EU GMP Annex 22: Artificial Intelligence (draft)

Context and history

Annex 22 would be the first EU GMP text written specifically for artificial intelligence. It is a new annex, so it replaces nothing. It was published in draft on 7 July 2025 for public consultation, alongside the revised Annex 11 and Chapter 4, and the consultation closed on 7 October 2025. EMA held a multistakeholder workshop on comments and generative AI from 30 June to 1 July 2026, and the final text is targeted to go to the European Commission around the end of 2026.

It adds to Annex 11 for systems with embedded AI models, and it is short, at about six pages. Its central idea is that a model's fitness comes from testing against a defined intended use, using independent, representative test data, and not from how the model was built. It is a draft: the final text may differ, and until then no EU GMP annex specifically governs AI.

Who and what it applies to

All types of computerised systems used in making medicinal products and active substances, where AI models are used in critical applications with direct impact on patient safety, product quality or data integrity, for example to predict or classify data.

It covers machine learning models trained with data, including models made up of several individual models. It covers static models, which do not adapt during use, and deterministic models, which give identical outputs for identical inputs. Requirements apply whether a model is built in-house or supplied by a vendor.

Not covered: Dynamic models that keep learning during use, probabilistic models, and generative AI and large language models. The draft says these should not be used in critical GMP applications.

What it contains

1. Scope

Defines which systems and model types are covered, and lists those that are not and should not be used in critical GMP applications.

2. Principles

Close cooperation between process experts, QA, data scientists, IT and consultants, documentation available whoever built the model, and quality risk management.

3. Intended use

Describe the task in detail, characterise the input data including rare variations and limitations, and divide inputs into subgroups. A process expert is responsible, and the description must be approved before acceptance testing. Human-in-the-loop use must state the operator's responsibility.

4. Acceptance criteria

Define test metrics such as a confusion matrix, sensitivity, specificity, accuracy, precision or F1 score. Criteria may differ by subgroup and should be at least as high as the process being replaced.

5. Test data

Data should represent the full intended use, including subgroups, and be large enough for statistical confidence. Labels need verification, and pre-processing and exclusions must be justified. Generated test data is not recommended.

6. Test data independency

Test data must not be used in development, training or validation. It should be access-controlled with an audit trail, its use recorded, and staff who have seen it kept apart from training, or paired with someone who has not.

7. Test execution

Check the model generalises well. Requires an approved test plan, investigated deviations, and retained documentation including the test data.

8. Explainability

During testing, systems should capture the features behind a decision, using techniques such as SHAP, LIME or heat maps. Reviewing those features is part of approving results.

9. Confidence

Log confidence scores where applicable, and set thresholds. Very low confidence may be better flagged as undecided.

10. Operation

Change control and configuration control before deployment, regular performance monitoring, and monitoring that inputs stay within intended use.

Glossary

Defined terms.

Key points

• Critical applications only: it deals with AI that directly affects patient safety, product quality or data integrity.
• Dynamic, probabilistic and generative models are excluded and should not be used in critical GMP applications.
• A model must perform at least as well as the process it replaces, so you must know that process's performance.
• Intended use must be defined and approved before testing, with a process expert accountable.
• Test data must be independent, and people who saw it should not train the same model, unless working in a pair with someone who has not.
• Explainability is expected during testing, not only in operation.
• Deployed models need change control, configuration control and monitoring of performance and input data.
• Non-critical use of generative AI needs qualified people responsible for checking outputs.

In practice

General reading, not advice. This is drafted with AI and may not fit your situation. It does not replace the official text, your own quality system or qualified advice, and an inspector may take a different view.

• Inventory where AI is used today, and sort uses into critical and non-critical, because the scope depends on that line.
• For a model replacing a person, measure the person's performance first, or you may have no baseline.
• Keep test data separate early, as independence is hard to prove afterwards.
• Ask vendors now what evidence they can give on training, testing and intended use, as the draft expects the documentation whoever built the model.
• Generative AI tools for non-critical work, such as drafting, still need a named qualified person to check outputs.
• Wait for the final text before formal commitments, because generative AI in particular was debated in consultation.

Read from the source document on 1 Nov 2026. This is our own description, not the official text. Check the full dated history in the change log below.