Enhancing Business Intelligence Systems with Large Language Models: A Structured Review of Approaches, Challenges, and Future Directions

Business intelligence (BI) has evolved from centrally produced reporting toward self-service, increasingly real-time, and conversational analytical environments. Large language models (LLMs) may further advance this evolution by translating natural-language questions into database queries, coordinating analytical tools, retrieving enterprise knowledge, and communicating analytical findings. However, linguistic fluency does not guarantee analytical correctness. This article presents a structured narrative review of 45 foundational, review, empirical, and system-oriented studies published between 1958 and 2026. The evidence is synthesized across six themes: the evolution and foundations of BI; natural-language interfaces and text-to-SQL; LLM-assisted analytics; retrieval and grounding; conversational and agentic BI; and trustworthiness, security, privacy, and governance. The synthesis indicates that LLMs can lower the technical barrier to data access and support multi-turn analytical exploration, yet dependable enterprise use remains constrained by schema linking, ambiguous business semantics, semantically incorrect SQL, numerical reasoning errors, hallucination, knowledge conflicts, security and privacy risks, limited verification, and weak accountability. At present, the evidence does not justify replacing governed BI engines with generative models. A more defensible approach, instead, is a hybrid architecture in which the LLM interprets user intent, orchestrates analytical tools, and communicates findings, while databases, semantic models, retrieval services, validation mechanisms, and authorization controls provide computation, grounding, and governance. Future research should prioritize enterprise-specific benchmarks, semantic and numerical verification, longitudinal field evaluation, human–AI decision quality, and graduated autonomy. The review contributes an integrated taxonomy and a governance-oriented architecture for evaluating and designing reliable LLM-enabled BI systems.