OpenAI: September 2026 review

ChatGPT Data plugin: reviewing business data and dashboards

Use the Data plugin in ChatGPT Work and Codex: connect sources, define metrics, verify analyses and share dashboards with appropriate access.

WERKVERSTAND / CONNECTING INTELLIGENCE

The essential answer

Introduced on 10 September 2026, the Data plugin helps analyse connected business data in ChatGPT Work and Codex and can produce reports or dashboards. Reliable results still require precise metric definitions and reconciliation with source systems. In particular, data copied into a published Site needs a separate access review before the dashboard is shared.

Check the plugin, source and permissions separately

Installing Data does not create access to company information by itself. The plugin, the necessary data connection and any administrative template must be available in the account. Provider guidance describes data warehouses and existing BI tools as possible sources; the actual connection depends on the supported service and permissions. Begin the pilot explicitly with @Data and one approved source. Record the signed-in identity, data scope and purpose of the analysis. A team member with restricted access should receive only the information intended for that role. A successful administrator test is not sufficient evidence for the people who will eventually use the workflow.

Define the metric before asking for an analysis

A question such as why is revenue falling contains more assumptions than it appears to. Does revenue mean received payments, issued invoices or closed orders? Which period and time zone apply? Are credits, taxes and cancelled transactions included? Write these decisions into a concise metric definition and tell Data which source is authoritative. This business context often matters more than an elaborate chart. If teams use different definitions, the report must state the difference. Ask the assistant to explain filters and calculations before interpreting causes or recommending action. Otherwise a persuasive narrative can conceal a basic mismatch between the figures being compared.

Example: investigating differences in a monthly report

An illustrative service-company pilot starts with a discrepancy between accounting and the sales report. Ask Data to break the same two months down by service line and invoice status. Reconcile the total with an approved export first. Then inspect the largest differences: posting dates, duplicate records, credit notes or missing classifications. A discrepancy can remain an open question. Build the dashboard only after the business team has resolved the relevant definitions. The accountable person records which figures are confirmed and which explanation remains a hypothesis. AI then supports investigation without turning correlations automatically into a claimed business success or failure.

A shared dashboard is a new disclosure of data

Provider guidance explains that Data can share analyses through ChatGPT Sites and that the underlying analysis data is copied into the published Site. Permission to run a query is therefore not sufficient review of the eventual audience. Decide whether the dashboard needs individual transactions or whether aggregated values will do. Remove unnecessary customer names and internal comments. Use an ordinary recipient account to inspect what is actually visible. Also define how updates and withdrawal should work. A dashboard showing stale figures without a clear source date can mislead decision-makers even when the calculation was correct when first produced.

Acceptance includes figures, access and maintenance

Use a small but meaningful acceptance checklist. Do totals and a sample of individual transactions reconcile? Do filters behave correctly with empty datasets and unexpected categories? Are restricted records unavailable to unauthorised roles? Does the report recognise a failed retrieval rather than presenting old values as current? Record the data date, query scope and owner alongside the result. Before automating a recurring analysis, establish that the source and metric are stable enough. Operational handover needs a named contact for questions and failures. An attractive dashboard does not prove the surrounding process already works reliably, nor does it eliminate responsibility for maintaining definitions.

Common questions about the Data plugin

Does Data replace our BI system? It can complement existing sources and tools; supported actions depend on the particular connection. Can I describe a discrepancy as a cause? Only when evidence and business review support that conclusion. Does installing the plugin give every employee permission? No, plugin access and permissions in the underlying system remain separate. How should a small company begin? Use a familiar report, a documented metric and a total that can be independently checked. What benefit should be measured? Time to a confirmed answer, plus rework and clarification effort. Do not extrapolate one faster analysis into a claim about every report or company revenue.

Keep it verifiable

Primary sources

The next sensible step

Find the right AI workflow with the System Check

Eight steps from a general interest in AI to a clearer decision for your business.

Start AI System Check
FreeProvider-neutralNo credentials