AI agents are moving out of demos and into regulated work. FDA runs them internally. Pharma companies are piloting them for regulatory intelligence, authoring and document analysis. So the question is no longer whether regulatory affairs will use AI agents. It is how much we should let them do.
Most of the agent hype answers that question with "as much as possible". Having built an agent for EU regulatory affairs, I have ended up somewhere else entirely.
THE USUAL AGENT PROMISE REGULATED WORK
Understand Retrieve
↓ ↓
Plan Understand context
↓ ↓
Decide Compare evidence
↓ ↓
Execute Cite
↓ ↓
Act autonomously Identify uncertainty
↓
Escalate
↓
Human decidesMore autonomy is not automatically better AI.
AI agents are no longer just developer tools
On 1 December 2025, FDA announced agency-wide agentic AI capabilities. Not a pilot in one division, the whole agency. The use cases it listed were meeting management, pre-market reviews, review validation, post-market surveillance, inspections and compliance, and administrative work. The announcement described agents as systems that plan, reason and execute multi-step actions, with built-in guidelines "including human oversight".
This did not come out of nowhere. FDA had rolled out Elsa, its internal LLM tool, in May 2025, and by December said more than 70% of staff were using it voluntarily. The models run in GovCloud and do not train on anything industry submits, which is exactly the detail a regulated company should be asking its own vendors about.
Industry followed through in 2026. In June, IQVIA published a detailed description of agentic regulatory intelligence, built as a small team of specialised agents:
Detection agent monitors regulatory sources, spots changes
↓
Classification agent decides relevance by region, therapy area, product
↓
Impact agent maps the change to affected documents
↓
Orchestrator coordinates, escalates findings to RA expertsIBM, meanwhile, has Regulate.AI, a multi-agent authoring accelerator co-developed and piloted with global pharma organisations for regulatory submissions and clinical reports.
Two years ago, regulatory teams were pasting guideline paragraphs into ChatGPT to see what would happen. This is a very different conversation.
An AI agent is not just a chatbot with tools
"Agent" is being stretched to cover everything, so a quick definition. A chat assistant waits for you at every step:
Human asks → AI responds → StopAn agent holds on to a goal across several steps and decides the next step itself:
Goal → Plan → Retrieve → Use tools → Evaluate result → Next step → Complete or escalateIn regulatory affairs the difference is concrete. A chatbot answers "what changed in the new guideline?" when you remember to ask. An agent does something closer to this, without anyone prompting each step:
Detect a new EMA document
↓
Work out what kind of document it is
↓
Compare it with the previous version
↓
Identify affected topics
↓
Map them to the products you hold
↓
Prepare a draft impact assessment
↓
Notify the RA professionalThat is genuinely useful. Monitoring is the kind of work that eats afternoons and gets skipped when a submission deadline lands. It is also exactly where the autonomy question starts to bite.
Regulatory work has a different failure model
When a coding agent gets something wrong, you usually find out quickly. The tests fail, the build breaks, someone complains in the pull request. The feedback loop is short and loud.
Regulatory errors are quiet. An agent can:
- use guidance that was superseded last year,
- treat a draft out for consultation as if it were adopted,
- apply centralised procedure guidance to a national procedure,
- blend requirements from two jurisdictions,
- invent a citation that looks entirely plausible,
- miss the one exception that decides the case,
- or build a convincing answer on evidence that was incomplete from the start.
None of those produce an error message. The output reads fine. You find out months later, from an assessor, which is the worst possible feedback loop. I went through why this happens in why AI gets regulatory affairs wrong when you remove the source documents. Agents make it more serious, not less, because now the wrong premise from step two feeds steps three to seven.
In regulatory affairs, an answer without provenance is incomplete
Take a sentence an agent could plausibly produce:
This change can be submitted as a Type IB variation.
Sounds useful. Now ask what a regulatory professional would ask before acting on it:
Which guideline?
Which version, applicable since when?
EU requirement or national?
Which procedure?
Which classification entry exactly?
Are all of its conditions actually met?
Is the source still current?If the agent cannot answer those, the sentence is not an answer. It is a guess with good formatting. In regulated work the conclusion is only half the output. The evidence chain behind it is the other half, and it is the half an assessor will look at.
The regulatory direction is becoming clearer
Regulators are saying something similar, if more politely. In January 2026 EMA and FDA published ten joint guiding principles for good AI practice in drug development. I covered them alongside the AI Act in what the EU AI Act means for AI agents in regulatory affairs, so only the themes that matter here: human-centric design, a risk-based approach, a clearly defined context of use, data governance, multidisciplinary expertise, lifecycle management and clear information about what the system does and does not do.
Notice what is missing from that list. Nobody wrote "maximum autonomy".
