A credit card processor monitors millions of transactions per day and wants a model that learns the normal spending pattern for each cardholder and flags any transaction that deviates sharply from that pattern for manual review, without being given a labeled dataset of past transactions marked fraud or not fraud. Which AI capability fits this requirement?

The correct capability is anomaly detection, which learns what a cardholder's normal transaction pattern looks like and then measures how far each new transaction departs from that baseline, flagging outliers for review. This matches the processor's requirement to work without a labeled fraud dataset. The classification approach that assigns each transaction to a fraud or not-fraud label using labeled training examples is incorrect because classification requires a training set of transactions already labeled fraud or not fraud, which the processor explicitly does not have; without those labels a classifier cannot be trained to draw that boundary. The clustering approach that groups transactions into segments based on similarity without flagging individual outliers is incorrect because clustering groups similar transactions into segments, which is useful for discovering natural groupings like spending categories, but it does not, on its own, flag any particular transaction as unusual relative to an individual's history. The regression approach that predicts a continuous numeric value such as an expected transaction amount is incorrect because regression predicts a continuous number, such as an expected purchase amount, but producing a predicted amount is not the same as deciding whether an actual transaction is suspicious. Anomaly detection is the capability purpose-built for spotting deviations from an established baseline without prior fraud labels.