Document AI that reads 30,000 invoices a month, 99.2% accurately

Northgate receives around 30,000 supplier invoices a month in every imaginable format: scanned PDFs, phone photos, spreadsheets, and emails. A team of eight spent their days retyping them into the ERP, with a 4% error rate that caused payment disputes downstream.

Off-the-shelf OCR tools captured the text but not the meaning. They could read the numbers but could not tell a line-item discount from a tax adjustment, so everything still needed manual review.

We built an extraction pipeline that combines OCR with an LLM layer that understands invoice semantics: supplier, line items, taxes, currencies, and payment terms, mapped directly to Northgate's ERP schema.

A validation layer cross-checks totals, flags anomalies like duplicate invoices or unusual amounts, and routes only genuine edge cases to a human review queue. The team of eight now reviews exceptions instead of typing everything.