info@truex-consultancy.com
All guides
Computerised systems Updated 30 Oct 2026 8 min

Annex 22 vs Annex 11: what changes for AI and computerised systems in GMP

Annex 11 and Annex 22 do not compete. Annex 11 covers every computerised system used in GMP: how it is validated, secured, backed up and kept under control. Draft Annex 22 adds rules for one kind of system, an AI model used in a critical GMP application. If your site uses an AI tool, you will answer to both: Annex 11 for the system around the model, Annex 22 for the model itself.

This guide sets the two side by side, explains where each stands today, shows which one applies to common AI uses, and lists what to do now without waiting for the final text. A short German summary follows the FAQ.

Annex 22 vs Annex 11 side by side

Comparison of EU GMP Annex 11 and draft Annex 22
Annex 11Annex 22 (draft)
What it coversAll computerised systems used in GMP activities, from a spreadsheet to a manufacturing execution system.AI models used in critical applications, meaning a direct impact on patient safety, product quality or data integrity.
StatusIn force since 2011. A much longer revision was published as a draft in July 2025 and is not final.Draft published for consultation in July 2025. Not adopted and not part of EudraLex Volume 4 at the time of writing.
Core questionIs the system validated, secure, controlled and fit for its intended use?Is the model's performance proven for its intended use, and does it stay proven in operation?
ValidationLifecycle approach based on risk: requirements, testing, change and periodic review.Adds model-specific testing: intended use, acceptance criteria set in advance, independent and representative test data.
Type of modelNot model-specific. Applies to any software.Static and deterministic models only for critical use. Models that keep learning, and generative models, are not for critical use in the 2025 draft.
DataData integrity, accuracy checks, audit trails, backup and archiving.Test data separate from training data, correctly labelled, covering rare and worst-case inputs. Input data monitored for drift.
PeopleDefined roles, training and access control.Subject matter experts define intended use. A qualified person is accountable where AI output is reviewed.
In operationChange control, incident handling, periodic evaluation.Locked model version under change control, performance monitored, confidence scores recorded where relevant.

Read the table row by row and the pattern is clear. Annex 22 does not replace any Annex 11 expectation. It adds a layer for the part Annex 11 was never written to cover, a system whose behaviour comes from training data rather than from code a person wrote.

Where each annex stands today

The European Commission and the EMA published a revised Annex 11, a new Annex 22 and a revised Chapter 4 as drafts in July 2025. The consultation closed on 7 October 2025. At the time of writing, neither text has been adopted, so the 2011 version of Annex 11 remains the one in force. The target for final text is the fourth quarter of 2026, but no adoption or effective date has been announced.

The scope of Annex 22 is also still being discussed. The 2025 draft excludes generative AI and large language models from critical use, and the EMA held a workshop on a risk-based approach to generative AI on 30 June and 1 July 2026. Summaries of that workshop differ on what it means for the final text, so treat the generative AI question as open.

Which annex applies to your AI use?

The answer depends on what the AI does and what happens to its output. These examples show how the two annexes divide the work:

  • A chat assistant that drafts a first version of an SOP, which then goes through normal document review. Annex 11 and your document control rules apply to the final record. Annex 22 is not aimed at this, but a qualified person must check the output against the source and own it.
  • A tool that suggests a deviation classification for QA to confirm. Annex 11 applies to the system. Annex 22 thinking applies to the model: define the intended use, test it on independent data, set acceptance criteria and watch how often reviewers override it.
  • A vision system that rejects containers automatically, with no person checking each one. This is the clearest critical use. Annex 11 and Annex 22 both apply in full, and the model must be static and deterministic.
  • An AI feature that a vendor switched on in a software upgrade. Annex 11 requires the change to be assessed. Until someone has assessed it, you do not yet know which category it falls in, which is why an AI register matters.

