Dutch Banks Set Out Responsible AI Framework for AML Transaction Monitoring
The Dutch Banking Association (NVB) has published a 39-page industry white paper setting out how banks can use machine learning in anti-money laundering and counter-terrorist financing transaction monitoring while maintaining human oversight, model governance and explainability.
Published on 27 August 2026, the paper draws on Dutch banks’ experience developing, implementing and operating machine-learning models for transaction monitoring. It argues that traditional rule-based systems remain transparent and easy to link to known risk typologies, but often generate large numbers of false positives and can struggle to identify more complex suspicious behaviour.
From fixed rules to risk-based models
The NVB says machine-learning models can analyse multiple behavioural dimensions simultaneously and identify transaction patterns or customer characteristics associated with money laundering and terrorist-financing risks. Models may use supervised approaches trained on outcomes such as internal escalations, suspicious activity reports or customer offboarding, as well as unsupervised approaches including clustering and anomaly detection.
The paper recommends that banks assess model effectiveness not simply by alert volumes, but through measures such as recall, precision, sampling and performance against specific ML/TF risks. It also discusses residual risk, transition from existing monitoring approaches and the need for feedback loops and ongoing model maintenance.
The NVB links machine-learning transaction monitoring to a broader move toward trigger-based ongoing due diligence. Its framework supports event-driven reviews and risk-differentiated customer reviews rather than relying by default on fixed periodic review cycles.
Human intervention remains central
The white paper stresses that decisions to report suspicious activity to the Financial Intelligence Unit must continue to involve meaningful human intervention. Dutch banks currently file unusual transaction reports; the paper notes that the terminology will change to Suspicious Activity Reports from July 2027 under the EU Anti-Money Laundering Regulation.
Responsible deployment also requires comprehensive model governance. The NVB highlights risks including implementation errors, unintended use of model outputs, deteriorating performance as customer behaviour changes, poor data quality and historical bias in training data. It recommends continuous performance monitoring, defined retraining cycles, documentation of assumptions and methodology, explainability at both model and individual-decision level, and controls to identify and mitigate unfair outcomes.
The association also emphasises fairness and privacy because AML decisions can materially affect customers. Models should be explainable, auditable and demonstrably effective, with banks remaining accountable for their outputs rather than treating automation as a substitute for professional judgement.
The paper adds to a wider shift in AML supervision toward effectiveness and risk-based controls. For financial institutions, the practical question is increasingly not whether machine learning can be used in transaction monitoring, but how its performance, governance and human oversight can be evidenced to regulators.
