Banks are using AI to cut alert volumes and speed up investigations, and regulators now encourage it. But a 2026 US enforcement case shows what happens when automation closes alerts nobody checked.

Anti-money laundering teams have a volume problem. Rules-based monitoring throws up thousands of alerts, and most turn out to be false positives. Analysts spend their days clearing noise instead of finding crime. AI in AML promises to fix that. The evidence so far shows real gains, but also a clear way to get it wrong.

Where AI is working

The strongest public case for AI in AML comes from HSBC. Working with Google Cloud, the bank built a system it calls Dynamic Risk Assessment. It screens around 980 million transactions a month. HSBC says it found “two to four times as much suspicious activity” while cutting alert numbers by 60%, according to Google Cloud.

That result matters because it reports both sides. Fewer alerts alone could mean the system simply looks less hard. Finding more crime with fewer alerts suggests better targeting. Other banks report smaller gains. BMO, Scotiabank’s Tangerine and FNBO have described fewer false alerts or faster investigations, though those figures are self-reported.

Vendor claims run higher, often promising 40% to 70% fewer false positives. Treat those as marketing until a firm publishes its own baseline and period. For more on where models help and where they fail, see our roundup on AI fraud models.

Where automation went wrong

The warning case for AI in AML is recent. In April 2026, the OCC signed a consent order against Community Federal Savings Bank, a sponsor bank for Wise and Crypto.com. Its automated triage system “auto-closed a very high percentage of all ingested alerts” that should have gone to a person. The OCC found flaws in its logic, data and method, and said thresholds were not tuned to the bank’s growing cross-border business.

UK cases point the same way. The FCA fined Starling £29 million in 2024 after its sanctions screening checked only part of the list. In 2025 it fined Monzo £21 million after its controls failed to keep up with growth. Neither was an AI failure. Instead, both show that regulators punish untested systems, not technology itself.

What regulators now say about AI in AML

The tone from regulators has shifted towards encouragement. In April 2026, FinCEN proposed a new AML programme rule. It names “effective use of artificial intelligence” as a factor in judging programmes, and says responsible testing will not on its own trigger enforcement, according to the FinCEN fact sheet. However, it encourages AI rather than requiring it, and the rule is not final.

In the UK, the FCA is testing AI with firms directly. Its second AI live testing group, announced in April, includes Barclays, Lloyds, GoCardless and Experian, with AML and KYC among the use cases. Then, on 19 October 2026, the FCA published good and poor practice on AI in financial services. It applies existing rules rather than new ones and expects a named senior manager to own AI risks, Freshfields noted.

In Europe, the AI Act’s high-risk rules were pushed back to December 2027. AML monitoring is not listed as a high-risk use, and fraud detection is expressly excluded from the credit category. Even so, the EU’s new AML rules require firms to keep approval of detection criteria in-house, even when monitoring is outsourced.

What good evidence looks like

Taken together, the cases suggest what regulators will ask for when firms use AI in AML. First, a baseline: how many alerts, false positives and suspicious activity reports there were before the change. Second, proof that detection held up, not just that volumes fell. Third, a record of how the model was tuned, tested and approved, and by whom.

Auto-closing low-risk alerts with AI in AML systems is not banned. However, firms need sampling of closed alerts, regular retuning as the business changes, and an audit trail a supervisor can follow. In addition, the model should be explainable enough that an analyst can say why an alert was closed. AI in AML works best when it supports analysts rather than replacing their judgement.

Where AI in AML fits next

Identity checks are moving the same way. Firms now use AI to read documents, screen adverse media and verify customers at scale. As agents start acting for customers, identity firms are adding tools to verify the agent too, as our report on Socure’s agent tools showed. Ownership checks are also tightening, as our roundup on beneficial ownership rules found.

Adoption is already broad. A Bank of England and FCA survey published in October found 86% of UK financial firms now use AI, up from 75% in 2024. The question is no longer whether to use AI in AML, but how to prove it works.

What to watch next

Several decisions will shape AI in AML over the coming year. FinCEN’s final programme rule would reward effective use of AI. The FCA’s second live testing group ends in December, with an evaluation due in early 2027. In Europe, AMLA’s standards will set how monitoring rules are approved.

Above all, the lesson from 2026 is balanced. AI can cut noise and find more crime, as HSBC shows. But a system that quietly closes alerts without testing can turn a cost saving into a consent order.

FintechBits covers payments, banking and fintech developments for readers in the US and UK. This article is for information only and is not financial or legal advice. Views expressed are those of the FintechBits editorial team.