Compliance 13 min read

AI and EU GMP Annex 11: What Regulated Teams Should Do

J

November 14, 2026

Search for "AI Annex 11" and you will get three different documents that have almost nothing to do with each other. One is EU GMP Annex 11, the rulebook for computerised systems in medicines manufacturing. Another is Annex XI of the EU AI Act, which sets out technical documentation for providers of general-purpose AI models. The third is a draft Annex 22 that the European Commission published alongside a revised Annex 11 to deal with artificial intelligence in GMP.

If you work in a pharmaceutical, biologics, or device quality unit, EU GMP Annex 11 and draft Annex 22 are the ones that decide whether your next audit goes smoothly. I want to separate the three, then spend most of this guide on what you actually do when someone proposes putting an AI model into a GxP workflow.

Which "Annex" Are You Actually Asking About?

The confusion is understandable, because the numbering collides. Here is how the three documents differ.

Document Legal basis Who it applies to What it covers
EU GMP Annex 11 (Computerised Systems) EudraLex Volume 4 guidance, interpreting Directive 2001/83/EC and Directive 2003/94/EC Manufacturers of medicinal products in the EU, and importers and third-country sites supplying the EU Validation, data integrity, audit trails, electronic signatures, supplier oversight for any computerised system used in GMP
Draft EU GMP Annex 22 (Artificial Intelligence) EudraLex Volume 4 (consultation draft, July 2025) The same GMP population Use of AI models in critical GMP applications affecting patient safety, product quality, or data integrity
Annex XI of Regulation (EU) 2024/1689 (AI Act) EU AI Act Providers of general-purpose AI models under Article 53(1), point (a) Technical documentation the model provider must prepare: evaluation strategies, training data, compute, and similar

The first two are GMP instruments. The third is product-safety style regulation aimed at whoever builds a foundation model. A drug manufacturer is almost never the "provider" of a general-purpose AI model. You are more likely a deployer who buys or licenses one, and that distinction matters when you write your supplier assessment, which I will come back to.

What Annex 11 Already Requires Today

Before any AI-specific text arrives, Annex 11 (in force since 30 June 2011) already applies to an AI-enabled system. Nothing in it exempts software because it is "smart." The clauses that bite hardest:

  • Clause 1, Risk Management: Risk management must be applied throughout the lifecycle, with the extent of validation and data integrity controls based on a documented risk assessment.
  • Clause 3, Suppliers and Service Providers: Where third parties supply or support a system, there must be a formal agreement, and the user must assess the supplier. Annex 11 clause 3.2 expects the need for an audit to be based on a risk assessment.
  • Clause 4, Validation: Validation documentation must cover the lifecycle steps. Clause 4.4 requires user requirements specifications to describe the required functions, and clause 4.5 expects the regulated user to take all reasonable steps to ensure the system was developed under an appropriate quality management system.
  • Clause 5, Data: Electronically exchanged data needs built-in checks for correct and secure entry and processing.
  • Clause 9, Audit Trails: Consideration should be given to building into the system the creation of a log of all GMP-relevant changes and deletions.
  • Clause 10, Change and Configuration Management: Changes must be made in a controlled manner per a defined procedure.
  • Clause 11, Periodic Evaluation: Systems must be periodically evaluated to confirm they remain in a valid state, including functionality, deviation records, incidents, problems, upgrade history, performance, reliability, and security.

Notice what clause 11 implies for machine learning. A model that was validated in March and quietly changes behaviour in September is exactly the scenario periodic evaluation was written for, even though the authors in 2011 were thinking about patched ERP software, not neural networks.

What the Revised Annex 11 and Draft Annex 22 Change

In July 2025 the European Commission, working with the EMA and PIC/S, released for public consultation a revised Annex 11, a new Annex 22 on artificial intelligence, and a revised Chapter 4 on documentation. The consultation period closed in October 2025. This guide treats revised Annex 11 and Annex 22 as consultation drafts, not final requirements. Check EudraLex Volume 4 on the Commission's website for the current status before you cite any of the following in a procedure, because final wording can differ from a consultation draft.

With that caveat, here is what the drafts signalled and why it matters.

The revised Annex 11 direction

The draft expands the lifecycle expectations considerably. It leans on quality risk management, treats the regulated user and the supplier as sharing responsibility, and gives much more space to identity and access management, audit trails, data governance, and security than the 2011 text did. The practical message to me is that "we validated it once" will carry less weight, and that the evidence of ongoing control is where inspectors will look.

Draft Annex 22 and the line it draws

Annex 22 is short, and its most important feature is what it excludes. In the consultation draft, the scope is limited to AI models that are static (they do not continue learning in use) and deterministic (the same input gives the same output). Dynamic, continuously learning models and probabilistic outputs are not meant to be used in critical GMP applications. Generative AI and large language models fall on the wrong side of that line for critical uses. For non-critical uses, the draft suggests a qualified human should stay in the loop and take responsibility for the result.

