Banks and fintechs now run AI fraud models on almost every payment, login and new account. Vendors promise that these systems learn fast and miss little. The people who run them tell a more mixed story.

Some results are real. The US Treasury says its machine learning tools prevented and recovered over $4 billion in fiscal 2024, including about $1 billion in check fraud. But fraud has changed too. Criminals now use the same AI to fake the documents and behaviour that AI fraud models were built to check.

So we asked industry leaders one question: how are banks and fintechs using AI in fraud detection in practice, and where has it failed? We wanted live results and failure cases, not vendor claims. A fintech founder and a software operations chief with AML experience shared what they see.

What the models do well

AI fraud models are good at two jobs. First, they spot patterns that analysts have already seen and labelled. Second, they flag behaviour that looks odd for a given customer, such as a new device or an unusual payee.

That combination works at scale. A model can score millions of card payments in milliseconds, which no human team could match. As a result, most large banks now treat AI scoring as a basic layer, not an add-on.

Those wins come from volume. The Treasury result, for instance, came from checking a huge number of payments for patterns that people had already seen before. That is exactly the kind of work machines do better than people.

The catch is in how those models learn.

“The weak spot in AI-based AML and fraud systems is how they’re built. They’re good at two things: flagging the kinds of fraud they were trained on, which means patterns an analyst has already found, and spotting odd behaviour. The trouble is that fraudsters change their tactics and tools much faster than a model can be retrained.

To teach a model a new type of fraud, a team first has to catch those transactions, break them down into patterns, build a dataset or generate synthetic data, split it into training and test sets, train the model, test it, and only then release it. On live systems that takes a long time. By the time it ships, the scheme it was built for has often moved on.

So in practice, the AI model is your layer for known patterns. Your real defence against new schemes is still the analyst who spots the first cases, and how quickly your team can get them into a retrained model.”

Timur Iusubaliev, COO, Plus8Soft

Fintechbits analysis

Why AI fraud models fall behind

Timur’s point is about time. A supervised model can only learn a scheme once someone has found it, labelled it and built a dataset around it. Meanwhile, the fraud ring has moved on to the next trick.

Not every system works this way, to be fair. Unsupervised anomaly detection and graph analysis can flag new patterns before anyone labels them. Still, those tools mostly raise a question for a person to answer. In other words, the analyst stays at the centre of the defence.

Vendors have started to target that bottleneck. For example, FIS now sells an AML investigations agent to any US bank, aimed at the casework that follows an alert. The pitch is speed for the analyst, not AI fraud models that replace the analyst.

When the document stops being proof

The second failure is newer. For years, a bank could ask a customer for proof: a payslip, a bank statement or a utility bill. Now anyone can produce a convincing fake in seconds with free AI tools.

That hits fintechs hardest at onboarding. Lenders, earned wage apps and cross-border accounts often ask new users to upload a statement or an ID and then let a model read it. If the upload is fake, the model is reading fiction with great accuracy.

“Everyone asks how AI catches fraud. The bigger story is how AI broke the way fraud gets reviewed. For years, the safest check was asking a customer for proof. Now anyone can make a convincing document in seconds, so paperwork on its own doesn’t prove much. A fintech that still treats a PDF as evidence is grading the fraudster’s best work. The teams getting it right have stopped collecting proof and started looking for signals that are hard to fake.”

Greg Reveret, CEO, Klydo Inc.

The numbers back him up. Entrust’s identity fraud research recorded a deepfake attempt every five minutes and a 244% rise in digital document forgeries in a single year. So AI fraud models that score a scanned payslip are judging work a fraudster may have made with AI too.

Document checks have not disappeared, though. Instead, buyers are paying for better ones. Stakk’s $63 million ParaScript deal shows document fraud has moved up the agenda.

Signals that are hard to fake

So what replaces the PDF? The teams getting it right feed their AI fraud models signals a fraudster cannot easily copy. These include device data, typing and swiping patterns, the history of the receiving account and live data pulled straight from a bank.

Shared data helps as well. Incode’s TrustGraph lets firms swap fraud signals without handing over customer records. In addition, Starling Bank built an AI scam detector to warn customers before they pay a scammer. That matters because scam victims pass every identity check, since the real customer is the one sending the money.

Where controls still fail

Even good AI fraud models cannot fix weak controls around them. In July 2025, the UK’s Financial Conduct Authority fined Monzo £21 million over failings in its financial crime controls. The lesson is simple. A model only works inside a process that acts on its alerts.

False positives are the other cost. AI fraud models tuned too tightly block good customers, freeze accounts and flood analysts with noise. Tuned too loosely, they let losses through. Every team has to choose where to sit on that line, and the right answer shifts as fraud shifts.

This is also where failure cases rarely become public. A bank will announce a model that cut losses. It is far less likely to say how many good customers that model blocked along the way.

What to watch next

Taken together, the message from people who build and run AI fraud models is clear. They catch known fraud well and at scale. However, they lag new schemes, and they can no longer take a document at face value. Over the next year, watch how fast teams retrain their AI fraud models. Watch how far they move from paperwork to hard-to-fake signals. And watch whether analyst tools such as AI agents cut the time between the first case and the fix.