05 · Simulated output

Challenge the AI output

BankFraud AI produced the outputs below. They are fictional and deliberately imperfect. An output is not evidence that the result is valid.

‘95% accurate’ is not a customer-harm conclusion.

When fraud is rare, a model can be highly 'accurate' while most of the payments it flags are genuine. Ask: accurate at what, for whom, and with what consequence when it is wrong?

Model scores and actions
Case	Fraud-risk score	Model action	Simulated AI rationale	What later emerged
Customer ATraveller paying for a hotel abroad	91	Blocked	New country, first-time merchant category and unusual time increase the risk score above the block threshold.	Genuine. She waited six hours for review and slept in the lobby.
Customer BCarer buying medication, hard of hearing	88	Blocked	New card-not-present merchant and no successful verification.	Genuine. He could not complete the phone check and the medication was delayed two days.
Customer CSmall business owner paying a supplier	72	Verify	Amount above customer baseline at an overseas merchant.	Genuine. The verification arrived 20 minutes after the payment was declined.
Customer DAccount takeover victim	34	Approved	Low amounts at familiar merchant categories; within normal spending range.	Fraud. £760 lost before the customer noticed.
Customer ERecently moved home	87	Blocked (3 times)	Burst of new merchants in a new location.	Genuine. Repeated blocks led her to move her salary to another bank.

Your assessment

Do you trust these outputs? Why or why not?

I would not rely on these outputs without… / I would investigate first…

Patterns across the cases

What do the errors have in common? What could be causing them?

Answers save automatically and sync to your account when you are signed in.