On September 10, 2026, OpenAI introduced Data agent in ChatGPT Work, describing a tool that can connect to company data, investigate questions in natural language, and build interactive dashboards. According to the company announcement, the aim is to let more employees find answers themselves to questions such as where sales slowed, where spending is rising, or which large accounts may be at risk of non-renewal. This matters not merely as a new feature. It signals that the interface between workers and data is moving quickly from pre-built reports toward contextual dialogue, inquiry, and checking.
But asking data questions does not automatically produce better decisions. When more people can retrieve figures, create charts, and interpret change within minutes, the constraint shifts from tool operation to three management questions: which data count as authoritative, who may see or share what, and how the organisation responds when a new answer conflicts with an existing report. A company that treats this only as a faster way to make dashboards may add another reporting layer. A company that designs the conditions for checking can substantially shorten the path from signal to action.
What is changing is the distance from question to evidence
In many organisations, a business question begins in a meeting but the answer arrives much later. A functional leader raises a concern, an analyst receives the request, clarifies definitions, extracts data, checks exceptions, and presents a result. The process remains necessary for complex analysis, but it creates delay for many questions that could be investigated earlier. OpenAI says Data agent is intended to let users investigate company data, create dashboards, and refine analysis within a conversation rather than requiring query skills or a separate report.
The management implication is not that data teams become less important. As the technical barrier falls, their role in designing a trust layer becomes more important: defining metrics, managing data models, setting access, monitoring quality, and creating checkpoints for analyses that will be broadly shared. Without that layer, the same revenue or customer question can produce different answers simply because people select different tables or time periods.
Three conditions for self-service data without self-service error
| Condition | What to design | Risk if absent |
|---|---|---|
| Trusted sources | Attach definitions, owners, and update times to critical metrics. | Many technically correct dashboards that conflict in meaning. |
| Access and scope | Grant data, actions, and sharing ability that fit each role. | Information leakage or analysis beyond legitimate authority. |
| Checks before action | Require cited sources, exception checks, and a decision owner. | A convincing chart becomes a conclusion too early. |
The first condition is a traceable layer of trusted sources. Metrics such as revenue, active customers, churn, or cost to serve need definitions that are public inside the organisation, named owners, and clear update times. When an agent produces a finding, the user should know the data source, filters applied, and refresh point. This does not slow the experience for its own sake; it makes it possible to catch an answer that sounds plausible but uses the wrong definition.
The second condition is contextual access. An account manager needs information that helps resolve a request, but does not necessarily need all financial or people data. Access is not merely permission or no permission to enter a data store. It also includes permission to create a file, share a dashboard, take a next action, or combine sources. When these capabilities are designed with an approval flow, an organisation can expand data use without choosing between speed and control.
Example: a large-account review
A commercial team is preparing a review of accounts with falling revenue. It uses a data assistant to bring together sales changes, recent support requests, contract dates, and product use. The assistant does not decide the cause or send an offer. It cites the source for each signal, flags missing fields, and creates questions for the account manager to confirm. In the meeting, the team selects three accounts for action, records assumptions, and sets a two-week review point. The value is not the chart produced quickly; it is the meeting having one evidence set and a clear follow-up step.
Measure the decision cycle, not the number of dashboards
When conversational data tools are introduced, query volume, dashboards created, or active users are early activity signals, not complete evidence of success. Companies should also track the time from a question to identifying the data that require checking, the share of analyses used in an actual decision, occasions when a wrong definition or missing data are detected, and the share of actions that must be reworked after new information emerges. These measures tie the project to operating quality rather than the novelty of the interface.
Tool boundaries also need definition. Fast analysis can be extremely useful for detecting a signal, but decisions about pricing, credit, hiring, sensitive customers, or compliance often need independent checks and clear accountability. A data assistant can make preparation faster; it does not replace agreeing definitions, assessing causes, or accepting responsibility for a choice and its consequences.
OpenAI’s announcement is a recent example of a broader trend: analytical capability is moving closer to decision-makers. What companies need to prepare is not only another data connection. They need a way of working in which every important answer has a source, every access right has a purpose, and every analysis becomes action through a named owner. When these conditions are present, self-service data can reduce waiting time without reducing decision discipline.
Source
OpenAI. (2026, September 10). Now everyone can put data to work. https://openai.com/index/put-data-to-work/
