The draft EU GMP Annex 22 shut generative AI out of GMP. Its scope statement says the document does not apply to generative AI and large language models, models with a probabilistic output, dynamic models and models which adapt performance — and that such models should not be used in critical GMP applications. A great deal of vendor and consultancy material has been written on that basis.
That position is being reconsidered, and the evidence is now public. EMA published the report of its multistakeholder workshop on AI guidance development for Annex 22 (reference EMA/156789/2026) in early October 2026, covering a workshop held on 30 June. It is worth reading directly rather than through summaries, because what it says about scope is more specific than most coverage suggests.
Nothing is decided. But the direction of travel has changed, and if you have been told that generative AI is prohibited in GMP, that was never quite right and is now out of date.
What the report actually says about scope
Three statements in the introduction do the work, and the sequence matters.
First, the consultation pushed back. The report states that comments received during the public consultation indicated support for allowing manufacturers to use large language models and generative AI systems in GMP applications — whether critical or non-critical — and attributes that position to the need to support innovation and encourage investment by pharmaceutical companies.
Second, the drafting group responded. The report says the drafting group has discussed the scope of Annex 22 to allow dynamic or adaptive models, probabilistic models and generative AI or LLMs in GMP applications, provided they comply with the requirements of the annex and are supported by a fully documented, robust, risk-based control strategy.
Third, the open question is not whether but how. The report is explicit that concerns remain about the elements needed for such a control strategy, which is why the drafting group asked to meet AI experts in the first place. The workshop existed to answer a narrower question than the headline suggests: can a risk-based approach be applied to generative AI and LLMs, and what control mechanisms would serve as mitigations.
The conclusion, in the report's own framing
The workshop conclusion is the part worth quoting carefully, because it is deliberately conditional. The report says the discussions suggest that subject to appropriate controls and a robust risk-based framework, there may be viable pathways for the use of adaptive, probabilistic and generative AI technologies in GMP-regulated applications.
May be viable pathways. That is not permission, and anyone selling it as permission is overstating it. What it is, is the regulator publishing a document that no longer treats the technology itself as the disqualifying factor.
The reasoning behind that is the genuinely significant part. The report records that participants consistently advocated a risk-based, technology-neutral approach in which the acceptability of an AI application is determined by its intended use, the risks it presents, and the effectiveness of the controls applied throughout its lifecycle rather than by the underlying technology itself.
Why technology-neutral is the whole story
A scope rule that names technologies ages badly and invites gaming. Draw the line at generative models and you get arguments about whether a particular architecture counts, rather than conversations about what the system actually does and what happens if it is wrong.
A rule based on intended use, risk and controls does not have that problem. It also asks more of you, which is the trade. Under a categorical exclusion, compliance is a classification exercise: establish that your model is static and deterministic and you are inside the permitted set. Under a technology-neutral framework you have to demonstrate a control strategy, and the report is clear that this is where the drafting group's remaining concerns sit.
Six topics structured the workshop: regulatory pathways for adaptive and probabilistic models, technical reliability and incident response, human oversight and accountability, validation and lifecycle management, strategic risk and compliance limits, and cybersecurity and outsourced activities. The emphasis throughout was the application of ICH Q9 quality risk management principles — which is to say, the framework already exists and the question is how to apply it here. The risk-based control strategy Annex 22 will ask for goes through what that control strategy has to contain.
What the report says has not changed
Worth as much attention as the scope news, and less widely reported. The report records that existing GMP principles — quality risk management, validation, lifecycle management, oversight and outsourced activities — remain relevant and applicable to AI systems.
In other words this is not a new regime bolted on beside GMP. It is GMP applied to a harder case. The common themes the report lists are recognisable to anyone who has run a validation programme: documented intended use, robust quality risk management, lifecycle validation and monitoring, risk-proportionate human oversight, effective control mechanisms, clear accountability, supplier oversight, and the maintenance of a demonstrable state of control.
Three of those are where we see most organisations fall short today, regardless of AI. The AI intended-use statement is absent more often than not. Monitoring exists as a chart nobody acts on, because no threshold was set in advance. And human oversight is a procedure rather than something the system enforces — why human-in-the-loop can't be a footnote covers why that distinction decides whether the control is real.
On suppliers, the answer was blunt
One exchange in the report deserves to be better known. Asked whether current Annex 11 requirements are sufficient for AI suppliers and service providers, the response recorded was that Annex 11 Chapter 7 is fully applicable because AI systems are computerised systems, though agreement types may vary with the service and intended use.
The key messages from that session are unambiguous: AI systems should be treated as computerised systems for cybersecurity and outsourcing governance; the regulated user remains accountable even when using third-party AI or cloud services; and supplier agreements must address responsibilities, evidence, configuration and change control.
That last point lands on anyone buying AI as a service, which is most people. Annex 11 Chapter 7 and your AI vendor covers it properly, because it is the obligation most likely to be discovered late.
What this means for Annex 22's status
Unchanged, and it is worth being precise because the field is full of material that is not.
Annex 22 is a consultation draft. It was published for consultation on 7 July 2025 and the consultation closed that October. It does not appear in EudraLex Volume 4. It is not in force, and nobody can be cited against it today. The earlier expectation of a final text in mid-2026 has passed without one.
The report's own next steps say the drafting group will review the workshop input, assess the key themes, common approaches, areas of divergence and identified challenges, and use the outcomes to inform the ongoing revision. No date is committed to. A revised draft is the next artefact to expect.
Annex 22 is also supplementary rather than freestanding: the report describes it as an annex to the GMP Guide intended to provide additional guidance to Annex 11 for computerised systems in which AI models are embedded. EU Annex 11 & Annex 22 readiness remains the governing text, and its own revision is still in consultation too.
What to do with this now
Four things, none of which depend on the final text.
- Stop planning around the exclusion. If your AI roadmap was shaped by the belief that generative models are barred from GMP, revisit it. The exclusion may not survive, and a technology-neutral framework would judge your use case on intended use and controls instead.
- Start building the control strategy, not the classification argument. Under the likely direction, effort spent proving your model is deterministic is effort misdirected. Effort spent on documented intended use, evaluation, monitoring thresholds, oversight and fallback transfers directly.
- Treat your AI vendor as a computerised system supplier. Annex 11 Chapter 7 already applies, which means this obligation exists today rather than arriving with Annex 22.
- Keep citing it as a draft. Presenting a draft annex as a binding obligation is the fastest way to lose the trust of people who know the subject — and the scope reconsideration is a good illustration of why. Material written confidently against the July 2025 scope statement is now describing a position the drafting group has moved on from.
AI validation services is the practice, the GxP AI hub covers the wider regulatory picture, and book a demo if you want a specific use case assessed against where this is heading rather than where the draft currently sits.
Where to go next
Explore GxP Copilot for AI-native validation, TraceDraft for source-traceable clinical documentation, or book a demo to see either on your own data.