EU AI Act 2026: what it means for AI agents in pharmaceutical regulatory affairs

Two years ago the question in most regulatory affairs departments was whether you could use generative AI at all. That question is gone. The one replacing it is harder: if we use it, how do we control it, validate it, trace its sources, protect confidential documents and keep a human responsible for the outcome?

I build software for this space, so I get asked some version of that question every week. I am not a lawyer and this is not legal analysis. It is the view from the person who has to turn "responsible AI" from a slide into actual design decisions, and who has made a few wrong ones along the way.

August 2026 changed the conversation

The EU AI Act became generally applicable on 2 August 2026. That sentence is true and also slightly misleading, which explains a lot of the confusion I hear.

The Act was never one switch. Prohibited practices and the AI literacy duty have applied since February 2025. General-purpose model rules came in August 2025. What August 2026 added is mostly enforcement machinery (national authorities, penalties) and the Article 50 transparency duties, which means telling people when they are talking to an AI system. Then, days before the deadline, the Digital Omnibus amendment entered into force and pushed the big high-risk obligations back: stand-alone Annex III systems to December 2027, AI inside regulated products to August 2028.

So "the AI Act applies now" means different things depending on what you build and who you ask. But the practical effect inside pharma companies is clear enough. Legal, quality and IT are all asking about AI at the same time, and "we are just piloting a chatbot" no longer ends the meeting.

One thing worth knowing early: if you fine-tune, rebrand or substantially modify an AI system, you can stop being a deployer and become a provider. That shift does not wait for 2027.

Pharma has a second AI layer on top

This is where a generic AI Act article stops being useful for anyone working with medicines.

In January 2026, EMA and FDA published a joint set of ten guiding principles for good AI practice in drug development. They are not binding guidance. They are, however, the clearest signal yet of how both agencies think about AI across the medicines lifecycle, and regulators tend to ask about the things they have already written down.

The themes will look familiar to anyone who has been through a GxP audit: human-centric design, a risk-based approach, a clearly defined context of use, data governance, multidisciplinary expertise, alignment with existing drug development standards, lifecycle management and plain-language information about what the system does. None of it is exotic. All of it is much easier to design in than to bolt on afterwards.

Put the two together and you get a fairly specific picture. The AI Act sets the legal floor. The EMA/FDA principles describe what "good" looks like to the people who eventually review your submissions.

Regulatory affairs is an unusual AI use case

A wrong restaurant recommendation is annoying. A confidently invented regulatory citation is a different category of problem.

Regulatory work runs on precision that general-purpose AI is not built for. It is never just "what does the guidance say". It is which source, which version, which jurisdiction, which procedure, as of which date, and with what legal weight. A Q&A document from CMDh does not carry the same authority as the Variations Regulation. A draft guideline out for consultation is not the current expectation. A national requirement in one member state may not exist in the next.

An answer can be 90% right and still be useless, because the 10% it got wrong was the part that decided the case.

Why generic chat is not enough

I wrote about this in more detail in why AI gets regulatory affairs wrong when you remove the source documents, so the short version here. When a general chat assistant answers a regulatory question, these are the failures I keep seeing:

• Citations that look perfect and point to sections that do not exist.
• Guidance that was superseded years ago, described in the present tense.
• Draft and final versions of the same document treated as interchangeable.
• EMA guidance blended with legislation until you cannot tell which is which.
• National requirements simply missing, because the model averaged across Europe.
• And the worst one: a full, fluent answer in a situation where the evidence was never there to begin with.

That last failure is the one the AI Act and the EMA/FDA principles are really circling around. The model did not know, and nothing in the system forced it to say so.

What we changed when building RafiHive

RafiHive is the regulatory affairs agent we have been building for EU pharma teams. I am bringing it in here not as a pitch but because it is the only system I can describe honestly from the inside, including the parts that are not finished.

