Position: Avoid Overstretching LLMs for every Enterprise Task

A new academic paper submitted to arXiv argues against the widespread deployment of Large Language Models (LLMs) for all enterprise tasks. The authors contend that using monolithic LLMs for deterministic, structured, and knowledge-dependent workloads is inefficient, unreliable, and misaligned with strict corporate constraints regarding cost, latency, and reliability. Theoretical evidence presented in the study suggests that finite-capacity models cannot fully capture the breadth of knowledge required for complex enterprise operations, leading to inherent limits in efficiency and interpretability. Instead, the researchers propose a modular architecture where language models serve primarily as interfaces for structured extraction within deterministic workflows. Computation and storage responsibilities should be externalized to dedicated knowledge bases and symbolic procedures. This approach aims to enhance system reliability, scalability, and transparency. The paper formally demonstrates that such modular frameworks are more maintainable than current monolithic solutions, offering a sustainable foundation for future enterprise AI systems by addressing the limitations of relying solely on large language models for diverse computational tasks.