Claims triage that explains its own decisions

The situation

Claims triage was manual and inconsistent. Two adjusters could route the same claim differently, and neither decision was recorded in a way that could be reviewed later.

Constraints

This is a regulated decision affecting individuals, so it was treated as high risk from the first design session. Every automated decision needed a logged reason and a human able to overturn it.

What we did

A triage model scoring 12,000 claims a day, with three model cards covering the components and a documented override path used by the claims team.

The override is not an escape hatch bolted on at the end. It is the primary interface: the model proposes a route, the adjuster confirms or changes it, and the reason feeds the next evaluation cycle.

Governance in practice

When the regulator asked how a decision was reached, the answer took an afternoon rather than a project, because the documentation had been written during the build rather than reconstructed afterwards.

Results

Measure	Before	After
Average handling time	14 days	9.7 days
Claims routed automatically	0%	62%
Decisions with a logged reason	Partial	100%
Time to answer a regulator	Weeks	Same day