Status on 20 October 2026: a consultation draft, not adopted guidance
Status as checked on the European Commission's pages on 20 October 2026. The Commission's consultation page for EudraLex Volume 4 Chapter 4, Annex 11 and new Annex 22 shows the consultation as closed, with an opening date of 7 July 2025 and a deadline of 7 October 2025. The page says the three documents were drafted by the EMA GMDP Inspectors Working Group in cooperation with PIC/S. The Commission's EudraLex Volume 4 page still lists Annex 11 in its January 2011 revision and does not list an Annex 22. Annex 22 is therefore a consultation draft, not EU GMP guidance in force, and its wording may change before adoption.
Everything below quotes or paraphrases the consultation draft (the six-page PDF linked in the sources), and clause numbers refer to it. Where this page gives our reading of what a clause means for an AI agent, it says so. This page is general information, not legal or regulatory advice. Decisions on GMP scope, criticality and validation belong to your quality and regulatory functions.
Atlas Bio's answer on how to validate an AI agent used in a GxP process covers validation steps. This page covers the draft annex itself: what it covers, what it excludes, and what each group of clauses asks for.
What the draft covers
Section 1 opens: 'This annex applies to all types of computerised 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, e.g. to predict or classify data.' It describes itself as 'additional guidance to Annex 11 for computerised systems in which AI models are embedded.'
Three conditions narrow that scope. The model must be machine learning that has 'obtained their functionality through training with data, rather than being explicitly programmed.' It must be static, meaning it does not 'adapt their performance during use by incorporating new data'; the glossary calls this a frozen model whose parameters are finally set. Its output must be deterministic: 'when given identical inputs, provide identical outputs.' A model may consist of several individual models, each automating a specific GMP process step.
Two scope points are easy to miss. The setting is manufacturing of medicinal products and active substances. That excludes discovery and the conduct of clinical trials, but not the manufacture of investigational medicinal products: making clinical trial supplies is GMP manufacturing, and EudraLex Volume 4 lists the GMP rules for it. In our reading, AI used in GMP quality control testing, for example to classify chromatograms or microscopy images, can also fall within the stated scope, because the draft's trigger is direct impact on product quality or data integrity, not the production line itself. The second point is the trigger: a critical application, defined by direct impact on patient safety, product quality or data integrity. Data integrity is listed alongside patient safety and product quality, so a model that classifies or alters GMP records can be critical even if it never touches product.
Advanced therapy medicinal products, including gene therapies, are a special case. Their EU GMP is set out in Part IV of EudraLex Volume 4, which states: 'Other documents developing GMP requirements for medicinal products which are contained in Volume 4 are not applicable to ATMPs, unless specific reference thereto is made in these Guidelines.' In our reading, Annex 22 would therefore reach ATMP manufacture only if Part IV refers to it. Part IV already refers to Annex 11 for validating computer equipment, and Annex 22 calls itself additional guidance to Annex 11, so the question is open. Check the current status of Part IV on the day you rely on this.
What the draft excludes, and why that matters for agents
Section 1 excludes three kinds of model and says each should stay out of critical GMP use. Dynamic models that 'continuously and automatically learn and adapt performance during use' are 'not covered by this document, and should not be used in critical GMP applications.' Models with a probabilistic output, which 'when given identical inputs, might not provide identical outputs', are treated the same way. Then: 'the document does not apply to Generative AI and Large Language Models (LLM), and such models should not be used in critical GMP applications.'
For non-critical uses, the draft continues: 'If used in non-critical GMP applications, which do not have 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, i.e. a human-in-the-loop (HITL) and the principles described in this document may be considered where applicable.'
Most AI agents in use today are built on a large language model that plans steps, calls tools and writes text. Our reading is that the draft gives such an agent two positions in GMP manufacturing. It can be used in a non-critical application, with a qualified, trained person responsible for its outputs. Or it is kept out of the critical application. The draft offers no third route of validating an LLM more heavily so that it can be used in a critical application. The table applies that reading to common agent set-ups. It is illustrative, and whether a given use is critical is a decision for your quality function.
