Pharma · · 8 min read
EMA AI guidance: what pharma data teams should do now
Acumen Editorial · Updated November 20, 2026
European regulators are moving from principles to paperwork. On October 5, 2026, the HMA-EMA group on AI met with industry stakeholders. The summary notes, published October 13, say a concept paper on AI in clinical development is being finalised, with a public consultation to follow. A horizontal guidance on machine learning is being developed in parallel.
A week ago, on November 12 and 13, EMA hosted its multistakeholder AI and Data Forum in Amsterdam. The title was blunt: "Build trustworthy AI on strong data foundations." Outcomes have not been published yet. The agenda alone tells you where scrutiny is heading.
For heads of AI, medical affairs, regulatory and clinical data science, this is the window to get ready. Once the consultation opens, the expectations are public. Data you build today should already meet them.
What is EMA actually working on?
Several tracks are running at once. Here is where each stands, based on EMA's own published notes and event pages, and on reporting about the GMP draft.
| Instrument | Status | Scope | What it asks of your data |
|---|---|---|---|
| FDA-EMA guiding principles of good AI practice | Published January 14, 2026 | AI across the medicine lifecycle, from research to safety monitoring | Multidisciplinary expertise, data governance and documentation, risk-based performance assessment |
| Concept paper on AI in clinical development | Being finalised; public consultation to follow | AI used in clinical development | Not yet public; industry has asked for clarity on validation and evidence |
| Horizontal machine-learning guidance | In development | Machine learning across regulatory use | Not yet public |
| Draft EU GMP Annex 22 (AI) | Consultation closed October 2025; EMA expected a final text by end of 2026 | AI models in critical GMP applications | Expert-approved acceptance criteria before testing; human review records |
| AI Guidance Coordination Group | New | Tracks and aligns AI guidance; does not draft it | None directly |
Two dates are worth putting on the calendar. The European Commission, EMA and FDA planned a joint discussion for November 2026. The HMA-EMA industry group meets again on February 2, 2027.
What do the FDA-EMA principles mean for training data?
In January 2026, FDA and EMA set ten shared principles for AI in drug development. Four of them land directly on the data your models learn from and are tested against.
Clear context of use. A model is validated for a purpose. Your evaluation set has to reflect that purpose: the therapeutic area, the user, the question types and the decisions the output feeds.
Multidisciplinary expertise. EMA's announcement says AI "needs to be expertly managed." In practice that means clinicians, MSLs, regulatory writers and statisticians shaping the data, not only engineers.
Data governance and documentation. You need to show where each item came from, who wrote or graded it, and what changed between versions.
Risk-based performance assessment. Testing depth should match the stakes. A literature triage tool and a safety signal tool do not need the same evidence.
What does draft Annex 22 tell us about validation?
Annex 22 covers AI in GMP manufacturing, not clinical development. It is still the most detailed signal of how European inspectors think about AI evidence, and it is a useful template for any regulated use.
According to Bioprocess Online's November 2026 analysis, the draft applies to static models with deterministic output in critical GMP applications. Generative AI and large language models "should not be used in critical GMP applications." For noncritical generative uses, qualified people must check outputs against the intended use.
The validation logic is the part to borrow. A process subject matter expert approves the intended use before acceptance testing. A subject matter expert sets the acceptance criteria before testing starts, and they must be at least as high as the performance of the process the model replaces. Reviewers examine the features driving model decisions. Low-confidence outputs can be flagged as undecided. EMA said in February 2026 that the draft drew about 1,300 comments, so wording may change. The principle of expert-set criteria, fixed in advance, is unlikely to.
Where does human oversight fit?
Everywhere regulators see real use. The October notes describe pharmacovigilance as a mature environment for AI, with uses that keep "scientific oversight and human decision-making."
That has a data consequence. If a human expert is the safeguard, the expert's judgment is part of the system. You need records of who reviewed what, how often they disagreed with the model, and how disagreements were settled. Draft Annex 22 goes further: depending on criticality, it may require reviewing or testing every model output.
Where should pharma data teams start?
