AI monitoring: using machine learning to flag suspicious VASP flows

Three ways to see it

1. 1

   AI in AML monitoring is a layer above rules, not a replacement for them. Rules are deterministic: if amount > X and counterparty in country Y and time of day in Z window, then alert. They are auditable, explainable, and required by regulators as a baseline. AI adds three things rules cannot do well: ranking alerts by likely seriousness, clustering related alerts that look unrelated under any single rule, and detecting drift in customer behaviour over weeks rather than instants. Used together, the rule layer says 'something might be off here' and the AI layer says 'and these are the ones to look at first'.

2. 2

   There are three model families that matter. Anomaly detection: a model learns each customer's normal monthly pattern (volume, counterparty mix, geographic spread) and scores each new month for distance from normal. The output is a number between 0 and 1. Customers with high anomaly scores are reviewed first. Network analysis: a model treats every transaction as an edge in a graph and looks for suspicious shapes such as star patterns, layering chains, and round-tripping. The output is a flagged subgraph the analyst can drill into. Risk scoring: a supervised model trained on past STRs and confirmed cases learns which combinations of features tend to precede a true positive. The output is a probability assigned to each new alert.

3. 3

   Way one to think about AI in AML: it is alert triage, not alert generation. The most defensible deployments use AI to reorder a queue that rules already produced. The rule says alert; the AI says 'work this one before that one'. This is regulatorily palatable because the underlying alert universe is unchanged and auditable. Banks that try to use AI to suppress alerts entirely have repeatedly run into trouble with regulators because the suppressed alert that turned out to be a real laundering case is impossible to defend after the fact.

The why-tree

Try this with Claude

AI-edge prompt: 'You are a PVARA-aware AML model risk reviewer. I am proposing to deploy a supervised risk-scoring model trained on three years of our STR labels. List the ten validation tests you would require before approving production deployment, including label-quality audits, bias tests, drift monitoring, and explainability acceptance criteria. Cite SR 11-7 and PRA SS3/18 patterns where useful.'