Case study
Ledgerbot — Invoice Automation Agent
An agent that reads inbound invoices, matches them to POs, flags anomalies, and posts clean entries — humans only see the exceptions.
01The problem
Meridian processed ~3,000 supplier invoices a month by hand: reading PDFs, matching against purchase orders, chasing mismatched line items. Three full-time staff, two-day average turnaround, and a steady drip of duplicate payments slipping through during month-end crunch.
02What we built
We shipped a document-understanding agent that extracts line items from any invoice layout, three-way-matches them against POs and receiving records, and posts clean entries directly to the ledger. Anything ambiguous — price drift over tolerance, unknown vendors, possible duplicates — lands in a review queue with the agent's reasoning attached. Every action is logged and reversible.
03The outcome
87% of invoices now post with zero human touch. Turnaround went from two days to twenty minutes. The review queue catches duplicates before payment instead of after — the system paid for itself in recovered duplicates within four months.
