**Key takeaways** - Treat Genie readiness like semantic model readiness for any AI consumer. - Certify a small gold surface; do not point chat at bronze or raw. - Make grain explicit; one row means one thing in the table description. - Put business names and glossary where the planner and the model both look. - Declare real relationships and single owned metric definitions. - Classify sensitive columns before you open conversational access.

Teams ask how good the data needs to be before conversational analytics helps. The honest answer: good enough that a careful analyst would trust the same tables without a chat UI. Prompts cannot rescue ambiguous grain, cryptic column names, or missing relationships. Conversational analytics reads metadata and structure first. System prompts are a thin overlay.

**Named model: Genie readiness is modeling readiness** - **Curated schemas:** certified gold surface with explicit, boring grain. Not bronze or raw for end-user chat. - **Names and glossary:** business-readable labels, catalog descriptions for codes, synonyms next to the data the tool can read. - **Relationships:** declared keys and joins, constrained exploration. Wrong joins produce plausible numbers. - **Owned metrics:** one definition, a named owner, and a place the agent is allowed to read. Sensitivity classified before chat opens. - **Contract:** prompts are a thin overlay on structure and metadata. Ambiguity underneath becomes fluent wrong answers.

![Four pillars of Genie readiness: curated schemas, names and glossary, declared relationships, and owned metrics under a thin prompt overlay](/assets/blogs/13-genie-readiness-is-a-modeling-problem/genie-readiness-as-modeling.svg) *Figure 1. Genie readiness is modeling readiness. Curated schemas, glossary next to the data, declared relationships, and owned metrics. Prompts stay thin.* ## Context and constraint I have seen the same recovery patte

Browse and find notes