For decades, catching financial crime meant hiring more people to review more alerts. Most of those alerts turned out to be nothing.
Speaking at Semafor’s Banking on the Future Forum in Washington last week, Revolut U.S. CEO Cetin Duransoy said the FinTech’s AI transaction-monitoring systems now perform “statistically significantly better than human reviews of the transactions.” Human investigators at the company now focus exclusively on higher-risk cases.
What the System Actually Does
Revolut’s compliance stack runs across 39 countries, with agentic AI handling both know-your-customer onboarding and ongoing transaction monitoring. The architecture separates work by risk level: AI handles the high-volume, lower-complexity screening layer while human investigators take cases that require judgment.
The operational logic is straightforward. Retail Banker International reported that traditional AML systems generate false positives on up to 95% of alerts. Every one of those false positives lands in a human queue, consumes investigator time and produces nothing. An AI system that cuts false positive volume frees investigators for work that static rule-based systems can’t handle. Global AML compliance costs have climbed above $274 billion annually, with much of that spending going toward handling low-quality alerts rather than catching actual criminals.
The economics of that model were already strained. Real-time payments made them worse. Banks processing euro transfers under SEPA Instant Payments rules must complete AML checks, sanctions screening, and fraud detection within a 10-second window, a requirement legacy compliance systems weren’t built to meet. When the Federal Reserve lifted FedNow® Service’s transaction limit from $1 million to $10 million last year, high-value instant wire transfers that once gave compliance teams until the end of the day for review began requiring real-time decisions.
How the Industry Is Rebuilding
Other institutions have run the same experiment. Nasdaq Verafin announced in July that beta testers of its agentic AI system reported more than an 80% reduction in sanctions-screening alert review workloads, with human investigators redirected to higher-risk cases. HSBC said its deployment of Google Cloud’s AML AI cut false positive cases by 60% while detecting two to four times more confirmed suspicious activity than its prior model.
Advertisement: Scroll to Continue
The gains reflect a structural shift rather than incremental improvement. Traditional compliance relied on static rules built around known patterns. Those rules produced the false positive problem. The systems replacing them build behavioral baselines over time, flagging deviations rather than threshold breaches, which means they get more accurate as transaction volume grows.
Where the Threat Is Moving
The compliance shift runs in both directions. Fraudsters now generate convincing synthetic identities complete with realistic documents and AI-generated images, building credit profiles over months before executing fraud. Those identities pass standard KYC verification because they’ve never failed a check before. Experian found that nearly 60% of companies reported an increase in fraud losses between 2024 and 2025, with agentic AI enabling bots to run complex scams without a human in the loop.
PYMNTS Intelligence found that 68% of financial institutions increased fraud-detection spending year over year. In April, FinCEN and the OCC jointly issued a proposed rulemaking to replace the existing process-driven AML compliance model with one based on risk-weighted effectiveness, a regulatory framework that rewards outcomes over paperwork volume.
For all PYMNTS AI and digital transformation coverage, subscribe to the daily AI and Digital Transformation Newsletters.
