Short answer: As of October 2026, Annex 22 is still a draft and not in force. The July 2025 text allows only static, deterministic models in critical GMP applications and keeps generative AI out of them. Today, use generative AI to draft, summarize, search and train, with a named person signing off, and keep it away from critical GMP decisions and records.
As of 4 October 2026, Annex 22 on artificial intelligence is still a draft. The EU GMP guide on EudraLex Volume 4 lists no Annex 22, and Annex 11 and Chapter 4 are still the 2011 versions. The July 2025 draft allows only static, deterministic machine-learning models in critical GMP applications and says generative AI and large language models should not be used there. Outside critical uses, generative AI is allowed if a qualified person is responsible for every output.
So, today: use generative AI to draft, summarize, search and train, with a named person signing off, and keep it away from anything that decides or records something affecting patient safety, product quality or data integrity. That line may move.
A workshop report EMA published on 2 October 2026 says the drafting group has discussed allowing generative AI in GMP applications under a documented, risk-based control strategy. The EMA inspectors’ group aims to hand a final text to the European Commission in Q4 2026. No publication or application date has been announced.
This is general information, not regulatory advice. Your QA function and, where needed, your inspectorate decide how it applies to your site.
What is Annex 22, and is it in force?
Annex 22 is a proposed new annex to the EU GMP guide, the first EU GMP text written specifically for AI. The EMA GMP/GDP Inspectors Working Group drafted it together with PIC/S, the international scheme of GMP inspectorates. The Commission published it for stakeholder consultation from 7 July to 7 October 2025, together with a revised Annex 11 (computerised systems) and a revised Chapter 4 (documentation). PIC/S ran the same consultation for its own GMP guide.
None of the three is in force. A draft annex signals where inspectors are heading; it is not a requirement. Today, AI in a GMP system is inspected under the current Annex 11 and Chapter 4 and the general rules on validation and quality risk management.
What does the draft cover, and what does it exclude?
The draft runs to six pages and ten sections. Its scope is precise, so we quote it directly.
Where it applies: computerised systems used to manufacture 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, e.g. to predict or classify data.”
Static models only: dynamic models “which continuously and automatically learn and adapt performance during use” are not covered and “should not be used in critical GMP applications.”
Deterministic output only: models that, “when given identical inputs, provide identical outputs.” Probabilistic models that “might not provide identical outputs” are excluded from critical use.
Generative AI and LLMs: the draft “does not apply to Generative AI and Large Language Models (LLM), and such models should not be used in critical GMP applications.”
The opening for non-critical use: where generative models are used in applications without direct impact on patient safety, product quality or data integrity, “personnel with adequate qualification and training should always be responsible for ensuring that the outputs from such models are suitable for the intended use.”
For the models it covers, the draft asks a lot. An expert must write and approve the intended use before testing starts. Acceptance criteria “should be at least as high as the performance of the process it replaces,” which means you need to know how well your current manual process performs.
Test data must be kept independent of training data, and people who have seen the test data should not train the model unless paired with a colleague who has not. Generating test data with generative AI “is not recommended.” Models should log confidence scores and flag low-confidence cases as “undecided” rather than guess. After go-live, the model, its system and the whole process sit under change control.
The link to the EU AI Act
One detail most summaries miss: the glossary’s definition of an AI system is word for word the one in Article 3(1) of the EU AI Act. The two regimes are separate, but they share vocabulary. Our guide to what the EU AI Act requires of companies covers that side.
What counts as a critical GMP application?
The draft does not list critical applications. Its test is “direct impact on patient safety, product quality or data integrity.” In practice we ask two questions, which echo the terms regulators and industry used at EMA’s workshop and in the FDA’s draft AI guidance: how much does the model’s output drive the decision (model influence), and how bad is it if that decision is wrong (decision consequence)?
A model that accepts or rejects tablets is critical. So is one that writes values into a batch record, because the record is GMP data. A model that drafts a paragraph an expert then checks against sources, rewrites and approves through document control has low influence on the final record. It becomes critical once its output flows into a decision or record without that independent check.
