The author, with over 20 years in analytics, highlights the ongoing evolution of data analytics, emphasizing the importance of accuracy in Text-to-SQL systems. While Large Language Models (LLMs) offer exciting opportunities for democratizing data access, a mere 80% or 90% accuracy is inadequate. Errors in SQL generation can lead to mistrust from users, which ultimately hampers adoption and risks significant business decisions based on incorrect data. Creating a reliable Text-to-SQL system involves a complex pipeline, including components like intent classifiers, vector databases, retrieval mechanisms, and SQL generation tailored to specific databases. The evaluation of such systems is often neglected but is essential for enterprise reliability. Current evaluation methods fall short, relying on outdated benchmarks like Spider 1.0, which doesn't reflect the messy realities of enterprise data. The author argues for a shift to more relevant metrics, such as Execution Accuracy (EX) and the Soft-F1 metric, which provide a better understanding of system performance. Spider 2.0 emerges as a significant advancement, designed to address the challenges in evaluating LLM performance in real-world scenarios. It introduces realistic complexities of large enterprise schemas, with some databases exceeding 3,000 columns. This new framework aims to bridge the gap between high accuracy in controlled environments and the unpredictable nature of production systems, ensuring that tools developed for analytics are not just theoretically sound but practically reliable as well.