AI invoice and document-processing pipeline for a logistics operator

A freight operator was keying around 4,000 supplier documents a month into its ERP by hand. We built an ai-automation pipeline that extracts fields from invoices, delivery notes and customs paperwork, cross-checks them against purchase orders and posts straight-through cases to SAP, routing only exceptions to people. Roughly 80% of invoices now post without a manual touch.

Challenge

Documents arrived as PDFs, scans and photographs from more than 300 suppliers, in four languages and dozens of layouts. Two accounts-payable staff spent most of their week typing, and month-end closes slipped because invoices sat in inboxes. Earlier attempts with template-based OCR broke every time a supplier changed its layout. The finance director wanted fewer late-payment penalties, an audit trail per document, and no change to the ERP itself.

What we built

• Intake from a shared mailbox and an SFTP drop, with de-duplication and document-type classification (invoice, credit note, delivery note, customs declaration).
• Layout-agnostic extraction: Azure Document Intelligence for tables and key-value pairs, with a GPT-4o pass to normalise supplier names, currencies, VAT lines and free-text references into a fixed schema.
• A validation layer that matches each invoice to purchase orders and goods receipts, checks totals and tax arithmetic, and assigns a confidence score per field.
• Straight-through posting to SAP Business One for high-confidence, fully matched documents; a review queue in n8n with side-by-side document and extracted fields for everything else.
• Every decision logged with the source page and the rule or model that produced it, so auditors can trace any posted line back to the scan.

Results

Around 80% of invoices post automatically ; the rest reach a reviewer within minutes with the mismatch highlighted. Typical processing time fell from about three days to same-day , and the field-level error rate after validation is under 1% on the monthly sample audit. The AP team was redeployed to supplier reconciliation and dispute handling, and month-end close now finishes two working days earlier. Related pillars: nlp for the multilingual normalisation and ai-integration for the ERP connector.

What made the difference

We built the review queue first and the automation second. Staff corrected extractions from day one, which gave us a labelled evaluation set within a fortnight and a clear, measured path from "review everything" to "review exceptions", one supplier group at a time.