That is why a rubber-stamp review does not help: if reviewers approve outputs unread, the human in the loop exists only on paper. Our guide to human-in-the-loop approval points covers how to design a review that holds up.
Which GMP tasks can generative AI do today?
The table below is our reading of the July 2025 draft, not an official list. Your QA function makes the call for each use and should record the reasoning. Each row pairs a task that fits non-critical use with a related task that falls under Annex 22 critical use.
Fine for generative AI with human review today | Annex 22 critical use (no generative AI, as drafted) |
|---|---|
Drafting an SOP revision that then goes through document control and approval | Deciding that a process change needs no revalidation |
Summarizing a deviation file for the investigator, with links to every source record | Classifying deviation severity or closing a deviation without a person deciding |
Training slides and quiz questions built from approved SOPs | Accept or reject decisions in automated visual inspection |
Searching SOPs and guidelines, with citations to the exact section | Reviewing batch records in a way that feeds QP certification |
First drafts of CAPA text, change descriptions and the narrative of an annual product review | Transcribing values from paper or instrument printouts into a GMP record |
Translating an approved document for a qualified reviewer | Adjusting a process parameter automatically based on a model’s prediction |
Two rules sit behind the left column: the output is a draft, never the record, and the reviewer has the sources in front of them and the competence to reject it. Under the draft Chapter 4, accountability for documents “produced or processed with artificial intelligence” rests with the regulated user either way.
How does Annex 22 connect to the revised Annex 11 and Chapter 4?
The three drafts were written as a set. Annex 22 says it “provides additional guidance to Annex 11 for computerised systems in which AI models are embedded.” Anything Annex 11 requires of a computerised system still applies to an AI system. Four parts of the revised Annex 11 draft matter most for AI:
Requirements even for software-as-a-service. A user requirements specification is needed whether a system is built in-house, bought off the shelf or “provided as-a-service” (6.1). That includes an LLM you reach through an API.
Supplier control. Relying on a vendor “does not change the requirements” and the regulated user “remains fully responsible” (7.1). Contracts should cover how new versions are released and whether you can test them first (7.5), which is the hard clause with a model vendor that updates monthly.
No added risk. Replacing a manual step with a system should cause no decrease in quality or data integrity and no increase in overall risk (2.8).
Signatures with meaning. Electronic signatures should prompt for their meaning, such as “reviewer or approver” (13.5). A human approval of AI output is a signature with a meaning, and it should be recorded as one.
Chapter 4 adds the documentation layer. Paragraph 4.23 requires full compliance, including Annex 11, where “decision making in manufacturing … is supported by automatic validation scripts or artificial intelligence (Annex 22).” Paragraph 4.24 puts accountability for the integrity of AI-produced records on the regulated user. Paragraph 4.25 requires AI support to be part of the quality system “regardless of the service located on premise or as a hosted service.”
Is the generative AI exclusion going to change?
Possibly. EMA held a multistakeholder workshop on 30 June and 1 July 2026. The event page says consultation responses suggested support for enabling generative AI in manufacturing and that EMA “is still considering the implications.”
The report, published on 2 October 2026, goes further. It says the drafting group has discussed widening the scope to dynamic, probabilistic and generative models, provided they meet the annex and are backed by “a fully documented, robust, risk-based control strategy.” It also says “concerns remain about the elements needed for such a control strategy.” Industry argued against any ban based on a technology label. Regulators pressed on hallucination, automation bias and how an inspector could verify layered controls.
Read the report as a record of discussion, not a decision. Three of its key messages hold whatever the final text says. Oversight “cannot compensate for inadequate validation.” Human in the loop is “one possible control, not a default requirement for all AI uses.” And a third-party model or API that cannot provide the outputs, logs and evidence needed for control cannot be adequately justified.
What should you document now?
Whichever way the final text goes, it will ask for evidence. The draft says its principles “may be considered where applicable” for non-critical uses. We would build this file for every generative AI use in a GMP area today:
Intended use. The task, the inputs and their variations, what is out of scope, and who reviews. Approved by the process owner before testing, as section 3 of the draft asks.
