Cutting Invoice Processing From 90 Seconds to 10 With Document AI

The problem

Around 900 supplier invoices arrived every month, almost all as email attachments. Formats varied enormously: clean digital PDFs from larger suppliers, phone photographs of paper invoices from smaller ones, and a few faxes that had been scanned twice.

Two people opened each one, found the supplier, invoice number, date, line items, VAT, and total, and typed them into the finance system. About ninety seconds each when the document was clean, several minutes when it wasn’t.

They had tried traditional OCR software twice. Both times it worked acceptably on the tidy digital invoices — the ones that were already easy — and failed on everything else, which meant staff had to check all of it anyway. Net time saved: approximately zero.

What I built

An extraction pipeline that watches the invoices inbox, and for each attachment identifies the supplier, pulls the header fields and every line item, and matches the invoice against the corresponding purchase order.

A review screen built for speed, not completeness. This is where most of the value came from. Rather than showing a blank form, it shows the extracted values next to the original document with every field pre-filled and low-confidence fields highlighted in amber. The reviewer’s job became confirming rather than typing. Ten seconds for a clean invoice, longer only when something is genuinely flagged.

Confidence-based routing. Invoices where every field is high-confidence and the total matches an existing purchase order within tolerance are auto-approved and posted. Everything else goes to a human. That threshold was set deliberately conservatively and tightened over the first two months as the accuracy data came in.

A discrepancy report for invoices that didn’t match their purchase order. This wasn’t in the brief. It turned out to catch supplier overcharging that had previously been absorbed without anyone noticing.

Decisions worth noting

A human stayed in the loop from the start, permanently. The system prepares the data; a person approves it. Not because the model is unreliable — measured field accuracy settled at 97.8% — but because the failure mode of unattended financial data entry is unacceptable at any accuracy rate. The time saving comes from eliminating typing, not from eliminating oversight.

Accuracy was measured before anything was trusted. For the first three weeks the pipeline ran alongside the existing manual process, extracting in parallel without touching the finance system. That gave a real per-field accuracy figure across their actual document mix — not a vendor’s benchmark — and made the automation threshold an evidence-based decision.

Failures are loud. If the pipeline can’t process a document, it lands in a visible queue and the team is notified. Silence never means success. The most dangerous automation is the kind that fails quietly for a month.

The results

Average handling time per invoice fell from about ninety seconds to roughly ten. Across 900 invoices a month, that’s approximately fifty hours returned to a two-person finance team.

Field-level accuracy measured 97.8% across the full document mix, including the phone photographs. Around 60% of invoices now clear the auto-approval threshold without human review at all.

The unplanned discrepancy report found around £14,000 in supplier overcharges in its first six months.

---

Note: this is a representative case study written to illustrate the shape of a typical engagement while real client write-ups are being prepared and approved. Get in touch if you’d like to talk through a comparable project.