Practical AI for AML: Where banks see results

Banks and other financial institutions have begun moving artificial intelligence from pilot projects into anti-money laundering operations in the past 12–18 months. Deployments cover customer onboarding, know-your-customer checks, transaction monitoring, alert triage and the drafting of suspicious activity reports. Institutions report the change is driven by targets to cut manual work, reduce false positives and speed investigations.

Live or near-live use cases include automated searches of adverse media, automated customer risk scoring at onboarding and during reviews, optimisation of screening rules to lower false alerts, machine-assisted alert prioritisation for investigators, and AI-assisted drafting of regulatory reports. Banks say these tools aim to shorten review times, raise the share of true positives and accelerate decision cycles for compliance teams.

The technology stack mixes traditional statistical machine learning with transformer-based language models. These models extract structured data from unstructured text such as news, filings and case notes, and link data across customer records and transaction histories. Agentic systems that can run multi-step routines are being trialled in controlled settings for routine investigative tasks and case management, with humans kept in the decision loop for final outcomes.

Institutions face a choice between building models in-house, purchasing vendor solutions, or using a hybrid of both. Building gives firms more control over design and data governance but requires investment in staff and infrastructure. Buying can shorten time to deployment and offers packaged compliance features. Hybrid approaches combine vendor tools with internal models and data pipelines to balance speed and control. Decision factors commonly cited include model explainability, ease of integration with legacy systems, the ability to tune models to an institution’s risk appetite, and total cost of ownership.

Governance and regulation are shaping how firms deploy AI. Compliance teams are prioritising explainability, audit trails and defined escalation paths. Supervisory expectations include documented controls for model development and validation, data quality checks and clear human oversight. Data protection rules limit how some external sources can be used, affecting the deployment of third-party LLMs and other models that rely on outside data.

Operational work often focuses on integrating AI outputs into existing systems. Many banks need to reconcile multiple data warehouses, legacy screening tools and separate teams before AI results can be acted on consistently. Teams that reported successful pilots described cross-functional governance, clear metrics for time savings and detection lift, and phased rollouts that keep humans in the loop during tuning.

Scaling AI in AML raises practical questions about where agentic capabilities can be applied safely, how to preserve explainability as models grow complex, and how to select build, buy or hybrid strategies. Firms continue to document model performance, strengthen data integration and formalise governance as prerequisites for broader operational use.