Classification rationale. Why the use is non-critical, in two or three sentences, signed by QA.
Test data. Real past cases with known good answers, not synthetic ones, kept separate from anything used to build prompts or retrieval.
Acceptance criteria. Set before testing and at least as good as the current manual process. Measure that process first.
Human-in-the-loop definition. Who reviews, their training, what they check, and a record of the AI output next to the approved version. Sections 3.3 and 10.5 of the draft ask for operator monitoring and records where human review replaces some testing.
Change control. Model version, prompts, retrieval sources and vendor release notes under change control, with a retest trigger when any of them changes (section 10.1).
Supplier evidence. The vendor assessment and contract terms Annex 11 section 7 describes.
Monitoring. Edit rate, rejection rate and escalations reviewed on a schedule, folded into the periodic review.
What is the exact status as of October 2026?
Date | Status |
|---|---|
January 2011 | Date of the current Annex 11 and Chapter 4, still the versions on EudraLex Volume 4 on 4 October 2026 |
27 March 2025 | EMA inspectors’ 2025–2027 work plan targets Q1 2026 for final Annex 11 and Chapter 4 texts |
7 July 2025 | Drafts of Annex 22, Annex 11 and Chapter 4 published for joint EU–PIC/S consultation |
7 October 2025 | Consultation closed |
14 January 2026 | EMA and FDA publish ten guiding principles for AI across the medicine lifecycle, including manufacturing |
23 March 2026 | New 2026–2028 work plan: final texts of Annex 22, Annex 11 and Chapter 4 to the Commission targeted for Q4 2026 |
30 June – 1 July 2026 | EMA workshop on generative AI and control measures for Annex 22 |
23 September 2026 | PIC/S GMP guide revised to PE 009-18, covering Annex 19 only |
2 October 2026 | EMA publishes the workshop report |
Q4 2026 (target) | Final texts due to the Commission; no publication or application date announced |
Note the slip: the 2025 plan aimed for Q1 2026 on Annex 11 and Chapter 4, and the 2026 plan moved all three to Q4 2026. That is the date for handing the text to the Commission, not for publication. The same plan schedules inspector training on the three documents “during 2026 - 2028.”
What about US and UK sites?
The FDA has no Annex 22 equivalent. Its January 2025 draft guidance on AI for regulatory decision-making excludes AI used “for operational efficiencies (e.g., internal workflows, resource allocation, drafting/writing a regulatory submission)” that does not affect patient safety, drug quality or study reliability. FDA’s AI in drug development page still lists it as a draft.
The CGMP rules already apply, though. In Warning Letter 320-26-58 of 2 April 2026, the FDA cited a firm that used AI agents to create specifications, procedures and master records without proper review. The letter states that “any output or recommendations from an AI agent must be reviewed and cleared by an authorized human representative of your firm’s QU.” The firm had not done process validation either; it told investigators the AI agent never said it was required.
For the UK and other PIC/S countries: Annex 22 is a joint EU–PIC/S text. When PIC/S adopted the revised Annex 19 in September 2026, it invited its non-EEA participating authorities to transpose it. Expect the same route for Annex 22.
What to watch for when the final text lands
The scope paragraph. Whether “should not be used in critical GMP applications” survives for generative AI, softens to a control-strategy requirement, or moves to an appendix.
A definition of critical. The draft has none. The workshop discussed model influence, uncertainty, complexity and detectability as factors.
The application date. The last EU GMP change, Annex 19, was published on 23 June 2026 and applied from 24 September 2026. Annex 22 may get a different lead time, so read the date on the final text itself.
Rules for existing systems. Whether models already in use need retesting under the new annex.
Cross-references. Annex 22 section 4.3 points to “Annex 11 2.7,” but in the Annex 11 consultation draft, 2.7 is security and the no-added-risk principle is 2.8. Numbering will likely change.
Reviewer competence. Brian Drapeau argued in PharmTech that the person and system should be qualified together. Watch whether the final text says how reviewer competence is shown.
