The setup

The carrier handles roughly 40,000 first-notice-of-loss claims a year across personal and small-commercial P&C lines. Triage — the decision about which adjuster a claim goes to and in what order — was running off a rule table that had been edited incrementally for a decade and a queue manager whose primary input was who was logged in. The result was the pattern any operations leader recognizes: simple claims that should have closed in two days were waiting four or five because they sat behind a complex bodily-injury matter on the same adjuster’s desk, and complex claims weren’t reliably getting to the adjusters with the right experience.

Leadership had two related concerns. Average cycle time was tracking ten to fifteen percent above regional peers, which was showing up in customer-satisfaction scores and renewal economics. And the rule table had become institutionally fragile: nobody could fully explain why specific routes existed, several rules contradicted each other, and changes required cross-team negotiation that took weeks per edit. The carrier was not asking for a model for the sake of having one — they were asking why their highest-volume operational decision was the part of the business that received the least analytical attention.

What we did

We treated the triage decision as a scored routing problem rather than a rules problem. The first phase was the unglamorous one: assembling a clean training set out of three years of historical claims, with the routing decisions, downstream cycle times, reassignments, and outcome dispositions joined into a single record per claim. That join surfaced the pattern that had been hidden in the queue: reassignment rates for certain coverage-and-severity combinations were running above thirty percent, which meant the system was paying for triage twice on those claims.

We then built a complexity-and-fit model that produces two outputs per incoming claim — a complexity score and a ranked list of compatible adjuster cohorts — and wired it into the existing claims platform via a thin service the platform already supported. We deliberately did not replace the platform, the assignment rules a regulator might ask about, or the supervisor’s authority to override. The model is a recommendation surface; the routing decision still belongs to the platform and to the supervisors who manage exceptions.

The harder design call was the supervisor surface. Adjusting routing logic in a regulated business requires that humans can see why the model proposed what it proposed, override it cleanly, and have those overrides feed back into model evaluation. We built a review surface that shows the top features driving each score, the adjuster cohorts being recommended and why, and a one-click override that captures the supervisor’s reasoning as structured data. Overrides are not friction — they are training signal.

We rolled the model out shadow-mode first for six weeks, comparing its recommendations against the existing routing without changing any actual decisions, and only flipped it to live routing once supervisors had calibrated against its behavior on their own desks.

What changed

Average cycle time across all lines fell by 38% within the first full quarter after live cutover. The larger move was on simple claims, where time-to-close fell by 51% — those claims stopped queueing behind complex matters because the model recognized them on intake and routed them to fast-track adjusters. Reassignment rates fell by 44%, which is the metric that most directly translates to internal cost: the carrier is no longer paying for the same triage decision twice.

Adjuster headcount did not change, and was not intended to. The carrier was not trying to reduce staff; they were trying to get more out of the staff they had, and to put the experienced adjusters on the claims that actually required experience. Customer-satisfaction scores on closed claims moved up materially over the same window, though the carrier is appropriately cautious about attributing that to a single intervention.

The rule table was retired in favor of a documented, version-controlled scoring model that supervisors and compliance can both reason about. Changes that used to take weeks of cross-team negotiation now ship in days, with a clear audit trail.

Why it stuck

Two reasons. First, the supervisor surface was treated as a first-class part of the system rather than a console bolted on at the end. Supervisors have not disengaged from triage — they are engaged on the exceptions where their judgement is highest-leverage, and the model handles the routine bulk that was crowding their attention. Second, the override channel feeds the evaluation loop, so the model gets better in the directions the business actually cares about, not the directions a generic accuracy metric would push it. The 38% number is the externally visible result; the internal result is that triage is now a decision the carrier can reason about, change deliberately, and defend.