OpenAI introduced a Data agent in ChatGPT Work on September 10, 2026, aimed at people who have a business question and no patience for a ticketed report. The agent connects to approved company data, investigates what changed, and builds interactive dashboards that teammates can edit, share, and refresh. OpenAI says users can keep the analysis in one conversation instead of writing queries or opening a separate analytics tool.

The first connector list includes Amazon Redshift, Datadog, Google BigQuery, ClickHouse, Databricks, MongoDB, and Snowflake, plus files from Google Drive and SharePoint. Metric definitions and joins come from semantic layers the company already trusts, including Databricks Genie Ontology, dbt, GitHub, Snowflake Horizon, and existing BI dashboards. Administrators choose which connections and roles are live. Queries keep the connected account’s table, row, and column restrictions.

The agent can also build and edit dashboards in Omni, Oracle BI, Power BI, Sigma, Tableau, and ThoughtSpot. After a finding, ChatGPT Work can recommend next steps, name who should be involved, and send the result through Slack or email once a person approves the action. OpenAI said it already uses the same stack internally: nearly all of its product team and more than two-thirds of its go-to-market staff run data agents against company data.

Warehouse and BI vendors lined up quotes for the launch. AWS, ClickHouse, Databricks, Snowflake, MongoDB, Redis, Tableau, Microsoft Fabric, Sigma, and ThoughtSpot all described the agent as a front door onto data they already govern. Snowflake’s Umesh Unnikrishnan said employees can use ChatGPT Work while OpenAI models also sit inside Snowflake CoCo and CoWork, so the same definitions travel with the user.

Decoded Take

Text-to-SQL has been promised for a decade. What is new here is OpenAI putting the query, the semantic layer, and the BI publish step inside ChatGPT Work, which is already where a lot of enterprise seats live. That is convenient, and it is also a distribution play against Databricks Genie and Snowflake Cortex as the place a non-analyst asks a question. The constraint that matters is permission inheritance. If row filters and Horizon or Genie ontologies actually hold when the agent writes a dashboard, this becomes default infrastructure. If they leak, it becomes a shadow BI tool that security teams shut off. Watch which warehouse shows up as a named, governed production deployment, not another launch quote.