Regulations and compliance · Practical guide

Does draft EU GMP Annex 22 cover generative AI?

The draft addresses certain static models with deterministic outputs in critical GMP applications. Embedding a language model in a validated software interface does not automatically bring it within that scope. Identify the model and the task it actually performs.

Table of contents

At a glance

What does the Annex 22 draft say about generative AI in critical GMP applications?

The 2025 consultation draft excludes generative AI and large language models from its proposed scope and says that these models should not be used in critical GMP applications. An EMA workshop report published in October 2026 discusses a possible extension under a documented, robust, risk-based control strategy. As of 17 November 2026 that discussion has not produced a final annex. Assess each use against the requirements currently applicable to it.

Scope and regulatory status

As of 17 November 2026, the text discussed here is the 2025 Annex 22 consultation draft. We checked its status against the European Commission’s official pages. Its proposals are distinct from requirements already applicable to a particular GMP process.

The EMA report published on 2 October 2026 records discussion of adding further model types. The revision is ongoing. A human signature alone does not establish either that a control works or that the proposed use is permissible.

Sources: [1] [2] [5]

Which models does the draft cover?

The draft covers static models with deterministic outputs. It excludes models that continue learning dynamically or produce probabilistic outputs. It explicitly places generative AI and large language models outside its proposed scope and advises against their use in critical GMP applications. This draft position does not provide general authorisation or establish a new binding prohibition.

Assess each proposed use against the requirements that already apply to its data, process, responsibilities and evidence. A model can fall outside the draft’s scope and still affect a GMP decision or record. Calling the application an “AI assistant” tells you neither how critical it is nor whether its controls work.

Sources: [1] [2] [3] [4]

Assess use cases by their actual effect

Identify the input, output, user and next process step. Drafting an internal email changes a process differently from automatically accepting a quality assessment. A summary can become critical if the reviewer stops reading the original and the missing information changes the decision.

The matrix is our assessment aid. It is not an official list of approved applications. Assess each use in context. Support tasks can directly affect data or quality, even if a human confirms the output.

Use cases, potential effects, and control questions
Use caseEffect to assessQuestion for your team
Suggesting wording Substantive modification of a previously reviewed document Is each substantive claim checked against its source?
Summary Omitted information or changed emphasis Does the summary retain everything needed for the decision?
Research aid Incorrect or inappropriate sources Are the original source and scope verified?
Classification Impact on processing or priority What are the implications of a misclassification?
Automated quality decision Direct process or product effect Is this use suitable within the applicable framework?

Sources: [1] [2] [3] [4]

Worked example of a summary that omits a critical exception

An LLM summarises an investigation report. The summary appears accurate but omits a paragraph about an unresolved exception. The reviewer reads only the summary and confirms the case. Human involvement did not catch the error in this workflow.

For your trial, use approved reference texts with known routine cases and exceptions. Specify the expected points and unacceptable omissions before testing. Check the output against the original. Asking the model a second question does not supply an independent source.

Test conflicting inputs, missing context and model changes. One successful case cannot establish suitability in every future situation. Limit use to tasks for which the controls work. If reviewers cannot reliably detect errors that matter, reassess the proposed use.

Test cases for a summary with a critical exception
Test caseInputExpected outcome of the control
Routine text Unambiguous reference with no exception Accurate main point traceable to the original
Critical exception Important limitation in the original Limitation remains visible or is identified during review
Contradiction Two conflicting statements The reviewer detects the conflict and avoids an unchecked conclusion
Missing source Assertion without credible evidence The claim is not accepted as a verified fact
Model change Same reference text with a new model version Assess how the change affects the defined use

Sources: [1] [2] [3] [4]

Does human review work as a control?

For non-critical GMP uses of generative models, the draft assigns responsibility for suitable outputs to people with adequate qualifications and training. An extra signature does not grant blanket approval for every use of an LLM. Define the intended use first.

A reviewer needs independent reference information, time and authority to reject an output. Asking someone to approve hundreds of long outputs without access to source material may leave the control effective on paper only. Test known errors under a realistic workload. Record which are caught and which are missed.

Define what happens when the reviewer is uncertain. The output can be rejected, corrected or referred for further assessment. A process that allows only approval cannot provide an effective review. Identify the tasks that remain outside permitted use.

Sources: [1] [2] [3] [4]

Control model versions, data access and changes

Record the service, model identification, input sources and relevant settings. An external service can change its behaviour without any software deployment by your team. Check the supplier’s change information and its impact on intended use. The same interface can conceal a change in output quality.

Decide which data may be entered and who needs access. Being able to upload a report does not establish that it is suitable for an external service. Address information protection, authorisation and contract requirements within your organisation’s framework.

Keep suitable reference cases and documented limits. Use those cases to assess relevant changes. They do not guarantee identical outputs from a probabilistic model. State what the assessment supports for the defined use and its controls.

Sources: [1] [2] [3] [4]

What to record before deciding on a generative AI use case

Connect the task, inputs, output, users, consequences, failure modes and controls. Identify the applicable requirements and distinguish them from proposals in the draft. Record open questions explicitly.

