FDA's AI Will See You Now
FDA built the AI inspector faster than the AI rulebook — and MedTech is already feeling the impact
Your next FDA submission has two readers now. The human reviewer assigned to your file, and the AI that reads it before they do, summarizes it, and tells them what they're looking at. On May 6, 2026, FDA's Commissioner acknowledged this from the FDLI stage, attributing facility selection in the new One-Day Inspectional Assessments pilot to AI. The press release that dropped the same day described the same pilot using "risk-based criteria" and never used the word AI. Both were technically true. They were describing two different agencies.
FDA built the AI that decides whether to inspect you before it finalized the rule for the AI you're shipping. That asymmetry, not the rules themselves, is now the structural fact behind MedTech compliance.
I’ve watched FDA roll out new tools before. I’ve never watched the agency build the inspector faster than it built the rulebook. That’s what’s happening right now, and almost nobody in our industry is naming it cleanly.
The 13-month sprint vs. the 16-month stall
Lay the timelines side by side. They tell two different stories.
Internal AI at FDA: Jeremy Walsh hired as the first Chief AI Officer in April 2025. Elsa 1.0 launches agency-wide in June 2025. Agentic AI deployment, including “inspections and compliance” workflow automation, by December 2025. One-Day Inspectional Assessments pilot live across four inspectorates by April 2026. HALO (Harmonized AI & Lifecycle Operations for Data) consolidates 40+ legacy data systems on May 6, 2026. Adoption above 80% across the agency. Some centers above 90%.
External AI policy for industry: the draft AI-Enabled Device Software Functions guidance published January 2025. Comments closed April 2025. As of FY26, the guidance has been demoted to CDRH’s “B-list” with finalization “as resources permit.” Best-case finalization 2027. More likely 2028.
FDA shipped Elsa to 80%+ adoption in roughly 13 months. The parallel AI-DSF rulemaking has spent 16 months in draft with no end visible. Same agency. Same fiscal year. Same talent pool.
Two FDAs, one mic
There are two FDAs operating right now, and you have to plan for both.
The press-release FDA describes “risk-based criteria,” validated workflows, bounded use cases, and human reviewers in every loop. The press-release FDA is careful, lawyered, and (I would argue) slightly behind where the agency actually is.
The conference-stage FDA is something else. At FDLI on May 6, Makary said it plainly: “The one-day inspections are a screening inspection in low risk facilities that our AI is identifying as low risk. The idea is we can do more inspections.” Bloomberg Law and RAPS captured the remark. The press release the same day described the same pilot using “risk-based criteria” and never used the word AI. The verbal version is what the agency’s senior leadership is willing to say to industry’s face.
If you build your quality program around the press-release FDA, you are designing for the wrong inspector. Plan for the conference-stage version.
The hallucinator is the inspector
In July 2025, CNN reported that Elsa was hallucinating studies and misrepresenting research to FDA reviewers. Walsh acknowledged it on the record.
HALO sits underneath that same family of architectures. HALO is now correlating registration records against MDR data, against 483 history, against import alerts, against complaint patterns. The pattern-matching engine reading your submission has already, on the record, generated content that wasn’t there.
This cuts two ways.
For industry, it’s leverage. A submission that is internally consistent (registration matches actual, supplier data matches DHF, MDR patterns match QMS narrative) is defensible against an unreliable correlator. Inconsistency, real or hallucinated, is what a hallucinating AI amplifies. The cost of any contradiction in your submission just went up.
For industry, it’s also risk. Your reviewer doesn’t have to be wrong for the AI feeding them to be wrong. The 483 you eventually receive may be a human-written paragraph based on what an algorithm told a reviewer about your DHF. Closing it out requires rebutting the algorithm’s read, not just the inspector’s.
The model under the hood keeps changing
Elsa originally ran on Anthropic’s Claude. After the February 27, 2026 directive halting federal use of Anthropic, HHS phased Claude out and FDA transitioned Elsa to Google Gemini in March. ChatGPT Enterprise is approved as an alternative for some HHS contexts. A White House workshop is reportedly working on a return path for Anthropic’s newer “Mythos” model.
