How banks use AI to flag unusual payments

Automated payment checks are one of the most common ways people meet AI without noticing. This article explains the general ideas behind them. It does not describe any particular bank, card or app, and it is not advice about financial products.

The problem banks are trying to solve

Banks and card networks process an enormous number of payments every day. A small share of them are fraudulent: a stolen card number, a taken-over account, or a scam that pressures someone into sending money. Checking each payment by hand is impossible, and checking after the fact is often too late.

So large payment systems run automated checks while a payment is being approved, often within a fraction of a second. Increasingly, those checks include machine learning models alongside traditional rules written by people.

What "unusual" means to a model

A model cannot know your intentions. What it can do is compare a payment with patterns. The information it considers may include:

• the amount, compared with your usual spending;
• the type of merchant and where it is located;
• the time of day, and how quickly payments follow each other;
• whether the card was present or the payment was made online;
• whether the device is one you have used before.

The model learned from large amounts of past payments, a few later confirmed as fraud and most confirmed as genuine. From those examples it picked up combinations that tended to come before fraud. A single detail rarely matters on its own. A large online purchase from a new device at 3 a.m., minutes after a tiny test payment, is a combination worth a second look.

Note that unusual is not the same as wrong. Your first purchase on holiday is unusual. So is a new hobby. That is why the next step matters.

Scores, rules and people

The model usually produces a risk score rather than a simple yes or no. What happens next depends on rules the bank sets.

• Low score: the payment goes through normally.
• Medium score: the bank may ask you to confirm the payment through its app, a text message or a call.
• High score: the payment may be held or declined, and the case passed to a person to review.

Human reviewers and customer conversations remain part of the process, especially for unclear cases. The model narrows down where people should look; it does not replace them.

The trade-off: false alarms and missed cases

Every alert system faces two kinds of mistakes. A false alarm, or false positive, is when a genuine payment is flagged. A missed case, or false negative, is when a fraudulent payment slips through.

The two are linked. Set the threshold low and the system flags more payments: it catches more fraud but bothers more honest customers. Set it high and fewer people are bothered, but more fraud gets through. There is no setting with zero mistakes. Banks choose a balance and adjust it as patterns change, because the people behind fraud also change their methods.

Models need regular retraining for the same reason. Spending habits shift with seasons, new payment methods and life events. A model trained on old patterns slowly becomes less accurate, a problem known as drift.

Fairness and explanations

Because these models learn from past data, they can treat some customers differently in ways nobody intended. Someone who travels often, shops at unusual hours for work, or has only a short history with the bank may see more alerts. Responsible institutions test for patterns like this and give customers a quick way to sort out a blocked payment.

In Canada, federally regulated banks must follow consumer protection rules and privacy law when they handle your information. If you think a decision about your account was handled unfairly, the first step is your bank's own complaint process. The Financial Consumer Agency of Canada, a federal government agency, explains how that process works.

What to do when you get an alert

1. Pause before you respond. Real alerts and scam messages can look very similar.
2. Use channels you already trust. Call the number on the back of your card, use the official app you already have, or type the bank's address yourself. Do not use links or numbers from the message.
3. Never share a password or one-time code. A genuine bank will not ask for them to release a payment.
4. Report problems quickly. If a payment you did not make goes through, contact your bank as soon as possible. Card agreements usually set time limits.

Payment checks are pattern spotting at scale: a model scores, rules decide, people review. Knowing that chain makes a declined card a little less mysterious, and helps you respond calmly when an alert turns out to be real.

This article explains general concepts. It does not describe any specific bank's system, does not compare financial products and is not financial advice.