• Proposes TRIAG, a novel framework coordinating three generative AI agents using MARL (Multi-Agent Reinforcement Learning). • The TRIAG artifact is designed to automate complex financial regulatory risk tasks. • Extends Generative Sociotechnical System theory by designing a MARL framework as a governance architecture. • Evaluation with FinTech (Financial Technology) compliance experts validates the artifact's practical relevance. Managing financial regulatory compliance (FRC) presents significant challenges for Financial Technology (FinTech) organisations. These challenges are exacerbated by a rapidly evolving regulatory landscape and the inherent limitations of existing computational support. Although FinTech organisations continuously seek contemporary methods for addressing the limitations, common issues persist with time-consuming and labor-intensive compliance processes. Previous Artificial Intelligence-driven approaches are typically fragmented, relying on static single-agent models that lack comprehensive support for the dynamic nature of regulatory requirements. To address these gaps through leveraging the increasing availability of regulatory and financial data, the study utilizes a design science research paradigm to design TRIAG (Tri-Reinforced Infused Generative Agents), an innovative computational framework grounded in Hierarchical Multi-Agent Reinforcement Learning (MARL). The TRIAG artifact can be viewed as a prototype that integrates three distinct LLM-based agents (Alpha, Beta, Gamma) that autonomously acquire, refine, and coordinate domain-specific expertise related to FinTech regulatory compliance. We evaluate by combining quantitative and qualitative benchmarking, demonstrating that TRIAG achieves a regulatory retrieval F1-score of 0.93 with 96% reduction of inference costs. The results confirm the enhancement of the efficiency and accuracy of compliance officers' decision-making. This work introduces a novel MARL-guided, multi-GenAI agent system for mitigating high-stakes financial regulatory risks. Overview details of the TRIAG Framework: (Left) The system ingests heterogeneous regulatory and financial data. (Center) The core architecture utilizes a Hierarchical Multi-Agent Reinforcement Learning (MARL) approach, where an Orchestrator (Alpha) coordinate specialized agents (Beta for Policy, Gamma for Horizon Scanning) via Fine-Tuning and a Hybrid RAG retrieval pipeline. (Right) The system delivers cost-effective, real-time compliance decision support with high accuracy (F1: 0.93), validated through industry workshops.
Sheikh et al. (Sun,) studied this question.