Automating supplier document handling with document AI

Around 1,200 supplier invoices arrived weekly across PDF, scanned paper, and email body text, in as many formats as there were suppliers. Three full-time staff keyed them into the ERP. Transcription errors surfaced weeks later as payment disputes, and the backlog grew each quarter. An off-the-shelf OCR product had been trialled and abandoned because it failed silently — wrong values entered the ERP with no indication anything was uncertain.

1. 01

   Started by measuring the actual distribution of document formats. Around seventy percent of volume came from twelve suppliers, so most of the value was reachable without solving the general case.

2. 02

   Built an extraction pipeline combining a document model with deterministic validation — totals must reconcile, dates must be plausible, supplier references must exist. The model proposes; the rules verify.

3. 03

   Made confidence explicit and routed on it. High-confidence, fully-reconciling documents post automatically; anything uncertain goes to a review queue with extracted fields shown against the source document.

4. 04

   Designed the review screen so correcting a field takes a keystroke, on the basis that the humans handling exceptions matter more to throughput than the model's raw accuracy.

5. 05

   Wrote back to the existing ERP through its supported interface, leaving the finance team's downstream processes and reports untouched.

About 92% of invoices now post without human involvement. The remaining 8% reach a review queue taking a fraction of the original effort and — critically — the system no longer fails silently, because anything it is unsure about it declines to post. The three staff previously keying invoices moved to supplier management and exception handling.

• Python
• Document AI
• PostgreSQL
• AWS
• React