How we automated 10,000 invoices a month

When 350–500 invoices a day land on a distributor’s finance desk, manual processing becomes the bottleneck. An accountant spends 4–6 minutes per document.

What we did

Built a pipeline on GPT-4 + RAG over the corporate contract base and 1C reference data. The LLM classifies the document, reconciles it with open contracts and publishes a draft posting into the ERP.

Numbers after 4 months

• 92% of documents pass without human input
• Average time from arrival to posting — 4 minutes vs. 47
• Cost-line classification accuracy — 96.4%

Stack

Vector DB: pgvector. LLM: GPT-4 + local Llama 3 for sensitive docs. Queue: RabbitMQ. 1C integration via REST. Deployed on-prem.