A Data-Driven Decision Support Framework for Business Intelligence Systems Leveraging BERT-Enhanced LLMs

Abstract

Many techniques were introduced to intimate the gap existing between raw organizational data of business intelligence systems and practical decision support. One of these techniques is to integrate the large language models (LLMs) into enterprise business intelligence (BI) architectures. The available BI frameworks suffer from bi-direction drawbacks. These drawbacks represented by treat natural language query interfaces as peripheral additions or by failing in providing empirically confirmed and decision-quality outcomes. This will put practicians without any principled basis for architectural investment. This paper addresses both limitations. We present a three-layer BI-Enabled Decision Quality (BIDQ) framework that formalizes the interdependencies among data readiness, analytical capability, and decision integration, and we introduce a BERT-based query understanding module trained on 180,000 enterprise query--response pairs drawn from five production systems (SAP ERP, Salesforce CRM, Oracle SCM, customer service ticketing, and financial reporting). The BERT model achieves 86.5% test accuracy, an F1-score of 86%, and a mean inference latency of 45 ms a 5.5-percentage-point accuracy improvement over the best LSTM baseline (p < 0.001, Cohen's d = 0.42). Training converged in 72 GPU-hours on eight NVIDIA A100 GPUs using AdamW optimization with linear warmup and decay. To ground the framework in measurable outcomes, we conducted a twelve-month field study across five industry sectors (retail, banking, healthcare, manufacturing, and logistics), finding an average 26-percentage-point improvement in decision quality and an 89% reduction in decision cycle time following BI deployment. Financial modelling indicates a typical break-even point at approximately month 12 of deployment. These results determine the quantitative criterions that can guide the organizations in formalization and structuring the investments and measuring of LLM-augmented BI.