Regulation In, Regulation Out: An Agentic LLM Architecture for AML

Traditional Anti-Money Laundering (AML) featureengineering manually translates regulatory guidance into detection logic, creating traceability gaps between compliance intentand operational systems. We introduce an agentic LLM architecture where specialized agents autonomously compile regulatorytext into executable detection features via a Perceive–ReasonAct loop with self-correction. The key artifact is the Regulatory Feature Specification (RFS), a structured intermediaterepresentation maintaining structural traceability from regulatory passages to deterministic feature logic. The architectureis schema-adaptive: the same regulatory text from FFIEC andFINTRAC sources compiles to different feature implementationson incompatible database schemas. Experiments on two AMLdatasets demonstrate 67.9–75.0% per-model compilation success(14 indicators × 2 schemas per model), with regulation-groundedfeatures achieving Precision@50 of 0.26 (95% CI: 0.146, 0.403)to 0.30 (95% CI: 0.179, 0.446) on IBM AML versus 0.00 forraw statistical features—while maintaining structural traceability(not semantic correctness) from regulatory passage to flaggedaccount. We use “compilation” throughout as a design metaphorfor structured, artifact-producing translation; the system doesnot provide formal semantic preservation guarantees. The architecture enables AML teams to answer: “Why did you build thisfeature?”