Language models in GMP document work: what the draft Annex 22 allows
The European Commission's draft Annex 22 to the EU GMP guidelines is the first GMP text written specifically about artificial intelligence. Its position on generative models is short and clear. This note sets out what it says, what it leaves open, and how we design document pipelines to stay on the right side of it.
Where the draft stands
The Commission consulted on three documents together: a revised Chapter 4 (Documentation), a revised Annex 11 (Computerised Systems) and a new Annex 22 (Artificial Intelligence). The consultation opened on 7 July 2025 and closed on 7 October 2025. The Commission says the documents were drafted by the EMA GMDP Inspectors Working Group in cooperation with PIC/S.
When we checked on 24 October 2026, the EudraLex Volume 4 page still listed Annex 11 as the January 2011 revision and did not list an Annex 22. Treat everything below as a draft that may change.
Sources: European Commission consultation page; EudraLex Volume 4.
What it says about generative AI
The scope section limits Annex 22 to static models with deterministic output, then says:
“Following the above, the document does not apply to Generative AI and Large Language Models (LLM), and such models should not be used in critical GMP applications. If used in non-critical GMP applications, which do not have direct impact on patient safety, product quality or data integrity, personnel with adequate qualification and training should always be responsible for ensuring that the outputs from such models are suitable for the intended use, i.e. a human-in-the-loop (HITL) and the principles described in this document may be considered where applicable.”Draft Annex 22, section 1, lines 20–25
Two things follow. A language model should not make or feed a critical GMP decision. And where one is used for non-critical work, a qualified person is responsible for every output.
What is left for a document pipeline
A large share of document work in a quality system is preparation, not decision. These are the uses we think fit the draft's non-critical category, always with a person approving. Whether a given use is non-critical is your quality unit's call, not ours.
- Finding the passage. Search over SOPs and specifications that returns the source text, so a person reads the rule itself rather than a summary of it.
- Drafting fields from free text. A model proposes values for a form from a free-text description; a person checks and submits. Nothing is saved without that step.
- Checking for gaps. A model flags where a document seems to be missing a required section; a person decides whether it is.
The controls we put around it
Human review is the floor, not the whole design. A reviewer who sees a plausible, complete-looking record is easily persuaded by it. So we make the model's output easy to doubt:
- “Missing” is a correct answer. The model is told that returning “missing” is right when a value is not in the text. That removes the pressure to invent one.
- A rule checks every field. A deterministic check confirms that every required field is present and in the expected form. A record that fails stops and names the field.
- The source travels with the value. Each proposed value carries the passage it came from, so the reviewer can check it in one look.
- Every step is logged. Input, model output, check result, reviewer and decision are recorded with a timestamp. That gives your quality team the record of operator actions that 21 CFR Part 11, §11.10(e) describes for audit trails.
We built and published the first two controls as a small lab project: an extractor that is allowed to say “missing”.
What this note is not. It is not a validation approach and not legal advice. We have not yet delivered a project inside a GMP-regulated validation. If you are planning one, we would work alongside your validation and quality teams, not instead of them.
Data integrity still applies
PIC/S guidance on data integrity expects data to be “attributable, legible, contemporaneous, original, accurate, complete, consistent, enduring, and available”, and it applies to electronic and paper records alike. A model in the pipeline does not change that. If anything, it adds one more step whose output has to be attributable and kept.