Document intake that reads carrier forms without a person.
Ninety-four percent field accuracy without review; ninety-nine with. Time per document from nine minutes to under one. They’ve decided to go ahead with the full build after renewal season, leaving out the two carrier formats that weren’t worth automating.
01 · The situation
What it was costing them before we started.
Enrollment and claims documents arrived as PDFs from dozens of carriers, each in its own format, and were re-keyed by hand. Renewal season doubled the volume and the error rate.
- 9 min to re-key one carrier document by hand.
- 2× the volume in renewal season, on the same team.
- Dozens of carrier formats, each one different, none of them going away.
02 · The approach
How we went about it.
Scoped deliberately as a four-week paid pilot rather than a build, because the honest answer to “can this read our documents?” is a number, and nobody had one. One process, one team, the top twelve carrier formats, on live documents. The decision at the end was a real decision, including the option to stop.
- Week 0
Twelve carrier formats chosen by volume. Success threshold agreed in writing before any work.
- Weeks 1 to 4
Extraction built and measured on live documents in the client’s own tenancy.
- Week 4
Findings presented with the two formats we recommend not automating, and why.
03 · What we built
The pieces, and what each one is for.
A four-week pilot: document extraction across the top twelve carrier formats, on a live subset, with a review screen for anything below a confidence threshold. Scoped to one process and one team. Built to OWASP practice with data kept in the client’s tenancy.
- Extraction across twelve formats
The top twelve by volume, measured individually rather than as an average that hides the bad one.
- Confidence review screen
Anything below the threshold goes to a person, with the source document beside the extracted field.
- In the client’s tenancy
Documents never left their environment. OWASP practice, per-person credentials, removed at the end.
- An honest recommendation
Two of the twelve formats were not worth automating. That is in the report.
04 · The result
What changed, and by how much.
- Field accuracy, no review Not tracked 94%
- Field accuracy, with review Not tracked 99%
- Time per document 9 min under 1 min
- Formats recommended for automation 12 10
Four-week pilot, March 2026.
Does this sound a bit like your week?
Tell us how it works at your place today, and we’ll tell you honestly what it would take.
Book a 15-minute callNot ready to talk yet? Take the build-or-buy check