Most of the design came from turning each failure above into a constraint:

Problem                        Design choice

Unknown source             →   Defined source scope
Outdated information       →   Version and status tracking
Hallucinated citation      →   Citation validation
Missing evidence           →   Stop and report the gap
AI making the final call   →   Qualified RA review
Black-box performance      →   Published evaluations
Confidential documents     →   EU processing, workspace isolation

In practice that means the agent works from a reviewed set of EU/EEA sources (EMA, European Commission, EUR-Lex, CMDh, EDQM, national authorities, ICH guidelines) and not the open internet. Every source is labelled as law, guidance, Q&A, draft or superseded, so the reader can see the weight behind each claim. Substantive claims need a citation. When the knowledge base does not support an answer, the agent returns "not found" instead of improvising one. Customer documents are processed in the EU and never used for model training.

None of this is clever. Honestly, the hardest part was saying no to features that would have demoed well.

Publish the evaluation, including the uncomfortable numbers

This is the part I think the whole industry should copy, whether they use our product or not.

"Our AI is accurate" is not a claim anyone can check. What can be checked is a set of narrow, boring questions:

• Did it retrieve the source a specialist would expect?
• Did it invent a source?
• Did it use superseded or unapproved guidance?
• Did it make a claim the evidence does not support?
• Did it know when to stop?

Our current published run, from 7 September 2026, uses a golden set of 16 regulatory questions covering variations, dossier structure, labelling, ICH quality topics and the blue box. The results: 100% retrieval hit rate, 75% answer pass rate, zero fabricated citations, zero superseded or unapproved sources retrieved. An answer only passes if every expected source is cited, nothing is invented, the required concepts are present and no forbidden claim appears.

75% is not a marketing number. I publish it anyway, because a vendor showing 100% on everything is either testing easy questions or not telling you something. The same page also says that none of the 16 cases carry a named RA sign-off yet, and that a golden set is a regression guard, not proof that your specific question will be answered correctly. Both of those things are true, and a buyer in a regulated company deserves to know them before signing anything.

If a vendor cannot tell you how often their system invents a citation, they have not measured it.

Does an AI agent need to be autonomous to be useful?

Most of the current excitement around agents is about autonomy. Plan, execute, decide, act. Fewer humans in the loop, more work done while you sleep.

In regulatory affairs I think maximum autonomy is the wrong goal. The architecture I trust looks more like this:

AI retrieves
    ↓
AI compares
    ↓
AI structures
    ↓
AI cites
    ↓
AI identifies gaps
    ↓
Human assesses
    ↓
Human decides

RafiHive does not submit anything to EMA or national competent authorities, and it does not make final classification or submission decisions. People sometimes read that as a missing feature. It is a system boundary, and a deliberate one. The agent does the reading, comparing and structuring, which is where most of the hours go. The qualified person keeps the judgment, which is where the accountability already sits.

That split also happens to line up with human oversight under the AI Act and the human-centric principle in the EMA/FDA document. I did not design it that way because of either. I designed it that way because every RA professional I showed an early prototype to asked the same question: "and who signs this?"

What kind of AI will pharma trust with regulated work?

In 2026 nobody serious is asking whether pharmaceutical companies will use AI. They already do. The real question is which AI systems they will trust with regulated work, and I think the answer is going to come down to a short, unglamorous list:

• Traceability. Every claim leads back to a document you can open.
• Provenance. You know the version, status and date of that document.
• Defined context of use. The system says what it is for and what it is not for.
• Measurable evaluation. Published, repeatable, with the bad numbers left in.
• Data governance. Where confidential documents go, and where they never go.
• Human oversight. A named, qualified person makes the decision.
• Stated limitations. In writing, before the sales call, not after the incident.

None of that requires waiting for December 2027. Most of it is just good engineering that regulated industries were always going to demand eventually. The AI Act made "eventually" a lot closer.

RafiHive is our attempt to apply these principles to one narrow problem: AI-assisted EU pharmaceutical regulatory affairs.