Generative AI-Empowered ERP Systems: Semantic Retrieval and Business Decision Support for Unstructured Data

Abstract

Enterprise Resource Planning (ERP) systems are effective in processing structured transactional data but remain limited in exploiting heterogeneous unstructured information such as financial reports, procurement descriptions, supplier notes, and internal textual records. Existing retrieval-augmented generation approaches improve information access, yet semantic retrieval, evidence utilization, and confidence control are often weakly coupled, which may reduce traceability in business decision support. This study proposes an ERP-oriented Semantic Retrieval and Decision Support (ERP-SRDS) framework integrating metadata-aware semantic encoding, hybrid lexical–semantic retrieval, cross-encoder evidence reranking, evidence-grounded generation, and confidence-based output control. Experiments on FiQA-2018 and FinQA show that ERP-SRDS achieves an nDCG@10 of 0.421 ± 0.008, a decision accuracy of 67.8 ± 1.3%, and an Evidence Coverage@5 of 92.4 ± 0.9%. Although BGE-M3 attains a slightly higher FiQA nDCG@10 of 0.429 ± 0.007, ERP-SRDS provides stronger downstream reasoning and evidence coverage. Under 30% irrelevant-context perturbation, its decision accuracy remains 63.9 ± 1.6%, with a 3.9-percentage-point decline compared with 7.0 points for RankRAG. These results indicate that combining retrieval relevance, evidence grounding, and confidence control can improve the traceability and robustness of generative AI-assisted ERP decision support without replacing managerial judgment.