Is FDA using AI to pick your inspection?
FDA says it uses its Elsa AI tool to find inspection targets. An inspection run by AI is still theoretical. I would check the records FDA already holds.
FDA and AI in inspections: what is confirmed
Status as of October 25, 2026. Events confirmed by FDA, then planned, then only reported.
Confirmed by FDA
- June 2, 2025FDA launches Elsa and says it already helps identify high priority inspection targets.FDA
- December 1, 2025FDA announces optional agentic AI, with inspections and compliance among its uses.FDA
- April 2026The one day inspectional assessment pilot begins.FDA, May 6, 2026 release
- May 6, 2026FDA launches Elsa 4.0 and announces HALO, which consolidates more than 40 data sources.FDA
Planned or announced
- Through fiscal year 2026The pilot is set to run through fiscal year 2026. No extension had been announced as of October 25, 2026.FDA
Only reported
- May 6, 2026Commissioner Makary says AI is identifying lower risk facilities for the pilot.Bloomberg Law
- May 7, 2026Musser describes difficulty integrating AI information during an inspection.Regulatory Affairs Professionals Society
- May 12, 2026Musser calls an inspection run entirely by AI theoretical.Skadden
- October 21, 2026A conference talk claims regulators use AI to check batch records before visits.GeneOnline
The FDA pilot release does not identify AI as the selection method. Sources: FDA press announcements, Bloomberg Law, Regulatory Affairs Professionals Society, Skadden and GeneOnline.
Yes, in part. The U.S. Food and Drug Administration (FDA) said on June 2, 2025 that it already used Elsa, its artificial intelligence (AI) tool, to identify high priority inspection targets. An inspection run entirely by AI remained theoretical according to Steven Musser, FDA’s associate commissioner for Human Food Research in the Human Foods Program, in May 2026 remarks reported by Skadden on May 12.
If you lead generative AI programs, FDA’s use of the Elsa language model to identify inspection targets is a reason to include record quality and human review in your work with quality assurance teams on inspection readiness.
My advice is to start by checking the records the agency already holds. I would review earlier inspection commitments and the evidence supporting their closure.
What FDA confirms about its Elsa AI tool
Elsa’s launch and wider deployment
On June 2, 2025, FDA described Elsa as a large language model tool for tasks such as summarizing documents, hosted in GovCloud. The same announcement included clinical protocol review among its uses.
On December 1, 2025, FDA announced optional agentic AI, meaning systems that carry out tasks across multiple steps, including inspections and compliance. It described human oversight and said more than 70% of staff voluntarily used Elsa, according to internal data.
Elsa 4.0 and HALO
On May 6, 2026, FDA launched Elsa 4.0 and announced HALO, Harmonized AI & Lifecycle Operations for Data. HALO consolidated more than 40 application and submission data sources across FDA centers. Integration with Elsa had begun.
That release describes a Google Cloud environment rated High under the Federal Risk and Authorization Management Program, or FedRAMP. It also says human subject matter experts verify inputs and check the analysis through to the use of its outputs.
The one day assessment pilot
FDA’s May 6 announcement of the pilot says it began in April 2026 and would run through fiscal year 2026. Selection considers product type and prior inspection outcomes, along with how a facility operates. Investigators can expand a visit’s scope or duration. Higher risk or more complex facilities are excluded. The release does not identify AI as the selection method.
As of October 25, 2026, FDA has not announced an extension beyond fiscal year 2026.
What officials reportedly said about inspection selection
At the Food and Drug Law Institute conference on May 6, 2026, Commissioner Marty Makary said AI was identifying lower risk facilities for the pilot, according to Bloomberg Law’s report.
According to Skadden’s May 12 account, Musser described an entirely AI driven inspection as theoretical and outlined uses for analysis before and after inspections.
The Regulatory Affairs Professionals Society reported on May 7 that Musser described difficulty integrating AI information during an inspection. I would be careful about treating his remarks from the Human Foods Program as a drug manufacturing inspection policy.
My reading of the record access claims
FDA’s HALO announcement describes application and submission data. It does not say FDA uses AI to check a site’s batch records before a visit. Skadden separately reports that FDA was applying AI to data already submitted to it, including adverse event reports.
GeneOnline’s October 21 account of Frances Zipp’s Milan conference talk claims regulators use AI to check batch records before visits. I treat that as unconfirmed conference commentary. I found no support in the FDA releases cited here for the claim that FDA uses AI to check executed batch records before a visit.
I would keep these evidence levels separate:
| Claim | Status | Source |
|---|---|---|
| Elsa helps identify inspection targets | FDA confirmed | June 2025 |
| HALO combines agency data sources | FDA confirmed | May 2026 |
| Pilot sites selected using risk criteria | FDA confirmed | Pilot release |
| AI selects lower risk pilot sites | Official's reported remark | Bloomberg Law |
| Entire inspection run by AI | Reported as theoretical | Skadden |
| AI checks batch records before a visit | Unconfirmed conference commentary | Conference account |
What I would ask quality assurance teams to check now
These are my recommendations for quality assurance teams. I would use them to improve record quality without promising a change in inspection probability.
-
Compare filed site information with actual operations. FDA identifies discrepancies between registered and actual operations as useful information for future risk models.
-
Review earlier inspection commitments and find the evidence supporting closure. FDA’s published drug surveillance selection policy includes compliance history among its risk factors. Keep each commitment linked to its supporting records.
-
Ask someone unfamiliar with an investigation to read it and trace the conclusion to the evidence. Record what needs explanation and fix the underlying record through the approved process. For records involving language models, see my writing on an audit trail for model calls.
-
Rehearse retrieving the site summary and supporting records promptly. Prepare for an investigator to extend a screening visit when concerns arise, as the pilot release allows.
-
Check human review of any procedures drafted with AI. My warning letter note is a starting point for that review. Also check agency correspondence against the underlying evidence, following Skadden’s advice.
Frequently asked questions
What does the FDA Elsa AI tool name stand for?
FDA’s June 2025 launch announcement calls the tool Elsa without expanding the name as an acronym. It describes a large language model tool that assists employees with reading and summarizing. I would use Elsa as a product name and retain FDA’s description of what it does.
What can FDA Elsa 4.0 do with scanned documents?
FDA lists optical character recognition, which converts scanned documents and images into searchable text, among Elsa 4.0’s features. It also lists document generation and improved search across large repositories. Those are tool capabilities. The announcement does not document their use on a particular site’s records.
Does connecting HALO mean Elsa trains on company data?
FDA said in December 2025 that models do not train on input data or information submitted by regulated industry. I would distinguish that training policy from a tool’s ability to query records. For a confidentiality review, I would ask about data access permissions separately.
Is FDA still doing inspections, and what are the four types?
Yes. FDA describes four basic types of inspections: surveillance, follow up, for cause and application based inspections. Surveillance examines ongoing manufacturing practices. Follow up checks corrections after earlier findings. For cause inspections address specific concerns, while application based inspections assess facilities before application approval.
What do reports about FDA Elsa hallucinations establish?
Skadden’s May 2026 report describes outputs varying between queries and Musser’s warning that AI does not assess the quality or nuance of scientific evidence. It gives no Elsa hallucination rate for inspection decisions. I would require the underlying evidence before accepting any disputed conclusion.
Want to prepare for an AI-informed inspection? That is the work I do.