We will update this post when the final text is published.
Where to start
Start with an inventory of every AI tool at your GMP sites, including AI features inside systems you already run. Classify each use as critical or non-critical and write down why. Build the documentation file above for the non-critical uses you keep, and pause anything in the right-hand column of the table.
Your QA decides how each use is classified. We build the drafting and review workflow around that decision, with the approval points and records your QA asks for. To check whether a workflow fits, book a fit check. For the oversight design, read our guide to human-in-the-loop approval points. For the wider legal picture, see what the EU AI Act requires of companies in 2026.
Frequently asked questions
Is Annex 22 in force?
No. The Commission published Annex 22 for consultation from 7 July to 7 October 2025, together with a revised Annex 11 and Chapter 4. None of the three is in force. Today, AI in a GMP system is inspected under the current Annex 11 and Chapter 4 and the general rules on validation and quality risk management.
Is generative AI allowed in GMP?
Not in critical GMP applications under the July 2025 draft. Outside critical uses, it is allowed if qualified, trained personnel are responsible for every output. EMA’s workshop report of 2 October 2026 says the drafting group has discussed widening the scope under a documented control strategy. Nothing is decided yet.
What counts as a critical GMP application?
The draft’s test is direct impact on patient safety, product quality or data integrity, and it gives no list. A model that accepts or rejects tablets, or writes values into a batch record, is critical. A draft that an expert checks, rewrites and approves through document control has low influence on the final record.
What should you document for generative AI in GMP now?
For each use: the intended use, why it is non-critical, real test data and acceptance criteria set before testing. Add who reviews and how, change control for model versions and prompts, supplier evidence, and monitoring of edit and rejection rates. Your QA function decides how each use is classified.
When will the final Annex 22 be published?
No publication or application date has been announced. The EMA inspectors’ group aims to hand a final text to the European Commission in Q4 2026. That is the date for the handover, not for publication, so read the application date on the final text itself.
Related reading
Sources
European Commission, Stakeholders’ Consultation on EudraLex Volume 4: Chapter 4, Annex 11 and New Annex 22, 7 July to 7 October 2025.
European Commission, Draft Annex 22: Artificial Intelligence, consultation version, July 2025.
European Commission, Draft revised Annex 11: Computerised Systems, consultation version, July 2025.
European Commission, Draft revised Chapter 4: Documentation, consultation version, July 2025.
European Commission, EudraLex Volume 4: Good Manufacturing Practice guidelines, accessed 4 October 2026.
PIC/S, Joint stakeholders consultation on the revision of Chapter 4, Annex 11 and the new Annex 22, 7 July 2025.
PIC/S, Revision of PIC/S GMP Guide (PE 009-18), 23 September 2026.
EMA, Multistakeholder workshop on expert contributions to artificial intelligence guidance development (Annex 22), event page, 30 June to 1 July 2026.
EMA, Report: Multistakeholder workshop on expert contributions to AI guidance development (Annex 22), EMA/156789/2026, published 2 October 2026.
EMA, 3-year work plan for the GMDP Inspectors Working Group, 2026–2028, published 23 March 2026.
EMA, 3-year work plan for the GMDP Inspectors Working Group, 2025–2027, EMA/INS/GMP/50356/2025, published 27 March 2025.
EMA, EMA and FDA set common principles for AI in medicine development, 14 January 2026, and the Guiding principles of good AI practice in drug development.
FDA, Considerations for the Use of Artificial Intelligence to Support Regulatory Decision-Making for Drug and Biological Products, draft guidance, January 2025 (PDF).
FDA, Artificial Intelligence for Drug Development, page content current as of 1 May 2026.
FDA, Warning Letter 320-26-58, Purolea Cosmetics Lab, 2 April 2026.
European Commission, AI Act Service Desk, Article 3: Definitions, Regulation (EU) 2024/1689.
Brian Drapeau, Europe Tried to Ban Generative AI From Critical GMP. The Ban May Not Survive. It Does Not Matter., PharmTech, 13 August 2026.