ChatGPT Work's Data agent is a data-analysis agent that OpenAI released on September 10, 2026. It can investigate metric changes across workspace-approved sources and turn the results into editable, refreshable interactive dashboards. With user approval, it can also send follow-up actions through Slack or email.
Prompting is not the hard part of adoption. Before relying on an answer, the team must be able to explain how each metric is calculated, who owns the data, and which tables, rows, and columns each account may read. If any of those answers is unclear, do not use the Data agent's numbers for a business decision.
If Data already appears in the workspace, start with one low-risk metric. Have an admin check plugin and source access, let the analyst connect their own account, then ask a scoped question with @Data and compare the answer with an existing report. If Data is missing from Plugins, ask an admin to follow the official setup instructions and check availability and role settings under Workspace settings > Plugins.
Confirm three prerequisites before connecting Data
Tie each prerequisite to a named owner or a setting that can be tested. If any item still exists only as an informal agreement, fix it at the source or metric-definition layer first.
| Prerequisite | What to confirm | If it is missing |
|---|---|---|
| Data owner | Who approves the source, explains metrics, and handles errors | Assign an owner for each data domain |
| Access control | A test account can see only approved tables, rows, and columns | Tighten permissions in the source system before connecting it |
| Metric definition | Each core metric has a name, formula, time zone, and deduplication rule | Fix the definition in a semantic layer or trusted document first |
As of September 11, 2026, the Data agent launch announcement lists Amazon Redshift, Datadog, Google BigQuery, ClickHouse, Databricks, MongoDB, and Snowflake as data sources. Files in Google Drive and SharePoint can also be included in an analysis.
The agent can use dbt, GitHub, Snowflake Horizon, Databricks Genie Ontology, and existing BI dashboards for business terms, calculations, and data relationships. That is the announced connection range, not confirmation that every connector is enabled in a particular workspace.
Install Data and constrain access
In a managed workspace, permission to open Data and permission to read a source are separate. The admin makes the plugin available; the source system still determines what each user's account may read. Based on the official Data agent setup instructions, the setup order is:
- Open the Plugins tab in ChatGPT, search for Data, open its details, and select Install plugin. If the account cannot install it, hand the task to a workspace admin.
- In Workspace settings > Plugins, review Data's availability and confirm that the intended roles can use its required plugins or apps.
- Make the required source connection available to those roles, such as Databricks or Snowflake. Enabling a plugin does not grant source-system access.
- Have each user authorize their own source account rather than sharing an admin connection.
- Validate with a restricted test account instead of using an admin account as a substitute for a normal user.
The Data agent launch announcement says queries respect the connected account's database, table, row, and column permissions. The OpenAI workspace-permissions guide draws the same boundary: the workspace controls plugin availability, while the source system controls which data the signed-in account can read.
Do not stop testing once the connection succeeds. Prepare two questions: one that should return data and one that must be rejected. If the second question returns records, aggregates, or enough information to infer restricted data, pause the pilot and inspect source permissions, the semantic layer, and connector configuration.
Start with one narrow analysis question
Choose a metric that already appears in an official report for the first task. Avoid starting with “analyze the whole company.” State the metric, time range, comparison window, breakdown dimensions, and evidence requirements. OpenAI Docs' KPI root-cause analysis workflow likewise fixes the metric definition, comparison window, and source-of-truth data before separating confirmed drivers from hypotheses.
For example, a weekly-active-user investigation could begin with this prompt. Replace each bracketed item with a definition the team has already approved:
@Data analyze how [weekly active users] changed during [August 24–30, 2026]
compared with the average of the previous four weeks.
Requirements:
1. Use the approved [product analytics] sources and metric definitions in this workspace.
2. Break down the change by [platform, region, and new versus returning users], and identify the largest contributors to the total movement.