For critical applications inside the scope, the draft asks for things a validation engineer will recognise:

  1. A documented intended use that describes the process the model supports, its limitations, and the characterisation of the input data.
  2. Defined acceptance criteria, at least as demanding as the performance of the process it replaces.
  3. Test data that is representative, independent from training data, and controlled so that it cannot leak into training.
  4. Controls over change and configuration, with re-testing when the model, its inputs, or its intended use change.
  5. Monitoring of input data drift and performance during operation.

Some people read this as hostile to AI. I read it as a clear instruction: use AI where you can freeze it, test it, and explain what it was for, and keep generative tools out of decisions that release product. That is a workable position, and it is far more useful than the ambiguity we had before.

How FDA Frames the Same Problem

If you sell into the US as well as the EU, you will be answering to both, and the vocabulary differs even when the intent matches.

  • 21 CFR Part 11: Section 11.10(a) requires validation of systems to ensure accuracy, reliability, consistent intended performance, and the ability to discern invalid or altered records. Section 11.10(e) requires secure, computer-generated, time-stamped audit trails. These apply to an AI tool that creates, modifies, or maintains a required record.
  • 21 CFR 211.68: Automatic, mechanical, or electronic equipment used in drug manufacturing must be routinely calibrated, inspected, or checked per a written program, and controls must exist over computer systems to assure that changes are made only by authorised personnel.
  • FDA's Computer Software Assurance guidance: FDA finalised its guidance "Computer Software Assurance for Production and Quality System Software" in September 2025 (check the FDA guidance page for the current text). Note that it was written for device software. It describes a risk-based approach where assurance effort scales with the risk that a software failure poses to product quality or patient safety. That sits comfortably with Annex 11 clause 1.
  • FDA's January 2025 draft guidance on AI: "Considerations for the Use of Artificial Intelligence to Support Regulatory Decision-Making for Drug and Biological Products" proposes a risk-based credibility assessment framework built around the model's context of use. Note that it is about AI supporting regulatory decisions on safety, effectiveness, or quality, and it is a draft.

Side by Side: Annex 11, Part 11, and ISO 42001

Quality teams often ask whether an ISO 42001 AI management system replaces any of this. It does not, but it fills a gap the GxP rules leave open.

Topic EU GMP Annex 11 / draft Annex 22 21 CFR Part 11 / FDA CSA ISO/IEC 42001:2023
Nature of the document Volume 4 is guidance that interprets the GMP directives; it is enforced through national competent authorities Binding regulation (Part 11) and non-binding guidance (CSA) Voluntary, certifiable management system standard
Risk basis Clause 1 risk management, QRM principles Risk-based assurance under CSA Clause 6.1.2 AI risk assessment, clause 6.1.4 AI system impact assessment
Validation of the system Clause 4 and draft Annex 22 test-data requirements 11.10(a) No product validation requirement, but requires controls across the AI lifecycle (Annex A)
Supplier oversight Clause 3 Expected through quality agreements and 211.22 responsibilities Annex A controls on third-party and customer relationships
Records and audit trail Clause 9 11.10(e) Documented information, clause 7.5
Ongoing monitoring Clause 11 periodic evaluation, draft Annex 22 drift monitoring Maintained validated state through change control Clause 9.1 monitoring, measurement, analysis and evaluation

My own view is that ISO 42001 is a useful governance wrapper, particularly for an organisation running many AI use cases. It will not get you through an Annex 11 inspection by itself, because inspectors ask about specific systems and specific records.

A Practical Path to Putting AI Inside a GMP Process

Here is the sequence I walk teams through. It is deliberately plain.

Step 1: Decide whether the use is critical

Ask one question: could a wrong output affect patient safety, product quality, or data integrity without a human independently catching it? If yes, treat it as critical and expect the full Annex 22 style evidence. For example, a visual inspection model that accepts or rejects vials for release is critical: a missed defect reaches patients and no one rechecks every unit. A deviation triage classifier that only suggests a category, which a qualified reviewer confirms before the record is saved, can reasonably be argued non-critical. It becomes critical if its output auto-closes or down-classifies deviations with no review. If a qualified person reviews and owns every output, you can reasonably argue non-critical and document that argument. Write the rationale down, because the inspector will ask for it.

Step 2: Write the intended use and user requirements

Annex 11 clause 4.4 already wants a URS. For an AI function, the URS should state the input data types, the decision or output, the performance threshold, the situations where the model must not be used, and who reviews the output. Vague statements such as "assist with deviation triage" will not survive a validation review.

Step 3: Classify the model type

Is it static or does it learn in production? Is it deterministic or probabilistic? A classifier trained once and locked is a very different object from a chat assistant. If the answer is "probabilistic and generative," keep it away from critical decisions, per the direction of draft Annex 22. General-purpose chatbots are a poor fit for quality records for exactly these reasons, as this piece on chatbots and quality management records explains in terms of Part 11 and Annex 11 clause 9.

