FDA warning letter on AI agents: what Quality Units now need to prove

What FDA actually observed

Warning Letter 320-26-58 was issued to Purolea Cosmetics Lab, a drug manufacturer. The AI passage matters, but it is not the whole case. FDA also documented insanitary conditions, missing microbiological testing, inadequate component testing and unapproved new drugs. Turning the letter into a pure AI story would omit most of the agency's findings.

Within that wider quality-system failure, FDA also recorded a clear AI observation. The firm had used AI agents to create drug product specifications, procedures, and master production or control records. Those AI-generated documents were not adequately reviewed to establish that they were accurate and actually CGMP compliant. FDA linked that failure to 21 CFR 211.22(c).

A second observation made the responsibility gap even clearer. The firm had not completed process validation before distributing its products. Its explanation to investigators was, in substance, that the AI agent had not told the firm this was required. A model may prepare work, but it cannot inherit the regulatory responsibility of the manufacturer or its Quality Unit.

Not a new AI rule — an existing duty applied to a new tool

The warning letter does not invent a special rule for ChatGPT, LLMs or AI agents. It applies established CGMP responsibility to a new tool. Under 21 CFR 211.22(c), the Quality Control Unit is responsible for approving or rejecting procedures and specifications that affect the identity, strength, quality and purity of a drug product.

21 CFR 211.100 is equally technology-neutral. Production and process-control procedures must be written, appropriately reviewed, and approved by the relevant organisational units, including the Quality Unit. Whether the first draft came from an employee, a contractor or a language model does not move that accountability.

If AI is used again for CGMP activities, FDA requires in this case that an authorised human representative of the Quality Unit review and clear every AI output and recommendation.

Quality teams therefore need to ask more than may AI write this? The defensible question is: can the company show what the draft was based on, who reviewed it professionally, and why it entered the controlled state?

AI can take on the writing. It cannot take responsibility for proving that nothing material is missing.

Three control gaps behind “the AI did not tell us”

The documented explanation exposes three organisational gaps that can also appear in much more mature companies.

1. Plausibility was mistaken for completeness

A fluent draft can sound convincing while omitting a material requirement. An LLM therefore cannot define for itself which sources or obligations are complete. The permitted source space and expected review criteria must exist before generation begins.

2. Review was not an auditable work step

Having a human somewhere in the process is not enough if nobody can later see what that person checked, corrected or deliberately accepted. Human review needs a visible object: source, suggestion, change, open points and an attributable decision.

3. The tool became the subject-matter owner

Once a team assumes that an AI agent will flag every relevant obligation, the role has shifted. A drafting tool has silently become a gatekeeper. A generative model must not hold that role in GxP work.

What a controlled AI process for GMP documents should show

The warning letter is not a universal AI validation framework. For document-level AI work, it does surface six practical questions that should be answered before productive use:

1. Intended use: which task may AI prepare, and which decision is explicitly outside its role?
2. Source boundary: which approved SOPs, policies, laws, guidance documents and project evidence must the draft consider?
3. Draft status: is it always clear that AI output is not yet a confirmed work state?
4. Expert review: can the reviewer check each statement against its source, context and expected completeness?
5. Decision trail: do changes, corrections, acceptances and rejections remain attributable to a person and rationale?
6. Procedure boundary: is it clear which process, system or method validation remains separately required?

A chat transcript can support parts of this process. It does not establish the controls by itself. The process becomes defensible when the controlled basis, generated draft and human decision remain linked.

What traqx controls here — and what it does not

The traqx System is built for this document-level part of the work. Sources are defined before wording is generated. Statements carry citations and are checked deterministically against the available sources. Suggestions remain visibly suggestions. For Word updates, traqx returns tracked changes and source or rationale comments. Review and decisions remain with the expert, while the audit trail keeps the path connected.

That addresses the document and review gap visible in the warning letter. It is not a guarantee against warning letters. traqx does not inspect manufacturing hygiene, replace microbiological testing or automatically perform process validation. It does not replace the Quality Unit either.

The product role remains bounded: AI takes on research, drafting and document updates. The company keeps subject-matter responsibility, process control and approval where they belong.

What Quality Units should do next

Quality Units do not need to stop AI indiscriminately. They need to prevent a useful tool from replacing source knowledge, completeness checks or professional judgement. That does not start with a long AI policy. It starts with one well-bounded work case.

Take a real document, define the permitted sources and expected review criteria, let AI prepare a draft, and then review the output and the entire evidence chain. When source, suggestion, correction and decision stay connected, faster writing becomes controlled support. When that connection is missing, it remains a good-looking answer without a defensible work state.

Frequently asked questions

Did FDA ban AI or AI agents in GMP?

No. Warning Letter 320-26-58 contains no blanket ban on AI in pharmaceutical manufacturing. In the specific case, FDA cited unreviewed AI-generated GMP documents, inadequate Quality Unit oversight and the transfer of subject-matter responsibility to an AI agent.

What did FDA cite about AI-generated GMP documents?

The firm used AI agents to create drug product specifications, procedures and master production or control records. FDA stated that the documents had not been adequately reviewed for accuracy and actual CGMP compliance and linked the failure to 21 CFR 211.22(c).

Must the Quality Unit review AI-generated documents?

Yes, when those documents fall within the Quality Unit's responsibilities. 21 CFR 211.22(c) assigns the unit responsibility for approving or rejecting quality-relevant procedures and specifications; an AI-generated first draft does not change that responsibility.

Can ChatGPT be used for GxP documents?

The product name does not determine GxP suitability. Intended use, data and source boundaries, proportionate qualification or validation, documented human review, change control and the audit trail matter. A general chat interface does not establish those controls merely because an employee reads the output.

Would traqx have prevented this FDA warning letter?

That would not be a defensible claim. The warning letter includes fundamental manufacturing, testing and quality-system failures in addition to the AI observation. traqx addresses controlled document work through source binding, visible suggestions, human review and an audit trail; it does not replace the Quality Unit, process validation or manufacturing controls.