Fraud models are getting sharper. The FTC says losses jumped 25 percent anyway.

In 2024, people reported losing more than $12.5 billion to fraud, a 25 percent jump over 2023, across 2.6 million reports to the Federal Trade Commission (FTC data). That figure is the reported floor, not the ceiling, because most fraud never gets reported at all. For a bank, the number matters less as a headline than as a map of where money leaves the building, and which of those exits a model can actually close.

The number, and what it actually measures

The FTC's $12.5 billion counts consumer-reported losses, and investment scams led it at $5.7 billion, followed by imposter scams at $2.95 billion. Those are not card-skimming losses or classic account breaches. They are payments the account holder authorized, under a false story. That distinction is the whole game for a bank buying fraud tools. A model trained to spot a stolen card behaving strangely is watching the wrong exit when the customer themselves sends the wire. So the first honest question about any fraud tool is not how accurate it is, but which of your losses it was even built to see.

The same money hides across the rest of the bank: where a Chicago bank's cost actually sits

What AI fraud detection in banking is genuinely good at

AI fraud detection in banking works by scoring a transaction against the account's own history and against patterns learned across millions of others. When a payment posts, the model reads hundreds of variables at once: amount, merchant category, location, device fingerprint, time of day, and how this account has behaved for months. It returns a risk score in milliseconds, before the money moves. That real-time scoring is where machine-learning models beat static rules cleanly. A rule says block any wire over $10,000 to a new payee. A model learns that this particular business wires exactly that every second Tuesday and lets it through, while flagging the same amount from an account that never wires at all. The payoff is fewer false positives on good customers and faster catches on genuinely novel patterns, the ones no analyst wrote a rule for yet. On card and account-takeover fraud, this is a solved-enough problem, and the banks still losing there are usually the ones running on rules alone.

Where the models still lose: checks and the customer's own hands

Here is the exit the models keep missing. Check fraud related suspicious activity reports filed with FinCEN went from over 350,000 in 2021 to over 680,000 in 2022, nearly double in a single year (FinCEN alert). A paper check stolen from a mailbox, washed, and re-deposited does not look anomalous to a transaction model, because it is a normal-looking deposit of a real instrument. And checks remain the format most targeted. In the 2025 AFP Payments Fraud and Control Survey, 79 percent of organizations were hit by attempted or actual payments fraud in 2024, and 63 percent of that ran through checks, more than any other method (AFP survey). The uncomfortable part is that more than three quarters of those organizations have no plan to cut check usage. So the highest-volume fraud channel is the one AI touches least, and the one operators are least willing to close. This is the same back-office discipline we argue for elsewhere, pointed at fraud instead of reconciliation.

The next number: the same AI, pointed back at the bank

Detection is one side of the ledger. The other is that fraudsters now hold the same tools. Deloitte's Center for Financial Services projects that gen AI could push fraud losses in the United States to $40 billion by 2027, up from $12.3 billion in 2023, a compound annual growth rate of 32 percent (Deloitte analysis). The mechanism is deepfaked voices on a verification call, synthetic identities that pass onboarding, and phishing written in flawless English at scale. A static defense loses ground here by standing still. That is the real case for adaptive models, and it is not that they are in fashion. It is that the attack is now adaptive, and a rule set frozen in a policy document is not.

The order we would run it

When we map a bank's fraud week, the losses do not sit in one system. For a Chicago community bank or a fintech running deposit accounts, they enter through four rooms, and each wants a different fix.

The order we would run it is deliberate. First, put a real-time model on card and digital-payment flows if one is not already there, because that is the settled ground and the fastest return. Second, and this is where most banks underinvest, treat check and mail-theft fraud as an operations problem, not a model problem: positive pay, payee-name verification, and image analysis on deposits, because a scoring model alone will not catch a washed check. Third, harden identity at the two moments AI attacks it, onboarding and step-up verification, with liveness and device signals rather than a voice a deepfake can now clone. Fourth, tune for the analyst, not the demo, because a model that floods the queue with false positives trains the team to click approve, which is worse than no model at all. This is the same order of operations we use to decide what to automate in claims, and the way we map an operating week before touching a tool. What an operator walks away with is a ranked list tied to where their money actually leaves, not a single tool bolted over everything.

The trap inside a good detection rate

A 95 percent catch rate reads well on a slide and can still hide a losing operation. If the model's precision is low, every real catch arrives buried in false alarms, and a tired analyst clearing 400 alerts a shift will approve the one that mattered. The metric that pays is not detection rate in isolation. It is losses prevented per analyst hour, net of the good customers you did not wrongly freeze. We would rather ship a narrower model the front line trusts than a wide one that teaches them to override it.

The read on AI fraud detection in banking is straightforward. It is necessary, and it works on the fraud it was built for, which is card and account takeover. It is close to useless on the channel growing fastest, which is checks and authorized-payment scams, and the attackers now hold the same generative tools the defenders do. A bank that buys one model and calls fraud solved has closed one exit and left three open. The banks that come out ahead are the ones that map every exit first, then match the fix to the door.