Explainable AML Triage with LLMs: Evidence Retrieval and Counterfactual Checks
Summary: arXiv:2604.19755v1 Announce Type: new
Abstract: Anti-money laundering (AML) transaction monitoring generates large volumes of alerts that must be rapidly triaged by investigators under strict audit and governance constraints. While large language models (LLMs) can summarize heterogeneous evidence and draft rationales, unconstrained generation is risky in regulated workflows due to hallucinations, weak provenance, and explanations that are not faithful to the underlying decision.
We propose an explainable AML triage framework that treats triage as an evidence-constrained decision process. Our method combines:
- Retrieval-augmented evidence bundling: This involves gathering information from policy/typology guidance, customer context, alert triggers, and transaction subgraphs.
- Structured LLM output contract: This contract mandates explicit citations and separates supporting evidence from contradicting or missing evidence.
- Counterfactual checks: These checks validate whether minimal, plausible perturbations lead to coherent changes in both the triage recommendation and its rationale.
We evaluate our framework using public synthetic AML benchmarks and simulators, comparing it against traditional rules, tabular and graph machine-learning baselines, as well as LLM-only and RAG-only variants. Our findings reveal that:
- Evidence grounding significantly enhances auditability and decreases numerical and policy hallucination errors.
- Counterfactual validation further improves decision-linked explainability and robustness.
- Our approach achieves the best overall triage performance, with a PR-AUC of 0.75 and an Escalate F1 score of 0.62.
- We also observe strong provenance and faithfulness metrics, including a citation validity of 0.98, evidence support of 0.88, and counterfactual faithfulness of 0.76.
These results indicate that governed, verifiable LLM systems can offer practical decision support for AML triage without compromising compliance requirements for traceability and defensibility. With the increasing complexity of financial transactions and associated regulatory requirements, the need for explainable AI systems in AML is more critical than ever.
As financial institutions strive to enhance their anti-money laundering efforts, integrating advanced AI methodologies such as our explainable triage framework could significantly streamline the process, ensuring that investigators can act swiftly and accurately in high-stakes environments. The implications of our research extend beyond mere compliance; they pave the way for more reliable and transparent decision-making systems in financial crime prevention.