The obvious next hire after our first engineer was a second engineer. We hired Sarah — a claims adjuster with twelve years of experience — instead. Her job: sit next to us for six weeks and tell us everything we were getting wrong.
What she found
She redlined 400 FNOL documents and flagged 23 extraction patterns that were wrong in ways we would never have caught: reserves set on multi-vehicle claims that should have been split, coverage dates read as loss dates, endorsements ignored in favor of the base policy.
None of these appeared in our accuracy metrics. They looked like successful extractions. They were extracting the wrong thing correctly.
What she changed
Our mental model. We had been building extraction flows. She showed us we should be building decision-support flows. The accuracy numbers went up as a side effect.
The lesson
For a product that lives in the details of a specific industry's documents, the domain hire is the thing that determines whether the product actually works. It feels like a luxury. It isn't.
%20(1).png)
