Deploying LLMs for self-service BI seems promising — until you encounter real-world data. This presentation shares concrete insights from several production and pre-production deployments of AI chatbots designed to interface with business data. We will cover how data teams approached prompt engineering, retrieval-augmented generation (RAG), metadata injection (from dbt, catalogs, lineage tools), and evaluation pipelines to mitigate hallucinations. You will also see how we measured performance (semantic accuracy vs SQL validity), managed user expectations, and iterated on UX and guardrails. Expect a no-fluff overview of architectural patterns, lessons learned, and why some promising PoCs still failed — and what to do about it!