Explainable AML Triage with LLMs: Evidence Retrieval and Counterfactual Checks

1. Problem Statement

Anti-Money Laundering (AML) transaction monitoring generates massive volumes of alerts, creating a bottleneck for human investigators who must rapidly triage them under strict regulatory constraints. The core challenges include:

• High False Positives: Traditional rule-based systems generate excessive noise, forcing analysts to spend disproportionate time on benign alerts.
• The "Black Box" of LLMs: While Large Language Models (LLMs) can summarize heterogeneous data (KYC, transactions, policies) and draft rationales, unconstrained generation poses severe risks in regulated environments. These include hallucinations (fabricating facts), weak provenance (citing non-existent sources), and unfaithful explanations (rationales that do not logically follow the decision).
• Auditability Gap: Regulators require decisions to be traceable, defensible, and grounded in specific evidence. Current ML approaches often lack the ability to provide human-readable, evidence-backed justifications, while raw LLMs fail to adhere to strict audit requirements.

The paper posits that AML triage must be treated not as a free-form text generation task, but as an evidence-constrained decision process.

2. Methodology

The authors propose a three-stage framework that integrates Retrieval-Augmented Generation (RAG) with counterfactual validation to ensure auditability and robustness.

A. Evidence Retrieval and Bundling

The system treats triage as a retrieval-grounded task. For every alert, a retriever constructs a compact "Evidence Bundle" ($B(a)$) from heterogeneous sources:

• Policy/Typology Corpus ($P$): Institutional rules and AML typology guides.
• Customer Context ($K$): KYC profiles, risk attributes, and beneficial ownership.
• Transactional Context ($T$): Transaction subgraphs, timestamps, and amounts.
• Case Memory ($H$): Historical similar cases and investigator notes.
• Constraints: Retrieval uses hybrid filtering (structured constraints like time windows and ACLs) and semantic search. It enforces deduplication, time-consistency (preferring latest policies), and access control to ensure only permissible data reaches the model.

B. Structured LLM Output with Provenance Constraints

The LLM is not asked to "write an explanation" freely. Instead, it operates under a strict Output Contract:

• Structured Disposition: Output includes a decision ($y \in \{dismiss, monitor, escalate\}$), confidence score, and next actions.
• Explicit Citations: Every claim in the rationale ($r$) must cite specific Evidence IDs from the bundle.
• Separation of Evidence: The model must explicitly distinguish between supporting evidence, contradicting evidence, and missing evidence.
• Verification Loop: An automated verifier checks for:
  • Provenance Integrity: All cited IDs must exist in the bundle.
  • Consistency: Numerical values (amounts, dates) and logical ordering must match the source text.
  • Repair: If checks fail, the system provides targeted feedback to the LLM to regenerate a corrected output.

C. Counterfactual Checks for Faithfulness

To ensure the explanation is faithful to the decision logic (and not just a decorative narrative), the system performs Counterfactual Validation:

• Decision Counterfactuals: The system generates minimal, plausible perturbations (e.g., removing a high-risk counterparty link, adjusting a time window) to see if the triage decision flips.
• Evidence Counterfactuals: Critical cited evidence items are removed from the bundle. The system verifies if the model:
  1. Stops citing the removed item.
  2. Adjusts its rationale and decision accordingly.
• Validation: A "Triage Validator" (rules engine or lightweight model) confirms if the perturbation actually changes the risk score, ensuring the counterfactual is meaningful.

3. Key Contributions

1. Evidence-Grounded Triage Formulation: Formalizes AML triage as a retrieval-grounded task where rationales are bound to specific, auditable evidence sources, moving away from unconstrained generation.
2. Counterfactual Robustness Layer: Introduces a novel mechanism to test explanation faithfulness by verifying that decisions and rationales change coherently under minimal, domain-relevant perturbations.
3. Reproducible Evaluation Blueprint: Establishes a rigorous evaluation methodology using public synthetic benchmarks (SynthAML10, AMLSim) to measure triage quality, evidence faithfulness, and robustness without relying on proprietary bank data.

4. Experimental Results

The system was evaluated on synthetic AML benchmarks (SynthAML10 and AMLSim) against baselines including Rules, Tabular ML (GBDT), Graph ML (GNN), LLM-only, and RAG-only.

• Triage Performance: The full system (RAG + Counterfactuals) achieved the best performance:
  • PR-AUC: 0.75 (vs. 0.69 for GNN and 0.52 for LLM-only).
  • Escalate F1: 0.62 (vs. 0.56 for GNN).
  • Insight: LLMs alone underperformed; their value lies in structured decision support when grounded in evidence.
• Auditability & Provenance:
  • Citation Validity: 0.98 (vs. 0.62 for LLM-only), indicating near-zero fabrication of evidence IDs.
  • Evidence Support: 0.88 (vs. 0.54 for LLM-only), meaning 88% of rationale sentences were directly supported by the cited text.
• Counterfactual Faithfulness:
  • CF-Flip Rate: 0.58 (vs. 0.41 for RAG-only), showing the system better identifies actionable "flip conditions."
  • Evidence Faithfulness (Removal): 0.76, demonstrating the model correctly updates its reasoning when key evidence is removed.
• Safety & Hallucination:
  • Numerical Inconsistency: Reduced to 0.03 (vs. 0.11 for LLM-only).
  • Policy Hallucination: Reduced to 0.02 (vs. 0.14 for LLM-only).

5. Significance and Implications

• Regulatory Compliance: The framework directly addresses the "explainability" gap in AI for finance. By enforcing citations and counterfactual checks, it produces decision artifacts that are defensible under Risk-Based Approach (RBA) supervision.
• Operational Efficiency: The system improves precision without sacrificing recall, potentially reducing the volume of false positives analysts must review while maintaining high detection rates.
• Trustworthy AI: The study demonstrates that LLMs can be safely deployed in high-stakes financial workflows if they are treated as components of a governed system (with retrieval, verification, and stress-testing) rather than standalone generators.
• Future Deployment: The authors suggest a "human-in-the-loop" model where the system handles the heavy lifting of evidence synthesis and initial triage, flagging uncertainty and providing actionable next steps, while human analysts retain final accountability.

In conclusion, the paper argues that governed, verifiable LLM systems are the key to unlocking the potential of generative AI in AML, balancing the need for advanced pattern recognition with the non-negotiable requirements of traceability and defensibility.