Regulatory Affairs AI Agent · EU human medicines
AI Agent for Regulatory Affairs
RafiHive is an AI agent for EU pharmaceutical regulatory affairs. It searches defined official EU/EEA regulatory sources and your own documents, assesses regulatory questions and dossier evidence, and produces source-cited outputs for qualified human review. It does not submit, decide or advise on its own.
RAFI stands for Regulatory Affairs Factual Intelligence. The agent exists to give EU pharma RA professionals more time to apply their expertise: it does the retrieval, comparison and first-pass drafting, and shows its sources so you can check them. You assess the evidence, resolve the uncertainty and own the decision.
What is a regulatory affairs AI agent?
A regulatory affairs AI agent is software that carries out multi-step regulatory work on request rather than answering from memory. Given a question or a document, it retrieves the applicable sources, assesses the material against them, and returns a structured, cited result a reviewer can verify. When evaluating a tool, check its source scope, whether it distinguishes law, guidance and drafts, and what it does when evidence is missing. These capabilities depend on the product and configuration.
RafiHive applies that definition to EU human-medicines regulatory affairs. The agent is maintained as assistive decision support: it prepares work for a qualified regulatory professional and never replaces one. What it does and does not do is written down on the AI governance page.
How the RafiHive regulatory AI agent works
Five steps, in the same order every time. The same path serves a free-text regulatory question and the preset dossier checks; the checks add a fixed scope and a fixed finding format.
- 01Scope
You ask a regulatory question, or set the submission scope for a preset check: procedure type, product, target markets and the evidence you have. Attach documents if you want them assessed; the Module 1 check runs from a declared inventory without any upload.
- 02Retrieve
The agent searches the reviewed knowledge base of official EU/EEA sources (EMA, European Commission, EUR-Lex, CMDh, EDQM, eSubmission and national authorities) together with the documents you attached. An optional web fallback can be switched on per conversation; its findings are labelled separately and never treated as reviewed sources.
- 03Assess
Retrieved evidence is separated from assumptions. Law, guidance, Q&A, draft and superseded material are distinguished so a reviewer can see the legal weight behind each statement, not just the statement.
- 04Cite, or stop
Every substantive claim carries a citation to the source it came from. When the knowledge base, your documents and any approved web evidence do not contain sufficient support, the agent returns a not-found result instead of generating an answer.
- 05Deliver for review
Preset checks return one row per finding: evidence status, severity, source locator and the reviewer action it needs. Research answers carry confidence, assumptions and a qualified-review boundary. Outputs export to DOCX, XLSX, CSV, Markdown and HTML for your own review trail.
The step that matters most is the fourth one. When retrieval finds no sufficient evidence in the knowledge base, your documents or approved web findings, the agent returns a not-found result and no answer is generated. Missing coverage is reported, never guessed. See a full sample output for what the finished answer looks like.
What the agent does today
Regulatory research
Ask about variations, procedures, dossier structure, labelling, ICH quality topics or national requirements. Answers cite the source, state assumptions, grade confidence and end with the review boundary.
Common RA workflowsEU eCTD Module 1 gap check
Set the submission scope, attach the dossier evidence you have or declare the inventory, and receive a cited first-pass deficiency list: each item present, missing, unclear or not applicable, with the reviewer action it needs. Runs on the Team plan; the scope brief is free.
About the Module 1 checkEU packaging & labelling check
Attach the SmPC, labelling text and package leaflet. EU-wide requirements are checked once, then the national requirements run once per declared market, from the published CMDh blue box text (the MHRA's Windsor Framework rules for Northern Ireland). Cross-document mismatches are reported on their own. Runs on the Team plan; the scope brief is free.
About the packaging checkWhich regulatory sources does the AI agent use?
A reviewed knowledge base, not the open web. 59 EU/EEA source documents and 76 ICH guidelines are listed from the current manifest, each with its version and review status. Sources come from these authorities:
The agent prioritises those official sources before anything else. An optional web fallback exists for questions the knowledge base cannot cover; its findings are labelled as web findings, are never treated as reviewed sources, and your private documents are never sent to it. Known gaps in coverage are published alongside the loaded sources on the source coverage page, so you can see what the agent cannot yet answer from.
Can an AI agent review regulatory documents?
Yes, and this is where an agent earns its place over a chat window. RafiHive assesses the documents you attach against the applicable requirements: dossier evidence against the EU Module 1 specification, product information against Directive 2001/83/EC, the QRD template and each market's national requirements. Every finding names the evidence found, the source that requires it and the action a reviewer should take.
Documents are processed in the EU under the GDPR, encrypted in transit and at rest, isolated to your workspace, never used to train foundation models, and deleted when you say so. RafiHive is operated from Germany. The full lifecycle of an uploaded file is on the data handling & security page.
Can an AI agent monitor EMA regulatory changes?
Honest answer: RafiHive is not a general regulatory change-monitoring feed, and this page will say so until it is. What it does today is narrower and useful. When a newly reviewed source is indexed into the knowledge base and semantically matches the context of one of your projects, that project receives a regulatory alert, and you can open a follow-up chat from it with the new source already in scope. Source freshness and version handling are described on the source coverage page.
How should you compare AI tools for regulatory affairs?
