p4mx health
← All impulses
Process applicationsNovember 14, 2026· 5 min read

Annex 22 sets out how AI in medicines manufacturing must be tested

With Annex 22, the EU GMP Guide is getting its first dedicated rules for artificial intelligence. The 2025 draft has not yet been adopted, and work on it continues. Its main features, however, are settled enough to plan for today: a precisely described intended use, independent test data, acceptance criteria fixed in advance and monitoring in operation. This article sets out what the draft requires, where it is still open and what you can prepare now.

Many manufacturers are trialling artificial intelligence where people inspect, sort or assess today: in the visual inspection of containers, in forecasting process parameters or in classifying deviations. As soon as such an application has a say in product quality or batch release, the question arises of how to demonstrate it to GMP standards. The EU GMP Guide will answer that question in a dedicated annex of its own: Annex 22.

The basis: a new annex, still in draft

On 7 July 2025, the European Commission put the draft of Annex 22 “Artificial Intelligence” out to consultation, together with revisions of Chapter 4 (Documentation) and Annex 11 (Computerised Systems); comments could be submitted until 7 October 2025. Annex 22 supplements Annex 11 for systems in which AI models are embedded. It does not replace it.

The annex has not yet been adopted. The Commission’s guide (EudraLex Volume 4) still lists Annex 11 in its 2011 version and does not include an Annex 22. In its work plan, the GMP inspectors’ working group at the European Medicines Agency (EMA) names the fourth quarter of 2026 as the target for providing the Commission with a final text. Minutes of an EMA working group from June 2026 refer to publication in early 2027. No date of application and no transition period have been announced so far.

What the draft covers

The draft applies to AI models in critical applications, that is, where they have a direct impact on patient safety, product quality or data integrity. It is aimed at machine learning models that have obtained their functionality through training with data rather than by being explicitly programmed. Conventional rule-based software remains a matter for Annex 11.

Two restrictions shape the draft. It covers only static models, which do not keep developing during use, and only models with a deterministic output, which always give the same answer to the same input. Under the draft, models that learn continuously, models with a probabilistic output, and generative AI and large language models should not be used in critical GMP applications. In non-critical applications they remain possible, provided that qualified personnel take responsibility for every output, with a human in the loop.

What the draft requires

At its core, the draft sets out a test procedure reminiscent of a well-run clinical trial: fix the rules first, then test with untouched data.

Whether the model is developed in-house or bought in, its documentation must be available to the medicinal product manufacturer and reviewed by it.

What is still moving

Drawing the line by technology is the most contested part of the draft. According to the EMA report on an expert workshop held on 30 June 2026, comments received during the consultation supported allowing language models and generative AI in GMP applications as well. The drafting group is therefore discussing whether to extend the scope to adaptive, probabilistic and generative models, provided a documented, risk-based control strategy supports them. Nothing has been decided. At the workshop, the regulators explicitly asked whether phenomena such as fabricated answers from language models are a distinct kind of failure that calls for specific measures.

For planning purposes, this means that the requirements on intended use, test data and monitoring will, in our assessment, remain. What is chiefly open is which types of model will be permitted in critical applications.

How to prepare today

  1. Take stock: which AI applications are in use or planned in manufacturing and quality control, and which of them touch product quality, patient safety or data integrity?
  2. Measure the baseline: how well does today’s, usually manual, process perform? Without that figure, the “at least as good” criterion cannot be demonstrated.
  3. Separate test data early: an independent, access-controlled test dataset can hardly be created after the fact. It belongs at the start of the project.
  4. Involve suppliers: for bought-in models, secure contractual access to training and test documentation.
  5. Plan for operation: define metrics for performance and data drift and integrate them into existing change control, as quality risk management under ICH Q9 already provides.

Where this does not apply

Annex 22 concerns the manufacture of medicinal products and active substances. AI in medical devices is governed by other rules, as is AI in clinical trials. Even within manufacturing, the draft applies only to critical applications: a language model that drafts training materials which a person then reviews does not, in our assessment, fall under the strict testing requirements. And a draft is not law in force. Until the final version, Annex 11 in its 2011 version remains the binding standard, together with the risk management across the entire lifecycle that Annex 11 already requires today. Anyone implementing individual requirements of the draft now does so voluntarily and should expect details to change.

Conclusion

Annex 22 translates familiar GMP principles into the language of machine learning: a defined purpose, independent evidence, controlled operation. Which types of model will ultimately be permitted remains open. The testing discipline the draft calls for pays off today, however, because it can hardly be established after the fact.

Sources: European Commission: Stakeholders’ Consultation on EudraLex Volume 4 – Good Manufacturing Practice Guidelines: Chapter 4, Annex 11 and New Annex 22, consultation from 7 July to 7 October 2025. European Commission: EudraLex Volume 4, Annex 22 “Artificial Intelligence”, consultation draft, 2025 — sections 1, 2.2, 3.1 to 3.3, 4.1 to 4.3, 5.1 to 5.6, 6.1 to 6.5, 7.2, 8, 9 and 10.1 to 10.5. European Commission: EudraLex Volume 4, Annex 11 “Computerised Systems”, 2011 (version in force). European Medicines Agency: The 3-year work plan for the Inspectors Working Group, GMDP Inspectors Working Group, January 2026 to December 2028, 2026. European Medicines Agency: Report, GMP Multistakeholder workshop on expert contributions to artificial intelligence guidance development (Annex 22), 30 June 2026, EMA/156789/2026. HMA/EMA Joint Network Data Steering Group: Minutes of the meeting on 16 June 2026, EMA/143308/2026. ICH: Q9(R1) Quality Risk Management, 2023.

Does this match your situation?

We begin with a process analysis and show, with evidence, what is possible.

Arrange a conversation