In April 2026, FDA told a manufacturer whose AI agents had written its drug specifications, procedures and master production records what it expects. Not a ban, but a person in the firm’s quality unit who reviews and clears whatever the AI produces. The violation FDA cited was the review that did not happen, under a rule that says nothing about AI.

The letter also shows why that review must do more than read the page. The firm did not know a legal requirement existed, and by its own account the AI agent never told it. A reviewer who checks only what is written cannot see what is missing.

What happened at the lab

FDA inspected the facility, a cosmetics lab in Livonia, Michigan, that also made drug products, in October 2025. When FDA finds what it believes are significant violations, it often notifies the firm by warning letter. This one, dated April 2, 2026, summarizes “significant violations of Current Good Manufacturing Practice (CGMP) regulations for finished pharmaceuticals,” the federal rules for making drugs.

Most of the letter is not about AI. It cites insanitary conditions, batches and components not adequately tested, a quality unit that did not oversee operations, and two products FDA considers unapproved new drugs. One testing finding, reliance on suppliers’ certificates of analysis without establishing their reliability, is covered in its own article.

The AI passage has its own heading, “Inappropriate Use of Artificial Intelligence in Pharmaceutical Manufacturing.” It records that “you stated to FDA investigators that you utilized artificial intelligence (AI) agents (b)(4) to help your firm comply with FDA regulations,” the “(b)(4)” marking text withheld from the published letter. “Specifically, you used AI to create drug product specifications, procedures, and master production or control records to be in compliance with FDA requirements.”

What FDA cited: the missing review, not the AI

The rule cited does not mention AI. Section 211.22(c) reads in full: “The quality control unit shall have the responsibility for approving or rejecting all procedures or specifications impacting on the identity, strength, quality, and purity of the drug product.” The quality control unit, which the letter calls the quality unit, is “any person or organizational element designated by the firm to be responsible for the duties relating to quality control” (21 CFR 210.3).

The drug rules already require a second reader for each kind of document the AI wrote. Production and process control procedures must be “drafted, reviewed, and approved by the appropriate organizational units and reviewed and approved by the quality control unit” (21 CFR 211.100(a)). A master production and control record, the master instructions for making a product, must be “prepared, dated, and signed (full signature, handwritten) by one person and independently checked, dated, and signed by a second person” (21 CFR 211.186(a)). AI changed who drafted the documents, not who must approve them.

Why reading the page would not have been enough

The letter continues: “Overreliance on artificial intelligence for your drug manufacturing operations was also documented during the inspection.” Investigators found that the firm “had not conducted process validation prior to distribution of your drug products, as required under 21 CFR 211.100.” Its answer, in FDA’s record: “You replied that you were not aware of the legal requirement, as the AI agent you used (b)(4), never told you it was required.”

Process validation is the evidence that a manufacturing process delivers the intended product consistently, not just once. FDA’s guidance defines it as “the collection and evaluation of data, from the process design stage through commercial production, which establishes scientific evidence that a process is capable of consistently delivering quality product.”

The word “validation” does not appear in 21 CFR 211.100. FDA’s guidance explains that process validation “is required, in both general and specific terms, by the CGMP regulations in parts 210 and 211,” and that its foundation “is provided in § 211.100(a),” the paragraph requiring written procedures “designed to assure” the quality a drug claims. The requirement lives in how the rule is read, so a draft can leave it out without leaving a visible gap.

A review that checks a document against itself, for clarity, consistency and format, will pass one that is missing a step. The gap shows only to a reviewer who starts from what the rules require and looks for each item. That takes knowledge, and the drug rules require everyone engaged in making a drug to be trained “in current good manufacturing practice” as it relates to their functions (21 CFR 211.25).

What FDA asked for instead

The firm had stopped making drugs. For a restart, FDA set a condition, not a prohibition: “If you plan to resume drug production, and use AI to help with CGMP activities, such as development of procedures and specifications, any output or recommendations from an AI agent must be reviewed and cleared by an authorized human representative of your firm’s QU in accordance with section 501(a)(2)(B) of the FD&C Act.”

That section of the Federal Food, Drug, and Cosmetic Act treats a drug as adulterated when the methods, facilities or controls used to make it do not conform with current good manufacturing practice. Neither it nor the two CFR sections cited alongside it mentions AI.

Elsewhere the letter advises the firm to engage a consultant qualified under 21 CFR 211.34, which requires “sufficient education, training, and experience,” and states that “FDA regards contractors as extensions of the manufacturer.” Neither point mentions AI, but read beside the AI passage they say the same thing: outside help leaves the responsibility with the firm.

Does this apply to cosmetics, supplements and food?

A warning letter speaks to one firm, and this one concerns drug manufacturing. It cites no cosmetics violation, though it reminds the firm of the Modernization of Cosmetics Regulation Act of 2022 (MoCRA), which requires FDA to establish good manufacturing practice regulations for cosmetics facilities.

Supplement and food rules already give someone the job the letter describes:

Both rules also require that person to be qualified (21 CFR 111.12(b) and 117.180(c)). For a maker of cosmetics alone the drug rule binds nothing, but its reasoning fits any AI-drafted document a quality team relies on.

In the EU, a draft GMP Annex 22, put to public consultation from July to October 2025, proposes requirements for AI used in manufacturing medicines, including human review where necessary. A separate article covers it.

How to review an AI-drafted procedure or specification

The letter does not say how to review. These steps follow from what it cites and fit any document a quality team approves but did not write.

• Record which documents an AI tool drafted or changed, and who approved each.
• Before reading the draft, list what the rules and your quality system require of that kind of document. For a drug master record, 21 CFR 211.186(b) is the list.
• Check the draft against that list item by item. A missing step is the error that reading alone does not catch.
• Ask what the document assumes is already done, such as process validation or supplier qualification, and confirm it.
• Trace every limit in a specification to a source, and every cited regulation to its current text.
• Have a named person with the authority and training approve it, and keep the signature and date.

Smart-Suited Tech’s document agent, Elara, applies the same division of work to supplier documents: it resolves ordinary conflicts between documents by rule, with the reason on the record, and surfaces gaps and compliance-critical changes for a person to decide.

The violation FDA cited was not that a program wrote the documents. It was that no one with the authority to approve them made sure they were accurate and compliant. The process validation exchange shows what that takes: a reviewer who knows what the rules require before reading what the draft says.