OpenAI shipped the data agent for ChatGPT Work on 10 September. It is a plugin, admins install it from Workspace settings, and from there people query company data in plain sentences and get interactive dashboards back.
It connects to Amazon Redshift, Google BigQuery, Snowflake, Databricks, MongoDB, ClickHouse and others. On the output side it works with Tableau, Power BI, Sigma and ThoughtSpot, so the BI tools you already pay for stay in the picture.
The sentence that matters comes later
Two things separate this from every “chat with your database” demo of the last two years.
First: the agent picks up business definitions and semantic layers from dbt, Databricks Genie and Snowflake Horizon. If “active customer” has a specific meaning at your company, the agent uses it - instead of the meaning it would guess from column names. That is exactly where these tools usually fall over: they return a plausible-looking number that does not match the one in the quarterly report.
Second: existing row-, column- and table-level permissions stay in force. That reads like fine print and it is the condition under which a tool like this gets into a company at all. An agent that can see everything is not a product in most organisations, it is a data protection incident with a chat interface.
The number from their own building
OpenAI says nearly everyone on its product teams and two-thirds of its go-to-market staff use data agents in ChatGPT Work internally. Alpha customers named: NTT Data, Thermo Fisher Scientific, ServiceTitan and Zipline.
Self-reported numbers are marketing, obviously. Still, the second one is the interesting half: two-thirds of sales and marketing means people without SQL are using it. That is who it was built for.
What it means if you use Claude
The comparison is obvious and a little uncomfortable. Everything this product hard-wires - database connectors, semantic layers, permission inheritance - is something you assemble yourself with Claude through MCP servers. More flexible, not owned by one vendor. Also work, and somebody has to do it and keep doing it.
For a small team that is fine, even preferable: an MCP server for your own database is an afternoon. For a group with forty data sources and a permission structure grown over a decade, the maths looks different. There, “the admin installs it and permissions stay as they are” is a very strong sales line.
I would still watch the semantic layer rather than the agent. If dbt or Horizon is well maintained, you can attach any of these tools to it. If it is not, every vendor will hand you the same pretty dashboards with the wrong numbers.