We ran into the same question while building RafiHive
RafiHive is the regulatory affairs agent we build for EU pharma teams. The first prototype was the architecture everyone starts with:
Ask a question → LLM answersIt demos beautifully. The trouble starts when people who do this work for a living use it. Not because the answers are always bad, but because nobody can tell which ones are. The further we got into the domain, the less acceptable that design became, and most of what RafiHive is today came from turning each failure into a rule:
| Problem | Design decision |
|---|---|
| General model knowledge goes stale | Work only from defined regulatory sources |
| Guidance changes and gets replaced | Track source status and versions |
| Models fabricate citations | Verify every source reference |
| Evidence is often incomplete | Surface the gap instead of inventing certainty |
| RA questions depend on context | Keep product and procedure context visible |
| AI sounds surer than its evidence | Separate evidence from interpretation |
| Some calls need professional judgment | Make human review explicit |
| AI quality is hard to judge from outside | Publish evaluation results |
Every row on the right cost us a feature that would have looked great in a sales demo. I would make the same trade again.
Regulatory AI needs bounded autonomy
IQVIA uses the phrase "bounded autonomy" in its piece too, which I took as a good sign rather than competition for a buzzword. My working definition:
Give the agent enough autonomy to remove repetitive work, but not enough authority to quietly turn uncertainty into a regulatory decision.
In RafiHive that boundary looks like this:
Regulatory sources
↓
Retrieve evidence
↓
Compare requirements
↓
Analyse supplied documents
↓
Identify gaps
↓
Produce first-pass assessment
↓
Cite supporting sources
↓
Flag uncertainty and missing evidence
↓
┌──────────────────────────────┐
│ Qualified RA professional │
│ reviews and decides │
└──────────────────────────────┘The agent does: research, retrieval, comparison, extraction, first-pass assessment, gap identification, drafting and citation.
The human stays responsible for: regulatory strategy, interpretation in ambiguous cases, risk acceptance, final classification, submission decisions and every interaction with the authority.
RafiHive does not submit anything to EMA or a national competent authority. That is not a missing feature. That line is the product.
What this looks like in practice
Say a company wants to add a manufacturing site to an authorised EU medicinal product. Depending on the site, the product and what exactly changes, that can land in different variation categories with different conditions and documentation. A well-bounded agent can take a lot of the legwork:
1. Identify the relevant variations framework.
2. Retrieve the current classification guidance.
3. Narrow down the classification entries that could apply.
4. List the conditions and required documentation for each.
5. Check the supplied dossier documents against those conditions.
6. Report what is missing.
7. Produce a first-pass assessment with citations.
8. Hand it to the RA professional.That can compress a substantial amount of manual research into a much shorter first pass, while keeping each step reviewable.
What it should never say is "I have submitted the variation for you". Almost as bad: "this is definitely Type IB", when the dossier does not show whether a key condition is met. The honest output in that situation is "Type IB if condition X is met; the supplied documents do not show it; here is the source". Less satisfying. Much more useful.
If we're going to trust agents, we need to test the boring things
A lot of AI benchmarking asks whether an answer looks correct. For regulatory AI that is the wrong test, because looking correct is precisely what these systems are good at. The useful questions are narrower:
Did it retrieve the expected source?
Did it use the current version?
Did it retrieve anything superseded?
Did it invent a citation?
Does the citation actually support the claim?
Did it recognise missing evidence?
Did it avoid overstating the conclusion?RafiHive's current published evaluation runs a 16-question golden set: 100% retrieval hit rate, 75% answer pass rate, zero fabricated citations, zero superseded or unapproved sources retrieved. The page also says plainly what those numbers do not prove. A small golden set is a regression guard, not a guarantee about your question.
These numbers will change. They should. If the benchmark never gets harder as the product improves, it has stopped measuring anything. I would take that over a vendor telling me their agent is "95% accurate" with no word about accurate at what.
Regulatory AI is moving from answers to workflows
RafiHive is not the only team landing here, which honestly is reassuring. IQVIA's regulatory intelligence design puts human oversight at the points where regulatory judgment is required, and keeps an audit trail of every inference. IBM describes Regulate.AI as keeping humans "firmly in the loop" while agents handle formatting and data processing. FDA's own deployment names human oversight in the first description of the technology.
Different companies, different problems, same shape. Regulatory AI is moving from answer generation toward workflow participation, and everyone serious is drawing a boundary somewhere inside that workflow.
The real design question is where the agent stops
"Agent or no agent" is the wrong debate. The useful one is which actions get which boundary:
LOW AUTONOMY HIGH AUTONOMY
Search Summarise Analyse Recommend Decide Act
│ │ │ │ │ │
└──────── useful automation ────────┘ │ │
│ │
regulatory judgment executionSearching and comparing sources can take a lot of automation, because the output is easy to check and cheap to redo. Recommending needs citations and visible uncertainty. Deciding and acting need a named, qualified human, because that is where accountability already sits, and no model is going to be the one explaining the decision to an assessor.
This is not an anti-automation position. I want agents doing far more of the reading than they do today. I just want them to stop at the right place.
Maybe the best agent is the one that knows when to stop
The first phase of generative AI was about how much work a model could do. Agentic AI moved the question to how much it can do without being asked at every step.
In regulatory affairs I think a third question matters more than both: does the system know where the evidence ends and human judgment begins? An agent that can answer that honestly is worth more than one that can do twice as many steps on its own.
Domain experts are the people who know where that line sits, which is why they belong in the product loop and not just in the acceptance test.
RafiHive is our attempt to build around that boundary for EU pharmaceutical regulatory affairs.