Calibrating AML models: when rules-based systems fail and machine learning takes over.

For more than a decade, transaction monitoring in regulated institutions has been dominated by static rule thresholds: flag any wire over €10,000, any rapid sequence of deposits, any deviation from a customer’s historical pattern. The problem is that these rules are both noisy and brittle. They generate thousands of false positives while missing sophisticated layering schemes that stay deliberately below each threshold.

1. The failure mode of static thresholds

Rules-based systems assume that suspicious behaviour looks the same across all customers. In practice, a £50,000 monthly transfer is unremarkable for a corporate treasurer and extraordinary for a retail saver. Static thresholds cannot encode this context without exploding into an unmaintainable matrix of exceptions.

2. Machine learning as a behavioural lens

Adaptive models learn the normal behavioural fingerprint of each customer or entity class, then flag deviations that are genuinely anomalous relative to that baseline. The result is typically a 40–60% reduction in false positives and a measurable improvement in true-positive detection rates.

3. Regulatory acceptance and model risk

Regulators now expect explainability. A black-box neural network will not pass scrutiny. The institutions we work with use gradient-boosted decision trees with SHAP-value explainability, allowing compliance officers to articulate why a given alert was raised in plain language.

Practical next steps

Begin with a parallel run: keep the existing rules engine live while training the adaptive model on historical data. Compare detection rates, false-positive ratios, and operational workload for at least two quarters before proposing a switch.