Spotting a fraudulent payment · Ceded (since 1992)

In 1992, HNC Software released Falcon, a neural network that scored payment card transactions for fraud as they were authorised, and the major card networks adopted it during the decade that followed.

Fraud used to be found afterwards, by an analyst reading a statement and noticing that a pattern of spending had changed. Scoring moved the decision into the authorisation itself, in the fraction of a second between a card being presented and the answer coming back, at a volume no review team could approach. The judgment that stayed human is what the bank does with the score: where to set the threshold between blocking a genuine customer at a petrol station and letting a stolen card through, how to handle the customer who was wrongly refused, and who is liable for the loss. Investigators still build the cases that lead to prosecution, and they still find the fraud types the model has never seen.