AI in AML: from pilots to operational deployment

Finextra and NICE Actimize hosted a recent webinar examining how artificial intelligence is being applied across the anti-money laundering lifecycle. The panel included Nicola Eschenburg, EMEA AML Strategy Leader at NICE Actimize; Ashley Bostel, Head of Economic Crime Analytics Strategy at NatWest Group; and Jessica Cath, Managing Director at Thistle Consulting. Teresa Connors moderated the discussion.

Panelists described specific use cases now in production or advanced trials: automated adverse media screening to flag reputational risk, enhanced customer risk scoring during KYC checks, optimisation of screening rules to reduce irrelevant matches, and alert triage to prioritise cases for human review. They reported work on investigation automation and tools that draft suspicious activity report narratives by summarising transaction histories and pulling together relevant evidence.

Speakers explained how machine learning models can group related transactions, surface links between entities and highlight unusual patterns. Large language models are being tested to extract and summarise information from unstructured sources such as news articles and investigative notes. The panel said these capabilities can help investigators see a broader set of signals that exist across different teams and systems.

The discussion addressed barriers to moving from pilots to full deployment. Panelists identified challenges including inconsistent or low-quality input data, integration with existing case management systems, and the need to align workflows so investigators can act on AI-generated leads. They said institutions require clear, measurable indicators to support rollout decisions, citing reductions in false positives, shorter investigation cycle times and improved detection of high-risk activity as examples.

Regulatory and governance issues featured throughout. The panel highlighted transparency, explainability, documentation of decision logic, model validation and ongoing performance monitoring. They raised risks specific to language models, including occasional inaccurate or fabricated outputs, and recommended maintaining human oversight and review steps before filings.

On build-versus-buy choices, speakers outlined trade-offs: building internally provides more control over data and tuning but requires specialist staff and longer development; vendor solutions can accelerate deployment; hybrid approaches combine third-party tools with internal models. Factors that influence the approach include data availability, internal machine learning expertise, cost and the speed at which compliance functions must scale.

Panelists framed agentic AI as a support mechanism rather than an autonomous decision-maker, giving examples such as auto-drafting routine sections of investigations, suggesting next investigative steps in guided workflows and auto-populating SAR templates while leaving final judgements and filing decisions to qualified staff.

Speakers positioned the session against growing regulatory scrutiny of AML effectiveness and efficiency and focused on use cases that can be governed, measured and integrated into existing risk-management frameworks.