Enterprise data assistants built on large language models (LLMs) aim to give non-technical staff direct access to organizational data. The authors report on a real-world deployment within the CFO unit's data lake at an industry partner, where they observed that business users' prompts present challenges. This observation motivated a simulation-based evaluation approach to test the assistant's robustness.

Using knowledge graphs as a structural layer, the evaluation simulates a range of imperfect prompts to understand failure modes. The paper argues that designing for imperfect prompts is critical for real-world adoption, and that simulation offers a scalable way to assess performance before deployment. The work underscores the gap between idealized LLM capabilities and the messy reality of enterprise usage.