The shared mailbox problem
Every claim arrived in one mailbox. Someone opened it, worked out what kind of claim it was, which team owned it, whether it was urgent, and forwarded it on.
Average time to first response was around two days, most of which was queue time before anyone had read anything.
Why the scope stayed narrow
The obvious pitch here is claims assessment. We did not scope it and we were against it from the start.
Cost of error on an assessment is high and the decision is regulated. The value of automating it is much lower than it looks, because the assessment is not where the time goes. The time goes in the triage, and triage errors are cheap and immediately visible.
So the scope was: classify, prioritise, route, and acknowledge, with nothing that touches the merits of a claim.
What was built
Inbound messages and attachments are read, classified by claim type and urgency, and checked against the policy record. They are routed to the owning team with a structured summary attached. The claimant gets an acknowledgement within minutes rather than days.
Anything the system is unsure about goes to a single human queue with the ambiguity flagged, rather than being guessed at.
Measurement before load
We built the evaluation set from six months of historical claims that were already routed. That let us score proposed routing against what the team did. That ran for three weeks before the workflow carried any live volume.
The first version sat at 91 percent. It went live at 96. We found that one claim type was systematically misread because of how one insurer formats their reference numbers.
Result
First response time dropped from around two days to about four hours. The operations team stopped starting each morning with an hour of mailbox sorting.
Every assessment is still made by a person.