What is the ChatGPT Data agent?
OpenAI introduced the Data agent on 10 September 2026 as a way for people to investigate business questions in plain language without first writing SQL or learning a separate analytics interface. The user asks a question, Data gathers information from the approved sources it can access, reasons over the company’s business context, and returns an analysis that can be refined in the same conversation.
The important distinction is that Data agent is the workflow; Data is the plugin you install. OpenAI currently documents the plugin for ChatGPT Work and Codex. It is designed to work with company data rather than acting as a generic public-data chatbot.
A useful Data task is usually outcome-oriented: explain why retention fell, compare revenue by region, investigate a cost increase, identify risks in a sales pipeline, or create an executive dashboard from the resulting analysis. The same conversation can move from question to evidence, then from evidence to a dashboard or report.
Official overview: OpenAI, “Now everyone can put data to work” and Using the Data plugin in ChatGPT Work and Codex.
Who can access the Data agent?
The current documentation places Data inside ChatGPT Work and Codex. Before you can use it, the Data plugin must be available in your account or workspace. In a managed workspace, an administrator can control whether Data is available, pre-installed, restricted to certain roles or groups, and which underlying data-source plugins users may access.
Installing Data and connecting a source are separate steps. You can have the Data plugin installed while a required Snowflake, Databricks, Google Drive, SharePoint, or other source still needs authorization or workspace approval.
| Requirement | What it means |
|---|---|
| ChatGPT Work or Codex | The current Data plugin experience is documented for these work surfaces. |
| Data plugin access | Data must be available or installed for your account or workspace. |
| Source access | The relevant warehouse, BI, document, or app connection must also be enabled and authorized. |
| Underlying permissions | Your connected account still controls which tables, rows, columns, files, dashboards, and other resources Data can use. |
Access controls: OpenAI Data plugin help.
How to set up the Data agent
For an individual user, the basic setup is short. The harder part in a company environment is usually making sure the approved data sources and permissions are already configured.
- Open Plugins. In ChatGPT Desktop, open Plugins. On the web, use the Plugins directory.
- Find Data. Select the OpenAI Data plugin and choose Install plugin if it is not already installed.
- Connect the sources you need. Complete any required authorization or organization setup for the warehouse, BI tool, file system, or app that contains the data.
- Start with
@Data. Open a new Work or Codex conversation and invoke Data directly. - Ask a concrete business question. Include the source, metric, period, comparison, and desired output when you know them.
For administrators, OpenAI directs workspace setup through Workspace settings → Plugins → Data. From there, admins can review installation policy, enable the required source plugins, and manage availability for specific roles or groups.
Setup steps: OpenAI’s Data plugin guide.
Connect the data and the business context behind it
Data can work with approved sources including Amazon Redshift, ClickHouse, Databricks, Google BigQuery, MongoDB, and Snowflake. OpenAI also lists Datadog in its launch announcement, and the broader data ecosystem includes BI, analytics, and enrichment tools. Files and documents can be brought into an analysis from services such as Google Drive and SharePoint when those connections are available.
Raw access is only part of the setup. OpenAI strongly recommends giving Data a semantic layer: authoritative metric definitions, custom calculations, known relationships between datasets, and trusted queries or dashboards. This helps prevent a common analytics problem in which two teams use the same word, such as “active customer” or “net revenue,” but calculate it differently.
Semantic context can come from sources such as Databricks Genie Ontology, dbt, GitHub documentation, Snowflake Horizon, and trusted BI dashboards. If your organization has one preferred definition, tell Data which source should control the analysis.
BeeVaults recommendation: before asking Data to explain a metric change, identify the authoritative definition of that metric. A sophisticated analysis built on the wrong definition is still the wrong analysis.
Connected data and semantic context: OpenAI launch announcement and Data plugin documentation.
How to ask Data a useful first question
OpenAI recommends including the source, metric, time period, and comparison when you know them. That makes the first pass easier to audit and reduces the chance that Data silently chooses a definition or timeframe you did not intend.
@Data Compare net revenue this quarter with last quarter using our approved finance definition. Break the change down by region and product. Identify the three largest drivers, show the source and filters used for each conclusion, flag missing or conflicting data, and finish with the next checks you recommend.
After the first answer, continue in the same conversation. Ask Data to isolate one segment, compare a different period, test an alternative explanation, or show the evidence behind a specific chart. OpenAI specifically encourages users to refine, validate, and publish analyses through follow-up questions rather than treating the first response as final.
Before relying on a result, review the source, time period, filters, and metric definition. If the answer disagrees with an existing report, ask Data to compare those details explicitly instead of assuming one side is automatically correct.
Prompting and validation guidance: Using the Data plugin in ChatGPT Work and Codex.