Define when use should stop or be reassessed. A change in task, a new data source or a missed critical error can trigger that decision. Identify who is responsible. Otherwise, a pilot can expand into a use that has never been assessed.

Sources: [1] [2] [3] [4]

Even a non-critical support tool needs controls

Even for a non-critical task, define how an unsuitable result will be detected. A draft can cite a wrong source, omit a requirement or introduce a plausible error. An “AI-generated” label does not identify who reviews and approves the content.

Define the specific use. Drafting internal wording has different consequences from automatically adopting a quality decision. Assess the data flow, consequences of error and existing controls. Calling a tool an assistant says little about its criticality.

Sources: [1] [2] [3] [4]

Worksheet for defining generative AI use before a pilot

Use the worked example for one named task. Specify permitted inputs and outputs, excluded uses and who can approve a change to those limits. The questions support a case-specific assessment and do not confer general approval for generative AI.

Worked decision worksheet
Decision areaExample entryAssessment question
Task Suggest wording for a specified draft. Which actual decision is affected?
Input Approved original report containing a known exception. Is the source complete and suitable?
Output The suggestion is not incorporated automatically. What use is expressly permitted?
Error Plausible false statement or omission. Can the reviewer spot the incorrect claim or omission?
Control Check against the original and retain the ability to reject the output. Can the reviewer do this under the actual workload?
Change A new source, task or model version. When is a re-evaluation required?

Sources: [1] [2] [3] [4]

Next steps

  • Describe the proposed AI use, its data flow and the consequences of error.
  • Check current regulatory status and model scope.
  • Define effective review and specify the limits of use.
  • Reassess changes to the purpose, model or process.
Further reading:

Project experience and collaboration

Define the permitted supporting use before the pilot

Before connecting a generative assistant to internal documents, we define its purpose with the process owners. Will it help a user find a source or generate a statement that informs a decision? Controlled examples let us test detection of omitted exceptions, incorrect versions and unsupported additions. Effective expert review may suit limited assistance. A more critical use requires reassessment of the limits of use. Connecting sources alone does not make the use low risk.

Frequently asked questions about Annex 22 and generative AI

Is Annex 22 already in force?

No. The EudraLex index did not list Annex 22 as an annex in force when checked on 18 October 2026. A further status check on 17 November 2026 reached the same conclusion. The available text was the 2025 consultation draft. Recheck before deciding on a particular use and record the version assessed.

Sources: [1] [2]

Does the draft also apply to AI used outside pharmaceutical manufacturing?

Section 1 covers systems used in the manufacture of medicinal products and active substances where AI models perform critical applications. It does not establish a general AI rule across industries. Assess the rules applicable to other activities and jurisdictions separately.

Sources: [1] [2]

Does Annex 22 already prohibit all generative AI use in GMP?

The 2025 consultation draft excludes generative AI from its proposed model scope and advises against its use in critical GMP applications. Publication of the draft did not create a new prohibition or confer general approval. Assess the requirements that already apply to the process.

Does a human approval click make a critical use non-critical?

No. Examine the actual impact and whether the control works. Can the reviewer identify significant errors using independent information and prevent the output from being used? A formal confirmation without enough time, expertise or information does not demonstrate effective risk control.

Why is the label “assistant” insufficient for an assessment?

It does not identify the information or decision being influenced. A summary can omit something important to a decision even if a human confirms it. Identify the input, output and subsequent action. Then assess whether the controls are suitable.

How should we record the draft’s status in project documents?

Record the document version, assessment date and specific draft position. Identify the applicable requirements and your own methods separately. Present a possible future framework as a proposal and revisit the assessment if it changes.

Sources and review date

The model scope comes from the 2025 consultation draft. The EMA workshop report published on 2 October 2026 documents the subsequent discussion. The regulatory status was checked on 17 November 2026. We developed the assessment questions and examples.

  1. European Commission: Annex 22 Artificial Intelligence – consultation draft 2025Consultation draft checked on 14 October 2026 for scope, data, human control and model testing. It did not establish a new requirement in force.
  2. European Commission · EudraLex Volume 4 indexThe index was checked on 14 October 2026 to distinguish texts in force from consultation drafts.
  3. European Commission: Consultation on Chapter 4, Annex 11 and Annex 22 (2025)Consultation ran from 7 July to 7 October 2025 and is closed. Publication for consultation does not establish entry into force.
  4. European Commission: EU GMP Annex 11, Computerised Systems (2011)EudraLex listed this version as current on 14 October 2026. It covers applications, risks, operation and the data lifecycle. The proposed tests are our own practical interpretation.
  5. EMA · Workshop report on the development of Annex 22EMA/156789/2026, published on 2 October 2026. See the introduction on p. 3 and the next steps on p. 14.

Sources checked: .

Daniel Herrmann, Daniel Herrmann Consulting

Daniel Herrmann Consulting

Defining the boundaries of your generative AI use case

We help you assess the intended use, establish who is responsible and put workable evidence requirements into operation. The right support depends on your system, your team and the decision you need to make.

  • AI system validation →

    Bring your AI system into GxP use with traceable evidence and keep it under control in operation. Once the system and intended GxP use are defined, we validate according to risk, addressing data, supplier, tests, release and change control.

Send a project enquiry Customer testimonials →