This presentation on AI in BI using the self-driving car analogy is the clearest framework I have found

High-accuracy requirements like text-to-SQL for business decisions where a wrong answer has material consequences sit differently than creative suggestion tasks like charting recommendations where an option that does not quite fit can be corrected by a human. Fully automated driving on a highway is fine. Fully automated navigation through a complex intersection is not.

MCP being flagged as a disruptive technology allowing LLMs to interact with applications with user permissions is the architectural shift worth watching. Semantic layers as guardrails for AI navigating messy data are the practical response.

The emphasis that foundational BI skills, data models, SQL and data quality, remain crucial because AI relies on good data to avoid delivering faster wrong answers is the honest counter to the AI-replaces-analysts narrative.

For data teams thinking about where to introduce AI in their BI stack: what is the first workflow you would automate versus the first one you would never fully automate regardless of capability?