A mid-sized regional property & casualty insurer, ~1.2M active policies

المشكلة

Claims arrived as unstructured attachments - PDFs, phone-photo scans, handwritten forms - into a shared inbox, where 14 clerks read, classified, and keyed each one by hand. First-touch to adjuster assignment averaged 4.3 business days, and roughly 9% of claims were misrouted at least once.

ما بنيناه

We began with a two-week audit rather than a model, sampling 1,800 submissions to map where time actually went. Two findings reshaped the scope: 60% of volume came from four claim types, and most misrouting traced to a data-entry step, not a reading problem. We redesigned the routing workflow first, then built NLP extraction against the four high-volume types, running it in shadow mode until accuracy held above 94% on a live holdout.