What is a regulatory affairs AI agent?

A working definition

A regulatory affairs AI agent is software that takes a regulatory task, works through it in steps — deciding what to search, retrieving evidence from defined sources, comparing that evidence against the question or the dossier — and returns an output that shows what it found and where. It supports a regulatory professional; it does not replace one.

The word 'agent' is used loosely in the market. The useful distinction is not how conversational a tool is, but whether it acts on a controlled body of evidence and exposes its work. A tool that writes fluent text from its training data is a writing assistant. A tool that retrieves from named, versioned sources, cites them, and stops when the evidence is not there is closer to what RA work needs.

How it differs from a general chat assistant

A general assistant answers from patterns learned across the open web. It does not know which version of a guideline it absorbed, whether that version is superseded, or whether the text it paraphrases is law, guidance, a Q&A answer or a consultation draft. It will usually produce an answer either way.

An RA agent should reverse each of those defaults. Its source scope is defined and published. Each document carries its authority, legal status and version. Retrieval comes before generation, so the answer is built from passages the reader can open. When nothing in scope supports an answer, the right output is 'not found', not a plausible guess.

• Defined source scope, published so users can see what is and is not covered.
• Version and status on every source: current, superseded, draft.
• Legal weight distinguished: legislation, guideline, Q&A, procedural advice.
• A citation on every substantive claim, pointing to the passage used.
• A not-found result when the evidence is not in scope.

What an RA agent can do well today

The strongest use is first-pass research: finding the provisions, guideline entries and procedural documents that bear on a question, and laying them out with their status so a professional can read the right things first. That is where hours go in practice, and it is the part a machine can do quickly and check itself on.

The second is structured comparison. Given a declared submission scope or a set of documents, an agent can check what a checklist or template expects against what is present, and return one row per finding with a source locator. Module 1 completeness and SmPC, labelling and leaflet consistency are good examples: the rules are written down, and a missed item is costly.

The third is drafting support — a cited summary, a first version of a classification note, a structured list of assumptions — written so a reviewer edits it rather than starting from a blank page.

What it should not decide

An agent can identify candidate classification codes and the conditions attached to them. It cannot know whether a condition is actually met at your site, whether an undeclared change has drifted from the approved dossier, or how an assessor in a particular Member State has handled similar cases. Those facts live outside the sources.

Regulatory strategy, the final classification, the decision to submit and the content of what is submitted are professional judgements with legal consequences for the marketing authorisation holder. A well-designed agent makes that boundary explicit on every output instead of leaving it to the fine print.

The regulators' own direction of travel

Regulators are building the same distinction into their thinking. On 14 January 2026 EMA and the US FDA published ten jointly identified guiding principles for good AI practice in drug development. They address AI used to generate evidence across the medicines lifecycle rather than RA research tools specifically, but the themes — human-centred design, a clear context of use, a risk-based approach, data governance, performance assessment and lifecycle management — are the right questions to ask of any AI in a regulated workflow.

EMA states the principles will underpin future AI guidance, building on its 2024 reflection paper. For RA teams, the practical reading is that 'what is this tool for, what data does it use, how is it measured and who reviews it' will increasingly be expected answers, not optional ones.

Questions to ask before using one

Treat an RA agent like any other system whose output informs a regulated decision. The questions below separate tools built for the work from general models with a regulatory landing page.

• Which sources are in scope, and is that list published with versions?
• How is superseded material handled — excluded, or labelled?
• What happens when the answer is not in the sources?
• Is there a published evaluation, and does it count fabricated citations?
• Where are your documents processed, and are they used to train models?
• Who in your organisation reviews and signs off the output?

Where RafiHive fits

RafiHive is built as this kind of agent for EU pharmaceutical regulatory affairs: a reviewed knowledge base of EU/EEA official sources, cited outputs, a not-found result when support is missing, and a stated requirement for qualified review. Its source coverage and golden-set evaluation results, including their limits, are published so the claims above can be checked rather than taken on trust.