Now everyone can put data to work

OpenAI launched a Data agent in ChatGPT Work that connects to approved company sources, investigates what changed, and turns findings into interactive dashboards people can share and refresh. The pitch is plain-language analysis for people who don’t write SQL.

Connectors span warehouses and ops stores — Amazon Redshift, BigQuery, Databricks, Snowflake, ClickHouse, MongoDB, Datadog, and more — plus Drive and SharePoint files. Semantic context can come from layers like Databricks Genie Ontology, dbt, Snowflake Horizon, and existing BI dashboards.

Admins pick which connections and roles get access; queries inherit the connected account’s table, row, and column permissions. That governance story is the difference between a demo and something enterprises will actually turn on.

We’ve been waiting for the moment when “ask your data” stops being a BI vendor slogan and becomes a default workplace habit. This is OpenAI’s bid. If the agent respects permissions and metric definitions, it can pull analysis out of specialist bottlenecks — sales, ops, and finance asking follow-ups in one thread instead of filing tickets.

OpenAI says nearly all of its product team and over two-thirds of its GTM org already use data agents this way internally, after their data team built shared definitions and safeguards. Alpha customers like NTT DATA, Thermo Fisher, and ServiceTitan report non-engineers building dashboards and catching reporting errors. We’re inclined to believe the useful case: governed natural-language analytics on trusted semantics is a real productivity unlock.

The downside is wrong answers at the speed of chat. A polished dashboard that silently misreads a join or a metric definition can ship a bad decision farther and faster than a cautious analyst. Permission inheritance helps; it doesn’t fix ambiguous business logic. And once Slack and email actions sit on top of those findings, a mistaken insight can become an automated ping to the wrong people.

I’m watching whether enterprises treat this as a second BI surface with audit trails — or as a shadow analytics layer that bypasses the data team until something breaks. The useful path is clear: semantic layers first, agent second. Skip that order and this becomes confident fiction in chart form.

The agent can also build and work inside Omni, Oracle BI, Power BI, Sigma, Tableau, and ThoughtSpot. It’s listed under Data in ChatGPT Work’s Plugins directory; admins install and configure source plugins from workspace settings.