3. For every conclusion, include the source, query period, filters, and calculation method.
4. Separate facts, hypotheses, and next checks. Write "cannot confirm" when the evidence is insufficient.
5. Return the analysis table first. Do not send messages, create tickets, or take any other external action.
Once the analysis table returns, validate the largest anomalies first:
@Data validate the two largest contributors. Check for data latency, tracking changes,
missing values, and duplicate records. Show the values before and after validation side
by side, and list explanations that still cannot be ruled out.
These prompts illustrate how to structure the task; AIGCDev did not run them.
Validate the answer and its evidence
An answer can read well and still contain the wrong number. Before publishing a conclusion, the analyst and data owner should check:
- Scope: Dates, time zone, business entities, and filters match the question.
- Definition: The formula, deduplication method, and denominator come from an approved metric definition.
- Permissions: The output contains no details or inferable data that the current account should not see.
- Evidence: Each key conclusion can be traced to a source, query condition, or existing report.
- Freshness: The data update time is sufficient for the decision and latency is not misreported as a decline.
- Anomalies: Missing values, duplicate records, tracking changes, and small samples are disclosed.
- Recalculation: At least one key number can be checked independently with existing SQL, a BI report, or manual sampling.
If Data and an existing report disagree, compare their time ranges, filters, and metric versions. Do not choose the result that better matches expectations. If the difference remains unexplained, keep both results and their query conditions, then give them to the data owner.
Lock dashboard definitions and actions before publishing
Generate a dashboard only after the key numbers have been checked. The Data agent launch announcement says dashboards can be edited, shared, and refreshed, and can work with Omni, Oracle BI, Power BI, Sigma, Tableau, and ThoughtSpot. Which tools are actually available depends on enabled plugins and the connected account's access.
Before choosing charts, decide who will read the dashboard, how often it will refresh, who handles anomalies, and which metrics have passed data-quality checks. OpenAI Docs' dashboard-planning workflow treats those details as prerequisites rather than chart options.
For a weekly product-leadership report, the request could set those boundaries this way:
@Data turn the validated analysis into a weekly dashboard for product leaders.
Keep the metric definitions, data update time, and filters visible, and show the
comparison period on every chart. Do not add unvalidated metrics. If refresh fails
or data is delayed, display a warning at the top of the dashboard.
Before publishing, test access with a recipient account. A dashboard that opens for its creator should not automatically expose the same data to every recipient. If Data will send follow-up actions to Slack, email, or another tool, approve each action separately. The ChatGPT Work guide recommends writing stop points, review steps, and approval conditions for important actions into the task.
Stop rollout when any of these checks fail
Pause expansion of the pilot if any of these checks fails:
- A restricted test account can see protected data or infer a sensitive aggregate.
- Data, the semantic layer, and the official report use different definitions for the same metric.
- A key conclusion cannot be traced to a source, time range, and filter set.
- Numbers change after refresh without a visible data version or update time.
- A user presents correlation as causation without additional evidence.
- Outbound messages or tool actions cannot preserve per-action approval and an audit trail.
FAQ
Why does Data not appear in ChatGPT Work?
Check the Plugins tab first. For a managed account, ask an admin to review availability and role settings under Workspace settings > Plugins. The official setup instructions direct admins to make Data and its related source plugins available there; each user must still complete their own account connection. If the admin confirms that Data is not available to the role yet, keep the existing SQL or BI report as the production analysis path and wait to start the pilot until workspace access is available.
Does enabling Data automatically grant access to Snowflake or other sources?
No. The admin enables the Data plugin, but the source system still applies the signed-in user's permissions. Each user should connect their own account, and the pilot should test both an allowed query and a rejected query with a restricted account.
Can the Data agent replace existing SQL or BI reports?
Do not replace them during the pilot. Keep an existing SQL query or BI report as a control and recalculate at least one key number independently. If the results disagree and the difference cannot be explained, do not publish the conclusion.
Can the Data agent send follow-up actions to Slack or email?
The OpenAI launch announcement says it can do so with user approval. In practice, confirm the recipient, content, and data permissions for each action; completing an analysis does not provide approval to send it elsewhere.