The directive itself was blocked on March 26 by a preliminary injunction from U.S. District Judge Rita F. Lin in the Northern District of California, who found Anthropic “likely to succeed” on its First Amendment retaliation claim. HHS terminated Anthropic use across the department within days of the original directive and acted on the ban regardless. The legal architecture under the model swap is contested and unfinished.
Watch who is holding the contract. Deloitte built Elsa as a custom retrieval-augmented generation system optimized specifically for Claude’s behavior: architecture, embedding models, vector databases, prompt templates, all tuned to Claude. On April 22, 2026, Deloitte announced an expanded alliance with Google Cloud focused on agentic AI transformation. Three weeks later, HALO launched on Gemini. The firm that built Elsa for Claude got a fresh Gemini partnership immediately before the model swap. Whether that’s coincidence, conflict, or both, it’s the kind of procurement detail that explains how a 30-day model migration actually happens inside a federal agency.
Then there’s the administrative record. Kimberly Chew and Michael Yang at Husch Blackwell argued in March that a submission reviewed partly under Claude-Elsa and partly under Gemini-Elsa can produce inconsistent AI-generated analysis inside FDA’s own decision record. Under the Administrative Procedure Act, internal inconsistencies of that kind are precisely what plaintiffs use to challenge agency decisions as arbitrary and capricious. The reasoning trail Elsa generated on a submission filed in February 2026 was built on Claude. The same submission reviewed in May was built on Gemini. Six months from now, it may be back on Claude.
If you sell into FDA’s AI workflow as a vendor, model-agnostic architecture is now a procurement requirement. If you submit into it as a manufacturer, you cannot assume the underlying model reading your filing is the same one that read your last one. And the lawyers in the room should be tracking which model touched which page of your administrative record.
The B-list isn’t an oversight. It’s the pattern.
The reasons the AI-Enabled Device Software Functions guidance is sitting on the B-list aren’t a mystery. The 10-for-1 deregulatory executive order has chilled new guidance publication agency-wide. CDRH lost AI/ML staff in the 2025 reduction in force. Comments on the draft (especially on generative AI and foundation models, which the draft barely addressed) require real rework. The volume is real.
Here’s the uncomfortable part.
The agency that should be writing rules for AI in your device is the same agency building the AI that inspects your facility. The talent, the contracts, the political bandwidth, the press cycles: all of it is currently being absorbed by the internal build. The B-list isn’t a parking lot. It’s where the rule-writing function got crowded out by the surveillance-building function.
Don’t expect that to reverse. Plan for years of continued uncertainty on AI device guidance, while FDA’s internal AI surface gets sharper every quarter.
What this means for the next 12 to 18 months
If you’re a VP of Quality, three things have changed.
First: your submission audience is now an AI-augmented reviewer, not a reviewer. Design for cross-document consistency the way you used to design for pre-submission meetings. The AI doesn’t read like a human and doesn’t forgive like a human. Internal contradictions a senior reviewer would have flagged as “we’ll discuss” now generate inconsistencies that surface as written questions. The Predetermined Change Control Plan guidance (final, August 2025) is the most stable AI-related document CDRH has shipped. Roughly 10% of 2025 AI clearances included a PCCP. That number should be higher.
Second: hospital procurement is the actual regulatory floor now. The January 2026 CDS and General Wellness revisions moved many AI tools out of FDA’s regulatory ambit. Joint Commission and CHAI moved into the gap with November 2025 responsible-use guidance. Hospitals will use those frameworks for AI tool diligence regardless of FDA jurisdiction. Quality teams that map their products against CHAI alongside FDA expectations will sell faster.
Third: ISO/IEC 42001 is the shadow QMS. It will start showing up as a customer requirement before FDA references it formally. Auditors who can speak both ISO 13485 and ISO/IEC 42001 will have a 12-to-24-month head start.
For investors evaluating MedTech compliance bets, the read is sharper. Generic AI features in compliance platforms commoditize fast. Vendors that mirror FDA’s AI surface, the correlation patterns HALO is now running, win. Auditable reasoning trails and human-in-command architectures (not wrapper-on-wrapper LLM stacks) command a defensible premium. And the B-list status of AI-DSF helps the current vendor cohort, not hurts it: continued draft-state means firms keep paying for clarity FDA hasn’t shipped.
The next 483 you receive won’t be written by a human reading your DHF. It will be written by a human reading what an algorithm wrote about your DHF. Plan accordingly.