| Agent set-up (illustrative) | Position under the draft, in our reading | Draft text it rests on |
|---|---|---|
| LLM agent drafting a deviation summary for a trained person who writes or verifies the content against the source records and approves it; the person, not the agent, determines the root cause, the product-impact assessment and the content of the approved record | Arguable as non-critical, with a qualified, trained person responsible for every output; your quality function may class it as critical where the agent's text shapes the product-impact assessment, batch disposition or QP certification | Section 1: non-critical use with a human-in-the-loop; the principles 'may be considered where applicable' |
| LLM agent entering values into, or changing, a GMP record | Likely critical, because of direct impact on data integrity; the draft says LLMs should not be used there. Adding a reviewer does not by itself make the use non-critical (see What to do now) | Section 1: critical means direct impact on patient safety, product quality or data integrity; LLMs 'should not be used in critical GMP applications' |
| LLM agent deciding whether a batch or test result is acceptable | Critical; the draft says LLMs should not be used here | Section 1, as above |
| Agent that calls a separate static, deterministic machine-learning model, for example an image classifier in visual inspection, whose output decides product acceptance | The classifier is the kind of model the draft covers, and in a critical use the draft expects Sections 2 to 10 to apply to it. The LLM agent should not sit in the critical path: it should not choose or transform the classifier's input, interpret or relay its output into the acceptance decision, or make that decision. Any agent in that path becomes part of the critical application | Section 1: scope limited to static, deterministic machine learning, and LLMs 'should not be used in critical GMP applications'; Section 10.1: change control of the model, 'the system it is implemented in, and the whole process' |
| Agent that updates or retrains its own model from new production data during use | Dynamic model; should not be used in critical GMP applications | Section 1: dynamic models are 'not covered by this document, and should not be used in critical GMP applications' |
Intended use: Sections 2 and 3
Section 2 sets three principles. Section 2.1 asks for 'close cooperation between all relevant parties during algorithm selection, and model training, validation, testing and operation', naming process subject matter experts (SMEs), QA, data scientists, IT and consultants. Section 2.2 says documentation 'should be available and reviewed by the regulated user irrespective of whether a model is trained, validated and tested in-house or whether it is provided by a supplier or service provider.' Section 2.3 bases all activities on 'the risk to patient safety, product quality and data integrity.'
Section 2.2 matters for anyone using a vendor's model. Buying the model does not move the documentation expectation to the vendor; the regulated user should still have the documentation available and review it.
Section 3.1 asks that the intended use and 'the specific tasks it is designed to assist or automate should be described in detail based on an in-depth knowledge of the process'. That description includes 'a comprehensive characterisation of the data the model is intended to use as input and all common and rare variations; i.e. the input sample space.' Limitations and 'possible erroneous and biased inputs should be identified.' A process SME should be responsible for the adequacy of the description, which 'should be documented and approved before the start of acceptance testing.' Section 3.2 says that, where applicable, the input sample space should be divided into subgroups, for example by decision output, site or equipment, material characteristics or defect type.
Section 3.3 covers the human-in-the-loop for in-scope models: 'Where a model is used to give an input to a decision made by a human operator (human-in-the-loop), and where the effort to test such model has been diminished, the description of the intended use should include the responsibility of the operator. In this case, the training and consistent performance of the operator should be monitored like any other manual process.'
Acceptance criteria and test data: Sections 4 to 7
Section 4 sets the performance bar. Test metrics should be 'case dependent' and measure performance 'according to the intended use'; for a classifier, examples given are a confusion matrix, sensitivity, specificity, accuracy, precision and F1 score (4.1). Acceptance criteria may differ by subgroup, and a process SME should be responsible for defining them, documented and approved before acceptance testing (4.2). Section 4.3, headed 'No decrease', says acceptance criteria 'should be at least as high as the performance of the process it replaces', which 'implies, that the performance should be known for the process which is to be replaced by a model.'
Section 5 governs test data. It should be representative of and expand 'the full sample space of the intended use', be stratified and include all subgroups (5.1). It should be 'sufficient in size to calculate the test metrics with adequate statistical confidence' (5.2). Labels should be verified 'following a process that ensures a very high degree of correctness' (5.3). Pre-processing should be pre-specified, with a rationale that it represents intended-use conditions (5.4), and any cleaning or exclusion 'documented and fully justified' (5.5). Section 5.6 bears directly on agents: 'Generation of test data or labels, e.g. by means of generative AI, is not recommended and any use hereof should be fully justified.' In our reading, an LLM agent should not label or generate the test set for a model the draft covers unless that choice is justified on the record.
