Why LLM analytics demos fail in production: metric definitions and permissions

A practitioner building governed conversational analytics on Looker argues that demos of LLMs querying clean warehouse schemas break down at company scale. Different departments define the same term, such as revenue or active customer, in conflicting ways, so a model querying raw tables guesses inconsistently and erodes trust. The author proposes a semantic layer where each metric is defined once in version-controlled code owned by accountable teams, with the model interpreting intent while the layer defines meaning. The article also warns that broad service-account access lets chat interfaces bypass permissions, and recommends that agents inherit the requesting user's access rights, enforced at the data layer rather than through prompts.