I contributed to the industry review of the EU GMP Annex 22 (AI) draft, and I keep this page updated as each step happens, with a link to the primary source for every line. The wider method for validating AI is in my guide to AI validation in GxP.
Status table
| Date | What happened | Who | Status |
|---|---|---|---|
| 7 July 2025 | Draft Annex 22 published for stakeholder consultation, together with revised Chapter 4 (Documentation) and revised Annex 11: Computerised Systems. EMA's annual report gives 10 July 2025; I use the Commission's date. | European Commission; drafted by the EMA GMDP Inspectors Working Group in cooperation with PIC/S | Draft |
| 7 October 2025 | Consultation closed at 23:59 CEST | European Commission | Closed |
| 22 October 2025 | The working group's 2025 annual report says the drafting group continues to work on a final guideline, taking the stakeholder comments into account | EMA GMDP IWG | Drafting |
| 2026 | The working group's 3-year work plan (January 2026 to December 2028) sets Q4 2026 as the target to give the Commission final texts of Annex 22, Annex 11 and Chapter 4 | EMA GMDP IWG | Target set |
| 30 June and 1 July 2026 | Multistakeholder expert workshop. Day 1 was open to experts presenting their views; day 2 was a closed session where the Annex 22 drafting group reviewed the contributions | EMA | Input gathered |
| 2 October 2026 | Workshop report published. It records that the drafting group discussed extending the scope to dynamic or adaptive models, probabilistic models and generative AI and LLMs, provided they meet the annex's requirements under a documented control strategy | EMA | Input to the revision |
| 12 November 2026 | No final text published | Still a draft | |
| Expected next | Final text from the working group to the Commission, then adoption, publication in EudraLex Volume 4 with an effective date, and PIC/S adoption into its own GMP guide | EMA, European Commission, PIC/S | Not yet happened |
I'll add a row when anything moves. If the Q4 2026 target slips, I'll record that too.
What the draft covers
The draft applies to computer systems used in the manufacture of medicines and active substances where an AI model is used in a critical application, meaning one with direct impact on patient safety, product quality or data integrity. It gives additional guidance to Annex 11, and it covers machine learning models that are static (they don't adapt during use) and give deterministic output (identical inputs give identical outputs).
The draft is organised around a handful of plain requirements:
- Intended use. A detailed description of the task and the input data, with the input sample space divided into subgroups, approved by the process subject matter expert before acceptance testing. Where a person makes the final decision, the operator's responsibility is part of the intended use.
- Acceptance criteria. Performance metrics defined and approved in advance by the SME, and at least as high as the performance of the process the model replaces (see acceptance criteria).
- Test data. Representative of the full intended use, with labels verified to a very high degree of correctness, and large enough to calculate the metrics with adequate statistical confidence.
- Test data independence. Test data kept apart from development and training, and from the people who train the model, with access control and audit trails to prove it. Where people can't be separated, a 4-eyes check applies. See test data independence.
- Test execution. An approved test plan, with checks that the model generalises well, and deviations recorded.
- Explainability. A record of which features drove each result, and a review of those features as part of approving the test results. See explainability.
- Confidence. A confidence score logged with each output where applicable, and a threshold below which the output can be flagged as undecided.
- Operation. Change control, configuration management, performance monitoring, and monitoring of the input data for drift away from the tested sample space.
What the draft excludes
The draft says dynamic models that keep learning in use, and models with probabilistic output, are not covered and should not be used in critical GMP applications. It says it doesn't apply to generative AI and large language models, and that they should not be used in critical GMP applications either. For non-critical applications it allows these models where qualified and trained personnel stay responsible for checking the output, a human in the loop, and says the annex's principles may be considered where applicable.
This is the part most likely to change. EMA's October 2026 workshop report says participants felt the annex should stay technology-neutral and principle-based, and that the drafting group discussed allowing dynamic, probabilistic and generative models in GMP applications under a documented control strategy. The report says this input will inform the revision, so the scope question stays open until the final text appears.
How it relates to Annex 11 and Chapter 4
The 3 drafts were written as a set, and they're meant to be read together.
