A health-claims bureau where a person re-checked every machine-read document, then typed it again into the next system

Documents arrive from clinics and payers all day: claim forms, receipts, supporting notes, small slips paid on a card. The bureau already ran them through machine reading, but the read was never trusted, so a person opened every document and checked it line by line. The recurring errors were small and constant: a date read in the US order instead of the local one, a field pulled in the wrong language, the wrong reference number chosen off a busy form. The smallest items, loose receipts and card-paid expenses, took the most thought of all.

Then the same data crossed a gap. Bookkeeping lived in one system, but the end-of-period hand-off went out as a verification balance that a second person re-entered into a separate tool to build the statements. Files moved by hand between a to-process folder and a processed folder on a shared drive, and when pieces were missing the team built the chase list by hand and emailed it. Nothing was automated end to end, and the team's own read was: there is always some re-keying, and no tool removes all of it.

We did not try to remove the human. We removed the second keying. The tool reads each document once and scores its own confidence field by field. Anything it is sure of passes clean. Anything it is unsure of, the dates, the reference number, the odd handwritten slip, surfaces in a short review queue with the original image beside it, so the processor confirms a handful of fields instead of re-checking the whole document.

Once a record is confirmed it flows straight into the billing system, no retyping at the boundary, and the end-of-period balance moves across the gap as structured data instead of a re-entered file. The to-process and processed split the team already used became live status the system tracks on its own, and missing-document chase lists are assembled from what is actually outstanding rather than compiled by hand. The extraction is AI-driven with a deterministic fallback behind it, so a messy scan still produces a usable record and never simply fails.

The work that used to mean opening and re-keying every document now means confirming the few fields the system was unsure about. Each processor gets back roughly twelve hours a week, and the same team clears more files without adding people. The errors that used to slip through, a date in the wrong order, a mismatched reference number, are caught at the point of entry instead of found later in a reconciliation.

The second keying between systems is gone. The end-of-period balance crosses as data, the chase lists build themselves from what is outstanding, and the status of every file is visible in one place instead of inferred from which folder a document sits in.