Insurance · mid-market broker

Quote turnaround, three days to twenty minutes.

How a mid-market broker turned a backlog of PDF submissions into a same-hour quoting pipeline — without hiring, and without a single silent error making it to a customer.

[ figure 1 — before/after: the submission queue ]
20 min
median turnaround
+40%
quotes / same team
0
silent errors shipped
3 wks
to production

The broker’s underwriters were not slow. They were buried. Every commercial submission arrived as a PDF — sometimes a scan of a scan — and someone had to read it, key the fields into three systems, and only then start actually pricing the risk. On a good day that was an hour of typing before any judgment happened. On a bad day, submissions aged out and business walked.

The problem, precisely

The temptation with a problem like this is to reach for a demo-ready chatbot and call it transformation. That would have failed here, loudly. Underwriting is a regulated, high-consequence workflow: a confidently wrong field isn’t a cute hallucination, it’s a mispriced policy. The real constraint was never “can a model read a PDF” — it was “can we trust it enough to remove a human, and where can’t we?”

The approach

So we didn’t try to automate the judgment. We automated the typing, and made the judgment faster and better-informed. Extraction runs on every submission, but every field carries a confidence score, and anything below threshold routes into a review queue instead of silently flowing downstream.

The goal was never a system with no humans. It was a system that spends human attention only where it changes the answer.

What we built

Five stages, deliberately boring, each independently testable:

# pipeline
intake   →  normalize PDF, split pages, OCR fallback
extract  →  field model + schema validation
score    →  per-field confidence, threshold gate
review   →  human queue for low-confidence only
bind     →  push to rating engine + audit log

How we knew it worked

Every claim above is backed by an eval suite that ran before a single underwriter touched it: a labelled set of 400 real (scrubbed) submissions, scored on field-level accuracy and, more importantly, on calibration — did low confidence actually correlate with being wrong? It did, which is the only reason the threshold gate is safe to trust.

What I’d do differently

We over-invested in the extraction model early and under-invested in the review UI. The queue is where the humans actually live, and a week spent there in month one would have paid for itself twice. It’s all in the book — including the parts that didn’t work.

// the implementation book

The Submission-Intake Playbook

Everything behind this case, published in full: architecture, the prompt suite, the eval harness and dataset spec, the review-queue design, and the rollback plan. Scrubbed of client data, free to read or fork.

Have a problem shaped like this one?