Prior title found rate, up 6 points in 12 weeks.

The problem

Title search requests arrived as free-text emails, faxed cover sheets, and scanned PDFs. Before an examiner could begin a search, someone read each request and re-keyed roughly 25 fields — parcel numbers, party names, prior deed book and page references — into the search platform.

The transcription step cost minutes per request, and worse, it cost accuracy: a mistyped parcel number or a missed prior-deed reference meant a search that started from incomplete data. The company's prior title found rate — how often a search surfaces the company's own prior policy, which shortens the search dramatically — sat at 32%.

What we built

An intake pipeline that reads every inbound request, classifies the document type, and extracts the 25 fields into the search platform — with a completeness score per request. Anything below threshold, or any field the model flags as low-confidence, routes to a human queue before the search begins.

Examiners verify flagged fields in a single review screen. Every extraction, edit, and approval is logged with a timestamp and reviewer identity — giving the team complete visibility into every decision the system made.

The result

Cleaner, more complete search inputs lifted the prior title found rate from 32% to 38% within twelve weeks of deployment — a six-point gain on the metric that most directly drives search cost. Intake transcription time fell by 2.4 minutes per request across roughly 350 daily requests.

Every one of those requests still passes human review. The examiners didn't get replaced; they got promoted back to the judgment work the job is actually about.