Current text-to-SQL benchmarks are fundamentally flawed because they fail to address the messy reality of enterprise data and the inherent ambiguity of business questions. Rather than relying on pure zero-shot SQL generation, experts emphasize the need for agentic workflows, robust semantic layers, and strict data governance to validate outputs and prevent dangerous hallucinations. Ultimately, accurate AI analytics depend less on prompt engineering and more on foundational data engineering to clean and structure underlying warehouses.