How to turn an analysis into an interactive dashboard
Once the analysis is useful, you can ask Data to turn it into an interactive dashboard. Describe the metrics, breakdowns, comparisons, filters, and visual hierarchy you want. You can also provide brand guidelines so the result fits your organization’s visual system.
- Start from a validated analysis. Make sure the important metrics and definitions are correct before you design the dashboard around them.
- Specify the audience. A leadership dashboard should emphasize decisions and exceptions; an operational dashboard may need more filters and detail.
- Name the required views. Ask for summary KPIs, trends, segment breakdowns, comparisons, or diagnostic tables explicitly.
- Ask for interactions. Request filters or drill-downs that make the dashboard useful rather than merely decorative.
- Refine the layout conversationally. Ask Data to change chart types, reorder sections, simplify labels, or apply your brand guidance.
- Interrogate the dashboard. Ask which information supports a chart or why a metric moved. The visual should remain connected to the underlying analysis.
@Data Turn this analysis into an interactive weekly revenue dashboard for leadership. Show total revenue, quarter-over-quarter change, target attainment, regional contribution, top positive and negative drivers, and a 12-week trend. Add filters for region and product. Keep the first screen executive-level, then place diagnostic detail below it.
OpenAI says Data dashboards can be edited, shared, and refreshed as needed. The dashboard is therefore not limited to a one-time picture of the analysis.
Dashboard workflow: OpenAI Data agent announcement and dashboard instructions.
Data can also work with the BI tools your company already uses
The Data agent is not designed to force every organization to abandon its existing business-intelligence stack. OpenAI says Data can work with connected tools including Omni, Oracle BI, Power BI, Sigma, Tableau, and ThoughtSpot.
That can be useful when the business already has governed dashboards, familiar charting standards, or approval processes in one of those systems. Data can use plain-language instructions to help build or interact with dashboards in the connected tool, subject to the actions that tool exposes and the permissions attached to your account.
The exact action set is connection-specific. A tool being connected does not guarantee that every edit, publish, refresh, or write operation is available. Check what the plugin supports before designing a workflow that assumes a particular action.
Supported BI examples: OpenAI Data plugin guide and Data intelligence for every team.
How to publish, share and keep a dashboard current
OpenAI recommends ChatGPT Sites when you want to publish and share a locally generated dashboard or report with people in your workspace. From the generated artifact, use Publish, provide the requested sharing information, and choose the intended audience.
There is an important permission detail: OpenAI says the data used in the analysis is copied into the published site. That means sharing the site is not merely sharing a live pointer back to the original warehouse. Review the audience and the included information before publishing.
Once a dashboard is published, it can be kept current with a cloud automation. Data can help set up that recurring refresh workflow. This is useful for weekly operating dashboards, monthly executive reports, pipeline views, or recurring financial analyses that should update on a schedule.
Publishing and refresh guidance: OpenAI Data plugin help.
The Data agent does not bypass your existing data permissions
OpenAI says Data queries use the connected account’s existing permissions, including applicable table, row, and column restrictions. Connecting a provider does not create new permission to see information the account could not already access in that provider.
This has two practical consequences. First, two employees can ask the same question and receive different accessible evidence if their source permissions differ. Second, a successful connection is not proof that Data can see the specific dataset, report, or column needed for the task.
Workspace administrators control which plugins and connections are available, while actions in external tools can depend on both plugin capabilities and approval requirements. For consequential actions such as sending results, changing records, or publishing sensitive information, review the destination and content before approving the action.
OpenAI also recommends using trusted business definitions and governed sources so that the system has both the permission and the context needed to interpret data correctly.
Permissions and actions: OpenAI Data plugin access guidance.
What to check if Data or a dashboard is not working
- Data is missing
- Confirm you are in the intended Work or Codex account and workspace. In a managed workspace, ask an administrator whether the Data plugin is available or pre-installed for your role.
- The source is not available
- Data and the source plugin have separate setup. Check whether the underlying Snowflake, Databricks, BigQuery, Drive, SharePoint, or other connection has been enabled and authorized.
- The connection works but data is missing
- Verify that your account can see the information directly in the provider. Then name the dataset, report, table, metric, or file more specifically in your request.
- The answer disagrees with an existing report
- Compare the source, date range, filters, joins, and metric definition. A semantic mismatch is often more useful to investigate than simply asking Data to try again.
- A dashboard action fails
- Check whether the connected BI or publishing tool supports the requested action, whether your account can perform it, and whether ChatGPT is waiting for an approval.
The most reliable recovery pattern is to make the source and definition explicit, narrow the question, then ask Data to show the evidence it used. That turns troubleshooting into an auditable analysis rather than a sequence of blind retries.
Troubleshooting: OpenAI Data plugin guide.