EU GMP Annex 22: what the draft AI annex asks of GMP manufacturers

What is EU GMP Annex 22?

Annex 22, Artificial Intelligence, is a proposed new annex to EudraLex Volume 4, the EU's good manufacturing practice guidelines. The European Commission published it for consultation in July 2025 together with a revised Annex 11 (computerised systems) and a revised Chapter 4 (documentation). It was drafted jointly with PIC/S, so the same text is expected to reach PIC/S member authorities outside the EU.

It is the first GMP annex written specifically for AI. Annex 11 continues to cover computerised systems in general, and Annex 22 adds requirements for the model itself: how it is selected, trained, tested and kept under control.

Is Annex 22 in force?

No. The stakeholder consultation closed on 7 October 2025. When this article was last updated, the final text had not been published and no date of application had been set. Check the European Commission's consultation page listed under Sources for the current status before you plan against a date.

Draft or not, the direction is clear enough to act on. Inspectors already expect computerised systems to be validated for their intended use under Annex 11, and the draft shows how regulators intend to read that expectation for AI.

Which AI models does the draft allow in critical GMP applications?

The draft applies to AI models used in critical applications: those with a direct impact on patient safety, product quality or data integrity. For those applications it accepts only static models, meaning models that do not keep learning once they are in use, and only models with deterministic output, meaning the same input gives the same output.

• Allowed in critical applications: static machine learning models with deterministic output, for example a classifier trained once and then locked.
• Not to be used in critical applications: dynamic models that continue to learn in operation, and models with probabilistic output.
• Not to be used in critical applications: generative AI and large language models.

What does the draft say about generative AI?

The draft states that generative AI and large language models should not be used in critical GMP applications. Where they are used in non-critical applications, it expects personnel with adequate qualification and training to be responsible for checking that the output is fit for its purpose, which is a human-in-the-loop arrangement.

That is the line most quality teams already draw. A language model can draft an investigation summary, propose a root cause hypothesis or flag a record for attention. A qualified person reviews the draft, decides, and signs. The decision and the accountability stay with that person.

What does a manufacturer have to do for a model in scope?

• Describe the intended use in full, including the inputs the model is expected to see and the limits of that range.
• Set acceptance criteria before testing, with performance at least as good as the process the model replaces.
• Keep test data independent of the data used for training and validation, and show that it represents the intended use.
• Show which features drive the model's output, and record a confidence score so that low-confidence results are flagged instead of acted on.
• Keep the model under change control and configuration control, monitor its performance in operation, and watch for input data drifting outside the range it was tested on.
• Where a person reviews the output, define that review in procedures and train the reviewers.

What should a quality team do now?

• List every place AI is used or planned in GMP activities, and mark each one critical or non-critical.
• For generative AI, write down what the tool does, what the reviewer checks, and where the reviewer's decision is recorded.
• For machine learning models in critical use, gather the intended use statement, the test data and the acceptance criteria, and identify the gaps against the draft.
• Read the revised Annex 11 alongside it. Its requirements for audit trails, supplier oversight and security apply to the systems that host these models.