The claims queue, governed.

A queue measured in weeks.

Claims arrived faster than adjudicators could review them. Routine claims — the clear majority — consumed the same senior attention as genuinely contested ones. Backlog grew, cycle time stretched, and the cost of review scaled linearly with volume.

A previous automation pilot had stalled in security review: no audit trail, no escalation design, no answer to "who is accountable for a denial."

Bounded autonomy, by design.

We started from the accountability question, not the model. The agent was given explicit authority boundaries: it can approve routine claims within policy terms; it can draft determinations; it cannot deny. Every denial path routes to a human adjudicator with the agent's full reasoning attached.

Policy documents were indexed into a permission-aware retrieval layer so every determination cites the exact clause it relies on.

Four components, one audit trail.

Intake parses and classifies the claim. The policy RAG layer retrieves the controlling terms. The agent drafts the determination with citations. The review console routes exceptions and records the human decision. All four write to a single immutable log.

Faster, and defensible.

Within one quarter of cutover, routine claims cleared in hours instead of days, contested claims reached senior adjudicators sooner, and audit had a complete decision trail for the first time.