Invoices that read themselves — inbox to structured books

What it is

The unglamorous one, and the one that saves the most hours. The part worth noticing is the last clause: when the model is unsure it raises its hand instead of writing a confident wrong number into the accounts. Done by hand, invoice processing is a quiet tax on every week. Supplier invoices arrive as PDF attachments, each supplier with its own layout, and somebody opens each one, finds the totals, and retypes them into the books. It is data entry at its worst: dull enough to rush, important enough that rushing costs you. One transposed digit and the books disagree with the bank statement, and month end turns into detective work. What actually runs here is a short pipeline. An IMAP watcher sits on the mailbox and notices when a supplier invoice lands. The PDF goes to Gemini, which reads the document the way a person would: who sent this, what is it for, what are the amounts. The extraction comes back as structured data and is written into Postgres, where the rest of the operation can query it alongside everything else. No folder rituals, no retyping. The failure path matters more than the happy path. AI invoice extraction is good and it is not perfect, so every read carries a confidence check. A clean extraction lands in the books automatically. A doubtful one is flagged and waits for a human decision instead of being quietly guessed. That one rule is the difference between automated invoice processing you can trust with your accounts and automation you end up auditing every Friday. This is the operator's own system, not a demo: it feeds the bakery platform elsewhere on this page, so the invoices going through it are real ones. The shape suits any small operation that receives supplier invoices by email and has nobody spare to do the typing: the pipeline does not care whether the PDFs come from a flour supplier or a software vendor.