Explainable AML Triage with LLMs: Evidence Retrieval and Counterfactual Checks

A new research paper from Dorothy Torres, Wei Cheng, and Ke Hu introduces an explainable anti-money laundering (AML) triage framework that leverages large language models (LLMs) under strict governance constraints. The system treats triage as an evidence-constrained decision process, combining retrieval-augmented evidence bundling from policy guidance, customer context, alert triggers, and transaction subgraphs. It enforces a structured output contract requiring explicit citations and separates supporting from contradicting or missing evidence. Counterfactual checks validate that minimal plausible perturbations produce coherent changes in both the triage recommendation and its rationale.

Evaluated on public synthetic AML benchmarks and simulators, the framework outperforms rules, tabular and graph ML baselines, and LLM-only/RAG-only variants. Key results include PR-AUC 0.75, Escalate F1 0.62, citation validity 0.98, evidence support 0.88, and counterfactual faithfulness 0.76. The approach substantially reduces numerical and policy hallucination errors while improving auditability. This demonstrates that governed, verifiable LLM systems can provide practical decision support for AML triage without sacrificing compliance requirements for traceability and defensibility.