An accounts-payable agent that processes 4,000 invoices a month with 99.6% match accuracy

A Series-C fintech was growing invoice volume 22% quarter over quarter. Their three-person AP team was at capacity. Each invoice required: (1) extract line items from PDF, (2) match against PO and goods-receipt, (3) code to GL, (4) route for approval, (5) post to NetSuite. Median time per invoice: 9 minutes. They were a quarter away from needing two additional FTE.

Existing OCR-only solutions extracted text well but failed on the actual matching logic — vendors with five different name formats, bundled line items, partial deliveries, retroactive POs.

Sprint (2 weeks) — extract + match. Textract for table extraction from invoice PDFs. Claude Sonnet 4 (vision) as a second pass for hard cases (handwritten notes, stamps, non-standard layouts). Built a vendor-name normalization service that maps the messy real-world strings to the canonical NetSuite vendor record. Three-way-match logic in code with a Claude-driven exception handler for partial matches.

Extension (6 weeks) — workflow + approval. Wrapped the whole pipeline in Temporal so every step is durable, retryable, and auditable. Built the approval router: invoices under $5k auto-post if they pass three-way match. Above $5k they queue for the appropriate approver based on department and GL code. Added a Slack inbox where approvers can approve/reject with one click, and the agent posts the decision back to NetSuite.

Every decision the agent makes is logged with the source documents, the match it found, and a human-readable rationale. Auditors love it.

4,000 invoices a month flow through the agent with 99.6% three-way-match accuracy. Month-end close compressed from 8 days to 3 because the AP queue is no longer the bottleneck. The two-person hire that was on the budget never happened.

The Controller's favorite metric: zero late-payment fees in the two quarters since launch, against $11k in fees in the four quarters before.