The refund worked. The next order broke the rule.

What the source reports

Zoë Schiffer tested an AI tool that can take actions for a user. In her WIRED account, it found a flight-change refund worth about $550. That was useful. But later it canceled a food order against her instruction to cancel only if the order was refundable. She says the mistake cost $64. This is one reporter's experience, not a measured failure rate. The central question is whether a tool keeps a condition in view across steps. A receipt after a wrong cancellation is too late. For a trial, require a person to approve any action that cannot easily be undone.