AI in GMP systems: what the EU GMP draft Annex 22 means for your QMS and LIMS

The consultation draft is aimed at AI models used in critical GMP applications — those that could affect patient safety, product quality or data integrity. Its expectations will feel familiar to anyone who has validated a computerised system:

• Intended use: a documented description of what the model does, the data it uses and the decisions it supports, agreed by subject-matter experts.
• Acceptance criteria: performance measures defined before testing, at least as good as the process the model replaces.
• Independent test data: test data that is representative and kept separate from the data used to build the model.
• Explainability and confidence: enough insight into how an output was reached for a person to judge it.
• Operation and oversight: change control, performance monitoring in use and qualified people accountable for decisions.

The draft favoured static, deterministic models for critical uses, and treated other kinds of AI more cautiously. Commentary on the revision process suggests the final scope may change, so treat the details as provisional — but the direction is clear: AI in GMP must be fit for a defined purpose, tested and supervised by people.

Annex 22 adds to existing rules; it does not replace them. Anything an AI model reads, writes or influences in a GMP system is still subject to:

• EU GMP Annex 11 and 21 CFR Part 11 — validation, audit trails, electronic signatures, access control.
• Data integrity and ALCOA+ — records must remain attributable, legible, contemporaneous, original and accurate.
• ICH Q9 quality risk management and ICH Q10 — the pharmaceutical quality system owns the decision.

The attributable principle is the hardest test for AI. If a CAPA root cause, a deviation classification or a batch-release recommendation came from a model, an inspector will ask who decided, on what evidence and when. “The system did it” is not an answer.

Lower risk: help with the work

Summarising a long investigation for a reviewer, finding similar past deviations, drafting training questions from an approved SOP — these support a person who still reads, edits and signs. They still need control, but the GMP decision stays human.

Higher risk: deciding the outcome

Classifying a deviation’s criticality, judging an OOS result, accepting a batch for release or changing a GMP record without a person’s review sit squarely in Annex 22’s critical territory. Here the burden of intended use, testing, acceptance criteria and oversight is at its heaviest.

The design question that matters

For every AI feature, ask: does it suggest, or does it decide? A suggestion that a named person accepts or rejects — with both the suggestion and the decision on the audit trail — keeps the record attributable. A model that silently decides does not.

1. Which functions use AI, and what is each one’s intended use?
2. Does any AI function create, change or approve a GMP record without a person’s review and e-signature?
3. Is every AI suggestion, and the person who accepted or rejected it, recorded in the audit trail?
4. What data was the model tested on, against which acceptance criteria — and can we see the evidence?
5. How are model changes controlled, and how will we know when the model behaves differently?
6. Can we switch AI functions off, site by site, without affecting validated workflows?

A vendor that answers these clearly is treating AI as a GMP function. One that cannot is asking you to carry the risk.

Swat iQua Suite does not use AI in GMP decisions today. Every GMP record — a CAPA, a deviation, a test result, a batch step, a release — is created by a named person, and every decision carries an electronic signature, under one audit trail with a reason for every change and segregation of duties enforced in 25 modules.

AI assistance is on our roadmap. It is being designed to suggest and never decide: any suggestion will be reviewed and accepted by a person, recorded in the same audit trail as everything else, and switched on only when you choose. That keeps your records attributable — and keeps you in line with the human-oversight thinking in the Annex 22 draft.