Fraud systems taught me one rule most agent builders are about to learn the hard way

What fraud teams know in their bones

In fraud, a false negative is a bad transaction you let through. It costs you that transaction. A false positive is a good customer you blocked. It costs you the customer, every transaction they would have made for five years, the complaint, the support ticket and occasionally a letter from the regulator. Serious fraud teams run a false positive budget: a negotiated number of good customers you may inconvenience per thousand, watched daily, with a named person accountable.

• An agent has the same two ways of being wrong: miss a real problem, or block something that was fine.

• 94% accurate means nothing until you know which 6% is which.

• Five bad rejections for every real catch is a machine that creates work, not removes it.

Two numbers, never one

Every agent we put live reports its miss rate and its over-flag rate separately, against a labelled set of real cases, every week. The business owner signs off on both thresholds before launch. When we built the grading agent for an agri marketplace the rule agreed up front was that it may under-grade cautiously but must never over-grade. The two mistakes have different prices.

Before you build, write down what a wrong yes costs and what a wrong no costs, in actual money. They are always different.

Before you build, write down what a wrong yes costs and what a wrong no costs, in actual money. They are always different.

Before you build, write down what a wrong yes costs and what a wrong no costs, in actual money. They are always different.

Confidence is a routing decision

A fraud score does not just describe a transaction, it decides what happens next. Below one threshold approve, above another block, in between a person looks. Agents need the same three-way gate. Most have only two: do it, or do not.

Beside the person before instead of the person

The highest return system we built at PayPal was not an agent that made decisions. It gave investigators the whole picture in one place and let them act without switching screens. Handling time roughly halved. Volume dropped 45% because the easy cases were automated around the humans, not instead of them.

If you are running an agent and tracking one accuracy number, we would genuinely be glad to look at how it is set up.