Invoice processing that used to take 3 days now takes 4 hours

The Challenge

The company processed around 2,000 supplier invoices per month in 6 different currencies. Their finance team spent 3 full days each month on manual data entry and purchase order matching, leading to delays, errors, and bottlenecks at month-end close.

The Solution

A mid-size freight forwarding company in Central Europe was drowning in supplier invoices — about 2,000 per month across 6 currencies. Their team was spending 3 full days just on data entry and matching. We built a document processing pipeline with a fine-tuned Llama model that extracts line items, matches them against purchase orders, and flags discrepancies. The system handles about 85% of invoices without human intervention. The remaining 15% get routed to a reviewer with the problematic fields pre-highlighted.

Technical Approach

We built a document processing pipeline using a fine-tuned Llama model for line-item extraction and PO matching. The system uses Azure AI for OCR pre-processing and Blue Prism for ERP integration. We trained the model on 6 months of historical invoice data to handle multi-currency formats and vendor-specific layouts.

Technology Stack