Engagement snapshot
- Mandate
- Reduce adjuster backlog without pretending coverage decisions could be fully automated.
- Timeline
- 14 weeks from discovery to statewide rollout.
- Team shape
- Claims ops director, 2 platform engineers, 1 ML engineer, and 3 senior adjusters.
The problem
Adjusters had to jump between policy packets, repair estimates, claim notes, and photo evidence before deciding whether a claim was ready for straight-through handling or senior review.
What we built
Built a retrieval-grounded claims copilot that assembled policy citations, summarized loss context, flagged missing evidence, and routed uncertain cases to senior adjusters.
Operating context
Storm season drove a sharp increase in inbound claims, but each file still required manual comparison across multiple policy versions, vendor estimates, and image evidence. The insurer needed faster review without losing compliance discipline or creating a black-box decision layer.
Key constraints
- No generated policy language could be shown unless it traced back to the exact policy edition and clause.
- Claim photos and personal data had to stay inside the insurer environment with auditable access controls.
- Adjusters needed the experience inside their existing claim screen rather than in a separate AI portal.
What we built
Version-aware policy retrieval
Indexed endorsements, riders, and state-specific policy forms so every answer pulled the exact coverage language tied to the claim.
Structured claim assembly
Combined repair line items, adjuster notes, and evidence status into a machine-readable case packet before any model step.
Confidence-based escalation
Only low-risk recommendations stayed in the standard adjuster lane; anything ambiguous or missing evidence was escalated automatically.
Delivery path
Failure-mode audit
Reviewed historical claim files to identify where junior adjusters slowed down, over-escalated, or missed relevant policy clauses.
Shadow-mode pilot
Ran the copilot alongside live adjusters for three weeks, comparing citation quality and escalation decisions before enabling recommendations.
Controlled rollout
Released by claim type and region with adjuster QA review, telemetry on citation usage, and rollback paths for retrieval drift.
Why it mattered
The system did not replace adjusters. It removed repetitive evidence assembly and policy lookup work so adjusters could spend their time on true judgment calls. Senior reviewers saw cleaner escalations, and frontline adjusters got faster, better-grounded starting points.
Implementation notes
- Trust rose materially once every recommendation exposed the exact policy source and claim evidence behind it.
- Escalation quality mattered more than raw automation rate because the insurer still had regulated decision boundaries.
- Retrieval freshness for endorsements and regional rule updates was a bigger reliability issue than model selection.