On September 10, 2026, OpenAI added a new plugin called Data agent to ChatGPT Work, its offering for businesses. It connects to a company's data platforms so employees can investigate metric changes and assemble interactive dashboards simply by asking questions in a chat. No SQL, and no need to learn a dedicated BI tool.
Built to Remove the Wait for a Report
Why sales slowed down, which department is driving up spending, which large accounts are at risk of churning. The data usually holds the answers, yet in practice the question goes to an analyst and then waits in a queue. Data agent is aimed squarely at that wait.
Using it means addressing @Data in a ChatGPT Work conversation and asking. You can layer on follow-up questions and inspect the evidence behind each finding as you refine the analysis. Directing and adjusting the work inside a single conversation is what separates it from conventional dashboard workflows.
Connections Span Major Data Platforms and Documents
OpenAI lists Amazon Redshift, Google BigQuery, Snowflake, Databricks, MongoDB, ClickHouse, and Datadog among the data sources it connects to. Files and documents stored in Google Drive or SharePoint can also be pulled into an analysis.
What makes the design interesting is that it reads meaning, not just numbers. Business vocabulary, metric definitions, custom formulas, and the relationships between tables are picked up from sources such as dbt, Snowflake Horizon, Databricks Genie Ontology, GitHub, and existing BI dashboards. Because the agent knows what revenue or churn rate means inside that specific organization, it is less likely to produce figures that disagree from one team to the next.
Dashboards Can Live Inside Existing BI Tools
The dashboards it produces come with built-in visualizations, and teams can edit, share, and refresh them. Hand it your brand guidelines and it will match the output to your organization's look and feel.
Data agent can also create and operate dashboards inside Omni, Oracle BI, Power BI, Sigma, Tableau, and ThoughtSpot. Rather than forcing a wholesale migration of the analytics stack, it acts as a natural-language front door to tools teams already run. After the analysis, it suggests next steps and the people who should be involved, shares findings through Slack or email, and carries out approved actions through connected tools.
Permissions Follow the Connected Account
Handling internal data raises the question of access control. According to OpenAI, every query the Data agent runs executes under the connected account's existing permissions, including table-level, row-level, and column-level restrictions. Data that account cannot see stays invisible to the AI as well.
Administrators decide from workspace settings which data connections are available and which roles may use them. Whether to enable the data-source plugins for Databricks, Snowflake, and others is also an administrator's call.
How OpenAI and Early Adopters Use It
OpenAI says it relies on the capabilities behind Data agent broadly in-house. Nearly everyone on the product team, and more than two-thirds of the go-to-market organization, now runs their own analyses through data agents in ChatGPT Work. The groundwork came from the company's data team, which built shared business definitions, set access rules, and put safeguards around sensitive data.
The alpha program includes NTT DATA, Thermo Fisher Scientific, and ServiceTitan, among others. The head of NTT DATA's global AI office says licensing costs, effort, and required expertise had made it hard to widen dashboard usage, but that non-engineers in sales and corporate functions can now build and update dashboards themselves in plain language. ServiceTitan used a dashboard to establish that users of its AI feature launched campaigns at roughly 3 times the rate of non-users, and fed that finding into a redesign of onboarding.
Summary
Data agent is less a way to hand analysis over to generative AI than a way to widen the circle of people who can ask questions, built on the definitions and permission models a company has already invested in. The flip side is that organizations with fuzzy metric definitions will struggle to get the accuracy they expect. Since administrators control where it can be deployed, the practical path is to start with well-defined domains and learn how much can safely be delegated from there.
