Metabase Shares Hard-Earned Lessons from Building Production AI Analytics Agents

Summary

Metabase has published candid insights from developing Metabot, their AI-powered analytics agent, revealing how real-world deployment challenges differ dramatically from controlled demos. The company's engineering team experienced a high-profile failure when parallel development without proper integration testing caused their agent to malfunction during a CEO demonstration. The incident highlighted a critical gap between building AI for ideal scenarios versus production environments with messy data and ambiguous user queries.

Unlike typical text-to-SQL tools that work well with clean, well-documented databases, Metabot aims to navigate the complexity of real customer data—hundreds of tables, legacy systems, and vague user questions. The team discovered that while SQL generation is relatively straightforward due to SQL's prevalence in LLM training data, teaching an AI agent to work with Metabase's visual query builder and understand implicit user context presents far greater challenges. The article emphasizes that production AI must handle ambiguity: when a user asks "How many customers did we lose?", the system needs to clarify time periods, definitions, and metrics.

The Metabase team presented their findings at the AI Engineering conference 2025 in Paris, advocating for a fundamental shift in AI development philosophy. Rather than optimizing for the "happy path" with perfect data and clear questions, they argue developers must build systems resilient to chaos—the messy, ambiguous, and inconsistent nature of real-world business data and human communication. Their experience underscores the importance of comprehensive integration testing, especially when multiple engineers are shipping features simultaneously.

• The company advocates building AI systems for 'chaos'—designing for messy data and ambiguous queries rather than optimizing only for ideal scenarios