AI in transaction monitoring: where it helps and where it does not

Every bank, exchange house and payment provider runs transaction monitoring to detect money laundering and terrorist financing. Most still rely mainly on rules: a transfer above a threshold, too many cash deposits in a week, a payment to a high-risk country. Rules are easy to explain, but they generate many alerts, and in most programmes the large majority turn out to be false positives.

Where machine learning helps

• Scoring alerts. A model trained on past investigation outcomes ranks alerts so analysts see the riskiest first and close obvious ones faster.
• Finding the network. Graph analysis links customers through shared devices, addresses, beneficiaries and counterparties, and exposes mule networks that look harmless one account at a time.
• Resolving identities. Matching the same person across spellings, scripts and documents, a real problem with Arabic and Latin names, improves both screening and monitoring.
• Spotting the unusual. Anomaly detection compares a customer with their own history and their peers, rather than with a single fixed threshold.

Where it does not

Machine learning does not replace the decision to file a suspicious transaction report. That remains a human judgement, made by a trained officer who can explain it. A model also cannot fix poor data: if customer risk ratings, occupations or transaction purposes are missing or wrong, the model will learn the gaps.

What regulators expect

Supervisors, and FATF guidance on new technologies, are broadly supportive of advanced analytics, but they expect:

• a documented model with clear purpose, inputs and limitations;
• validation before use and monitoring for drift afterwards;
• explanations an investigator can read for each score;
• the ability to show that the model did not lower detection.

Start beside the rules, not instead of them

The safest path is to run models alongside existing rules, compare outcomes over several months, and only then let scores change how alerts are handled. Our fintech team builds monitoring and screening platforms with this approach, and keeps every model decision traceable for audit.