Invoice fraud succeeds when a fake invoice looks real - and that is exactly where machines hold the advantage, because they check pattern, not appearance. The common schemes are four: a fake vendor invoicing for services never rendered, IBAN substitution where a fraudster impersonates your known supplier and announces a “new” bank account, a near-identical duplicate invoice with a small change, and an urgent transfer order in an executive's name. None of these are new; what is new is that automated bookkeeping includes every invoice's context (the supplier's history, accounts, amounts, rhythm) so a broken pattern surfaces before payment, not during an audit.
Which signals does the machine catch?
The five strongest. A changed bank account: familiar supplier, but the IBAN differs from every previous invoice – the most dangerous and the most detectable signal. Amounts or frequency outside history: a supplier who bills monthly around a thousand euros suddenly sends two invoices a week. Duplicates: same amount, date and reference as an already-processed invoice. A first-time supplier: not suspicious in itself, but always worth human eyes. And timing outside the pattern: the Friday-evening invoice before holidays is a classic for a reason. Every signal means the same thing: the invoice stops and comes to a person, with the reason why it was flagged.
Is AI enough? Well: here are the human rules
Three procedures no tool replaces. Second-channel verification of account changes: when a supplier “announces” a new IBAN, call them on the number you already know; not the one on the invoice. Separation of rights: whoever approves an invoice should not be the only one who releases the payment; four eyes is an old rule for good reasons. And limits: an amount above which payments always require a second approver. AI makes these procedures rarer (fewer exceptions) but not unnecessary. The technical precondition for seeing patterns is clean data: a reconciled bank gift coded invoices are the history every new invoice is checked against.
What to do tomorrow morning
Three steps, free. Agree that bank account changes are always verified through a second channel. Review who can release payments alone. And if your bookkeeping already runs automated, switch exception flagging on: the pattern exists; use it.