Databricks Genie's table list curates answers, not access

The setup

Genie is Databricks' conversational analytics feature, aimed at a business user who wants an answer from company data without writing SQL or filing a ticket with a data analyst. Setting one up means creating a Genie agent, sometimes called a space, and adding specific tables, views or metric views, up to fifty per agent according to Databricks' own setup documentation, plus optional example queries that subject-matter experts write to steer how Genie answers domain-specific questions. Access requires the data already be registered in Unity Catalog, Databricks' governance layer, and a Pro or serverless SQL warehouse to run against.

What the documents show

Databricks' product page markets Genie as self-serve conversational analytics from scoped, domain-specific agents curated by subject-matter experts, stating Unity Catalog integration ensures compliance with existing governance policies. Its documentation index distinguishes this feature, Genie Agents, from a coding assistant, Genie Code, and a workspace hub, Genie One. The setup documentation adds a more precise claim: tables added to a space are what Genie uses by default and what its query suggestions build around, but they are not a hard boundary. It states plainly that Genie can query tables beyond those explicitly added, because access is controlled by Unity Catalog permissions, not the agent's own configuration.

The friction

That single line matters more than the marketing description of a scoped agent suggests: adding tables to a Genie space is a curation and accuracy aid, steering which data Genie treats as relevant, not a security control in itself. The actual access boundary is whatever Unity Catalog already grants the querying user's role, so an agent nominally scoped to a sales table can still reach other data that role can see. The documentation offers a real control for finer-grained cases, row filters and column masks defined in Unity Catalog, enforced per user automatically regardless of how the agent is shared, but a team has to set those up separately; they are not a Genie feature.

What changed in the work

Where a data team has already done the Unity Catalog governance work, Genie's documented design lets a business user ask a question in natural language and get scoped, curated answers without a new permissions system, the setup documentation's intended workflow. What it does not support is treating the list of attached tables as a security fence: a team that assumes adding tables also restricts access, rather than relying on Unity Catalog roles, is trusting a curation feature to do a governance feature's job, an editorial reading the documentation's wording invites.

• Does the team's Unity Catalog role structure already match what each Genie agent is meant to be scoped to?
• Are row filters or column masks needed for any table a Genie space treats as sensitive?
• Who is responsible for writing and maintaining the example queries that shape Genie's precision?

Databricks' own setup documentation is more precise than its marketing page about what scoping a Genie agent does: it steers relevance and accuracy, while the real access boundary continues to run on Unity Catalog permissions set up independently of Genie itself.