On September 10, 2026, OpenAI announced "Data agent," a tool for ChatGPT Work that analyzes company data through conversation. It connects to internal data platforms and documents, and supports everything from answering questions to building interactive dashboards and carrying out approved actions. The product aims to change the way analysis works: instead of requesting a report and waiting, the person doing the job keeps asking questions and moves forward on their own. But the figures that conversation draws on, including what they mean and who can access them, depend on how the company has configured things. Deciding what to hand to the agent and which judgments people keep for themselves will determine how far adoption can spread.
From a question to a dashboard and the next action
Data agent connects to approved data sources such as Amazon Redshift, Google BigQuery, and Snowflake. It can also pull in documents stored in Google Drive and SharePoint, so it can refer to tables of figures alongside materials that explain the business. OpenAI cites uses such as investigating a slowdown in sales or a rise in spending, and looking for problems that keep customers from renewing their contracts.
Users do not have to stop at the first answer. They can narrow the scope with follow-up questions, check the basis for a result, and revise the analysis within the same conversation. For example, after asking how sales changed, a user might compare results by region or customer segment. OpenAI's explanation of its data work offering likewise lists comparisons across segments and identifying possible causes as use cases. The candidate causes it surfaces do not mean a causal relationship has been established.
The analysis can be compiled into an interactive dashboard that a team can edit, share, and update. OpenAI also highlights integration with the BI tools companies use, meaning tools for analyzing and visualizing business data, and described features for creating and working with dashboards in Power BI, Tableau, Sigma, and others. The data work page also lists reports and presentations among the deliverables.
The aim of extending this kind of use to non-technical staff also shows in customer feedback. According to a comment from Yuji Shono of NTT DATA Group, included in the announcement, many non-technical employees, mainly in sales and administrative departments, were able to create and update their own dashboards in natural language. This is one participating company's account. It does not show how many people use the tool across the company, or that other companies will see the same results.
Conditions for moving from analysis to action
Company administrators choose which data connections Data agent can use and which roles can use them. When the agent reads data, the permissions the connected account already holds apply. In the access management section of the announcement, OpenAI states this explicitly:
"Queries respect the connected account's existing permissions, including table, row, and column restrictions" (translated)
This condition needs to be considered separately from the approval of actions taken after analysis. OpenAI explains that ChatGPT Work can suggest next actions and relevant people, share results via Slack or email, and then carry out approved actions through connected tools. Being able to read information is not the same permission as using that information to operate external tools.
Classifying the September 10 official announcement and the data work page checked on September 11 by role (input, deliverable, and external action) shows how features and conditions correspond.
| Stage of work | What can be done in conversation | Prerequisite settings and decisions |
|---|---|---|
| Investigating data | Refer to data and documents, and narrow the analysis with follow-up questions | Administrators provide the connections. Existing permissions apply to what can be referenced, and business definitions are used |
| Shaping results | Create interactive dashboards, then edit, share, and update them | Subject to the terms of the data connections and BI tools used |
| Moving to the next action | Suggest relevant people and next steps, and operate through connected tools | Approval is required for the actions to be carried out |
What Data agent links together is analysis, deliverables, and approved actions. It is not a mechanism that automatically decides which sources to connect or who gets access. The table sorts the official descriptions by the role each plays in the work. It should be kept distinct from a table showing the internal processing order, or showing that every integration works the same way.
Furthermore, approving an action does not guarantee that the analysis is correct. If someone is contacted because of a candidate cause for declining sales, both the data supporting that candidate and the recipient and content of the message need to be checked. Even when the conversation flows as one continuous thread, two judgments remain: whether the numbers can be trusted, and what to do with them.
The data team's work that comes before natural language
Data agent interprets data using a company's own terminology and metric definitions. Its own calculation methods and relationships between datasets are also referenced. OpenAI says this information comes from sources the company trusts, such as Databricks Genie Ontology, dbt, and Snowflake Horizon.
The semantic layer that appears here refers to a mechanism for sharing the meaning of metrics, how they are calculated, and the relationships between data. For example, the answer to "this month's sales" changes depending on whether it counts the amount of orders received or the amount actually paid. This is a hypothetical illustration of differing definitions, but it shows why the same question in Japanese can return different numbers if the definitions referenced differ. The appearance of a chart makes it hard to tell the difference.
In the in-house data agent OpenAI introduced on January 29, 2026, the context supporting answer quality was also an issue. In addition to table descriptions written by people, it used information derived from the code that produces the data. Even when column names are similar, what was excluded from aggregation and what range of data is included can differ. Code and business annotations are used to close that gap.
However, it cannot be confirmed that the model, memory, and other implementation details of this internal tool are used as-is in the new external product. What is useful as a reference is that understanding natural language alone did not settle what the numbers meant, and that work was needed to supply the agent with internal knowledge.
In this announcement, OpenAI said that nearly all of its product team, and more than two-thirds of the departments handling sales, go-to-market, and similar functions, use the ChatGPT Work data agent. At the same time, it said this usage was made possible by the data team's work on shared business definitions, access rules, and protections for sensitive data. The percentage reflects usage within OpenAI itself. It is not an accuracy rate or a productivity gain.
If non-specialist departments carry out more analysis themselves, data teams may be able to spend less time answering individual questions. In exchange, more of their work shifts to maintaining which definitions are shared and how far each person is allowed to reference data. Handing out the agent does not by itself complete that preparation.
What to check before expanding use
The entry point is "Data" in the plugin list in ChatGPT Work. Administrators make it available to teams or install it from "Plugins" in workspace settings, and also enable and configure plugins for the data sources they need. Users complete the necessary account connections, then call @Data in a conversation and send their business questions.
Adding Data and giving access to the necessary data are separate tasks. Which integrations need to be checked also depends on where dashboards are built and with whom they are shared. At rollout, it is realistic to limit the tasks users can ask for, based on the tools the company uses and users' permissions.
When trying out the analysis, business questions whose answers are already known are a useful starting point. OpenAI's earlier in-house tool compared the execution results of correct, hand-written SQL against SQL generated by the agent. SQL is a language for instructing a database, and different queries can return the same answer, so this method evaluates the output data as well. It is not a guarantee of this product's accuracy, but it is a useful reference for testing with a company's own questions.
For monthly sales, for example, align the internally settled definition and the target period, then check against known aggregate results. Moving on to follow-up questions with changed comparison conditions, or to actions that share the results, makes it concrete where judgment is needed. If companies can set up the definitions and permissions the agent refers to, and build result checking and action approval into their workflows, sales and administrative staff can turn time spent waiting for answers into time spent checking their next questions themselves.