AI Agents Move Into AML and KYC Workflows as Banks Test Compliance Automation

AI agents are moving from experimental pilots into practical financial-crime and compliance workflows, with banks increasingly testing them across onboarding, screening, due diligence and ongoing monitoring.

McKinsey has estimated that agentic AI could produce productivity gains of roughly 200% to 2,000% in financial-crime operations, with a human practitioner potentially supervising 20 or more specialised AI agents. The consultancy distinguishes this from conventional generative AI, which often assists investigators with individual tasks but does not fundamentally redesign the end-to-end operating model.

Adoption is still at an early stage. Capgemini Research Institute reported in its World Cloud Report in Financial Services 2026 that only 10% of financial institutions had deployed AI agents at scale. Among banks, fraud detection was identified by 64% as a leading large-scale use case, while 59% pointed to customer onboarding.

Agents are being assigned specific compliance roles

Duna, an AI-native business identity and compliance platform, has outlined how specialised agents can be distributed across the customer lifecycle rather than used as general-purpose chatbots.

An onboarding agent can read KYC documents, conduct sanctions, PEP and watchlist screening and prepare source-of-funds summaries. A due-diligence agent can map ultimate beneficial ownership structures, review adverse media and draft risk narratives. Screening and monitoring agents can investigate first-line alerts, assess false positives and prepare SAR narratives, while perpetual-KYC agents can monitor registry changes and re-score customer risk after onboarding.

Duna describes the model as policy-driven: a policy engine first determines what evidence is required, gathers available information from registries, documents and screening sources, and then assigns unresolved work to specialised agents. Human analysts remain responsible for judgement-heavy cases and escalations.

In one customer example cited by Duna, a European e-commerce platform reduced onboarding time from eight days to less than one minute. The same customer said it had completed more than 20,000 screenings through the platform while consolidating information that previously had to be gathered from multiple systems.

Governance remains the limiting factor

The technology also creates a new control problem. Duna notes that an agent can drift over time—for example, clearing a case it previously would have escalated—so recommendations need to remain traceable to underlying evidence and organisational policy. McKinsey similarly stresses clear agent roles, hand-off protocols, quality-assurance agents and human oversight for complex exceptions.

The emerging model is therefore less about replacing compliance teams than changing how work is allocated. The strongest use cases are likely to be those where repetitive evidence gathering and alert investigation can be automated while risk decisions, exceptions and accountability remain subject to human control.