An executive types "what were our top customers last quarter" into a Cortex Analyst-style query box and gets a confident answer back. It is wrong, not because the generated SQL was malformed, but because "customer" in that warehouse means at least three different things depending on which table answers the question, and "top" could mean revenue, order count, or margin. None of that ambiguity was visible in the question. It lives in years of internal convention that sit in analysts' heads, not in the schema.
This is the same wall every prior wave of natural-language query tools hit, going back well before language models existed. The current generation, built on top of Snowflake's Cortex Analyst, Databricks' Genie, or similar assistants, is genuinely better at producing syntactically correct queries than anything before it. Syntax was never the actual problem.
Analytics teams are the ones absorbing the gap. Rather than being replaced by tools that let executives query data directly, many have found their workload shift toward maintaining semantic layers, explicit, versioned documentation of what each field and metric actually means, in the tradition dbt Labs has built much of its current product strategy around and that Looker's older LookML modeling layer proved out years earlier. That work is less visible than writing queries and does not show up in a product demo.
What breaks is the pitch that a plain-language interface removes the need for someone who understands the data model. Organizations that skipped the semantic-layer work and went straight to a chat box in front of their warehouse have generally found the same failure mode: confident, fluent, wrong answers that are harder to catch than an obviously broken query, because they read like something a person could plausibly have written.
What is opening up is a market for tooling that treats semantic definitions as a first-class, versioned asset, closer to how source code has long been managed than how business metrics have historically been handled. It is a far less glamorous category than natural-language query, and it is the one actually determining whether the glamorous layer on top produces trustworthy answers.
Snowflake and Databricks are both selling the chat box. dbt is selling the dictionary underneath it. Which one a company buys first says a lot about whether it understood the actual problem, or just the demo.