SAN FRANCISCO — OpenAI on September 10 announced a new Data agent inside ChatGPT Work that lets employees query authorized company data sources in plain language, investigate why a metric moved, and generate shareable interactive dashboards — no SQL, and no waiting in an analyst's queue.
It is the most significant capability added to ChatGPT Work since the product launched on July 9, and it moves OpenAI squarely into an enterprise data-analytics market long occupied by Snowflake, Databricks and Microsoft's Power BI. The people most directly affected work the report backlog: sales, marketing and management teams that used to file a request with the data team and wait can now do the asking themselves.
[1]What it connects to, and what it does
According to OpenAI's blog post, the Data agent's first wave of supported data platforms includes Amazon Redshift, Datadog, Google BigQuery, ClickHouse, Databricks, MongoDB and Snowflake, and it can pull files from Google Drive and SharePoint into an analysis. Once the analysis is done, it can build or operate dashboards in Omni, Oracle BI, Power BI, Sigma, Tableau and ThoughtSpot.
OpenAI is emphatic that the tool reads a company's existing business terms, metric definitions and data relationships — context drawn from semantic layers such as Databricks Genie Ontology, dbt, GitHub and Snowflake Horizon. In other words, it is trying to understand not just what is in the tables, but how the company defines "revenue" and "customer."
On permissions, OpenAI says administrators decide which data connections are available to which roles, and every query enforces the connected account's existing table-, row- and column-level restrictions. The agent can also suggest who to contact next and, with user approval, take follow-up actions through connected tools.
[1]A year of eating its own cooking
The product was not built from scratch. VentureBeat reports (as relayed by Chinese financial outlet Wallstreetcn) that OpenAI's internal version of the data agent has served more than 3,500 employees across roughly 70,000 datasets and over 600 petabytes of data. Arpan Shah, OpenAI's head of enterprise technology, told VentureBeat that about a year ago only a small fraction of the company's data questions could be resolved end to end without human help; OpenAI now says nearly all product-team members and more than two-thirds of its go-to-market staff use the tool.
Those figures are OpenAI's own and have not been independently verified. But they establish at least one thing: this is a workflow the company productized after heavy internal use, not a lab demo.
[2]The number that wasn't published
The launch leaves a conspicuous blank: OpenAI released no public accuracy or retrieval benchmark for outside customers. According to VentureBeat, the company did run internal comparisons, including against its existing data tooling, but it did not publish the results.
That does not mean the product is inaccurate. It means enterprise buyers currently have no public, comparable metric for judging how often the agent is right inside a messy corporate data environment — and corporate data is the hardest environment there is: answers are scattered across tables, dashboards and documents, and different teams often define the same metric differently.
Competitors, meanwhile, are choosing to show their work. Databricks recently published research on its Adaptive Instructed-Retriever claiming answer quality on certain enterprise retrieval tasks comparable to Claude Sonnet 5, GPT-5.6 Luna and DeepSeek-V4-Flash, with an average response time of about 5.8 seconds. Those results come from Databricks' own testing and are likewise unverified.
[2]Our customers have built a trusted foundation for enterprise data and context in Snowflake. With the Data agent in ChatGPT Work, employees can tap into that data and access controls to investigate business questions—while OpenAI models simultaneously bring intelligence to experiences like Snowflake CoCo and CoWork. Teams share a common, governed understanding of their business wherever they choose to work.
OpenAI is not alone on the board
ChatGPT Work launched on July 9 alongside GPT-5.6, pitched as an agent for long, multi-step office tasks. The Data agent pushes it further — from drafting documents for you to querying data, running analysis and taking action on your behalf.
The partner list itself is worth a beat. AWS, ClickHouse, Databricks, Snowflake, MongoDB, G2 and Tableau all appear in the announcement as partners. The data-warehouse vendors plainly do not see OpenAI as a replacement — for now, OpenAI is building the front door, not the foundation.
For enterprise buyers, the launch settles the question of whether everyone can ask the data. What remains open is whether everyone can trust the answer — and until OpenAI publishes a reproducible public benchmark, each company will have to answer that with its own pilot.
[1]