OpenAI's Data agent borrows the one thing that makes it work

OpenAI launched a Data agent inside ChatGPT Work on September 10. It connects to Amazon Redshift, Datadog, Google BigQuery, ClickHouse, Databricks, MongoDB and Snowflake, pulls files from Google Drive and SharePoint, and builds dashboards it can write back into Omni, Oracle BI, Power BI, Sigma, Tableau and ThoughtSpot. OpenAI says nearly all of its own product team and more than two-thirds of its go-to-market organisation already use data agents internally.

Chat with your database has been shipped and abandoned roughly once a year since 2016. The reason is never the SQL. It is that "revenue" means six things at any company of size, and a model that guesses which one hands you a confident, wrong number that somebody then takes into a board meeting.

Read how OpenAI got around that, in its own words: the agent "uses your organization's business terms, metric definitions, custom calculations, and data relationships," and that context "comes from semantic layers and trusted sources such as Databricks Genie Ontology, dbt, GitHub, Snowflake Horizon, and BI dashboards."

So the part that makes the product trustworthy is the part OpenAI did not build and does not own. Whose product is this, exactly?

Count the quotes

Twenty-two. AWS, ClickHouse, Databricks, Snowflake, MongoDB, Redis and G2 on the data side; Tableau, Power BI, Sigma and ThoughtSpot on the dashboard side; eleven customers from NTT DATA to a firm called Doeren Mayhew. Four business intelligence vendors supplied warm words for a product whose entire pitch is that an employee never has to open a business intelligence tool (Sigma's SVP of product is quoted calling ChatGPT Work "the fastest place to ask a question").

Snowflake's quote hedges inside its own sentence, praising the Data agent "while OpenAI models simultaneously bring intelligence to experiences like Snowflake CoCo and CoWork." That is a partner reminding you it ships a competing surface.

Two absences. Looker is not mentioned anywhere on the page, though Google's warehouse is a connector, so OpenAI has BigQuery without Google's semantics. And there is no price. An enterprise launch with twenty-two testimonials and no number is probably a launch aimed at procurement later.

Our read

This is a land grab on someone else's land. It works for as long as Databricks and Snowflake leave the ontology endpoints open, and both of them sell an agent of their own. We would expect at least one of the seven named data platforms to meter, price or restrict the ChatGPT Work connector within twelve months, and we would expect OpenAI to build or buy a semantic layer rather than keep renting one. Databricks and Snowflake deepening the integration instead, with no metering, would say we have the power balance backwards.

The strongest argument against us has nothing to do with semantics. It is this line, which is the thing every previous version of this product got wrong:

Enterprise administrators choose which data connections are available and which roles can use them. Queries enforce the connected account's existing permissions, including table, row, and column restrictions.

Row and column level enforcement through the connected account is why a security team says yes (it is also the sentence your CISO will read, and the only one). If that holds in practice, adoption will run ahead of the ownership question for a year or two, and the ontology fight happens at renewal.

One number to watch, and it is OpenAI's own: two-thirds of its go-to-market organisation. What share of your own commercial staff writes their own queries today, honestly? If the honest answer is nearer a tenth, then this is either a real step change or a very good demo, and the difference will show up in how many of those eleven customers are still being quoted a year from now.