Human–AI Collaborative Retrieval-Augmented Decision Intelligence for Enterprise Knowledge Bases in Financial Services

Abstract

Financial service institutions increasingly require knowledge systems that can retrieve policy-relevant information, synthesize organizational evidence, and deliver explainable decision support under strict governance constraints. This study proposes a Retrieval-Augmented Decision Intelligence framework for enterprise knowledge bases in financial services by integrating semantic retrieval, metadata-aware re-ranking, grounded augmentation, and explainable response generation into a unified managerial intelligence pipeline. The framework was evaluated using enterprise-style query scenarios covering compliance clarification, product guidance, risk review support, and service resolution. The results showed that the proposed framework outperformed a baseline generative model across all major dimensions, achieving Precision@5 of 0.86 compared with 0.71, NDCG of 0.88 compared with 0.74, grounding score of 0.84 compared with 0.68, usefulness score of 0.87 compared with 0.70, and an overall effectiveness index of 0.86 compared with 0.71. Category-level analysis indicated that hybrid re-ranking improved ranking effectiveness in all query types, with NDCG increasing from 0.84 to 0.89 for compliance queries, from 0.87 to 0.91 for product guidance, from 0.82 to 0.87 for risk review support, and from 0.79 to 0.85 for service resolution. Grounding performance remained strong across categories, reaching 0.90 for compliance clarification, 0.88 for product guidance, 0.82 for risk review, and 0.79 for service resolution, demonstrating that retrieved enterprise evidence substantially constrained unsupported generation. Expert evaluation further confirmed high managerial value, with average scores of 4.4 for clarity, 4.5 for actionability, 4.3 for trustworthiness, 4.4 for interpretability, and 4.5 for decision value on a five-point scale. Failure analysis identified outdated policy retrieval, cross-document ambiguity, terminology mismatch, insufficient escalation signaling, and partial evidence coverage as the main residual weaknesses, with outdated policy retrieval accounting for 18 observed cases and cross-document ambiguity for 14. These findings indicate that retrieval-augmented architectures can move beyond information access and function as decision intelligence systems that support traceable, evidence-grounded, and operationally meaningful knowledge work in regulated financial environments.