Artificial Intelligence
Draft EudraLex Volume 4 annex setting the first GMP-specific expectations for artificial intelligence in regulated manufacture. Published 7 July 2025 by the European Commission jointly with PIC/S and consulted alongside revisions to Chapter 4 (Documentation) and Annex 11 (Computerised Systems); the consultation closed 7 October 2025. It remains a draft with no adoption date — plan against it, not to it. Until it is adopted there is no GMP-specific instrument for AI: Annex 11 remains the applicable text for computerised systems, read with ISPE GAMP 5.
What this does not cover
stated in the document's own scope- It is a draft. It creates no obligation until the Commission adopts it into EudraLex Volume 4.
- It does not replace Annex 11: computerised systems that contain an AI model remain computerised systems.
- It does not address the EU AI Act’s classification and conformity requirements, which come from that regulation.
- In the draft, dynamic models that keep learning in use, models with probabilistic output, and generative AI and large language models should not be used in critical GMP applications; for non-critical uses of generative AI, qualified people must remain responsible for the output.
Always verify against the current published text before relying on it for a submission or inspection.
Overview
Annex 22 is a draft annex to the EU Guide to Good Manufacturing Practice on the use of artificial intelligence in the manufacture of medicinal products and active substances. The European Commission published it on 7 July 2025, jointly with PIC/S, for a consultation that ran alongside proposed revisions of Chapter 4 on documentation and Annex 11 on computerised systems, and closed on 7 October 2025. It applies to machine-learning models used in critical applications with a direct impact on patient safety, product quality or data integrity, and only to static models with deterministic output. As of the Commission’s EudraLex Volume 4 page updated on 6 October 2026, the list of adopted annexes ends at Annex 21, so Annex 22 has not been adopted.
Scope & applicability
Draft scope covers AI and machine-learning models used in GMP-critical applications. The draft confines itself to static, deterministic models — parameters frozen, identical inputs producing identical outputs — and excludes dynamic self-learning models, generative AI and large language models from GMP-critical use.
Legal basis & how it acquires force
EudraLex Volume 4 contains guidance for interpreting the principles and guidelines of GMP laid down, for human medicines, in Commission Directive (EU) 2017/1572 of 15 September 2017. An annex acquires that standing only when the Commission adopts and publishes it in Volume 4. A consultation draft has no legal effect: until adoption, computerised systems used in GMP, including those that incorporate AI, are assessed against the adopted Annex 11 and Chapter 4, and against the PIC/S equivalents in participating authorities.
Document structure
| Part | Covers |
|---|---|
| 1 Scope | Machine-learning models in critical GMP applications; static models with deterministic output only; generative AI and large language models excluded |
| 2 Principles | The principles that frame the use of AI models in GMP |
| 3 Intended use and 4 Acceptance criteria | Defining what the model is for, including human-in-the-loop use, and the performance metrics it must meet |
| 5 Test data, 6 Test data independency, 7 Test execution | Quality and independence of the data used to test the model, and how the test is run |
| 8 Explainability and 9 Confidence | Understanding the basis of the model’s output and the confidence attached to it |
| 10 Operation | Change control, performance monitoring and human review once the model is in use |
Key requirements
- Models in GMP-critical applications are static and deterministic — parameters frozen, identical inputs producing identical outputs
- Dynamic self-learning models, generative AI and large language models are excluded from GMP-critical applications
- Test data is independent of training data
- Artificially generated test data is advised against in critical applications unless justified and documented
Three questions to ask of any AI in a GMP-critical use
For manufacturers deploying AI, and for the companies building it for them: a customer’s QA team can ask these of any model before it touches a GMP-critical decision. Annex 22 is a draft, so treat each answer as preparation, not compliance.
Same input, same output, every time?
What EU GMP Annex 22 saysModels in GMP-critical applications are static and deterministic — parameters frozen, identical inputs producing identical outputs
Why it matters · SPEQ’s readingValidation proves that the system you tested is the system you run. A model whose answer changes from one run to the next leaves QA nothing fixed to validate.
Is it a generative model or a large language model?
What EU GMP Annex 22 saysDynamic self-learning models, generative AI and large language models are excluded from GMP-critical applications
Why it matters · SPEQ’s readingIf it is, the draft keeps it out of GMP-critical applications. This is the provision most likely to move before final text, so plan against it rather than to it.
Where did the test data come from?
What EU GMP Annex 22 saysTest data is independent of training data
Artificially generated test data is advised against in critical applications unless justified and documented
Why it matters · SPEQ’s readingA test set the model has already seen cannot show how it performs on data it has not. Independent test data is the difference between a demo and evidence.
The questions and “why it matters” lines are SPEQ’s framing. Only the quoted requirements come from EU GMP Annex 22.
Revision notes
No previous version — this is the first EU GMP annex addressing artificial intelligence. Draft published 7 July 2025; consultation closed 7 October 2025; no final text published as of 4 September 2026. Annex 11 remains the applicable text for computerised systems in the meantime. Treat every requirement here as provisional until the Commission adopts final text.
Where this control fails
live FDA enforcementLive FDA recalls SPEQ maps to this standard’s topics — a SPEQ interpretation, not an FDA classification.
International alignment
The draft was released jointly by the European Commission and PIC/S. It sits beside Annex 11, which governs computerised systems generally, and Chapter 4, which governs documentation; revisions of both were consulted on at the same time. The Commission’s consultation page summarises the draft as setting requirements for the selection, training and validation of AI models, with emphasis on intended use, performance metrics, training-data quality and test data, and continuous oversight. Requirements under the EU AI Act are a separate legal regime.
EU GMP Annex 22: frequently asked questions
Quick answers to common questions about EU GMP Annex 22.
Is EU GMP Annex 22 in force?
No. The Commission published it as a consultation draft on 7 July 2025, and the EudraLex Volume 4 page updated on 6 October 2026 still lists annexes only up to Annex 21.
What governs AI in GMP until Annex 22 is adopted?
The adopted texts: Annex 11 on computerised systems and Chapter 4 on documentation in the EU and PIC/S GMP guides, applied to the AI-enabled system as a computerised system, with the quality risk management the GMP guide already requires.
Was Annex 22 consulted on its own?
No. It was released together with proposed revisions of Chapter 4 and Annex 11, and the consultation on all three closed on 7 October 2025.
This standard in practice
Recall domain is a SPEQ mapping of this standard’s topics, not an FDA classification.