The Problem

A mid-sized logistics company came to us with a straightforward but painful problem. Their accounts payable team was processing between 800 and 1,200 supplier invoices per week. Each invoice required a team member to open the document, extract the key data, match it against a purchase order in their ERP system, flag discrepancies, and either approve or route for exception handling.

The process was taking approximately 40 person-hours per week across two team members. It was error-prone, with a reconciliation error rate of around 2.5 percent. And it was a bottleneck: suppliers were waiting an average of 14 days for payment confirmation, which was straining relationships with key partners.

What We Built

We used Google Document AI to handle the extraction layer. The model was trained on a sample of the company's historical invoices to recognize their specific supplier formats. It extracts invoice number, date, line items, amounts, tax codes, and supplier identifiers with an accuracy rate of 97.8 percent on standard formats.

Extracted data flows into a matching engine built on n8n that pulls the corresponding purchase order from their ERP via API. The system checks line items against the PO, validates amounts within a configurable tolerance threshold, and confirms that the supplier ID matches the approved vendor list.

Invoices that pass all validation checks are automatically approved and pushed to the payment queue. Invoices with discrepancies are flagged with a specific reason code and routed to the appropriate team member with all context already assembled: the original invoice, the matched PO, and a clear description of the discrepancy.

The Results After 90 Days

Processing time for a standard compliant invoice: from 4 minutes to under 10 seconds. Weekly human hours spent on invoice processing: from 40 hours to 2 hours. Error rate: from 2.5 percent to 0.3 percent. Average payment confirmation time: from 14 days to 2 days.

The two team members who previously handled invoice processing are now focused on vendor relationship management and financial analysis, work that directly creates value rather than processing volume.

What Made It Work

Three things made this project succeed where similar attempts had failed before. First, we spent three weeks mapping the current process in detail before building anything. Second, we built the exception routing before the automation went live, so there was never a situation where the team did not know what to do with a flagged invoice. Third, the team was involved in the testing phase and had direct input on the exception reason codes and routing logic. Adoption was immediate because the people using the system had helped design it.