Governed LLM-assisted AML case triage: a non-authoritative, auditable decision support architecture

Abstract

Purpose. This study aims to identify an optimal decision-support architecture for regulated AML case triage that enables non-authoritative LLM assistance while maintaining auditability, regulatory compliance, and analyst accountability. The focus is on how evidence volume and type, bundle construction and validation, and the cost of reproducible audit artifacts affect workflow effectiveness. Method. The research specifies and analyzes a governed gateway architecture with a sequential workflow comprising evidence selection, canonicalization and hashing, minimal evidence bundle construction, structured draft generation, output validation, quarantine handling, and human acceptance with diff binding. The approach enforces policy and context constraints and binds all outputs to tamper-evident audit events. Qualitative evaluation examines narrative consistency, provenance coverage, and review readiness, alongside common failure modes such as prompt injection, policy leakage, and unintended actionable language. Results. The findings show that a governed gateway using minimal evidence bundles and explicit human acceptance gates improves audit readiness, consistency of outputs, and control over context leakage. While the architecture mitigates several governance risks, residual issues persist, including automation bias and subtle prompt injection, indicating the need for stronger human-in-the-loop safeguards and ongoing validator monitoring. Scientific novelty. The work operationalizes AI governance as a set of enforceable, checkable invariants–such as template pinning, bundle-only context, claim-level provenance, structured non-delegation, and tamper-evident logging–and integrates them into a full lifecycle for LLM-assisted AML artifacts. Practical importance. The proposed architecture provides compliance and engineering teams with a concrete, auditable pattern for deploying LLM drafting assistance in AML triage, enabling productivity gains while preserving analyst responsibility and minimizing uncontrolled disclosure under regulatory scrutiny.