AI fraud detection just had a busy few weeks. Three fraud-detection deals landed inside a few weeks of each other this quarter: Visa folded Featurespace’s models into Enhanced A2A Protect for bank-to-bank transfers, Flagright joined Nacha’s preferred partner list for ACH fraud monitoring, and Navan bolted DataVisor onto its travel-and-expense platform. None of the three are connected, but together they point to where fintech infrastructure spending is actually going in 2026: into AI systems built to catch financial crime before it clears, not after.

Each deal targets a different failure point. Visa’s Enhanced A2A Protect is aimed at real-time bank transfers, a category regulators have pushed banks to speed up even as faster settlement leaves less time to catch a bad payment before it becomes irreversible. Flagright’s Nacha status covers ACH, the batch-based rail that still moves the bulk of US payroll and bill payments and has historically leaned on delayed, rule-based reviews rather than real-time scoring. Navan’s addition targets expense fraud, a smaller-dollar but high-volume problem that most travel-and-expense platforms treated as a policy issue for employees to follow before treating it as a detection problem for software to catch.

The scale behind that spending is not subtle. Industry estimates put global financial crime losses above 579 billion dollars in 2025, and about two-thirds of banks and fintechs say fraud rates are still rising rather than leveling off. Roughly 90 percent of financial institutions now use AI somewhere in their fraud detection stack, and the reported results at the largest banks explain why the number is that high.

HSBC has said its AI systems catch two to four times more financial crime than its prior tools. DBS Bank has reported a 60 percent improvement in detection accuracy. JPMorgan credits its AI fraud system with preventing an estimated 1.5 billion dollars in losses a year at a claimed 98 percent accuracy rate. Those are self-reported figures from the banks running the systems, worth treating as directional rather than audited, but the direction is consistent across all three.

What AI Fraud Detection Actually Changed

The underlying shift is from static, rule-based monitoring, flagging a transaction because it crossed a fixed dollar threshold, to models that learn what normal looks like for a specific customer or a specific payment rail and flag deviations in real time. That is why bank-to-bank transfers, ACH batches and corporate expense platforms are among the first systems getting these upgrades: they are high-volume, low-friction rails where a human reviewing every transaction was never realistic, and where a bad payment is difficult to unwind once it clears.

The Other Side of AI Fraud Detection

The same underlying technology cuts both ways, and that is the honest caveat behind the adoption numbers. Deepfake voice cloning is now convincing enough to pass some voice-authentication checks. AI-generated synthetic identities can clear document verification that was built to catch forged paper, not generated pixels. Adversarial inputs crafted specifically to slip past a known model’s blind spots are turning fraud detection into an ongoing contest rather than a problem that gets solved once and stays solved. Buying an AI fraud model is not the same as staying ahead of the people trying to beat it.

AI detection is also reshaping the economics of running a fraud team, not only the accuracy. Reported figures put cost reductions from AI-assisted detection at around 60 percent, alongside accuracy rates in the mid-90s percent range, which is the kind of combination that makes a vendor contract easy to justify before a company has even totaled up the losses it prevents.

Fintechbits analysis

Why Fintechs Are Buying AI Fraud Detection Instead of Building It

The pattern across the Visa, Flagright and Navan deals is that none of them built their detection models in-house. A payments company is good at moving money, not necessarily at building specialized anomaly-detection models, and buying that capability from a specialist that sells it across multiple platforms means the model sees more fraud patterns, and gets smarter faster, than any single company’s in-house fraud team could manage alone. That division of labor is becoming the default, and it is a reasonable one: the vendors get more data to train on, and the platforms get a detection system that improves without a dedicated data science team on staff.

Worth watching next is whether the vendors powering this wave, Featurespace, DataVisor and their peers, can keep improving as fast as adversarial techniques evolve. The value of an AI fraud model depends entirely on how current its blind spots are, and that is a race with no finish line, only a widening gap between whoever updates fastest and everyone else.

Fintechbits covers fraud prevention, financial crime infrastructure and fintech risk technology. Nothing here constitutes financial or investment advice. All analysis represents the editorial views of Fintechbits.