Data agent in ChatGPT Work: from a question to a dashboard

What changed on 10 September 2026?

OpenAI announced its Data agent on 10 September 2026. Available through the Data plugin in ChatGPT Work and Codex, it gives supported users a way to investigate connected business data and create dashboards from a conversation. This is more specific than asking a general chat to comment on numbers you paste.

A useful starting question is why a weekly operating metric changed. The job includes choosing the right records, applying the same definition to both periods and explaining the result. The dashboard is the final presentation, not the evidence that the analysis is correct.

Who can use it?

OpenAI’s announcement and release notes place this feature in ChatGPT Work and Codex with the Data plugin. Do not assume that an ordinary personal Free or Plus account includes it. You need the relevant workspace, supported connections and permission to use the source data.

If your workspace needs setup, an administrator can check Workspace settings, then Plugins. Install the Data plugin through the available directory and connect the authorised sources. Follow the current official setup instructions if the controls differ. A connection is permission to access particular information, not a reason to connect everything.

Worked example: why did completed jobs fall?

Consider a fictional service company. Its weekly report shows 180 completed jobs last week and 150 this week. That is 30 fewer jobs, or a fall of about 16.7%. Those figures are a worked example, not a result from a live business or a tested Data-agent run.

Before asking for causes, define completed: does it mean the technician finished, the customer signed off or the invoice was issued? Choose one definition and the same cutoff time in Australia/Sydney for both weeks. A change in reporting rules can look like a change in performance.

Investigate before building the dashboard

1. Connect only the approved dataset needed for this question. Ask which source, date fields and job-status definitions are available.
2. State the two exact periods and the reporting timezone. Confirm how cancelled, reopened and duplicate jobs should be handled.
3. Ask for the headline counts and the calculation behind the change before requesting an explanation.
4. Break the movement down by useful dimensions, such as branch, job type or technician availability. Ask which records support each finding.
5. Reconcile the totals with the source system or an independent calculation. Treat an unexplained difference as unfinished work.
6. After checking the analysis, request a small dashboard showing the headline change, supporting breakdowns and data limitations.

Try this original Data-agent brief

Replace the fictional details below with an approved scope. Do not paste a customer export just to make the example work.

@Data Investigate the change in completed jobs between [period A] and [period B] using [approved source]. Use Australia/Sydney time. First confirm the definition of completed and how duplicates, cancellations and reopened jobs are treated. Show the two counts and percentage change. Then break the change down by branch and job type, naming the source evidence. Separate observed differences from possible explanations. Do not infer cause from correlation. Once I confirm the totals, draft a dashboard with the key figures, breakdowns, date range and unresolved limitations.

What a useful answer should contain

Look for a chain you can follow: definition, selected records, calculation, breakdown and interpretation. “Branch A fell most” may be an observation. “Branch A fell because staff were absent” needs separate evidence about staffing and the timing of the missed jobs.

Check the denominator as carefully as the headline. Thirty fewer completed jobs has a different meaning if incoming jobs also fell. Ask for relevant context without letting the report drift into every available metric.

A dashboard should state its period and where the data came from. Do not assume a generated dashboard refreshes automatically or has the same sharing controls as the source system. Confirm those behaviours before using it as a recurring management report.

Bring it into a team workflow

Save the agreed metric definitions and nominate someone to check the first reports. If the same question recurs, document the inputs and review steps before automating it. Measure the effort required to reach a checked result; this article makes no speed claim.

For help turning a real reporting question into a practical team exercise, use the workshop enquiry form on the 27 AI homepage.

Common questions

Is the Data agent included in every ChatGPT account?

No such assumption is safe. OpenAI describes it in ChatGPT Work and Codex through the Data plugin. Check workspace access, supported connections and administrator settings.

Does a dashboard prove the analysis is right?

No. Reconcile the underlying definitions, records and calculations first. Good presentation can hide a wrong filter or denominator.

Will it update automatically next week?

Do not assume so. Confirm the specific dashboard’s refresh, source and sharing behaviour before adopting it as a recurring report.