NEWS: Dutch Banking Association (NVB) urges lenders to embrace AI for transaction monitoring

Banks should increase their use of machine learning models to improve the efficiency and effectiveness of monitoring customers’ transactions, the Dutch Banking Association (NVB) said.

Financial institutions file hundreds of thousands of suspicious activity reports to authorities each year, but many turn out to be false alerts, as their systems are often based on transaction-amount thresholds and other strict parameters.

“Machine learning‑based [transaction monitoring] can improve the accuracy of risk detection, reduce false positives (enhancing both efficiency and client experience), and detect complex behaviour that rule-based systems struggle to capture,” the NVB wrote in a white paper.

Machine learning models are algorithms that make predictions based on patterns they have learned from past data.

The NVB called for these models to be transparent, interpretable and explainable to regulators, analysts and compliance officers. Users need insight into a model’s overall behaviour, as well as into the specific features that influenced individual predictions.

The use of machine learning could also apply to ongoing due diligence, the NVB said, by giving priority to reviews of customers’ data following important events rather than performing periodic reviews.

Unlike rule-based models, these systems cross-analyse data across multiple features to reveal patterns that can raise money laundering suspicions. They can be used either to generate alerts or to reduce noise, by scoring each alert and automatically discarding those that fall below a certain threshold.

“Such models can simultaneously analyse multiple behavioural dimensions and identify complex transaction patterns or client characteristics associated with ML/TF risks.”

The paper sets out recommendations for the effective and responsible use of these models, which still require meaningful human oversight despite the growing use of artificial intelligence.

The models will never fully replace the human component, the paper said. Banks and financial institutions remain accountable for the decisions these models make, as required by regulation, and the NVB stressed that human oversight remains key to controlling and monitoring AI systems.

“If, in the future, models would automatically submit SARs to the FIU, human oversight will become even more critical due to the potential impact on individuals. Strict quality checks would be essential,” the lobby group said.

The new EU anti-money laundering regulation coming into force in 2027 introduces even stricter rules, requiring meaningful human intervention to ensure a model’s accuracy and appropriateness.

The paper outlines two types of solution banks can adopt: a general model, which provides a framework to spot financial crime, with different features representing different risks; or focus models, smaller models tailored to detect individual financial crime risks.

The NVB recommended that banks ensure these algorithms remain fair and avoid biases that could target specific groups of customers. The paper lists techniques, such as adversarial debiasing and subgroup calibration, to prevent models from extracting patterns based on sensitive attributes such as gender and age.

If you would like unlimited access to AML Intelligence premium articles, newsletter delivered twice a week, access to our Global Bank Fines and Penalties database, free access to Boardroom Series events and much more, select one of our subscription options and become a subscriber!