AML's False Positive Problem: What HSBC, Barclays, and Banco do Brasil Actually Measured

Will Drewes

Founder, Fern Strategy · 3 min read

The alert queue is not a compliance program

Legacy AML transaction monitoring systems generate false positive rates of 90 to 98 percent. That means compliance teams spend the overwhelming majority of their time clearing noise, not investigating actual risk. Regulators know this. So do the banks that have started measuring something different.

Three of them have published numbers worth examining.

HSBC: 20x improvement in detection accuracy

HSBC deployed network intelligence AI to map connections across its entire customer portfolio, linking entities and transactions that rule-based systems treat as unrelated. The result was a 20x improvement in suspicious activity detection accuracy.

That number matters because it reframes the problem. Most AML conversations focus on reducing bad alerts. HSBC's outcome is about finding real ones that were previously invisible. Network-level analysis surfaces typologies, shell structures, and layering patterns that no static ruleset can anticipate. The accuracy gain is not incremental; it is structural.

Barclays: 40% fewer false positives, more genuine referrals

Barclays applied machine learning to transaction pattern analysis and achieved a 40% reduction in false positives while simultaneously increasing the number of genuine cases referred for investigation.

This is the combination that actually changes compliance operations. Reducing false positives alone could mean the model is simply less sensitive. Reducing them while surfacing more real cases means the signal-to-noise ratio improved in both directions. Investigators spend less time on dead ends and more time on cases that warrant scrutiny.

Banco do Brasil: 90% reduction in false positive alerts

Banco do Brasil's technology transformation initiative for sanctions operations produced a 90% reduction in false positive alerts, an outcome recognized with the Celent Model Risk Manager Award.

A 90% reduction at scale means a compliance team that was previously processing, say, 10,000 alerts per month is now processing roughly 1,000 with equivalent or better coverage of genuine risk. That is not a productivity improvement. It is a different operating model. It also changes the economics of hiring, training, and retaining AML analysts, which is its own regulatory and operational pressure point for most institutions.

What this means for compliance and operations leaders

These three outcomes share a common thread: the gains came from replacing static, rule-based alert logic with models that learn from actual transaction behavior and entity relationships. None of them are marginal improvements on the existing approach. They represent a different architecture.

For banking leaders evaluating this space, a few things follow directly from these results:

• The false positive rate is a governance metric, not just an efficiency metric. Regulators increasingly expect institutions to demonstrate that their AML programs are risk-based, not volume-based. A 95% false positive rate is hard to defend as risk-based.
• Detection accuracy and false positive reduction are not the same objective. HSBC optimized for finding more real activity. Barclays optimized for both simultaneously. The design question matters before the build question.
• Auditability is non-negotiable. AI-driven AML models that cannot explain why a specific alert was generated, or suppressed, create examination risk. The outcome numbers above are only sustainable if the underlying models can be interrogated, validated, and documented for regulators.

The build-vs-buy decision in AML is real, but it is secondary to a more fundamental question: does your current monitoring infrastructure have a measurable baseline, and do you know what you are optimizing for? Without that foundation, any new tooling, however capable, enters a governance vacuum.

That is where a structured assessment earns its cost before a single model goes into production.