Databricks put Genie inside ChatGPT, Claude and two other assistants so they answer business questions from governed data.
The hook is Genie Ontology, the certified layer for definitions, relationships and sources. Databricks says Genie works out what a term means, finds tables and documents the user is allowed to see, writes and runs the SQL, then returns an answer with citations. For specialized work it routes to a domain Genie Agent, like marketing for attribution or finance for costs. That follows its claim that Genie hits 91% on business tasks, described here.
Permissions do the quiet work here.
Agents get two paths and they matter for how much you trust the output. They can hand Genie a full question, why is activation low or which accounts are at risk, and let Genie run its own loop. Or they can keep the larger job and ask Genie only for the missing fact, then go back to editing files, drafting a plan or posting to Slack. Anything without a native plugin uses the same context through the Genie MCP server, and clients that support the MCP Apps extension get charts and Ontology citations inline.
You can see the split in the examples. A ChatGPT dot watches feature adoption by plan tier against target, using the certified active user definition, then flags slips in Slack. Claude pulls certified revenue, order volume and deal size to explain a Northeast miss, then checks pricing memos and call notes before writing the QBR summary. Microsoft Copilot Cowork pulls usage and commit data from Azure Databricks to build the Word brief, Excel charts and Thursday deck. Cowork already sends emails, schedules meetings, creates documents, posts in Teams and manages calendars on your behalf, per Microsoft Learn, so Databricks adds a rule that it waits for approval before anything is sent. A Grok Bot teammate runs the Monday loop on cost per qualified opportunity and drafts budget shifts.
The failure mode is obvious if you have run this kind of setup. You build the meaning once in Ontology and every agent reuses it, which is great until a definition goes stale. Then ChatGPT, Claude, Cowork and Grok all cite the same wrong metric with confidence. The fix is governance, not more connectors, and that is the real test for teams turning this on.
