Purchase orders, invoices and customs paperwork arrived as PDFs and email attachments, and a team of eleven re-keyed them into the ERP by hand. Every error was found downstream, usually at payment, and every correction cost more than the original entry.

We built a pipeline that reads each document, extracts the fields the ERP needs, validates them against vendor master data, and posts only what passes. Anything ambiguous lands in a review queue with the source page highlighted, so a human decides in seconds instead of retyping minutes.

Straight-through processing now covers seven of every ten documents. The team moved from data entry to exception review, and month-end close went from nine working days to four without adding headcount.

Before touching the model we mapped every document type and its failure history. That map decided the order of work: highest volume, lowest ambiguity first, so the team banked time savings while the harder templates were still being tuned.