OpenAI's new Data agent inside ChatGPT Work connects to a company's existing data stack, letting business users ask questions and build dashboards without filing a ticket with the data team.
A revenue operations lead who notices renewals slipping in one segment usually has one move available: file a request with the data team, wait for a slot in the queue, then wait again once the first chart raises a second question. OpenAI is trying to close that loop with a new Data agent built into ChatGPT Work, announced September 10. It connects directly to a company's existing data stack and lets someone ask the question in plain language, then get an answer, a dashboard, or both, without opening a ticket.
What changed
The Data agent plugs into a company's approved data sources, including Amazon Redshift, Google BigQuery, Snowflake, Databricks, MongoDB, ClickHouse, and Datadog, plus files pulled from Google Drive and SharePoint. It also reads the semantic layer many companies already maintain, things like Databricks Genie Ontology and dbt models, so an answer uses the company's own metric definitions instead of guessing at what counts as an "active customer" or an "at-risk account." A person can ask a follow-up question in the same conversation, review the evidence behind a finding, and turn the result into an interactive dashboard a team can keep editing, sharing, and refreshing. The agent can also work inside BI tools people already use, including Tableau, Power BI, Sigma, ThoughtSpot, Omni, and Oracle BI, updating views there instead of only inside ChatGPT.
Enterprise administrators still decide which data connections exist and which roles can use them, and OpenAI says queries respect a connected account's existing table, row, and column permissions rather than opening up new access. That detail matters: the pitch isn't that ChatGPT can suddenly see everything, it's that someone can ask questions of the data they were already allowed to see, without waiting on a person to run the query for them.
The examples OpenAI included speak more to the workflow shift than to the underlying model. NTT DATA said licensing costs and technical expertise had limited how many people across the company could build a dashboard, and that the tool let non-engineers in sales and corporate roles build and update their own. CookUnity's growth team said it used the agent to build and check a seasonal conversion dashboard against internal reports, and that doing so cut the time needed to plan acquisition spending. A strategic project lead at micro1 said the team rebuilt its performance tracking dashboards in half an hour and caught errors in the original along the way. Zipline's co-founder said the agent surfaced findings that would otherwise have taken one of the company's best people hours of manual digging to uncover. OpenAI describes these as customers in its alpha program, not a general release.
Why it matters
For a marketing, sales, or operations leader, the promise here isn't a new model capability, it's a shorter path between having a question and acting on the answer. Most business questions aren't hard to state. They're hard to route: find the right person, describe what's actually needed, wait for the report, then discover the report answers a slightly different question than the one that was asked. A tool that plugs into data a company already trusts, and respects the access rules already in place, removes a step in that routing rather than replacing the judgment at the end of it. That's a smaller claim than "AI replaces analysts," and it's the more believable one.
The honest caveat
Everything here comes from OpenAI's own announcement and the partners and customers it chose to quote, which makes it a best-case tour, not an independent evaluation. OpenAI doesn't say how many Data agent users hit friction, misread a metric, or got an answer nobody trusted enough to act on. The tool also depends entirely on a company having already done the less visible work: clean data connections, agreed-upon metric definitions, and access rules that are actually correct, which is exactly the kind of groundwork a lot of businesses haven't finished. A dashboard that ChatGPT builds still needs someone who understands the business well enough to notice when an answer looks wrong.
Closing observation
The interesting part of this release isn't that a chatbot can now query a database. It's that the bottleneck OpenAI is targeting was never really about the query, it was about who got to ask and how long they had to wait for someone else to ask it on their behalf. Removing that wait doesn't remove the need for someone who still knows which question was worth asking in the first place.