AIViewer editorial system · September 24, 2026
OpenAI’s Data agent, announced September 10, brings company-data analysis and interactive dashboards into ChatGPT Work. OpenAI lists connections including BigQuery, Snowflake, Databricks, Redshift, ClickHouse and MongoDB, with documents from Drive and SharePoint also available through connections. It describes using business definitions and data relationships to interpret results. Official announcement
The practical opportunity is to ask a business question and inspect the supporting analysis in one conversation. The practical challenge is deciding what the question means. “What were our sales last week?” can produce several defensible numbers if nobody specifies the dates, refund treatment or duplicate-handling rule.
This article explains the product’s documented scope, then gives you a small dataset with an answer you can calculate yourself. The exercise works as an evaluation task for any suitable assistant; it does not require access to a company database.
What Data connects and produces
OpenAI says Data can create interactive dashboards and work with BI tools including Power BI, Tableau, Sigma and ThoughtSpot. Administrators control available connections and roles; queries respect the connected account’s table, row and column permissions. Documented capabilities and controls
A semantic layer is a shared description of business meaning: which field represents revenue, which records count as customers, and how datasets relate. In practice, a team might define whether sales include tax, when a cancellation takes effect, and whether a refund belongs to its processing date or the original sale period. Writing those choices down makes disagreement easier to diagnose.
A dashboard can be visually consistent while using the wrong definition. The first useful output is therefore a short statement of the metric, time window, source and exclusions. Build the chart after that statement is clear.
Availability and setup
The current Help Center calls the product the Data plugin in ChatGPT Work and Codex. Data and the required source plugins/apps must be available to your account or workspace; installation alone does not grant database access. In a managed workspace, an administrator may need to enable connections or configure app templates. To begin, find Data in Plugins, complete the relevant setup, then address a question to @Data. Data setup guidance
OpenAI’s Work documentation specifies eligible paid plans for web and mobile; desktop access also depends on plan and workspace settings. These sources do not establish a universal Data entitlement for every plan or every free account. Check the actual plugin listing and workspace policy before planning a rollout. Work availability
A fictional sales exercise
Download the synthetic CSV. It contains seven rows and no real customer information. Amounts are Canadian dollars, before tax. Dates are already normalized calendar dates in one business timezone; there are no timestamp conversions to perform in this exercise.
| Event | Order | Date in September 2026 | Type | Amount CAD |
|---|---|---|---|---|
| E1 | O1 | 1 | Sale | 100 |
| E2 | O2 | 2 | Sale | 200 |
| E2 | O2 | 2 | Sale | 200 |
| E3 | O2 | 3 | Refund | 50 |
| E4 | O3 | 7 | Sale | 150 |
| E5 | O4 | 8 | Sale | 300 |
| E6 | O0 | 4 | Refund | 20 |
Here is the metric contract for the exercise:
- Include events dated September 1 through September 7, inclusive.
- Count each event ID once. The repeated E2 row is an exact ingestion duplicate.
- Refund amounts are positive in the file but must be subtracted.
- Subtract all refunds processed during the period, including E6 for an older order.
- Count distinct orders with a sale event in the period separately from net sales.
These are authored rules for this example, not an accounting standard or a claim about Data’s defaults. A different business definition would require a different calculation.
A useful request would be: “Analyze this fictional CSV using the metric contract above. Show included and excluded event IDs, gross sales, refunds, net sales and distinct sale-order count. Explain your calculation before proposing a dashboard. Do not publish or share anything.”
The calculated answer key
| Check | Expected result | Why |
|---|---|---|
| Raw rows | 7 | Includes the duplicate and the out-of-period sale. |
| Unique events | 6 | Keep E2 once. |
| Included unique events | 5 | Exclude E5, dated September 8. |
| Gross sales | CAD 450 | E1 + E2 + E4 = 100 + 200 + 150. |
| Refunds | CAD 70 | E3 + E6 = 50 + 20. |
| Net sales under this definition | CAD 380 | 450 minus 70. |
| Distinct sale orders | 3 | O1, O2 and O3. |
The statement you want is precise: “For September 1–7, after removing one exact duplicate event, gross sales are CAD 450 and period refunds are CAD 70, leaving CAD 380 net sales under the supplied definition.”
Now inspect the denominator. Gross sales per sale order are CAD 450 divided by 3, or CAD 150. Net sales divided by those same three orders are approximately CAD 126.67. The second ratio includes a refund for O0, whose original sale is outside the supplied period. Calling it simply “average order value” would hide that mismatch. Label the ratio explicitly or obtain the information needed for the team’s preferred definition.
Use wrong answers to locate the problem
The following are deliberately constructed errors, not outputs observed from the Data agent. With all other rules applied correctly, forgetting to remove duplicate E2 produces CAD 580 net sales. Including September 8’s E5 produces CAD 680. Dropping the older-order refund E6 produces CAD 400. Those differences point to specific checks rather than a vague request to “be more accurate.”
A production source may contain conflicting versions of an event, not exact duplicates. In that situation, do not arbitrarily keep the first row. Ask which timestamp or source is authoritative, preserve the conflict, and recalculate after the rule is settled.
For a second exercise, extend the window through September 8. The correct results become CAD 750 gross sales, CAD 70 refunds, CAD 680 net sales and four distinct sale orders. Changing a date filter should update the metric label, totals and denominator together.
What a checkable dashboard should show
For this example, a proposed dashboard would show three separate cards: gross sales CAD 450, refunds CAD 70 and net sales CAD 380. A subtitle would state the inclusive date window and currency. An evidence table would retain the five included unique events, while a short note would explain the duplicate and excluded September 8 sale.
This is a proposed design, not a Data screenshot. Its value is that another person can reproduce the result. A refresh should also identify when the underlying data was read and whether the definition changed. When a figure differs from an existing report, compare the inputs and rules before debating the visual presentation. Our guide to verifying an AI answer develops that source-checking habit.
Sharing changes the audience
OpenAI’s Data guidance says that analysis data is copied into a published Site. It also distinguishes source access from plugin installation and says connected actions depend on permissions and approval requirements. Query permissions therefore should not be treated as proof that every intended dashboard recipient may see the exported analysis. Publishing and access guidance
For a real team, inspect the actual dashboard contents and recipient list before sharing. Consider whether detailed rows are necessary or whether aggregate figures are enough for the intended audience. A person authorized to read a database may be producing an artifact for a different group.
Our Muse WhatsApp article describes the appeal of asking for several formats in one conversation. Data applies a related conversational approach to business analysis. In either case, the useful result is one whose numbers, sources and intended audience can be checked. Start with the small CSV above, establish the correct answer, and carry that standard into larger datasets.