their mission - but it also means a lot of the headline numbers are riding on how well one agency is actually implementing the oversight requirements that come with that designation.

The overall use case leaders are HHS, NASA, VA, DOE, and DOJ. HHS and VA holding their spots tracks. NASA showing up in the top five is new from 2024 - I’m curious what’s driving that and will probably look closer separately.

The COTS Picture

The consolidated commercial-off-the-shelf dataset covers roughly 570 entries across 47 agencies. The short version: Microsoft won.

• Some version of Copilot shows up in 102 use cases. Microsoft 365, Teams, and Azure fill out the rest of the top slots. If you want to understand what “federal AI adoption” actually looks like day-to-day, the honest answer is: Microsoft’s AI features getting switched on inside tools agencies were already paying for.
• GitHub Copilot is tracked separately from M365 Copilot, which is interesting - code generation is getting its own line item.
• ChatGPT/OpenAI appears, though less than the public attention around it would suggest.
• About 310 of the ~570 COTS entries are active. The other ~260 are inactive - licenses that aren’t being used or tools that got discontinued.

The Microsoft dominance doesn’t surprise me. Agencies already have enterprise agreements in place. Turning on Copilot inside an existing M365 environment clears a much lower procurement bar than standing up something new.

What the Individual Use Cases Show

The individually reported data is messier and more interesting. A few things jumped out:

The “human-in-the-loop” language is everywhere. Nearly every high-impact entry is documented with something like “AI does not serve as the principal basis for decisions” or “human review required.” That’s the language you’re supposed to use. Whether the review process actually holds up at volume is a different question - I wrote about that here.

The data governance gaps from 2024 are still there. Empty PIA URLs, incomplete impact assessment fields, missing appeal process documentation. Agencies are getting the use cases logged, but the compliance documentation around them is still uneven. OMB granted extensions for risk management requirements in 2024 - some of those gaps presumably haven’t closed.

Vendor dependency is heavy. NEC, Clearview AI, Dataminr, and others show up repeatedly across DHS and DOJ entries. That raises a practical question that nobody seems to have a clean answer to yet: when a vendor silently updates their model weights, does the agency’s ATO still cover the system? Most current interpretations say yes. I’m not sure that’s the right answer, but that’s a longer post.

My Take

Going from 41 to 56 agencies reporting is real progress - more of the government is at least tracking what it’s doing with AI. And 1,800+ deployed or piloted systems is a meaningful operational number, not just a planning document.

But two things I keep coming back to:

The VA concentration. Nearly half of all high-impact use cases at one agency means the overall “high-impact AI” story in this inventory is really a VA story. That’s not inherently bad - the VA has real scale and has been investing in this longer than most. But if you’re reading the 445 high-impact figure as evidence of broad federal AI maturity, it isn’t quite that.

What 3,611 use cases actually includes. 102 of them are Microsoft Copilot. A lot more involve meeting transcription, document drafting, and internal search. Those tools can be genuinely useful, but they’re a different category than the rights-and-safety-impacting applications. The inventory doesn’t make that easy to see at a glance.

What Would Make This More Useful

Same ask as last year: cleaner status breakdowns. The inventory distinguishes between initiated, in development, and operational - but you have to dig to find it and the definitions aren’t applied consistently. A clear deployed vs. not-deployed split upfront would make these numbers a lot more useful.

And at some point I’d love to see outcome data alongside input data. Not just “we deployed AI for X” but what changed - processing time, accuracy, error rate, something. Right now the inventory tells you what agencies are using, not whether it’s working. For a government accountability document, that’s a real gap.

Further Reading

• 2025 Federal Agency AI Use Case Inventory on GitHub
• Agencies report over 3,000 AI use cases in 2025 - Nextgov/FCW
• 2024 Federal AI Use Cases Inventory review - my notes from last year
• ATO Wasn’t Built for Systems That Change at Inference Time - on the vendor model update question