Underwriting agents with a human review gate

Placeholder. Every figure on this page is invented. Replace before launch.

The situation

Underwriters were reading the same document types by hand, all day. Volume had roughly tripled since the Series A, and the team's answer had been to hire — which worked until it didn't, because the bottleneck was attention, not headcount.

The constraint that shaped everything: this is a regulated product. A decision needs a person on the record and a trail showing how it was reached. Anything that removed the human entirely was not shippable, regardless of accuracy.

What we built

A pre-scoring pipeline that runs before a human opens a case:

• Extraction pulls the fields that matter out of submitted documents, with a confidence score on every field rather than a single score for the document.
• Scoring ranks the case and, crucially, explains itself — the underwriter sees which evidence drove the score.
• Routing resolves the unambiguous cases automatically and pushes anything below the confidence threshold into a human queue, ordered by how much the decision is worth.

The review gate was built in week one, not bolted on at the end. That ordering matters: teams that add the human checkpoint last tend to discover their whole data model assumed nobody would ever need to intervene.

What we'd do differently

We spent the first two weeks building extraction before establishing a baseline for how the manual process actually performed. When the numbers came in, we could show improvement — but we were reconstructing the "before" from memory. Measure first, even when it feels like a delay.