Annex 22 readiness checklist for GMP sites (EU GMP draft)
What a GMP site can do now for EU Annex 22 while it is still a draft, what to wait on, and where generative AI stands after EMA's 2026 workshop.
Status checked October 26, 2026. Annex 22 is a draft. EudraLex Volume 4 did not list a final Annex 22 on that date. Nothing below is a requirement yet, and the checklist is my reading of the draft and the current rules.
Annex 22 is still a draft, consulted on from July 7 to October 7, 2025. Sites can start now on the parts current GMP already expects: an inventory of AI uses with their intended use and criticality, controlled and independent test data, acceptance criteria set before testing, and change control for models. Wait on generative AI policy. The draft keeps large language models (LLMs) out of critical GMP use, and EMA’s June 2026 workshop discussed relaxing that without deciding.
Where EU GMP Annex 22 stands today
The European Commission published draft Annex 22 on artificial intelligence for consultation in July 2025. On October 25, 2026 the EudraLex Volume 4 page still did not list a final version. I keep the full status, the dates and what could change in AI validation in GxP has no settled playbook, so this page sticks to what a site can do while it waits.
The draft scope is narrow. It covers computerized systems “used in the manufacturing of medicinal products and active substances, where Artificial Intelligence models are used in critical applications with direct impact on patient safety, product quality or data integrity”. It applies to static, deterministic machine learning models.
Annex 22 readiness checklist: start now or wait
The table splits the work in two. The first group holds whatever the final text says, because current GMP already asks for most of it. The second group depends on wording that may still move.
| Item | Draft section | Already expected by |
|---|---|---|
Start now | ||
| Inventory of AI uses, with GMP impact and criticality | Draft section 1 | Annex 11, clause 4.3 |
| Intended use written and approved before acceptance testing | Draft section 3 | Annex 11, clause 4.4 |
| Acceptance criteria set by a process expert before testing | Draft section 4 | Annex 11, clause 4 (validation) |
| Test data that covers the intended use, with its rationale | Draft section 5 | Annex 11, clause 4 (validation) |
| Test data kept apart from training, under access control | Draft section 6 | Annex 11, clauses 7 and 12 |
| Approved test plan, with deviations investigated | Draft section 7 | Annex 11, clause 4 |
| Change and configuration control for the model | Draft section 10 | Annex 11, clause 10 |
Wait for the final text | ||
| A site policy on generative AI in critical use | Draft section 1 | Not applicable |
| Explainability tooling choices | Draft section 8 | Not applicable |
| Confidence thresholds as a site standard | Draft section 9 | Not applicable |
Each Annex 22 checklist item and the draft section behind it
The inventory comes first because everything else hangs off it. Current Annex 11 already says “An up to date listing of all relevant systems and their GMP functionality (inventory) should be available.” Adding each AI use to that list, with a note on whether it touches patient safety, product quality or data integrity, is the step that tells you which uses the draft would cover. I wrote up how I build that list in Start with the inventory: AI use cases and risk tiers.
Draft section 3 asks for the intended use to be “described in detail”, including “all common and rare variations” of the input data, and to be “documented and approved before the start of acceptance testing”. Annex 11 clause 4.4 already asks for user requirements “based on documented risk assessment and GMP impact”. A site that writes the intended use as part of its user requirements satisfies both.
Draft section 4 puts a process subject matter expert in charge of acceptance criteria and adds a “No decrease” rule: the model’s criteria “should be at least as high as the performance of the process it replaces”. That means you need a measured baseline for the current manual process. Many sites do not have one, so starting the measurement now is cheap insurance.
Draft sections 5 and 6 cover test data. The data should be “representative of and expand the full sample space of the intended use”, and data used for testing “is not used during development, training or validation of the model”. Section 6 goes further and asks for staff independence, with a “4-eyes principle” fallback where a small team cannot keep people apart. Access control and audit trails on a test data store are ordinary Annex 11 controls, so this is setup work you can do today.
Draft section 10 puts the model, the system and the process under change control and configuration control, with monitoring of performance and of input drift. Annex 11 clause 10 already says changes “should only be made in a controlled manner in accordance with a defined procedure”. How deep each control goes should follow the risk, and I cover sizing that in How much validation does your AI need.
Generative AI: what the draft says and what the workshop discussed
These are two different documents, and I keep them apart.
“the document does not apply to Generative AI and Large Language Models (LLM), and such models should not be used in critical GMP applications.”
European Commission, draft EU GMP Annex 22, section 1, consultation July to October 2025
The same section allows non critical use. Where LLMs are used in “non-critical GMP applications, which do not have direct impact on patient safety, product quality or data integrity”, the draft says “personnel with adequate qualification and training should always be responsible for ensuring that the outputs from such models are suitable for the intended use”. It calls this a human in the loop. Drafting an SOP or summarizing a document would usually sit here, as long as a qualified person owns the result.
The EMA multistakeholder workshop report (EMA/156789/2026, dated June 30, 2026) records a different mood. Its key messages say “No categorical prohibition was supported by industry; use and risk should be assessed in context.” That is industry’s position as recorded in a meeting report. It is not a decision, and the draft text has not changed.
What to wait on
A generative AI policy for critical GMP use is the item most likely to be wasted work. If you write one that assumes the ban stays, and the final text moves to a risk based control strategy, you rewrite it. If you write one that assumes the ban goes, you may have built controls for a use the final text still rules out. In my view the safer move today is a short rule that keeps LLMs to non critical uses with a named reviewer, and nothing more.
Explainability tooling and confidence thresholds are the other two. Draft section 8 names feature attribution methods such as SHAP values, and section 9 asks classifiers to log a confidence score and flag low scores as “undecided”. Both apply to the classification models the draft covers. Buying tools for them before the final text settles the scope would get ahead of the rule. Security work is different, because the draft revised Annex 11 already pushes it, so I would not hold that back.
If you are building the inventory now and want a second read on which uses sit in which group, get in touch.
Frequently asked questions
Is Annex 22 final?
No. On October 26, 2026 Annex 22 was still a draft. The Commission consulted on it from July 7 to October 7, 2025, and the EudraLex Volume 4 page did not list a final version on October 25, 2026. Treat every clause as a signal of direction, and check the EudraLex page again before you rely on a date or a wording.
Does Annex 22 ban LLMs?
The draft says generative AI and LLMs “should not be used in critical GMP applications”, meaning uses with direct impact on patient safety, product quality or data integrity. It allows non critical use when qualified and trained staff are responsible for checking that the outputs suit the intended use. That is draft wording, and EMA’s 2026 workshop discussed relaxing it.
When will Annex 22 be final?
No date has been published. I track the status and the expected next steps in AI validation in GxP has no settled playbook, and I update that page rather than this one when the Commission moves, so the two never disagree on dates.
Does Annex 22 apply to AI tools outside manufacturing?
The draft scope is computerized systems “used in the manufacturing of medicinal products and active substances”. A medical information assistant or a commercial tool sits outside it. Tools in quality systems that touch batch release, deviations or GMP records can fall inside it when the model’s output has direct impact on product quality or data integrity.