Annex 22 is the EU’s first proposed GMP annex written for artificial intelligence, and its message is narrow: only static, deterministic machine learning models belong in critical GMP uses. It was published on 7 July 2025 as a consultation draft, alongside a revised Annex 11 and Chapter 4, and the consultation closed on 7 October 2025. It is not adopted, and no adoption or effective date has been announced. Check EudraLex Volume 4 for the current status before you rely on any of it.
That makes it guidance to prepare against, not a rule to comply with today. What is in force for computerized systems is Annex 11 (January 2011). This post walks through what the draft Annex 22 expects, in the order it is written.
What Annex 22 Covers and What It Excludes
The draft applies to computerized systems used in manufacturing medicinal products 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, such as predicting or classifying data. It is described as additional guidance to Annex 11 for systems with embedded AI models. It addresses machine learning models whose behavior was learned from training data, not explicitly programmed.
Three limits define the scope:
- Static models only. A model that keeps learning and adapting during use is not covered, and the draft says such a model has no place in critical GMP use.
- Deterministic output only. A model that can give different outputs for identical inputs is not covered, and the same exclusion from critical use applies.
- No generative AI or large language models in critical applications. In non-critical uses, those with no direct impact on patient safety, product quality or data integrity, the draft expects qualified, trained personnel to stay responsible for confirming that outputs suit the intended use: a human in the loop.
Read together: a frozen classifier that inspects tablets can sit inside the draft’s scope. A chatbot that drafts a deviation summary cannot be a critical-use component, though it might be used under human review where the impact is not direct.
Annex 22 Principles: People, Documents, Risk
The three principles in section 2 are familiar from Annex 11. Personnel: process experts, QA, data scientists, IT and consultants, among others, cooperate from algorithm selection through operation, with defined responsibilities and appropriate access. Documentation: the regulated user has and reviews the documentation whether the model was built in-house or supplied by a vendor. Risk: the activities scale with the risk to patient safety, product quality and data integrity.
The second one matters for purchased AI. A supplier’s data science team does not move the review duty off your desk.
Intended Use and Subgroups in Annex 22
Section 3 starts where validation always starts. The intended use should be described in detail, based on deep knowledge of the process the model sits in. That includes a thorough characterization of the input data, with common and rare variations (the input sample space), known limitations, and possible erroneous or biased inputs. A process subject matter expert is responsible for the description’s adequacy, and it is approved before acceptance testing starts.
Where it applies, the input space is split into subgroups. The draft’s examples are the decision output (accept or reject), site or equipment, material or product characteristics, and task-specific features such as defect types and severity. Subgroups matter because a model that scores well overall can fail on one line, one product or one defect type.
Where a human makes the final decision on a model’s input and the testing effort was reduced, the intended use should state the operator’s responsibility, and the operator’s training and performance are monitored like any manual process.
Acceptance Criteria and the Cross-Reference Slip
Section 4 asks for test metrics suited to the case, and for an accept-or-reject classifier it lists a confusion matrix, sensitivity, specificity, accuracy, precision or F1 score as candidates. Acceptance criteria are set for those metrics, may differ by subgroup, and are defined and approved by the process expert before testing starts.
Then comes 4.3, “No decrease”: the model must be held to criteria no lower than what the current process achieves, so you have to measure that current performance first. If you have no measured baseline for a manual inspection, getting one is real work.
A drafting detail to know about: 4.3 points to “Annex 11 2.7” for this principle. In the July 2025 draft Annex 11, 2.7 is the security principle. The no-risk-increase principle is 2.8. It looks like a numbering slip in the consultation text, and we cite 2.8 and note the discrepancy. Do not copy the wrong number into your procedures.
Annex 22 Test Data and Independence
Sections 5 and 6 are the heart of the draft.
Test data (section 5). It should represent the full sample space, be stratified, include all subgroups and reflect rare variations, with the selection rationale documented. It should be large enough for each subgroup to calculate the metrics with adequate statistical confidence. Labels should be verified to a very high degree of correctness, for example by multiple independent experts, validated equipment or laboratory tests. Pre-processing is pre-specified and justified. Cleaning or excluding data is documented and justified. Using generative AI to create test data or labels is discouraged, and any such use needs a full justification.
