AI-Driven Underwriting Assistant

The lookups were the obvious problem and the least interesting one. Underwriters moved between property databases, claims history, credit reports, regulatory filings and geospatial maps, and every hop meant a context switch and a retype.

The allocation underneath it was the real finding. Simple applications were 80% of the volume and took 60% of the time. Complex applications were 20% of the volume and 80% of the revenue, and got what was left over.

Skepticism about black-box scoring came up in almost every conversation, and it was specific rather than general — people wanted the sources, the method, and a way to disagree on the record. Nobody I spoke to wanted the AI making the call. They wanted it to do the reading.

Data aggregation came first: one submission ID, 15+ third-party lookups fired in parallel, property through geospatial.

On top of that, a visual risk signal — red, yellow, green — chosen because it was already the vocabulary on the floor rather than because it tested well. Underneath it, the full breakdown with sources for anyone who wanted to argue with the signal.

The layer people asked for last and used most was the pattern work: which risk factors actually correlate with claims, which submissions convert. That only becomes useful once somebody trusts the first two layers, which is why it shipped in that order rather than first.

Usability testing went badly in a useful way. The model was identifying risk factors correctly and it did not matter — a score of 7.2 with nothing behind it read as a guess with a decimal point.

So the score got taken apart on screen. Prior claims history weighted 40%, property age 25%, geographic risk 20%, and so on down. Every data point links back to where it came from. Underwriters can adjust the weights or drop a factor entirely when they know something the model doesn't.

Adoption moved once people could see the weights. The overrides kept coming, which was the point — they stopped being complaints and started being corrections that went back into the model.

Property cares about flood zones. Liability cares about claims history. Workers comp cares about industry classification and safety records. A single fixed risk model would have meant picking one line and telling the others to wait.

So each line configures its own: which sources to query, how to weight them, what trips an alert, which indicators show on the surface view.

That is a considerably harder thing to build and it was a commercial decision more than a design one. It is also the reason the same platform could be sold into several lines instead of one.

- Configurability is the price of entry in B2B. Every customer wanted the same core product set up differently. Building for the first one would not have scaled to the second. - In a data-dense tool, the visual work is load-bearing. Hierarchy, color coding and spacing were what made a screen of aggregated risk data readable at all. That is not decoration, and it was the part that took longest to get right.

I validated initial usability and then stopped. What I would want now is three to six months of watching daily use: whether people drifted into accepting recommendations without checking, and what workflows they invented that nobody designed for. Both of those only show up over time, and neither was instrumented.