AI agent flags $1,411 in bad invoices, but hard-coded rules did the real work

A developer tested an AI-assisted invoice review system by feeding it 13 invoices, three of which were deliberately flawed — a padded total, an unauthorized purchase, and a duplicate. The system correctly blocked all three problematic invoices, totalling $1,411.25, while approving the remaining ten worth $9,049.05. The workflow used a vision model to convert invoice images into structured JSON data, while a separate code node applied three strict rules: matching purchase orders, exact amount verification, and checking for duplicate invoice numbers. The entire process cost roughly one cent per image and completed in 28 seconds, but the developer stressed that the AI only extracted data — all financial decisions were made by deterministic code. The experiment also exposed a key limitation: the vision model misread an ambiguous date with full confidence, highlighting that AI-extracted values should never directly drive consequential financial logic.