A The invoice that arrives as a photo in WhatsApp
- How it runs today
- A supplier photographs an invoice and sends it to whoever they have a number for. That person forwards it to accounts. Someone reads the image and keys the vendor, date, line items and total into the accounting system. Someone else checks it against a delivery order in a different folder. If a digit was misread it surfaces weeks later during reconciliation, and by then nobody remembers the case.
- What it actually costs
- Not the typing. The cost is that the invoice has no status: nobody can answer "has this been paid" without asking a human, so the question gets asked constantly and each answer costs an interruption. Errors are expensive not because they happen but because they are found late.
- How I would rebuild it
- One intake point that accepts the photo without asking the supplier to change anything — that constraint is real and fighting it fails. Extract the fields, and where confidence is low, show a human the image and the extracted values side by side rather than guessing. Match against the delivery order automatically, and route only genuine mismatches to a person. Every invoice then has a state, so the status question answers itself.
- What I would need from you
- Thirty real invoices, including the ugly ones, and one person who can tell me which mismatches genuinely matter. Access is usually the slow part, not the build.
- What could go wrong
- Extraction accuracy on bad photographs is the honest limit. The answer is not a better model, it is designing for the failure: a human review queue that is fast to clear, and never silent auto-approval of something the system was unsure about. If your volume is under about twenty invoices a week, careful manual entry may simply be cheaper — I will tell you that rather than sell you a build.