FROM THE EDITOR
FDA issued its first warning letter citing AI overreliance as a cGMP compliance failure in April 2026. The letter did not name the manufacturer. It did not need to. Every pharma company using AI in any compliance function read it the same way: this is the theory inspectors are now carrying into facilities.
For two years I have watched legal functions treat AI compliance as an IT problem. The vendor selects the tool, procurement reviews the contract, and compliance writes a policy. Nobody owns the regulatory exposure in that sequence. The warning letter tells you who FDA thinks owns it: the quality unit, the company, and by extension the legal function that approved the contracts and the oversight framework.
This issue is about the specific legal work the warning letter starts. The renegotiation it forces. The documentation gap it exposes. And what needs to be in place before the next inspection, not after the next 483.
The AI system failed. That is not the finding. The finding is that the company let it.
— John LaRocca
Founding Editor | LaRocca Advisory Group
A drug manufacturer used an AI tool to generate its drug product specifications, procedures, and master production records. FDA inspected and found cGMP violations. The company’s defense: the AI failed to flag the gaps. FDA’s response: that is not a defense.
This is the first FDA warning letter to cite a company’s use of AI as a compliance instrument — not whether the AI itself is a regulated device, but whether the company maintained adequate human oversight of what the AI produced. The distinction matters because it applies to every pharma company using AI in any compliance function, right now.
DOMAIN 1 WHAT THE WARNING LETTER ACTUALLY SAID
The company’s quality unit failed to adequately review and validate AI-generated documents for accuracy and cGMP compliance before those documents governed manufacturing. When inspectors found the gaps, company representatives pointed to the AI system’s failure to flag missing requirements.
FDA’s position: the quality unit’s obligation to review and validate is non-delegable. It does not transfer to the AI vendor. It does not transfer to the algorithm. It stays with the company.
The practical problem is that AI-generated compliance documents look authoritative. They are well-formatted, internally consistent, and written in regulatory language. A quality professional reviewing a document the AI produced has to know what’s missing — which is exactly what the AI was supposed to help with. When the AI fails to flag a gap and the reviewer doesn’t catch it either, the document enters the quality system undetected.
That is the failure pattern FDA just named and cited.
FDA doesn’t care what the vendor promised. The company’s non-delegable obligation to review stays with the company, not the algorithm.
DOMAIN 2 THE VENDOR AGREEMENT GAP
Most enterprise AI agreements disclaim fitness for any particular purpose and cap liability at annual fees. That disclaimer protects the vendor in a contract dispute. It does not protect the pharmaceutical company in the regulatory relationship.
The vendor’s contract excludes liability for the failure pattern FDA just described. The pharmaceutical company bears the entire regulatory consequence. The gap is now a fact pattern: AI-generated compliance document, inadequate human review, cGMP violation, warning letter. Your vendor’s indemnification section covers none of that sequence.
If your AI vendor agreements were drafted before March 2025 — before FDA and EMA issued their joint guidance on AI in regulated manufacturing — they were almost certainly written without the indemnification language your company now needs. The renegotiation this warning letter forces is not optional.
DOMAIN 3 HUMAN OVERSIGHT PROTOCOL
The company in this warning letter probably had people who reviewed the AI outputs. They either didn’t catch what was missing, or they didn’t document that they looked. FDA doesn’t distinguish between them.
A policy saying “employees must review AI-generated documents” is not sufficient. The quality system needs to specify who reviews, what the review must cover, how the reviewer certifies the review was substantive, and what happens when an AI-generated claim cannot be independently verified. That protocol needs to exist before the next inspection.
FDA inspectors now have a named theory for AI-related cGMP failures. Every facility using AI in any compliance function — specifications, procedures, production records, batch records, investigation reports — should expect inspectors to ask how AI outputs are reviewed. “We reviewed it” is not the answer. The answer is the documented protocol and evidence it was followed.
DOMAIN 4 THE RENEGOTIATION
The renegotiation objectives are specific: validation documentation commitment, 21 CFR Part 11 data integrity, and a meaningful liability carve-out for regulatory-related losses caused by AI system failures. Add to that a specific indemnification provision covering FDA warning letters, consent decrees, and import alerts that result from documented failures in AI-generated compliance outputs.
That last provision is what most AI vendors will resist hardest. It is also what your company most needs. Most vendors will agree to enhanced validation commitments before they agree to regulatory indemnification — use the validation commitment as the opening position and treat indemnification as the objective.
Start with your three highest-risk systems: those generating specifications, SOPs, or production records. Those are the document types FDA cited. They are also the documents most likely to produce a 483 observation if the AI-generated content contains a gap the review process didn’t catch.
CLO WORKSTREAMS
Priority actions for the legal function before the next inspection.
■ Compliance infrastructure inventory. Identify every AI system in your compliance infrastructure — specifications, SOPs, production records, investigation reports, change control documentation. Map what human oversight currently exists and whether it is documented in the quality system.
■ Gap assessment against FDA’s fact pattern. Commission a review against the pattern FDA just described: AI-generated document, inadequate human review, undetected gap. That sequence is the inspection theory now. Assess every system against it before your next inspection date.
■ Quality system documentation. The oversight protocol must be in the quality system, not just in a policy memo. Who reviews. What they must cover. How they certify the review was substantive. What happens when a claim cannot be independently verified.
■ Vendor agreement audit. For every AI system in your compliance infrastructure, pull the vendor agreement and assess: validation documentation, 21 CFR Part 11 data integrity obligations, liability cap and disclaimer language, absence of regulatory indemnification. That gap is what the warning letter just confirmed.
■ Renegotiation priority list. Rank your AI vendor agreements by regulatory risk. Begin renegotiations with the vendors whose systems generate specifications, SOPs, or production records — the specific document types FDA cited. Validation commitments first; indemnification is the objective.
For interested readers I have created an Ai Compliance Oversight Assessment. It can be found at https://laroccaadvisory.gumroad.com/l/sbbfnj.
John LaRocca
John LaRocca has served as general counsel and CLO in the pharmaceutical and biotech sectors for 26 years. He advises boards and executive leadership on regulatory risk, M&A, and governance.
laroccaj.substack.com
THE GC PLAYBOOK
Intelligence for pharma CLOs and GCs who can’t afford to be surprised.
laroccaj.substack.com


