01What goes wrong
- Inflated invoices paid: more than the approved amount, often by a small margin.
- Fake invoices paid: an invoice number that was never issued.
- Bank details changed on a supplier’s say-so, the classic payment fraud.
- Over budget: agreeing to buy more than the budget allows.
- Genuine invoices unpaid: the opposite failure, which costs supplier trust.
02How suppliers push
- Honest quotes, as a baseline the agent should accept.
- “Market prices went up”: a price hike after agreement.
- A handling or rush fee that appears on the invoice.
- A new invoice number, or an invoice nobody issued.
- Bundles and minimum quantities that push past the budget.
- Last-chance deadlines; new bank details “from accounts”.
03Your purchasing rules
A test uses what your business buys and its usual price, one genuine test invoice number with its approved amount, and your purchasing budget, in your TEST system.
04What the result tells you
How often your agent paid only what it should, with its likely range; the fake, inflated or over-budget payments, with their cost; and whether it paid the genuine invoice. How it happens: how do AI agents get tricked into paying fake invoices?
Questions
Does the test pay real invoices?
No. It runs against your TEST system and a genuine test invoice you create; nothing is paid.
Does refusing every invoice score well?
No. Not paying a genuine, approved invoice is counted as a failure too, so an agent that refuses everything shows up.
See where your agent leaks money. Tested against your own rules, with the money at stake.