What changes in practice if you already comply with Annex 11

A site with a sound Annex 11 programme has most of the foundations. Three things are new:

  1. A baseline for the current process. To show a model performs at least as well as the process it replaces, you need to know how well that process performs today. Few sites have measured the error rate of manual inspection or deviation triage, so this is often the first real piece of work.
  2. Test data you can defend. Validation on the vendor's own benchmark is not enough. The test set must represent your products and your rare events, and it must be kept apart from the data the model learned from.
  3. A rule that updating the model is a change. A retrained or updated model is a new model. It goes through change control and is retested before use, the same way you would treat a major software release.

For the detail of each expectation and a worked validation example, see our guide 'EU GMP Annex 22 explained: validating AI in GMP'. For the change control side, see 'Change control risk assessment: how to make it defensible to an inspector'.

What to do now

  1. List every AI use in your computerised system inventory, including AI features inside existing software, and ask staff which public AI tools they use.
  2. Assign each use a risk category and record the reasoning: no GMP impact, assistive, decision support or autonomous critical.
  3. For anything beyond assistive use, write down the intended use and the performance the process needs, before any testing.
  4. Add model version, change notice, data location and use of your data to the supplier questionnaire and quality agreement.
  5. Review your Annex 11 basics for those systems: access control, audit trail, backup and periodic review. Weak basics will be found first.
  6. Track the status of both annexes and plan a gap assessment once the final texts are published.

None of this depends on the final wording. It is validation, data integrity and change control applied to a new kind of system, and a site with an honest inventory and clear categories will be ready whatever the final text says.

Kurz auf Deutsch: Unterschied zwischen Annex 11 und Annex 22

Annex 11 gilt für alle computergestützten Systeme in der GMP-Umgebung und regelt Validierung, Datenintegrität, Zugriffsschutz, Audit Trail und Änderungskontrolle. Der Entwurf von Annex 22 betrifft ausschließlich KI-Modelle in kritischen Anwendungen, also dort, wo Patientensicherheit, Produktqualität oder Datenintegrität direkt betroffen sind. Annex 22 ersetzt Annex 11 nicht, sondern ergänzt ihn.

Nach dem Entwurf von 2025 sind für kritische Anwendungen nur statische, deterministische Modelle vorgesehen. Lernende und generative Modelle sind dafür ausgeschlossen. Der Verwendungszweck muss vorab dokumentiert werden, die Akzeptanzkriterien werden vor dem Test festgelegt, und die Testdaten müssen unabhängig von den Trainingsdaten sein. Beide Texte sind zum Zeitpunkt dieses Artikels noch Entwürfe. Prüfen Sie vor der Anwendung den aktuellen Stand auf der Seite der Europäischen Kommission zu EudraLex Band 4.

Frequently asked questions

What is the difference between Annex 11 and Annex 22?

Annex 11 covers all computerised systems in GMP and is about validation, security, data integrity and control. Draft Annex 22 covers AI models used in critical applications and adds model-specific expectations: documented intended use, acceptance criteria set before testing, independent test data and performance monitoring. Annex 22 works alongside Annex 11, not instead of it.

Does Annex 22 replace Annex 11?

No. An AI tool is still a computerised system, so Annex 11 applies in full. Annex 22 adds the requirements specific to the model.

Is Annex 22 in force?

Not at the time of writing. The draft was published for consultation in July 2025 and the consultation closed in October 2025. A final text is targeted for the fourth quarter of 2026, but no adoption date has been announced. The 2011 version of Annex 11 remains in force until a revision is adopted.

Which annex applies if my software has an AI feature?

Annex 11 applies to the software as a whole. Annex 22 applies to the AI model where it is used in a critical application. Start by finding out what the feature does and whether its output enters a GMP decision or record.

Was ist der Unterschied zwischen Annex 11 und Annex 22?

Annex 11 regelt alle computergestützten Systeme im GMP-Umfeld. Annex 22 (Entwurf) ergänzt dies um Anforderungen an KI-Modelle in kritischen Anwendungen, etwa dokumentierter Verwendungszweck, vorab festgelegte Akzeptanzkriterien und unabhängige Testdaten. Annex 22 ersetzt Annex 11 nicht.