Faster claims triage for a specialty insurer
We turned scattered PDFs into a triage copilot that keeps adjusters in control.
The client handled complex claims that arrived as long PDFs, photos, and email threads. Adjusters spent hours chasing missing details and retyping summaries. We built a triage copilot that pulled the right evidence together without losing auditability.
What we shipped
-
Ingestion with a document contract. We extracted key fields - policy number, loss date, jurisdictions, counterparties - and dropped any file that missed the contract. Missing fields created tasks for intake, not silent failures.
-
Retriever tuned for regulatory clauses. We biased the retriever toward exclusions and endorsements, not just recent documents. Every answer cited source pages and showed “last indexed” metadata.
-
Decision checklist generator. The copilot produced a checklist per claim: coverage questions, missing evidence, and suggested next actions. Adjusters could accept or edit each item before sending.
-
Audit-friendly controls. The UI kept a side-by-side view of the generated summary and the cited snippets. Adjusters could pin or exclude sources, and every override was logged to the case file.
Results
- First-touch decisions dropped from days to hours because adjusters stopped hunting for exclusions across attachments.
- Regulatory review time shrank thanks to the explicit citation list and “only use cited sources” toggle.
- Intake accuracy improved because missing contract fields created tasks instead of slipping through.
The stack was intentionally small: a rules-light retriever, a citation-aware prompt, and a feedback loop tied to weekly evals. The team kept ownership of the workflow while the copilot handled the tedious discovery work.