Intelligent automation and claims triage for a United States insurance carrier

First notice of loss arrived through four channels in inconsistent formats, and every claim was triaged manually regardless of complexity. A straightforward low value auto glass claim received the same handling as a complex liability matter, which meant simple claims waited behind complex ones and adjusters spent their expertise on work that did not need it. Claim cycle time was above the carrier's competitive benchmark and adjuster attrition was high, with exit interviews consistently citing administrative burden.

The carrier had been pitched an end-to-end AI claims platform by two vendors and rejected both, on the grounds that neither could explain what the system would do when it was wrong.

We built a narrower system deliberately. Intake across all four channels is normalised and classified. Claims are scored for complexity and routed accordingly, with straightforward claims following a largely automated path and complex claims routed to adjusters with the file already assembled and summarised. Document extraction handles the supporting evidence. Nothing in the system makes a coverage or settlement decision. Every routing decision is explainable, logged and reversible, and adjusters can override any classification with one action and no justification required, which is what made the system acceptable to the people who had to use it.

We ran the implementation against United States data residency requirements, with processing configured to remain in the United States and the Ontario team working under a documented cross-border access model reviewed by the carrier's compliance function. Bilingual English and Spanish handling was built in for the carrier's southern states.

Twenty-two week implementation followed by a managed operation covering monitoring, exception handling, model maintenance and continuous improvement, with an Ontario-based team.

Cycle time on straightforward claims reduced by 54 percent. Adjuster time per complex claim fell measurably as file assembly and summarisation moved upstream. Routing accuracy exceeded the carrier's acceptance threshold from month three, with override rates falling steadily as the model improved on real overrides. Adjuster satisfaction rose, and the carrier attributed a reduction in adjuster attrition partly to the change. No coverage or settlement decision was ever made by the system, which was the carrier's condition for proceeding at all.

The two rejected vendors had promised more. Promising less, and being precise about exactly where the system stops and a human starts, is what got this one into production.