EU GMP Annex 11 dates from January 2011. It was written before cloud was a default, before SaaS was how most software is bought, and well before anyone was putting a machine learning model anywhere near a batch decision. A revision has been in consultation since 2025 and has not been adopted — the 2011 text is what applies today.
That creates a planning problem rather than a compliance one. You cannot be cited against a draft, and you also should not design a system in 2026 as though the 2011 assumptions still hold. This is how we read the direction, and what we think is worth changing now regardless of what the final text says.
Where the current text shows its age
Annex 11 as it stands is a short, sensible document whose gaps are all in the same place: it assumes you control the computer.
- Suppliers and service providers. The text contemplates a supplier who sells you software you then install. It handles far less well a provider who runs the application, holds the data and updates it on their schedule.
- Cloud and shared infrastructure. There is no real treatment of multi-tenancy, data residency, or qualification of infrastructure you cannot inspect.
- Change control you do not own. The periodic evaluation and change control expectations assume changes originate with you.
- Artificial intelligence. Absent, which is why Annex 22 was drafted separately rather than folded in.
None of this makes the current text wrong. It makes it incomplete in a way that inspectors already work around by asking questions the annex does not explicitly require.
What the revision is expected to address
Treat this as direction rather than detail — until the final text is published, specifics can change, and quoting a draft clause as a requirement is how organisations lose credibility with people who know the subject.
The consistent themes across the consultation material and the wider EU GMP programme are: a fuller treatment of suppliers and service providers including cloud, data integrity expectations expressed more explicitly rather than inferred, clearer expectations around periodic review, and a more worked-through position on electronic signatures and audit trail review. Chapter 4 on documentation is being revised alongside it, which matters because the two are read together.
The thing to resist is building a compliance programme against a draft. The thing to avoid is the opposite mistake of waiting, because most of what the revision points at is good practice already.
What to change now, regardless
Everything in this section is defensible under the 2011 text and will remain so under any plausible revision. That makes it low-risk work.
Document the shared responsibility boundary for every cloud system. Who qualifies the infrastructure, who owns configuration, who owns the data, what the vendor tests and what you test. Most organisations have this knowledge distributed across several heads and no document. cloud & SaaS validation is where we write it down, and supplier qualification is where the provider side gets assessed.
Settle the change notification question with each SaaS vendor. How much notice before a production update, whether you can defer or stage it, whether a sandbox receives it first, and what regression evidence they provide. If the answer is little notice and no deferral, your mitigation is monitoring and a defined response — which is acceptable, and has to be a documented decision rather than an unexamined dependency.
Start reviewing audit trails. This is already expected and is the control most often present and unused. A targeted review — changes after approval, changes outside working patterns, repeated invalidations — produces a record that answers the question directly. audit readiness is where we plan it and data integrity (ALCOA+) is the wider practice.
Bring periodic reviews current. An undefined interval is a finding on its own. A defined interval that is routinely missed is worse than a longer one that is kept, which is why periodic review is the obligation we see skipped most.
Where Annex 22 fits
Annex 22 is the separate draft annex written specifically for artificial intelligence. Published for consultation on 7 July 2025, consultation closed that October, and EMA's work plan targets delivering a final text in late 2026. It is not in EudraLex Volume 4 and it is not in force.
Its scope is the part most likely to move, and it moved visibly in October 2026. As drafted it places dynamic, probabilistic and generative models — large language models included — outside its scope and outside critical GMP applications. EMA's workshop report (EMA/156789/2026), published in early October 2026 and covering a workshop held on 30 June, records that consultation comments supported permitting them, that the drafting group has discussed widening the scope provided such models are supported by a fully documented, robust, risk-based control strategy, and that there may be viable pathways. Participants advocated a risk-based, technology-neutral approach in which acceptability turns on intended use, risk and the effectiveness of controls rather than the technology itself. The report's next steps commit to no date, and a revised draft is what to expect next. what EMA's Annex 22 workshop report says goes through it.
The planning implication: if you are deploying AI in a GMP context, design for the assurance elements the draft describes — Annex 22 assurance is how we hold ourselves to them — because they are unlikely to reverse, and do not build a programme that depends on the scope boundary staying where the draft puts it. EU Annex 11 & Annex 22 readiness and the GxP AI hub go into it properly.
How to talk about a draft without overstating it
This is worth being disciplined about, because it is where vendor material routinely misleads. A draft annex is not a requirement. Saying a system must comply with Annex 22 today, or with a revised Annex 11, is inaccurate, and the audience for that claim is exactly the audience that will notice.
The accurate formulation is straightforward: the current Annex 11 text of January 2011 applies; a revision is in consultation and has not been adopted; here is what the draft indicates; here is what we are doing now and why it is defensible under the text that applies.
That version is more persuasive as well as more accurate, which is a happy coincidence rather than a rule. book a demo if you want a system assessed against what actually applies.
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.
