AI reshapes AML: banks deploy ML, LLMs and agentic tools

Financial institutions are shifting AI projects from pilots to live anti-money laundering operations. Banks and payment firms apply machine learning, large language models and agentic systems to transaction monitoring, alert triage and the preparation of suspicious activity reports.

Applications include customer onboarding, know-your-customer checks, ongoing KYC, investigations, case management and regulatory reporting. Firms link transaction histories, account relationships, screening results and media checks to present a more complete view of customer risk to investigators and compliance teams.

Near-term use cases in production include automated adverse media searches to surface negative press about customers, automated customer risk scoring at onboarding and during review cycles, optimisation of sanctions and name-screening to reduce false matches, and AI-assisted alert triage that ranks and routes potentially suspicious activity to investigators. Investigations teams increasingly use LLMs and automation to summarise transaction histories and draft narratives for regulatory filings, reducing time spent on routine analysis.

Operational deployment requires integration across teams and legacy systems. Firms report work to standardise taxonomies, improve data quality and make systems interoperable so that alerts and risk signals flow between compliance, risk, IT and business units. Organisations list data access, mapping of fields and end-to-end workflows as practical prerequisites for wider use.

Agentic systems, which can perform multi-step tasks under rules, are being trialled to automate repetitive investigation steps, escalate high-priority alerts and populate case files. Deployments keep human investigators in the loop: outputs are reviewed before decisions that affect customers are implemented, and firms apply limits so automated actions do not exceed predefined authorities.

Compliance teams face build-versus-buy decisions when scaling AI. Some groups develop models in-house to retain control of training data and explainability. Others adopt vendor platforms to speed deployment and obtain specialist analytics. Many combine internal models with third-party components. Choice factors include availability of clean data, internal AI expertise, regulatory reporting obligations and the resources needed for ongoing model validation.

Governance and regulatory requirements inform deployments. Firms put in place model governance frameworks, documentation, audit trails and performance monitoring to show how AI-driven processes operate. Explainability remains a priority so that automated outputs can be justified and errors or biases identified and corrected. Review processes and escalation paths are used to ensure human judgment is applied to high-risk cases.

Machine learning is used primarily for pattern detection in transaction data. Large language models are applied to natural-language tasks such as summarising documents, extracting entities from media articles and generating draft narratives. Agentic systems coordinate multiple steps but are operated with stricter controls and oversight.

Institutions track measurable outcomes from deployments, including lower volumes of alerts requiring manual review, faster time to close investigations, improved prioritisation of higher-risk cases and clearer audit records for regulatory review.