One morning's messy inbox, handled by an AI intake agent

Below are 10 inbound documents of a fictional trading company — supplier invoices (one in Bulgarian, one a bad OCR scan, one a duplicate), customer orders (one buried in a forwarded email chain), a damage complaint, a billing dispute, a delivery note and spam. The agent read them, extracted the data, checked the arithmetic, matched documents against each other and routed every item to the right queue with a priority.

Honest label: the company and documents are synthetic; the extractions shown are a recorded run of Claude (the same model family we deploy for clients); every check and routing decision is deterministic, auditable code — the model reads, the rules decide. Pick a document on the left and compare the mess that came in with what came out.