Practical AI for AML: How banks scale models

A report published this month by industry analysts and former bank compliance officers outlines how practical AI can be deployed in anti-money-laundering (AML) operations. It focuses on concrete uses such as alert triage, entity resolution and case prioritization and sets out technical, process and governance steps to move from pilots to production-scale systems.

The authors recommend prioritizing a small number of high-value use cases and integrating models into existing investigator workflows rather than replacing investigators. They advise starting with tools that reduce low-risk alerts, enrich cases with linked-entity data and surface the highest-risk investigations for human review. The report states those steps can shorten investigation times and allow compliance staff to focus on complex cases that require judgment.

The paper sets out a phased implementation plan: define clear success metrics; clean and centralize transaction, customer and watchlist data; develop explainable models for specific tasks; run parallel evaluations with current systems; and iterate using investigator feedback. It describes combining machine learning techniques such as graph analytics for network detection and natural language processing to extract details from customer notes with engineered rules kept during transition phases. The authors call for continuous monitoring of model performance and periodic retraining tied to labeled investigator outcomes.

Practical changes include building a central data layer to resolve customer identities across products, deploying triage models that assign numeric risk scores to alerts, and creating investigator dashboards that summarize model reasoning and supporting evidence. The report notes these changes typically require work on legacy IT, investments in data engineering and defined processes for model approval and audit logging. New roles mentioned include data engineers, model risk officers and investigators trained to validate model output and provide corrective labels.

The analysis identifies three main barriers to scaling AI in AML: fragmented and poor-quality data; regulatory and audit expectations for explainability and documentation; and organizational resistance when tools change daily processes. To address regulatory expectations, the authors recommend documenting model purpose, inputs, performance metrics and decision paths and keeping human oversight at the point of suspicious-activity report submission. They also suggest pilot timelines of three to nine months for targeted use cases and cross-functional teams to speed adoption.

The report includes case examples from international banks that moved from isolated proofs of concept to broader rollouts after focusing on measurable gains such as reduced time per case, fewer false positives reaching investigation and higher-quality referrals to law enforcement. Mark Thompson, head of financial crime at a mid-sized European bank that participated in case studies, noted early deployments reduced repetitive work for analysts and produced clearer audit trails for model decisions while requiring comparable investment in data plumbing and governance as in modeling.

Sarah Nguyen, lead analyst on the study, wrote, “Practical AI for AML is about shifting investigator effort, not removing it.” Nguyen added that when models handle routine enrichment and prioritization, investigators can spend more time on complex network analysis and prosecutable referrals.