Industry told regulators in October that its top priorities include validation and evidence requirements, and validation of AI systems in GxP environments. You do not need to wait for the answers to build the evidence base. We would start in three places.
Inventory the models that touch regulated work. List every model used in medical information, literature review, regulatory writing, pharmacovigilance and clinical operations. For each, write the context of use in one paragraph. Most teams find at least one tool in production with no written purpose at all.
Rebuild evaluation sets around that context. A general medical benchmark tells you little about whether a medical information assistant handles off-label questions in your therapeutic area. Evaluation items should be written and graded by the experts who do that work today, against criteria they set before the model is tested.
Make the documentation a deliverable, not an afterthought. Every training and evaluation item should arrive with its author, reviewer credentials, review history and version. If your vendor cannot supply that, you will be reconstructing it later, under deadline.
The checklist below turns this into questions your team can answer this quarter.
Checklist: is your AI data ready for EMA scrutiny?
- Each model has a written context of use: purpose, users, therapeutic area, decisions it informs.
- Evaluation sets match that context, not a generic medical benchmark.
- Acceptance criteria are set by named experts before testing begins.
- Criteria are benchmarked against the current human or manual process.
- Every training and evaluation item has provenance: source, author, date.
- Every reviewer has documented, verified credentials matched to the task.
- Each item carries its review history, including disagreements and how they were settled.
- Datasets are versioned, with change logs tied to label or guidance updates.
- Low-confidence and out-of-scope outputs have a defined escalation path to a human expert.
- Human review of live outputs is logged, with reviewer, decision and rationale.
- Testing depth is documented as proportionate to risk.
- Your quality team has reviewed the documentation package once, before any regulator asks.
How Acumen helps
We build medical affairs and regulatory AI data with verified MSLs, medical information specialists, medical writers and regulatory experts. Each item ships with the reviewer's credentials, its provenance, its review and adjudication history and a version record. Our experts can also set and document acceptance criteria for a specific context of use before your evaluation runs.
See medical affairs and regulatory AI data. For how to vet any expert data provider against these expectations, read what to ask a scientific expert data vendor.
FAQ
Has EMA published AI guidance for clinical development yet?
Not yet. EMA's October 2026 summary notes say a concept paper on AI in clinical development is being finalised and will go to public consultation. A horizontal machine-learning guidance is being developed in parallel. No dates were given. Teams should follow the consultation and prepare their evidence base now, since the direction is already set by the joint FDA-EMA principles.
What are the FDA-EMA guiding principles of good AI practice?
Published January 14, 2026, they are ten shared principles for AI in drug development. They include a human-centric design, a risk-based approach, adherence to standards such as GxP, a clear context of use, multidisciplinary expertise, data governance and documentation, risk-based performance assessment and life cycle management. They are principles, not detailed rules.
Does EU GMP Annex 22 apply to large language models?
The July 2025 consultation draft says generative AI and large language models should not be used in critical GMP applications, meaning those affecting patient safety, product quality or data integrity. Noncritical generative uses are allowed with qualified people checking outputs. The draft drew about 1,300 comments, and EMA expected a final text by the end of 2026, so wording may change.
Why does training data matter to regulators?
Because a model's behavior comes from the data it learned from and was tested against. The joint FDA-EMA principles name data governance and documentation directly, and draft Annex 22 requires expert-approved acceptance criteria before testing. If you cannot show where data came from and who reviewed it, you cannot fully explain how the model was validated.
What documentation should ship with AI evaluation data?
At minimum: the context of use the set was built for, reviewer credentials matched to each task, item provenance, the acceptance criteria and who approved them, review and adjudication history, and version logs. That package lets quality and regulatory teams trace any result back to the experts and decisions behind it.
Who should set acceptance criteria for a pharma AI model?
Subject matter experts who understand the process the model supports, set before testing begins. Draft Annex 22 asks for exactly this and requires criteria at least as high as the performance of the process being replaced. For medical affairs tools, that usually means medical information specialists, MSLs or regulatory writers in the relevant therapeutic area.