Step 4: Assess the supplier under clause 3

For a bought model or an AI-enabled SaaS tool, your supplier assessment should ask: how are model versions controlled, and will you be told before one changes? What data is the model trained on, and is any of your data used to train it? How are outputs logged? Where does the data reside? Can you get the training and test documentation? If the supplier is a general-purpose model provider, the Annex XI technical documentation under the AI Act is a useful thing to request, but it describes the model, not your validated use of it.

Step 5: Build the test strategy

Under a CSA mindset, put your effort where the risk is. For a critical AI function that means:

  • A test set independent from training data, with documented provenance.
  • Acceptance criteria tied to the existing process, for example the error rate of the manual inspection or review that the model replaces.
  • Challenge cases including edge conditions and known failure modes.
  • Documented review of any failures with root cause.

For lower-risk supporting functions, unscripted or exploratory testing with a short record is often enough, which is the point of CSA. If you want a view on how AI can draft the protocols themselves, see validation and qualification protocol generation using AI QMS tools. The protocol still needs a human owner who approves it.

Step 6: Lock it down and control change

Freeze the model version, its configuration, and its input preprocessing. Put them under change control per Annex 11 clause 10. Define in advance which changes trigger re-validation: a new model version, a new data source, a shift in the intended use, or a measurable change in input data distribution.

Step 7: Monitor and periodically review

Set up routine checks on performance and input drift, assign an owner, and feed the results into your clause 11 periodic evaluation. Decide in advance what threshold triggers a deviation, for example accuracy on a periodic check sample falling below the acceptance criterion set during validation, or a drift statistic on a key input exceeding a limit you defined and justified. A model that degrades without anyone noticing is the failure inspectors will find hardest to forgive.

Step 8: Keep the human accountable

Training records matter here too. The people who review AI outputs need documented training on the tool's limits, and the procedure should say who decides when the output is overridden. Annex 11 clause 2 requires that there be close cooperation between all relevant personnel, including those responsible for IT.

Where Teams Get It Wrong

Common patterns include:

  • Treating the vendor's validation package as your validation. Clause 4.5 allows you to leverage supplier documentation, but you remain responsible for fitness for your intended use.
  • Letting a pilot become production. A proof of concept run on live batch data with no change control is a data integrity finding waiting to happen.
  • Ignoring the audit trail for AI-generated content. If a model drafts a deviation narrative that a person then edits, the record should show what was generated and what was changed. Part 11 section 11.10(e) and Annex 11 clause 9 both point this way.
  • Calling everything "non-critical." It is tempting. If the output feeds a release decision even indirectly, think twice.

Many of these failures are less about technology than about ownership.

A Short Readiness Checklist

  • An inventory of AI-enabled functions, each with a critical or non-critical classification and a written rationale.
  • A URS and intended-use statement for each function.
  • Supplier assessments referencing Annex 11 clause 3 and quality agreements that cover model change notification.
  • Validation evidence with independent test data for critical uses.
  • Change control rules that name model version, data, and intended use as triggers.
  • A monitoring plan feeding periodic evaluation under clause 11.
  • Training records for reviewers, and a procedure for overriding outputs.

If most of those boxes are empty, you are not behind in any unusual way. Many organisations are exactly there, and the honest question is which AI use case you can bring under control first.

Last updated: 2026-10-18

Frequently Asked Questions

Does EU GMP Annex 11 apply to AI systems?

Yes. Annex 11 applies to any computerised system used in GMP activities, including AI-enabled software. Clauses on risk management (1), suppliers (3), validation (4), audit trails (9), change control (10), and periodic evaluation (11) all apply today, and draft Annex 22 adds AI-specific expectations.

What is draft Annex 22?

Annex 22 is a proposed EU GMP annex on artificial intelligence, released for consultation in July 2025 alongside a revised Annex 11. In the draft, it covers static, deterministic models in critical GMP applications and excludes dynamic or probabilistic models such as generative AI from critical use. Check EudraLex Volume 4 for the current status.

Is Annex XI of the EU AI Act the same as Annex 11?

No. Annex XI of Regulation (EU) 2024/1689 lists technical documentation that providers of general-purpose AI models must prepare under Article 53(1)(a). EU GMP Annex 11 governs computerised systems in medicines manufacturing. They are separate documents with different audiences.

Can I use generative AI or an LLM in a GMP process?

Under the direction of draft Annex 22, generative and probabilistic models should not be used in critical GMP applications. For non-critical uses, a qualified person should review and take responsibility for every output, and the use should still be documented, risk assessed, and controlled.

How does FDA treat AI in regulated quality systems?

FDA applies existing rules such as 21 CFR Part 11 and 211.68 to AI tools that touch GxP records. Its September 2025 Computer Software Assurance guidance supports risk-based validation, and its January 2025 draft guidance proposes a risk-based credibility framework for AI used in drug regulatory decision-making.

J

Jared Clark

Principal Consultant, Certify Consulting

Jared Clark is the founder of Certify Consulting, helping organizations achieve and maintain compliance with international standards and regulatory requirements.