Napier AI’s Michael Joseph attended a recent ACAMS New Jersey chapter discussion on the shift from threshold-based transaction monitoring (TM) to AI-driven systems, alongside PwC’s Samrat Jain, Remitly’s Murat Dagli and moderator Chris Phillips.
The panel debated a persistent question in financial crime circles: how much of AI’s promise in TM is being realised in practice, versus what remains aspirational.
Speaking on the topic, Joseph noted that most institutions using AI in TM today are applying it for efficiency, helping investigators clear alert queues faster, pull customer context, or draft first-pass Suspicious Activity Report (SAR) narratives. That is useful work, but in each case detection has already occurred; the rules fired the alert and AI simply processes it.
A smaller cohort is applying AI to detection itself, running models alongside rules to learn what is normal for each customer and surface activity rules were never built to catch, a shift accelerated since FinCEN’s April Notice of Proposed Rulemaking (NPRM) targeting reform of AML/CFT programmes under the Bank Secrecy Act.
Joseph noted that FinCEN has named AI directly as a factor in enforcement decisions, shifting the industry’s question from how much time AI saves to how much crime it actually helps identify, a move he called overdue, echoing a broader regulatory pivot from box-ticking compliance to outcomes.
On SAR quality, Joseph argued precision beats volume; “super SARs” stuffed with every transaction bury the real story, whereas AI-generated audit trails can support narratives that are sharper and more useful to law enforcement.
On build versus buy, Joseph said the calculation should account for the whole lifecycle, not just build cost. Vendors see typologies across many institutions and jurisdictions, a dynamic often called the consortium effect and difficult to replicate in-house, while ongoing costs like model validation, drift monitoring and retraining outlast any go-live date.
Joseph warned that measuring success purely by processing speed misses the point: false positive rates above 90% likely mean real crime is slipping through undetected. He also cautioned that alert volumes typically rise, not fall, in year one of AI deployment, as models surface new categories of activity, meaning cost savings should not be promised prematurely.
For more, read the full story here.
Copyright © 2026 FinTech Global
Copyright © 2018 RegTech Analyst