Section 6 keeps test data independent. Effective technical or procedural controls should ensure that data used to test a model is not 'used during development, training or validation of the model' (6.1). Where test data is split from a pool before training, staff who developed the model must never have had access to it, and it should sit under access control and audit trail with no copies outside the repository (6.2). It should be recorded which data was used for testing, when and how many times (6.3). Where test data comes from physical objects, the objects used for the final test should not have been used to train or validate the model, unless features are independent (6.4). Controls should keep staff with access to test data out of training and validation of the same model, or, where that is impossible, have them work in a pair with a colleague who has not had that access, the four-eyes principle (6.5).
Section 7 sets out test execution. The test should show the model is 'fit for intended use' and 'generalising well', including detecting over- or underfitting (7.1). A test plan, developed with a process SME, should be prepared and approved before testing. It should set out the intended use, the metrics and acceptance criteria, the test data, a test script and how the metrics are calculated (7.2). Deviations, failures against criteria and omission of any test data should be documented, investigated and fully justified (7.3). Test documentation should be retained with the intended-use description, the characterisation of test data and the test data itself, and access-control and audit trail records for the test data should be retained 'similarly to other GMP documentation' (7.4).
Explainability, confidence, operation and human review: Sections 8 to 10
Section 8.1 asks that, during testing of models in critical GMP applications, 'systems should capture and record the features in the test data that have contributed to a particular classification or decision (e.g. rejection)', naming feature attribution techniques such as SHAP values or LIME, or heat maps. Section 8.2 says a review of those features should be part of approving test results. Section 9.1 asks that, when a model used to predict or classify data is tested, the system logs its confidence score for each prediction or classification where applicable. Section 9.2 asks for an appropriate threshold so that predictions or classifications are made only when suitable. Where the confidence score is very low, the draft says it should be considered whether the model should flag the outcome as 'undecided' rather than make a potentially unreliable prediction; it does not require the flag.
Section 10 covers operation. The model, the system it sits in and the whole process should be placed under change control before deployment. Any change to the model, the system or the process 'should be documented and evaluated to determine if the model needs to be retested', and 'any decision not to conduct such retest should be fully justified' (10.1). A tested model should go under configuration control, with 'effective measures' to detect any unauthorised change (10.2). Performance against the defined metrics should be monitored regularly (10.3), and so should whether input data stays within the model's sample space, with metrics for drift (10.4).
Section 10.5 is the operational human-review clause: 'When a model is used to give an input to a decision made by a human operator (human-in-the-loop), and where the effort to test such model has been diminished, records should be kept from this process. Depending on the criticality of the process and the level of testing of the model, this may imply a consistent review and/or test of every output from the model, according to a procedure.'
In our reading, Sections 3.3 and 10.5 together allow reduced testing of an in-scope (static, deterministic) model to be balanced by a defined, trained, monitored and recorded human review. For an LLM agent in a non-critical use, the Section 1 human-in-the-loop is a separate expectation that applies in any case. In our reading, the logging, review records and change triggers in Sections 9 and 10 are the principles an agent operator is most likely to borrow, because they translate directly to model versions, prompts and tools. Our validation answer works through that translation.
How Annex 22 relates to Annex 11
Annex 22 does not replace Annex 11. It describes itself as 'additional guidance to Annex 11 for computerised systems in which AI models are embedded', so an AI model in GMP use sits inside a computerised system that Annex 11 already governs. The current Annex 11 (January 2011 revision) applies 'to all forms of computerised systems used as part of a GMP regulated activities.' Its principle that where a computerised system replaces a manual operation 'there should be no resultant decrease in product quality, process control or quality assurance' carries the same idea as Annex 22 Section 4.3. Section 4.3 cross-refers to 'Annex 11 2.7'; in the revised Annex 11 draft, 2.7 is the security clause and the matching 'No risk increase' clause is 2.8, so treat the cross-reference as one the final text may correct. Section 1 of Annex 11 asks for risk management throughout the lifecycle, 'taking into account patient safety, data integrity and product quality', the same three factors that define a critical application in Annex 22.
