Claims sat in queues, not in review
Nordwind processes 3,000 claims a month. Each one crossed five systems before an adjuster saw it: the intake portal, a document store, the policy system, a spreadsheet of exceptions and an email thread. Adjusters spent most of their day assembling context, not making decisions.
The backlog grew 18% year over year. Hiring more adjusters was the standing plan — and the reason Nordwind came to us instead.
A claims copilot inside the adjusters' own tool
We built a claims workbench: one screen that pulls the claim, the policy, the documents and the history together, with an LLM layer that reads the file first — extracting the loss details, checking coverage against the policy and drafting a recommendation with citations back to the source documents.
Adjusters stay in charge: every recommendation shows its evidence, and one click accepts, edits or rejects it. The design work — ten years of interface craft — is why adoption hit 95% in week one, not month three.
A day of assembly became a minute of review
Measured over the first 90 days in production against the quarter before, from Nordwind's claims system — their numbers, not ours.
In their words
“The recommendations show their sources, so adjusters trust them. The adoption fight we budgeted for never happened — by week two nobody wanted the old five screens back.”
What we would do next
Subrogation is the next queue with the same shape: scattered context, repeatable judgement. The agent layer and workbench extend to it directly. Scoped at $40K to $60K, 8 weeks.
Nothing here is committed. It is the honest answer to the question the client asked at handover, and the estimate we gave for it.
Team on this project
The same people who scoped it built it. No handover to a delivery pod.
Related work
Questions about this project
Twelve weeks from the scoping call to production: two of audit and design, eight of build with weekly demos in the client stack, two of hardening, training and handover.
The agent layer is model-agnostic behind one interface — it shipped on a frontier LLM and was benchmarked against two alternatives. Swapping models is a config change plus an evaluation run, not a rebuild.
Every recommendation carries citations to source documents, and an evaluation suite of 1,200 historical claims runs on every change. Below-threshold cases route to a human by default.
Everything runs in Nordwind's own cloud account. No claim data leaves their tenancy; model calls use a zero-retention endpoint. The security review is part of the audit week.
Nordwind does — repository, infrastructure and evaluation suite, documented and handed over. Eazetech keeps no access after the 30-day support window unless asked to.
The workbench pattern fits any queue where people assemble context from several systems before deciding. The audit week answers it precisely — if the answer is no, you keep the map of your process.