For years, banks have struggled with one of compliance’s biggest inefficiencies: sky-high false-positive rates in transaction monitoring. Traditional rules-based systems generate volumes of alerts that swamp investigators, while genuinely suspicious activity risks slipping through the cracks.

Machine learning (ML) is beginning to change that.

Carsten Helm shared a concrete result:

“Integration of Machine Learning components have meanwhile proven effective. A significant false positive reduction is possible  without losing any SARs”

This is not just about efficiency. It’s about effectiveness: freeing investigators to focus on genuine risk while strengthening assurance that critical red flags are not missed.

Carsten also emphasised integration:

“The component could also deliver information to the KYC process… not just ‘low/medium/high risk profiles’ but a more truthful probability of the client’s behaviour.”

Oonagh van den Berg highlighted a common blind spot:

“Firms routinely collect ~1,400 onboarding data points but operationalise less than 10% day-to-day.”

Why It Matters

• False positives waste resources. Industry data suggests that over 90% of AML alerts are false positives (FATF/Banking Policy Institute). That means billions in compliance spend is effectively misallocated.
• Fintechs have an edge. As Helm notes, fintechs typically operate on cleaner, unified data stacks, enabling faster feature engineering and deployment. Large universal banks, by contrast, struggle with thousands of fragmented legacy systems, limiting model quality.
• Models improve more than monitoring. The same ML features can enrich KYC, sanctions screening, and fraud detection, creating a more holistic view of customer behaviour.

What Good Looks Like

• Catalogue, cleanse, and govern data. Build a robust feature inventory across all customer and transaction touchpoints.
• Apply feature engineering to drive insight. Move beyond static “low/medium/high risk” labels to probabilistic risk scoring grounded in live behaviour.
• Close the loop. Push high-signal features back into KYC, case management, and investigator workflows — ensuring models don’t sit in isolation but shape real-world decision-making.
• Monitor model performance continuously. Use feedback loops from investigator outcomes to refine and recalibrate ML models over time.

The Bottom Line

Machine learning is not a silver bullet — but it is the most credible path to materially improving both efficiency and effectiveness in financial crime detection. By combining cleaner data with behavioural risk scoring, firms can finally move the needle from “rules overload” to risk intelligence that works.