A $483,000 payment request for about $48,000 of gloves and dressings. Asked to draft the CFO's approval email, two leading AIs passed it along forty times in forty.

Finance teams are handing AI the paperwork around payments. If a payment request carries an extra zero and nothing attached contradicts it, will the AI notice before it drafts the approval?

Both models pass the amount along at least four times in five. With one written rule, both question it at least four times in five.

An invented home health agency's accounts payable request for medical supplies, with the amount ten times the supplier's real invoice. We started with the invoice attached too: with both documents in front of them, both models caught the extra zero every time, and caught a doubled amount as well. Then we removed the invoice, the way a request often reaches an approver.

• Neither model questioned the amount, in any of the emails.
• More than half of one model's emails added that unit prices had not been checked. None said that $483,000 looked high for 345 cases and boxes of gloves, cuffs and dressings.

One written rule, about plausibility: an amount is judged against what is usual for the item, not only against other numbers in the document. With it, every email from one model put the amount in question, though mostly as could not be verified rather than looks too high. The other said it looked unusually high three times in four, and still passed it along one time in four.

The prediction was right about the failure and too hopeful about the fix. The AI checks numbers against other numbers. Noticing that a number is simply implausible is something it has to be told to do, and even then one model missed it one time in four. An approval email is exactly where a wrong number becomes a payment.

Two models, one invented request. Each model was used directly, without the chat app around it.