Annex 11 is the base. Every AI system is a computer system first, so everything in Annex 11 applies to it as well: supplier management, access control, audit trails, backup, periodic review and the rest. Annex 22 adds what's specific to the AI model on top. The revised Annex 11 draft is much longer and more detailed than the 2011 version, and it's also still a draft.
Chapter 4 covers documentation across GMP, and the working group's plan ties its revision to data integrity. It matters for AI because training and test data are GMP data, and the records an AI system produces are GMP records.
In practice, I'd build one control framework that covers all 3, and run a single gap assessment against it.
What a quality team should do now
You don't need to wait for the final text to start. The parts of the draft that are least likely to move are the ones that take the most time to put in place.
- Build an AI inventory. List every AI use in GxP areas, including the AI features inside tools you already own. You can't govern what you haven't found.
- Classify each use. Is it critical or non-critical under the draft's definition? Is the model static or self-learning, deterministic or probabilistic, generative or not? This tells you which uses the draft would allow today.
- Write intended use statements for every critical use, with an SME named as owner and the input range described.
- Lock down test data. Check that test sets for existing models were kept apart from training, and put access control and audit trails on them now. This is the hardest thing to fix after the fact.
- Add confidence and explainability records where your models support them, and find out from suppliers what they can provide.
- Set up drift monitoring for any model already in use, with defined triggers for investigation.
- Put human review on record for every generative AI use in non-critical GxP work. FDA's warning letter 320-26-58 of 2 April 2026, issued under the heading "Inappropriate Use of Artificial Intelligence in Pharmaceutical Manufacturing", shows what happens when AI-written GMP documents skip quality unit review.
- Map your SOPs against the draft Annex 11 and Chapter 4 at the same time, so you change each procedure once.
None of this is wasted if the final scope widens. If generative AI is let into critical uses, I'd expect it to come with a stricter control strategy, and these steps are the floor for it. If you'd like an outside view of where you stand, my AI and CSA gap assessment maps your AI uses and procedures against the drafts. For the underlying discipline, see my guide to computer system validation in pharma.
Frequently asked questions
Is EU GMP Annex 22 final?
No. On 12 November 2026 Annex 22 is still a draft. The consultation ran from 7 July to 7 October 2025, EudraLex Volume 4 lists no Annex 22, and the EMA inspectors working group has set Q4 2026 as its target to give the European Commission a final text.
Does the draft Annex 22 allow generative AI?
Not in critical GMP applications. The draft says it doesn't apply to generative AI and large language models, and that they should not be used in critical GMP applications. In non-critical uses a qualified person must be responsible for checking the output. EMA's October 2026 workshop report shows the drafting group discussed widening the scope, so this may change.
Does Annex 22 replace Annex 11?
No. The draft says it gives additional guidance to Annex 11 for computer systems with embedded AI models. Every AI system is a computer system first, so Annex 11 still applies in full.
What should we do before Annex 22 is final?
Build an AI inventory, classify each use as critical or not, write intended use statements, lock down test data, add confidence and explainability records, monitor for drift and put human review on record. Those steps hold whichever way the final scope goes.
Map your AI uses against the draft
If you'd like help mapping your AI uses against the draft, book a 30-minute call and we'll go through your inventory together. For the full method, my book covers Annex 22 alongside Annex 11, GAMP 5 and FDA's guidance.
Sources
Last checked: 12 November 2026. I checked the sources listed below on that date.
- European Commission, consultation on Chapter 4, Annex 11 and new Annex 22 (7 July to 7 October 2025)
- European Commission, draft Annex 22 consultation document
- European Commission, EudraLex Volume 4 (no Annex 22 listed on 12 November 2026)
- EMA, GMDP Inspectors Working Group annual report 2025
- EMA, GMDP Inspectors Working Group 3-year work plan 2026 to 2028
- EMA, Annex 22 multistakeholder workshop, event page
- EMA, Annex 22 workshop report (first published 2 October 2026)
- FDA, warning letter 320-26-58, 2 April 2026