What is a fraud engine?

What it can see in eight hundred milliseconds

Inside the two seconds of an authorization, a model considers the amount, the merchant, the time of day, the location, the device, the velocity of recent transactions, and how all of it compares both to your own established pattern and to fraud seen across millions of other cards.

Older systems ran on written rules: block anything above this amount at this merchant type after midnight. Modern systems layer machine-learned models on top, producing a probability rather than a verdict.

The trade nobody can escape

Every fraud engine faces exactly one trade-off: catch more fraud and block more honest people, or wave through more honest people and absorb more fraud. There is no setting labelled "only catch criminals."

So the decline at your own regular shop is not a malfunction. It is the false-positive rate, and somebody chose it deliberately.

Notice too whose costs are whose. Missed fraud costs the bank money directly. A false decline costs you dignity and the merchant a sale, and costs the bank almost nothing unless you complain or leave. That asymmetry tilts the dial, gently and permanently, toward inconveniencing you.

Why it improves as it learns you

When you respond to a "was this you?" message confirming a transaction, you are not just unblocking a purchase. You are providing labelled training data. The model learns your shape a little better.

Which explains a pattern you have probably lived: declines cluster in the early life of a new card and fade over months. You did not become more trustworthy. You became more legible.

It is not judging you. It is judging everyone shaped like you, right now, with the information available in eight hundred milliseconds.

Where you meet it

Every decline you did not expect. Keep notifications on and answer challenges quickly; a confirmed "yes, it was me" is the strongest signal you can send. Before travelling, make one small local transaction early to introduce the new pattern gently.