The Commission consulted on a revised Annex 11 and a revised Chapter 4 (Documentation) at the same time and on the same dates as Annex 22. The consultation page says the revised Annex 11 strengthens lifecycle quality risk management, supplier and service-provider oversight, data integrity, audit trails, electronic signatures and security. It describes the three documents 'taken together' as a framework for IT in pharmaceutical manufacturing. The revised Annex 11 is also a draft. In practice an AI model or agent in a GMP process has to meet Annex 11 as a computerised system. Where it is a static, deterministic model in a critical use, it would be expected to meet Annex 22 if Annex 22 is adopted in this form.
How Annex 22 relates to the EU AI Act
Annex 22 is GMP guidance under EU medicines law. The EU AI Act, Regulation (EU) 2024/1689, is a separate regulation with its own scope, roles and dates. The draft's glossary defines 'AI system' using the same wording as Article 3(1) of the AI Act: 'a machine-based system that is designed to operate with varying levels of autonomy and that may exhibit adaptiveness after deployment'. The draft does not otherwise refer to the AI Act, and meeting one does not show compliance with the other.
Under the AI Act, high-risk status comes through Article 6: AI that is, or is a safety component of, a product covered by the Union harmonisation legislation in Annex I and needing third-party conformity assessment, or AI in an Annex III area. Annex I includes the medical device and in vitro diagnostic regulations but does not list EU medicines legislation. Pharmaceutical manufacturing is not one of the Annex III areas. So, in our reading, an AI model used in GMP manufacturing is not high-risk under the AI Act merely because it is critical under Annex 22; the two classifications answer different questions. Other routes can still apply. Annex I also lists the Machinery Regulation (EU) 2023/1230, which the AI Omnibus moved to Section B of Annex I, so AI that is a safety component of manufacturing machinery may reach the AI Act that way, with the requirements applied through the Machinery Regulation rather than directly; how that works for a given system is a question for your legal function. AI in a companion diagnostic, or in a device placed on the market under the medical device regulation in its own right, can reach Annex I through the IVD or medical device regulations; whether the device part of an integral drug-device combination does is unsettled and a question for your legal function. And if AI were used for the operator monitoring that Annex 22 Section 3.3 asks for, AI that monitors or evaluates staff can fall under the Annex III employment area whatever the sector. Article 4, which requires providers and deployers to take measures to support the AI literacy of staff using AI systems in scope, has applied since 2 February 2025, is not limited to high-risk systems, and since the AI Omnibus took effect on 27 July 2026 states that it does not require anyone to guarantee a specific level of AI literacy for any individual.
Atlas Bio's answer on whether the EU AI Act applies to AI in drug discovery covers the Act's scope, its research exclusion and the dates set by the AI Omnibus. Record the two positions separately for each system: whether it is critical under Annex 22, and its provider or deployer role and risk class under the AI Act.
What to do now
Because the draft is not adopted, Annex 22 itself is not yet EU GMP guidance in force. That does not leave AI unregulated in GMP: any AI agent or model used in GMP activities is already part of a computerised system under the current Annex 11 (2011), and Chapter 4 documentation and data integrity expectations already apply to the records it touches. The draft shows how the GMP inspectors drafting the guidance think about AI, and it is a reasonable basis for planning. For each AI agent or model in or near GMP manufacturing, record what it is built on (LLM or other generative model, dynamic model, or static deterministic model) and whether its use is critical as the draft defines it. For any LLM-based agent, record who the qualified, trained person responsible for its outputs is.
Where an LLM agent's use would be critical, the draft's position is not to use it there. Either keep the agent out, or redesign the use so that a trained person makes the decision and creates or verifies the record from the source data, with the agent's output only as a drafting aid. The draft does not say that adding a review step turns a critical use into a non-critical one. Whether a redesigned use is still critical is for your quality function to decide, and in our reading a review that only confirms the agent's output is unlikely to change that. Where a static model does the critical work, start the Section 3 intended-use description and the Section 6 test data controls early, because they are hard to reconstruct later. Atlas Bio's answer on what AI agent governance in life sciences covers sets out the wider governance this fits into. Its answer on what an AI agent inventory should record includes fields for model and version, autonomy level and human checkpoint.
