The draft EU GMP Annex 22 on artificial intelligence was published for consultation in July 2025 and is not yet in force. It applies to AI models used in critical GMP applications and allows only static models with deterministic output there; generative AI and models that keep learning in use are kept out of critical uses. Where a model is in scope, the draft expects a detailed intended use, acceptance criteria set before testing, independent and representative test data, explainability and confidence where relevant, and change control, monitoring and human review in operation. Below is what each section asks for, what the draft does not say, and how to check a vendor against it.
What the draft covers, and what it leaves out
Annex 22 sits alongside Annex 11, which still governs the computerised system as a whole. Annex 22 adds expectations for the AI model inside it when that model is used in a critical GMP application, meaning one with a direct impact on patient safety, product quality or data integrity.
For those uses the draft limits itself to static models (frozen after training, not adapting from new data during use) with deterministic output, so the same input gives the same result. Dynamic models, generative AI and large language models are not to be used in critical applications. In non-critical uses they can be, with a qualified person responsible for checking what the model produces.
That line may move. EMA's workshop report, published on 2 October 2026, records that the drafting group has discussed allowing generative models under a documented, risk-based control strategy; our summary of the workshop report covers what changed and what did not.
Intended use
The draft expects the intended use to be described in detail and approved by the process subject-matter experts: what the model does, the input data it will see including its expected variation and limitations, and any subgroups of input that need separate attention. Everything later in the annex is tested against this description, so a vague intended use makes every later step weaker.
Acceptance criteria
Performance metrics and their acceptance criteria are set before testing, not after. The draft gives examples such as a confusion matrix, sensitivity, specificity, accuracy, precision and F1 score, and expects the model to perform at least as well as the process it replaces.
Test data and its independence
Test data must represent the full range of the intended use, be large enough to support the conclusions, and be correctly labelled, with the labelling itself checked. It must be independent of the data used to train and tune the model, and the draft expects controls that keep the test data separate from model development.
Test execution
Testing follows an approved plan, deviations are recorded and assessed, and the results are documented against the pre-set criteria. This is ordinary validation discipline applied to the model, and it should feel familiar to anyone who has run an OQ.
Explainability and confidence
Where it matters for the decision, the model should be able to show which features drove its output, for example through feature attribution or highlighting. The model should also record a confidence score with each result, and low-confidence results should be flagged rather than passed through as if they were certain.
Operation
Once in use, the model and its configuration are under change control, with retesting when either changes. Performance is monitored, and so is the input data, so that a shift away from what the model was tested on is noticed. Where a person reviews the model's output as a control, the draft expects that review to be recorded, and treats it as part of the system rather than an informal check.
What the draft does not say
Several claims about Annex 22 circulate that are not in the text. It does not use the term model card, though a model card is a sensible way to hold the intended use and test evidence. It does not require a documented fallback for when the AI is unavailable; that comes from Annex 11's business continuity expectations. It does not certify vendors or tools, and no software is "Annex 22 compliant" on its own, because the expectations attach to your use of the model.
It also does not apply to every AI feature in every system. A tool that drafts a document which a qualified person then reviews, edits and signs is a different case from a model that decides a batch disposition. Your own risk assessment decides which side of the critical line a use falls on.
How to check an AI tool against it
- Which decisions does the AI make, and which does it only propose for a person to decide?
- For any model in a critical use, is it static with deterministic output, and can the vendor show its intended use, acceptance criteria and test results?
- Is the test data independent of the training data, and how is that controlled?
- Does each output carry a confidence indication, and are low-confidence outputs flagged?
- How are model and configuration changes controlled and retested, and how is performance monitored?
- Is human review recorded in the system, with the reviewer, the decision and the time?
Our free Annex 22 readiness check turns the draft's sections into questions you can answer for one model.
How GxP Copilot is built against it
In Cybroscape GxP Copilot the AI drafts and quotes facts from your source documents; it does not decide. GxP impact, GAMP category and risk are computed deterministically from facts a person has confirmed, and every AI draft goes through Human-in-the-Loop reviews before it counts, with approval by electronic signature. Prompts, checks and model settings are fingerprinted, so a change fails the build until it is re-approved and the fixed test set is re-run. How the platform governs its own AI is described on the AI validation page.
Where to go next
Explore Cybroscape GxP Copilot for AI-native validation, Cybroscape TraceDraft for source-traceable clinical documentation, or book a demo to see either on your own data.