A regulatory affairs lead at a pharma company told us last quarter that a hallucinated citation in a regulatory submission isn't a software bug. It's a misrepresentation. Misrepresentations in pharma submissions trigger investigations, withdrawals, and potential criminal liability. The agent's tolerance for hallucination is precisely zero.
This shapes the entire architecture. Pharma agents don't generate text and then add citations. They retrieve citations and then generate text grounded in them. The order matters.
Sourcing-first architecture
A working pharma agent looks like a retrieval system with an LLM as the synthesis layer. Specifically:
- The query is decomposed into sub-questions.
- Each sub-question hits a curated corpus — internal documents, peer-reviewed literature, regulatory archives.
- The retrieval results are presented to the LLM with explicit instructions to quote and cite and refuse to assert anything not in the retrieved context.
- The output includes structured citations to every claim.
If the retrieval doesn't surface evidence for a claim, the agent doesn't make the claim. It says so. This is operationally different from "use the LLM and hope it cites correctly." The hope-and-pray pattern fails in pharma every time.
What gets retrieved matters more than what gets generated
The corpus the agent retrieves from is the most important asset:
- Curated. Hand-selected papers, regulatory documents, and internal records — not "everything we could find."
- Up-to-date. Pharma moves; the retrieval index has to track new submissions and approvals.
- Versioned. Yesterday's retrieval and today's retrieval may differ. Both have to be reproducible if a regulator asks.
- Provenance-tagged. Every document has metadata about its source, peer-review status, and regulatory standing.
The LLM is downstream of this corpus discipline. Most pharma-AI projects under-invest in the corpus and over-invest in the prompt. The trade-off is wrong.
What pharma agents help with
Literature review. Given a research question, surface the relevant peer-reviewed papers, summarise their findings, and cite specifically. The medical writer reviews the synthesis and uses it as input.
Adverse-event narrative drafting. Given a case report, draft the adverse-event narrative for safety reporting. The pharmacovigilance specialist reviews and signs.
Regulatory-document drafting. Given a study report and the regulatory requirements, draft sections of the submission with clear traceability to source documents. The regulatory affairs specialist reviews and signs.
Internal Q&A for medical-affairs teams who get product questions from sales and need to research the answer in compliant, citable form.
In every case the agent drafts. The licensed specialist reviews and signs. Same pattern as legal and clinical agents.
What we won't ship
Anything that interacts directly with patients. Patient-facing pharma communications are tightly regulated; they don't tolerate the variance an LLM introduces.
Promotional material drafting. "On-label" promotional language requires legal and medical review at every step; agents don't shortcut that workflow.
Any system that auto-publishes regulatory content. A human signs every regulatory submission. Always.
The reviewer workflow
Every output from a pharma agent flows through a reviewer of record:
- The reviewer sees the draft, the retrieved sources, and the model's reasoning.
- The reviewer can accept, edit, or reject.
- Edits and rejections feed back into the eval set.
- The reviewer's signature attaches to the final output.
This makes the agent's output defensible. It also bounds the agent's authority — nothing leaves the reviewer's desk without their signature.
The escalation tree
Some queries should never be answered by the agent:
- Questions about off-label use.
- Questions about specific patients or competitors.
- Questions that touch on legal interpretation of regulations.
The agent's first job, on every query, is classification: does this query fall in the agent's scope, or does it need to escalate? Misclassification here is a regulatory event waiting to happen. The eval set focuses heavily on the boundary cases.
How to start
Pick one workflow with high specialist time and clear documentation requirements. Internal Q&A for medical affairs is a good starter — high volume, lower stakes than regulatory submissions, strong existing documentation requirements.
Build the curated corpus first. Build the retrieval layer. Build the citation discipline. Run the agent in shadow mode for a quarter. Only after the citation discipline is provably tight does the agent's output go directly to specialists.
Close
Pharma agents work when they're retrieval engines with an LLM synthesis layer — not the other way around. The corpus is the load-bearing asset. The reviewer of record signs every output. Citation discipline is the design, not a feature. The teams that take this seriously ship something useful and defensible. The teams that lead with the LLM end up with a regulatory issue.
Related reading
- Agents in legal: contract review with receipts — same citation discipline, different domain.
- Agents in healthcare: scribe yes, nurse no — same human-decision boundary.
- RAG is a public library — the retrieval discipline pharma agents need.
We build AI-enabled software and help businesses put AI to work. If you're shipping a pharma agent, we'd love to hear about it. Get in touch.