Databricks Brings Genie Into ChatGPT to Query Company Data Instantly

Databricks has released a native Genie plugin for ChatGPT, allowing users to query governed lakehouse data in natural language without opening the Databricks interface. The integration can generate and execute SQL, summarize dashboards and return answers with source citations.

The release also includes native Genie plugins for Claude, Microsoft Copilot Cowork and Grok Bot. Other compatible agents can connect through the Genie One Model Context Protocol server. MCP is a standard that lets AI clients discover and call external tools, giving Databricks one integration layer across several assistant interfaces.

How Genie turns prompts into governed queries

Genie sits between the assistant and a customer’s Databricks environment. Its semantic layer, Genie Ontology, stores certified definitions, relationships and sources so the system can map business language to the appropriate data.

1. The user asks a question in an assistant such as ChatGPT.
2. Genie resolves business terms through the organization’s ontology.
3. Unity Catalog permissions determine which tables, documents and other sources the user can access.
4. Genie generates and runs SQL inside the Databricks environment.
5. The assistant returns the result with citations and, where supported, charts.

For a request such as “Why is activation low this quarter?”, Genie applies the company’s certified definition of activation when selecting data and constructing the query. That shared definition helps keep answers aligned with approved dashboards and reports.

Two routes into Genie

Identity and policy travel with each query

Access depends on controls in both the assistant workspace and Databricks. For ChatGPT and Claude, default OAuth applications let employees sign in with their Databricks accounts without requiring an administrator to register a custom OAuth app. Unity Catalog then applies the user’s existing permissions to every request.

For ChatGPT, an administrator creates the plugin from a workspace template, enters the Databricks workspace address and publishes it. Employees can then find the plugin and connect their accounts through OAuth.

• Latency: Responses are asynchronous because Genie may need to search sources, generate SQL and wait for query execution.
• Network access: Restricted Databricks workspaces must permit traffic from ChatGPT or Claude.
• Data handling: Teams should review each assistant provider’s retention, residency and training policies before exposing production data.
• Validation: Pilot answers should be checked against approved dashboards, including their citations and applied metric definitions.

Recurring agents reuse the same metrics

Databricks also positions Genie as a data layer for autonomous agents that run recurring tasks. A ChatGPT Dot, described as an always-on agent with its own computer environment, could monitor feature adoption, compare results with targets, collect related support themes and post a summary to Slack. Routing each step through Genie keeps the agent’s active-user calculation tied to the same certified definition used elsewhere in the company.

Databricks widens Genie’s remit

The plugins arrive alongside the second generation of Genie, which Databricks describes as an autonomous agent for ongoing enterprise work. Users can assign it a goal and allow it to continue gathering data, running analyses and updating outputs over time.

Databricks reports a 91% score on enterprise tasks, a 1.6x result against a frontier agent and half the token use, with requests governed through Unity Gateway. Evaluators will need the task set, baseline model, latency measurements and token-accounting method to assess how those figures translate to production workloads.

The announcement followed Anthropic’s release of Claude Sheets and Meetings for office workflows. By supporting several competing assistants, Databricks is positioning Genie as a shared governed-data layer instead of requiring customers to build a separate connector for each client.

What is available now

A careful path to production

For developers, the release packages authentication, context retrieval, query execution and citations behind native plugins and MCP. Production work still includes ontology design, catalog permissions, network controls, provider-policy review and systematic evaluation of generated answers.

Answer quality will track the quality of the underlying ontology. Citations establish where an answer came from, but they cannot repair an incorrect metric definition or a poorly modeled relationship. A narrow pilot should begin with one well-understood metric, representative user roles and a known dashboard result before expanding access to additional teams and datasets.