The problem
Finance teams were manually keying invoice data — vendor, amount, line items, PO references — from PDFs and scanned images into the approval system. This was slow, error-prone, and created bottlenecks whenever invoice volume spiked at month-end close.
The approach
Built an end-to-end pipeline that removes manual entry from the process entirely, using Microsoft's AI Builder for document understanding rather than a custom-trained model — the right tradeoff for a well-defined document type where a managed model handles the extraction reliably out of the box.
What it does
- Ingests invoices from email attachments and manual uploads
- Uses AI Builder's prebuilt invoice-processing model to extract vendor, amount, date, PO number, and line items from PDFs and scanned images
- Validates extracted data against PO records and flags mismatches for review instead of silently auto-approving
- Routes valid invoices through a Power Apps approval flow with manager sign-off
- Stores the original document plus structured metadata in SharePoint, keeping a full audit trail for compliance
Outcome
Cut manual data entry out of the invoice intake process and gave Finance a consistent, auditable approval trail instead of email threads and spreadsheets. Exceptions (mismatched amounts, missing POs) are now surfaced automatically rather than discovered during reconciliation.