Independence (section 6). Test data must not be used in development, training or validation of the model. You can capture it only after training and validation finish, or split it from the pool before training starts. If you split early, developers must never have had access, and the data sits in a repository with access control and an audit trail, with no copies outside it. Keep a record of which data went into each test, the date, and the number of times it was used. For physical objects, items used in the final test should be ones the model never saw in training or validation, unless the relevant features are independent.
Section 6.5 adds staff independence: people who have had access to test data should be kept out of training and validation work on that model. Where an organization cannot keep that separation, the draft allows a four-eyes arrangement, in which such a person works only in pair with a colleague who never had access.
Test Execution, Explainability and Confidence
Section 7 asks that testing shows the model is fit for intended use and “generalizing well”, including checks for over- or underfitting. Before testing, an approved test plan should hold the intended use summary, predefined metrics and criteria, the test data reference, a test script and the metric calculation. Deviations, failures against criteria, and any test data left unused are documented, investigated and justified. Everything is retained like other GMP documentation, along with the access-control and audit-trail records for the test data.
Section 8, explainability, expects systems to record which features in the test data drove a classification or decision, using techniques such as feature attribution (SHAP, LIME) or heat maps where applicable. A review of those features is part of approving the test results, to check the model decides on relevant, appropriate features. A model that rejects tablets by looking at the tray, not the tablet, would pass accuracy and fail this review.
Section 9, confidence, expects the system to log a confidence score for each prediction where applicable, and the model to have a threshold. If confidence is very low, consider an “undecided” output instead of a shaky prediction. An undecided result can then be routed to a person.
Operating a Model Under Annex 22
Section 10 covers life after release:
- Change control. The model, its system and the process around it go under change control before deployment. Any change, including a change to physical objects the model reads, is evaluated for retest, and a decision not to retest is justified.
- Configuration control. The tested model is under configuration control, with effective measures to detect unauthorized change.
- Performance monitoring. Regularly monitor metrics for shifts in the system, such as a change in lighting.
- Input monitoring. Regularly check that input data still falls within the model’s sample space, with defined drift metrics.
- Human review. Where a person acts on model output and testing effort was reduced, keep records, which may mean reviewing every output according to a procedure.
The design logic is that a frozen model is only safe if you can prove it stays frozen and its world stays the same.
Preparing Without an Adoption Date
You can act on the parts that are good practice today. Inventory every AI model in a GMP process and classify its use as critical or not. Retire or restrict anything probabilistic or generative from critical use. For each critical model, write the intended use, subgroups and acceptance criteria, build independent test data, and set up monitoring. Fit all of this into your existing risk-based validation under Annex 11 (2011) and the EU GMP Annex 11 work you already do, and keep requirements and tests linked as the traceability matrix post describes.
The AI Act is a separate regulation with its own scope. Whether a GMP model also falls under it is a legal question this post does not answer; our guide to EU AI Act Article 15 covers that regulation’s accuracy, robustness and cybersecurity duties for high-risk systems. Read the drafts themselves on the EU consultation page, and see the 21 CFR Part 11 requirements for the US side of the same systems.
For the paperwork side, the 21 CFR Part 11 Toolkit includes an AI Models in GMP Applications Procedure, an AI Model Intended Use, Test Plan and Report, and an AI Test Data and Performance Monitoring Workbook, mapped to the draft Annex 22. It holds 105 templates (68 Word and 37 Excel) for $99. They are templates to adapt, and they do not make any system compliant.
Annex 22 FAQ
Is Annex 22 in force?
No. It is a July 2025 consultation draft, and no adoption date has been given. Annex 11 (2011) remains the in-force text for computerized systems. Use the draft as a guide to where expectations may go, not as a legal duty.
Can we use a large language model in GMP?
Not in critical applications, according to the draft: generative AI and LLMs are ruled out wherever they directly affect patient safety, product quality or data integrity. In non-critical uses, the draft expects a trained person to remain responsible for the output.
What counts as a critical application?
The draft defines it by impact: direct effect on patient safety, product quality or data integrity. Decide per use, record the reasoning, and have QA approve it.
Does a model that learns on the job qualify?
Not under the draft. Dynamic models that adapt during use are outside its scope and are ruled out of critical applications. A retrained model is a change to the model, which the draft expects under change control and evaluated for retesting.
Do we need explainability for every model?
The draft expects feature capture and review during testing of models in critical applications, using techniques where applicable. Plan for it from the start rather than adding it afterward.


