AI Document Processing Development

Description

Our accounts team keys in supplier invoices by hand — around 1,200 a month, from roughly 200 suppliers, in every format you can imagine. We want to extract line items, totals, tax and PO references automatically and push them into our ERP. The interesting part is not extraction accuracy on a clean invoice; it is what happens on the messy 15%. We want a confidence threshold and a review queue, so a human checks the uncertain ones instead of the system silently posting a wrong total. We would rather review 200 invoices a month than find one wrong payment a quarter later. You will need to handle multi-page invoices, invoices where line items run across a page break, and suppliers who change their template without telling anyone.

• Build extraction for line items, totals, tax and PO reference across varied supplier templates
• Handle multi-page documents and line items that break across pages
• Implement a confidence score per field and route anything below threshold to a human review queue
• Build the review screen so a person can correct a field and have the correction feed back into accuracy
• Push approved invoices into our ERP through its existing API

• Extraction service with a documented API
• Review queue interface for the accounts team
• Accuracy report broken down by supplier template, measured on a held-out set of real invoices
• ERP integration, tested end to end in their sandbox