GxP AI means using artificial intelligence or machine learning in work governed by the Good Practice regulations: manufacturing (GMP), clinical trials (GCP), laboratories (GLP), distribution (GDP) and pharmacovigilance (GVP). It is not a product category or a separate regulation. The AI is a computerised system like any other in that work, so it has to be validated for its intended use and its outputs have to be attributable, reviewable and fit for the decision they support.
What makes it harder than ordinary software is that some AI is probabilistic, can change when it is retrained, and depends on its training data. The rest of this guide covers what that means in practice and what regulators have actually published.
Why AI is harder to validate than ordinary software
Most software is deterministic: the same input gives the same output, so a test run once stays valid until the software changes. Some AI breaks that assumption in three ways.
Probabilistic outputs. Generative models in particular can give different answers to the same input. A test that expects one exact output cannot cover that. You test against acceptance criteria instead: how often the output is right, how it fails, and whether a person catches the failures.
Change through retraining. A retrained model is a new version of the software, even if nothing else changed. Treat it that way under change control, with the same tests repeated against the same acceptance criteria.
Training and test data. The data used to build and test a model shapes what it does. Keep its source, selection and labelling documented, and keep test data separate from training data, or the test results tell you little.
What regulators have published so far
There is no single AI regulation for GxP. These are the documents that matter today, and their status.
- EU GMP Annex 22 (draft). Published for consultation in July 2025 and not yet in force. It applies to AI models used in critical GMP applications and limits them to static models, meaning frozen models that do not learn during use, with deterministic outputs. It keeps generative AI and large language models out of critical uses and expects a qualified person to check their output elsewhere. Its sections cover intended use, acceptance criteria, test data and its independence, test execution, explainability, confidence and operation. Our Annex 22 readiness check turns those sections into questions.
- EU GMP Annex 11. The existing rule for computerised systems, which covers AI used in GMP work today. A revised Annex 11 was also published for consultation in July 2025.
- FDA. The final Computer Software Assurance guidance (reissued February 2026) sets a risk-based approach for production and quality system software. In January 2025 FDA published draft guidance on using AI to support regulatory decisions for drugs and biologics, built on a credibility assessment matched to the risk of the use. For devices, FDA, Health Canada and the MHRA published Good Machine Learning Practice principles in 2021.
- ISPE GAMP. GAMP 5 Second Edition added Appendix D11 on AI and machine learning, and the ISPE GAMP Guide: Artificial Intelligence followed in July 2025. Neither is a regulation, but together they are the industry's working consensus.
What GxP-ready AI needs
- A risk classification for each use. A tool that summarises literature and a model that influences batch release both use AI, but they need very different assurance. The classification decides how deep the testing goes and how much human review is required. In Cybroscape GxP Copilot, the Classification Engine assigns this with a deterministic rule engine, not a language model.
- Human review at defined points. Write down where a person must review AI output, who that person is, and what they are deciding. Record the review in the audit trail. Human-in-the-Loop reviews shows how GxP Copilot enforces this.
- Traceability. For every AI output, you should be able to show which model version produced it, from what input, and what a person did next. The audit trail should be tamper-evident, as it is for any 21 CFR Part 11 record.
- Ongoing monitoring. Track how the AI performs in use and set the thresholds that trigger review or retesting. Validation does not end at go-live.
- Explainability matched to risk. The higher the risk of the use, the more you need to understand why the model gave its output. The draft Annex 22 asks for this for critical uses.
Where AI is used in GxP work today
- Drafting validation documents (requirements, risk assessments, IQ, OQ and PQ protocols) for qualified people to review and approve. Cybroscape GxP Copilot does this.
- Drafting clinical and regulatory documents from source data, with each statement traced to its source. Cybroscape TraceDraft does this.
- Reviewing batch records and quality records to flag gaps for a person to check.
- Pharmacovigilance case intake and signal detection, with human assessment of each case.
- Monitoring regulatory publications and guidance changes.
In each of these the AI drafts or flags and a qualified person decides. That pattern is also what the draft Annex 22 expects for generative AI.
Getting started
- Write an AI policy first. Decide who can propose an AI use, who classifies its risk, who approves it, and who monitors it. Without that, teams adopt tools on their own and quality finds out at the next audit.
- Start with a low-risk use where people already review the output. Validation document drafting is a common first step: the work is slow, every output is reviewed before it is used, and the risk to product and patients is indirect.
- Plan monitoring before go-live. Decide what you will measure in use and what triggers retesting, so the evidence exists before an inspector asks for it.
book a demo if you want to see Cybroscape GxP Copilot on one of your own systems.
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.