Anyone who has run a transaction monitoring queue knows the problem. Most alerts turn out to be nothing, analysts spend their days clearing noise, and the genuine case can sit unreviewed behind it. It is tempting to treat this as a tuning exercise, adjusting thresholds until the queue is manageable.
Machine-learning anomaly detection does help. In delivery I have led, applying AI and ML models reduced transaction-monitoring false positives by 30 percent and improved review turnaround by 35 percent. Those numbers matter to a compliance team's capacity. They are not what an inspector asks about. An inspector asks why a specific alert was closed, why another pattern never produced an alert at all, and who approved the change that altered either outcome.
Three controls answer that. First, a documented rationale for every threshold and every model change, with a named approver and a date. Second, testing of what the system suppressed: sampling transactions below the alerting line to confirm that real risk is not being filtered out, which is the only way to show that fewer false positives did not simply mean more missed positives. Third, a recorded human decision for every escalation or closure above a defined risk level, so that accountability sits with a person.
Third-party platforms add a further layer. RegTech tools for blockchain analytics, sanctions screening and risk scoring are useful, and a licensed institution still owns the outcome. The compliance function should understand what each tool scores, which inputs it uses, and what evidence it retains, and should be able to explain that to its regulator in plain terms.
Singapore's regulator has set out principles for the responsible use of AI in finance covering fairness, ethics, accountability and transparency, and they apply to monitoring models as much as to credit models. An institution that can show how its model is governed is better placed than one that can only show how well it performs.
The reporting habit follows from this. Never report a falling false-positive rate on its own. Report it alongside the result of the suppression testing, so the two numbers are read together. A metric optimised in isolation measures how quiet the queue is, not how well the institution is managing risk.
