From ranking to action
A risk score may rank transactions without being a calibrated probability. A threshold translates that output into an action such as review or additional verification. The right action depends on consequences and operational constraints, not only on which cutoff produces the most attractive validation chart.
Write the cost assumptions
Separate expected fraud loss, legitimate-customer friction, review expense and recoverability. These values may vary across transaction types. Keep the assumptions visible rather than converting every concern into one unexplained number. A policy that looks efficient under invented costs is still based on invented costs.
A review example
In a fictional queue, each additional review costs €4 and is expected to avert €3 of loss on average before considering customer impact. Expanding that queue is not justified by those two figures alone. Better evidence, different actions or lower-cost review could change the decision.
Test beyond the cutoff
Compare policies on held-out, appropriately labelled data and monitor actual outcomes after a controlled release. Include queue capacity and delays, because an overloaded review team changes the customer experience. These teaching concepts are not a claim that historical banking guidance imposes identical requirements on every fintech product.