Use of Large Language Models in GMP-Regulated Work

Why this exists

In April 2026 FDA issued its first cGMP warning letter about AI. A manufacturer had used AI agents to write its specifications, procedures and master production records, with no quality unit review, and skipped process validation because, in the letter’s words, the AI agent “never told you it was required.”

Nobody in that firm set out to break GMP. The tool was simply used without a procedure. There’s plenty of discussion about AI in regulated work and very few documents you can actually adopt. This is one.

What’s inside

• Three tiers, named for what the output becomes: Scratch (never enters the QMS), Draft (enters only through a person), Decision (would replace a person’s judgment, so not permitted).
• Severity, frequency and detectability for the borderline cases, tied to ICH Q9.
• A review procedure built for fluent text: check against the source, verify every citation, look for what’s missing.
• One recorded statement per use, placed in the records you already keep, plus an optional AI-Assistance Register.
• A “Why” under every control, so you can adapt it on purpose instead of by accident.
• Every reference shows where it’s used, and the ones used only as a framework are marked.
• Good and bad use cases for training.

Using it

Download the Word file, adapt anything shaded yellow to your own procedures, and approve it through your own document control. Appendix C walks through the adaptation. Cross-references are live fields: after editing, press Ctrl+A, then F9.

This is a template, not advice. Your company must adapt and validate it for its own products, systems and risks. Two of its key references were drafts when it was written; check them for final versions.