Compare AI tools on source traceability, evidence gaps, data handling and human oversight. Capabilities vary by service, plan and configuration. Use these checks on a representative regulatory question and verify the supporting documents yourself.
| Property | What to verify in any AI tool | RafiHive regulatory affairs AI agent |
|---|---|---|
| Source scope | Check which sources are searched and whether document versions and approval status are visible. | Reviewed EU/EEA official sources with version and status, listed publicly on the source coverage page. |
| When evidence is missing | Test a question outside the available evidence and check whether the tool identifies the gap. | Instructed to report missing evidence; retrieval without sufficient support returns a not-found result. Outputs still need verification. |
| Citations | Open the cited document and confirm that the passage supports the claim and applies to your procedure. | Citations are required by the answer instructions; fabricated citations are measured on the published benchmark, not ruled out for every answer. |
| Legal weight | Check whether binding law, guidance, Q&A, draft and superseded material are distinguished. | Law, guidance, Q&A, draft and superseded material are marked as such. |
| Your documents | Verify processing location, training use, access controls and deletion terms for the specific service and plan. | Processed in the EU under the GDPR by a German-based operator, never used to train foundation models, isolated to your workspace, deletable at any time. |
| Output shape | Ask for a sample that shows evidence, missing information and the action a reviewer should take. | Evidence status, severity, source locator and reviewer action per finding; confidence and assumptions per answer. |
| Human review | Confirm which decisions stay with qualified professionals and whether the tool can act externally. | Outputs require qualified RA review. RafiHive does not submit to authorities or make final regulatory decisions. |
| Measured accuracy | Inspect domain-specific test cases, scoring, review status and limits; a benchmark is not a guarantee on new questions. | Golden-set retrieval, citation and safety results are published on the evaluation page. |
How the agent is evaluated
Every release is scored against a golden set of 16 regulatory questions with known expected sources. The numbers come straight from the last harness run on 2026-09-07.
An answer passes only if every expected source is cited, no citation is invented and no forbidden claim appears. Method, limits and RA sign-off status are on the evaluation page. A golden set is a regression guard, not proof of correctness on your question.
Built to support RA professionals, not replace them
Human oversight
Every substantive output carries a qualified-review requirement. The professional who uses it remains responsible for verification and the final decision.
No autonomous action
The agent does not submit to EMA, NCAs, CESP or IRIS, does not make final classification or submission decisions, and does not give legal or medical advice.
Grounding first
Official knowledge-base evidence is prioritised, citations are required where evidence exists, and unsupported gaps are stated plainly rather than filled.
Built in Germany by EU RA specialists
A European product. EU regulatory-affairs specialists shape the sources, workflows and evaluations alongside the engineers in Germany, so the EU framework is in the product rather than layered over a generic model. Hosted in the EU under the GDPR.
The intended-use statement, prohibited uses and AI Act posture are on the AI governance page; who builds RafiHive and why is on the about page.
Frequently asked questions
Who is RafiHive for?
RafiHive supports regulatory-affairs professionals and teams working on EU human medicines. It helps with official-source research, first-pass drafting, EU eCTD Module 1 gap checks and Module 1.3 packaging and labelling reviews. Its defined scope is EU/EEA pharmaceutical regulatory affairs; it does not establish coverage of every jurisdiction or replace qualified professional judgment.
What is a regulatory affairs AI agent?
A regulatory affairs AI agent is software that carries out multi-step regulatory work on request: it retrieves regulatory sources, assesses a question or a document against them, and returns a result for review. Capabilities depend on the product and configuration. RafiHive supports EU human-medicines regulatory affairs with a defined source scope, legal-weight distinctions and instructions to report missing evidence. Every output requires qualified human review.
Is RafiHive an autonomous AI agent?
No. RafiHive runs multi-step research and dossier checks, but it does not submit documents to EMA, national competent authorities, CESP or IRIS, does not make final classification or submission decisions, and does not give legal or medical advice. A qualified regulatory professional reviews and signs off every output. This boundary is documented on the AI governance page.
Which regulatory sources does the AI agent use?
A reviewed knowledge base of official EU/EEA sources: EMA, the European Commission, EUR-Lex, CMDh, EDQM, eSubmission and national competent authorities, plus ICH guidelines. Each source is listed with its version and review status on the source coverage page, and known gaps are listed alongside them. Your own documents are retrieved in the same step when you attach them.
Can the AI agent review regulatory documents?
Yes. You can attach dossier evidence, SmPC, labelling text and package leaflets, and the agent assesses them against the applicable requirements with a citation for each finding. The EU eCTD Module 1 gap check and the EU packaging and labelling check are preset document reviews with a fixed finding format. Uploaded documents are processed in the EU under the GDPR, isolated to your workspace, never sent to the web fallback and never used to train models.
Can an AI agent monitor EMA regulatory changes?
RafiHive is not a general change-monitoring feed. What it does today: when a newly reviewed source is indexed into the knowledge base and semantically matches the context of one of your projects, that project receives a regulatory alert and you can open a follow-up chat from it. Source freshness and version handling are described on the source coverage page.
How is the accuracy of the AI agent measured?
RafiHive is evaluated on a defined benchmark of regulatory questions with known expected sources. The harness measures retrieval hit rate, answer pass rate, fabricated citations and retrieval of superseded or unapproved sources for those cases. The latest run, method and review status are published on the evaluation page. These results do not measure every customer answer or establish correctness on a new question; qualified review still applies.
Can I try RafiHive, and which plan includes dossier checks?
RafiHive offers 3 free questions with no payment card, including private document analysis in one conversation. The public Module 1 and packaging scope briefs are free without an account. The cited preset Module 1 and Module 1.3 checks require Team or Enterprise access; Individual RA includes research chat and private document analysis. Team starts at 5 seats. Current prices and plan details are on the pricing page.
Verify product fit against the current plans, source coverage and gaps, and benchmark method and limits.
See the agent work on a real regulatory question.
Three free questions, no card. Your questions stay in the EU and every answer is traceable to a named source for your review.