Knowing which of your AI uses sit near GMP manufacturing, and which are critical, comes first. Atlas Bio's Secure AI for Life Sciences starts with an AI risk assessment of current AI use and data exposure. Status, October 2026. In use in Atlas Bio's own agent platform today: a separate verification step that checks agent work, and a hold that stops a task until a person decides. Built and tested on synthetic data, not yet used with a client: the AI risk assessment tooling (AI discovery, data-flow and vendor-terms mapping, and a NIST AI RMF gap report) and a record of what data an AI used, what it produced and who reviewed it. In development: a customer-controlled AI environment in your own Azure tenant, an AI gateway, agent permissions, evidence linkage and notifications. Each engagement states which controls are built. Decisions on GMP scope, criticality, validation and compliance stay with your quality and regulatory functions.
Is EU GMP Annex 22 in force?
No. As checked on 20 October 2026, the Commission's consultation page shows the consultation closed, after running from 7 July to 7 October 2025. The EudraLex Volume 4 page lists Annex 11 in its January 2011 revision and no Annex 22. It is a consultation draft and may change before adoption.
Can an LLM-based AI agent be used in GMP manufacturing under Annex 22?
Not in a critical application. The draft says generative AI and large language models should not be used where there is direct impact on patient safety, product quality or data integrity. In non-critical uses, personnel with adequate qualification and training should always be responsible for ensuring the outputs are suitable for the intended use.
Does Annex 22 apply to AI in drug discovery or clinical trials?
Its stated scope is computerised systems used in the manufacturing of medicinal products and active substances. Discovery and the conduct of clinical trials fall outside that scope, and other guidance is more relevant there, such as EMA's reflection paper on AI in the medicinal product lifecycle and EMA's guideline on computerised systems in clinical trials. Manufacturing investigational medicinal products for a trial is GMP manufacturing, so AI used in that manufacturing is within the draft's stated scope. For gene and other advanced therapies, EudraLex Part IV applies other Volume 4 documents only where it refers to them.
Can generative AI be used to create test data under Annex 22?
Section 5.6 says generation of test data or labels, for example by generative AI, is not recommended, and any use should be fully justified. Section 5.3 asks that test labels be verified by a process that ensures a very high degree of correctness.
Is an AI model that is critical under Annex 22 high-risk under the EU AI Act?
Not for that reason alone. AI Act high-risk status comes from Article 6, through Annex I products such as medical devices that need third-party conformity assessment, or the Annex III areas. EU medicines legislation is not in Annex I, and manufacturing is not an Annex III area. Other routes can apply, for example AI that is a safety component of machinery (Annex I Section B since the AI Omnibus, where the requirements apply through the Machinery Regulation), or AI in a companion diagnostic or a stand-alone medical device. The two regimes are assessed separately.
How does Annex 22 relate to Annex 11?
It describes itself as additional guidance to Annex 11 for computerised systems in which AI models are embedded. Annex 11 continues to govern the computerised system; Annex 22 adds model-specific expectations on intended use, acceptance criteria, test data, explainability, confidence and operation. A revised Annex 11 was consulted on at the same time and is also a draft.
- European Commission: Stakeholders' consultation on EudraLex Volume 4 - Chapter 4, Annex 11 and new Annex 22 (status closed; opened 7 July 2025, deadline 7 October 2025; checked 20 October 2026)
- European Commission: Annex 22 Artificial Intelligence - consultation draft (PDF)
- European Commission: EudraLex Volume 4 - Good Manufacturing Practice guidelines (lists Annex 11, January 2011 revision; no Annex 22 as of 20 October 2026)
- European Commission: EudraLex Volume 4, Annex 11 - Computerised Systems (in operation since 30 June 2011)
- European Commission: Revised Annex 11 Computerised Systems - consultation draft (PDF)
- European Commission: EudraLex Volume 4, Part IV - Guidelines on Good Manufacturing Practice specific to Advanced Therapy Medicinal Products (PDF)
- AI Act Service Desk (European Commission): Article 3 - Definitions
- AI Act Service Desk: Article 4 - AI literacy
- AI Act Service Desk: Article 6 - Classification rules for high-risk AI systems
- AI Act Service Desk: Annex I - List of Union harmonisation legislation
- AI Act Service Desk: Annex III - High-risk AI systems referred to in Article 6(2)
- European Medicines Agency: Reflection paper on the use of artificial intelligence in the medicinal product lifecycle (September 2024)
- European Medicines Agency: Guideline on computerised systems and electronic data in clinical trials (EMA/INS/GCP/112288/2023)
Talk to Atlas Bio.
Sponsors, researchers and clinicians: tell us the decision you are trying to make and we will show what the platform can and cannot answer.
Contact Us →