A scalable knowledge base — evidence on demand

A global pharma's market access function runs the same clinical and real-world evidence through a different argument in every market — different comparator, different payer priorities, different local data. Before this engagement, that evidence lived in whatever format each function last used it in: a literature-review export here, a biostatistics output there, a prior dossier's appendix somewhere else. Every new market request meant reassembling the same underlying facts from scratch, and because no two analysts reassembled them the same way twice, two dossiers answering the same clinical question could cite different numbers. That's not a productivity problem a faster process fixes — it's a governance problem, and it's the one MLR reviewers actually push back on.

Deployed inside the pharma's own environment, not a hosted SaaS tool the evidence passes through. Market access and HEOR evidence is commercially sensitive long before a dossier is public, and a regional team in one APAC market needed to generate from the same base as global HQ without either side re-exporting data to the other. That ruled out the common shortcut of a shared cloud workspace with broad access for "whoever needs the evidence this quarter" — the access model had to follow the same market-by-market boundaries as the regulatory submissions themselves. See data residency for healthcare AI for why that constraint shapes the architecture rather than just the contract.

A governed evidence base — not a document repository, a structured, versioned collection where each source carries its provenance and is tagged against indication, population, comparator, endpoint and study design. On top of it: a continuous systematic literature review layer that keeps the base current rather than re-running from scratch per request (see AI for systematic literature review); a dossier-assembly agent that drafts the global dossier and its Asia-regional localisation from the same source material; and a localisation layer that adapts argument and language per market's label and comparator set without duplicating the underlying evidence (see localising payer value stories across markets). Six local value stories plus two regional dossiers now generate from the one base.

Every claim in every generated document carries a reference back to the specific source passage that supports it — not a bibliography at the end, a passage-level link per statement, the same discipline described in AI value dossier automation. Before a market access professional ever sees a draft, automated checks confirm every citation resolves, every cited passage actually supports the claim it's attached to, and nothing in the draft falls outside that market's approved label. The reviewer's job changed from reconstructing an argument to verifying one that already arrives with its sources attached — which is also what makes MLR review faster, since the reviewer is checking evidence that's already traceable rather than tracing it themselves. Where a claim's supporting passage doesn't exist in the governed base, the system is built to produce nothing for that claim rather than a plausible-sounding substitute — the same principle covered in AI hallucination mitigation in regulated settings, applied here to a dossier rather than a chat response.

Reused across 8 APAC markets from a single source of truth. For this workflow, dossier and value-story creation moved from a span of weeks and months down to days, and every claim stayed traceable for MLR review throughout — the speed and the auditability improved together, not one at the other's expense, because both came from the same underlying change: evidence assembled once, with provenance attached, instead of reassembled per market without it.

The market access team now operates the evidence base directly — adding a market is largely configuration (that market's label, comparator set and local sources), not a new build. systematic_lit_review · dossier_assembly · localisation · reference_linking · heor_agent — all registered modules now carried into later engagements rather than rebuilt.