Your fraud model has 99% accuracy. Why might that be worthless?

Opened the ML screen. They were after class imbalance, then precision/recall trade-offs, then what threshold I'd actually ship and who decides the cost of a false positive.