What the FDA Is Actually Building With AI: Inside the FY25 HHS Use Case Inventory

Under Executive Order 13960, every federal agency has to publish an inventory of how it uses AI. HHS just filed its FY25 return. It is one of the clearest windows we have into how the government's largest health agency is deploying machine learning, not in press releases, but in a line-item spreadsheet.

The numbers

HHS reported 446 AI use cases in FY25. The biggest builders are NIH with 124 and CDC with 103. CMS follows with 73 spread thin across eleven offices. The FDA is fourth at 67, but 40 of those sit in a single center, CDER.

By maturity:

• 167 already deployed

• 142 pre-deployment

• 88 pilots

• 49 retired

By topic, the largest buckets are:

• Administrative functions: 117

• Health and medical: 89

• Information technology: 87

• Science: 48

The high-impact count is the number that stopped us. Fourteen use cases were flagged as potentially high-impact. Thirteen were reviewed and reclassified. Exactly one, out of 446, carries the label today. Most of this is operational plumbing, not headline autonomy.

Source: the FY25 HHS AI Use Case Inventory. The full inventory CSV is downloadable.

What the FDA is actually building

Strip out the administrative tools and the FDA's 67 entries start to look like one project. The agency is using AI to structure unstructured regulatory records and score facilities against them.

Tool	Center	AI type	Status	What it does
356H ML Facility Supply Chain Role Classification	CDER	Predictive ML	Deployed	Classifies facilities by their role in the supply chain
FAR-based Facility Signal Detection Tool	CDER	NLP	Deployed	Mines Field Alert Reports for facility-level signals and clusters
Risk-based FAR Review and Decision Support*	CDER	Predictive ML	Deployed	Prioritizes which Field Alert Reports get attention first
Quality Surveillance Dashboard	CDER	NLP	Deployed	Surfaces quality signals across the manufacturing base
DMF (Drug Master File) Facilities	CDER	NLP	Deployed	Extracts facility identity from Drug Master Files
MedWatch Dashboard	CDER	NLP	Deployed	Flags product risk signals from MedWatch reports using time series analysis and topic modeling
Drug Shortage Predictive Model*	CDER	Predictive ML	Pilot	Forecasts shortages before they hit
LLM-Assisted VAERS Analyses	CBER	Generative AI	Pre-deployment	Ad hoc querying of vaccine adverse event reports
Supply Chain Resilience Program	CDRH	Predictive ML	Deployed	Models supply chain risk for medical devices

*Name not independently confirmed against the FY25 CSV this pass, worth a final check before publishing.

Several of these carry a "Renamed" flag in the file, meaning they existed in FY24 under different names. Post-market Surveillance Reports Signal Detection became the FAR-based Facility Signal Detection Tool. These are not new starts, they are programs that have been running long enough to get rebranded.

Why this matters for industry

Not all of the FDA's 67 use cases are aimed at industry. Plenty of it is internal review and casework routing, the agency doing its own paperwork faster. What's left once you set that aside is the part that matters from the outside: tools built to structure the enforcement and inspection record. You can't rank a facility's risk or catch a signal cluster until the underlying documents are queryable, and that's exactly what most of this list is for.

Read the tool list again as a supplier or a quality lead. This is the regulator building the same risk infrastructure a quality team would: pulling facility identity out of documents, ranking facilities against each other, flagging quality signals across the manufacturing base.

That is the same discipline KeyPedia gives industry, pointed the other direction.

And it is not a side project. Brookings estimates the potential value of HHS's AI contracts grew from $27 million in 2024 to $138 million in 2026, a 448% increase in two years.

If the FDA is scoring your facility's signal, you should be able to see the same picture first. Structuring the enforcement and inspection record, ranking facilities by their real history, predicting the observations most likely to apply. Most of what is on that list is already deployed, not on a roadmap.

The gap is not whether this analysis happens. It is who sees it before an inspection.

---

Want the same view of your own facilities? Open KeyPedia or book a demo.