Home / Case Studies / Regulatory AI
Pharma

PharmaRegAI: cutting regulatory dossier preparation by 70%

Regulatory affairs teams spend their weeks assembling dossiers from documents that already exist. We built an AI assistant that finds the right source, drafts the section and shows exactly where every sentence came from.

70%
less time preparing a dossier
100%
of drafted text traceable to a source
0
documents leaving their tenancy
3 wks
from kick-off to first live draft
The challenge

Where it started

Every submission meant re-reading hundreds of pages — specifications, stability data, previous dossiers, guidance documents — to assemble sections that were 80% the same as last time.

The team could not use a general AI tool: an unattributed sentence in a regulatory document is a serious problem, and nothing could leave their tenancy.

Client: iCretegy

What we built

The solution, in four parts

Retrieval over their own library

Approved documents, previous submissions and guidance indexed with permissions respected, so the assistant answers only from content the user is allowed to see.

Drafting with citations

Every generated paragraph carries a link to the source passage. Reviewers check the citation, not the whole document.

Change and gap checks

The assistant flags where a new submission differs from the approved precedent and where a required section is missing.

Human sign-off, recorded

Nothing is submitted without a named reviewer approving it, and the approval is part of the audit record.

Technology

What it runs on

Mainstream, well-supported technology — chosen so the client’s own team can own it afterwards.

Azure OpenAIHybrid retrievalPythonReactPostgres + pgvectorEvaluation harness

They understood our regulatory constraints before they wrote a line of code — and the citations are what convinced our QA team.

Head of Regulatory Affairs, iCretegy

More work

Other outcomes we’re proud of

Want an outcome like this one?

Tell us the number you need to move. We’ll tell you honestly what it takes and how quickly you’d see the first release.