Financial crime compliance has passed through three distinct eras. First came human-led investigation, then rules-based automation, and now a third wave is beginning to take shape: agentic AI. Unlike its predecessors, agentic AI does not simply execute fixed logic.
According to ComplyAdvantage, it plans, retrieves information, and moves through a sequence of decisions independently. For anti-money laundering (AML) teams, the question is no longer whether the technology exists but whether it can be trusted to operate inside a regulated environment.
Three converging forces are driving adoption faster than any previous technology cycle. The first is the industrialisation of financial crime itself. Criminals are increasingly deploying AI to evade legacy detection systems, a trend ComplyAdvantage’s research has linked to a 900% rise in AI-enabled financial crime in recent years. The second is the compression of time.
Real-time payments, embedded finance, and instant onboarding have collapsed the window available for compliance teams to assess an alert from days to mere seconds. The third is technological maturity. Agentic systems have arrived at a capability threshold that earlier iterations, including early machine learning models and large language models flagged for hallucination risk, had not reached.
Crucially, agentic architectures can deploy sub-agents to verify their own outputs, addressing the accuracy concerns that long made banks reluctant to hand over meaningful decisions to AI.
The clearest early return on investment lies in alert triage and case investigation. False-positive rates remain a chronic drain on compliance resources, with more than 70% of firms reporting rates above 30% and screening alert false positives well above 90% in some cases. Agentic AI can absorb that noise at scale, freeing investigators to concentrate on the alerts that genuinely require human judgement.
Demand is clearly there: ComplyAdvantage’s State of Financial Crime survey found that 99% of respondents are either using or evaluating AI for customer screening and transaction monitoring. Yet early returns have often underwhelmed.
A Bain & Company survey of 951 companies found that while 37% had targeted cost reductions of between 11% and 20% from AI, nearly 40% of those who measured outcomes landed in the 0–10% range instead. The gap, industry practitioners argue, comes down to focus: value accumulates when agentic AI is concentrated on high-volume, low-value first-line work rather than dispersed broadly and thinly.
As deployments move from proof-of-concept into production, the regulatory bar is rising in parallel. The Anti-Money Laundering Authority (AMLA) has been operational since 2025, with a Single Rulebook coming into force in 2027 and direct supervision of the highest-risk entities beginning in 2028. Regulators are increasingly requiring firms to demonstrate effectiveness rather than simply document their intentions. The compliance gap between what a governance document describes and what a system actually does in a live environment is where firms face the greatest exposure.
Closing that gap demands decision-level traceability across every step from data ingestion to final output, calibrated human oversight that keeps investigators genuinely in control rather than merely nominally in the loop, and continuous model validation over time.
Progress is being made: 98% of firms surveyed by ComplyAdvantage report an AI assurance programme either in place or under development, covering effectiveness, auditability, and model risk governance. Underpinning all of it is data quality, which has emerged as the single most important preparatory factor in making agentic AI function reliably.
Reaching a defensible deployment requires more than technical capability. Research firm Celent defines a genuinely capable agent as one that can execute multi-step processes, pursue objectives proactively rather than wait to be prompted, adapt in real time as circumstances shift, and act autonomously in ways that humans can subsequently audit. Getting there consistently is what distinguishes a controlled demonstration from a dependable compliance function.
Organisations that succeed tend to treat the challenge as much as an organisational one as a technical one. Building meaningful collaboration between AI engineers and compliance officers who must ultimately answer for the system’s decisions requires clear leadership from the top and a shared understanding of the business case for getting it right.
For most firms, the practical path forward is not to build from scratch but to select a technology partner carefully: one whose AI is native rather than bolted on, whose data foundations are sound, whose governance controls are demonstrable rather than theoretical, and whose architecture is flexible enough to accommodate a risk-based approach.
Read the full ComplyAdvantage post here.
Copyright © 2026 RegTech Analyst
Copyright © 2018 RegTech Analyst
