ChatGPT Work Can Now Turn Your Company’s Data into Answers, Interactive Dashboards, and Action

OpenAI introduced the Data agent on September 8, 2026, giving ChatGPT Work users a way to analyze company information, create interactive dashboards, and initiate follow-up work through a conversation. Instead of moving data into a separate AI product, organizations connect the systems and business context they already use.

The setup starts by adding the Data Plugin in ChatGPT Work and connecting approved sources. A user can then ask a business question, including a loosely defined one, and the agent will identify relevant data, plan an analysis, run queries or code, and present the findings with supporting evidence.

OpenAI’s “everyone” framing needs some qualification. This is not a feature for every free ChatGPT account. It is an organizational tool inside ChatGPT Work, subject to plugin configuration, company permissions, and access to the underlying data. Its purpose is to make analytical work accessible beyond dedicated data teams without discarding the controls those teams have established.

OpenAI's Data Agent Turns Questions Into Dashboards

The Data agent is designed for business questions that require more than one database query. A product manager might ask why adoption changed among a customer segment, for example. Answering that properly could require product usage events, account records, sales context, customer feedback, and an existing dashboard’s metric definitions.

OpenAI says the agent first maps the relevant sources and creates a research plan. It prioritizes authoritative information, weighs the reliability of different signals, and cross-checks conclusions against qualitative material such as support conversations or customer feedback. It can also write and execute SQL and Python during the analysis.

The output can be an answer, a table, a chart, or a multi-page interactive dashboard with filters and widgets. Because the conversation retains its analytical context, users can request another view, narrow the cohort, challenge an assumption, or change the time period without reconstructing the entire analysis. OpenAI says dashboards can also refresh when their supporting logic or data changes.

2098065296968011853

This conversational workflow is more useful than simply placing a chat box in front of a warehouse. The agent is supposed to preserve the research process, including the sources selected, methods used, generated queries, code, and intermediate artifacts. That gives an analyst or decision-maker something concrete to review when an answer looks surprising.

The Data Agent Connects to the Existing Analytics Stack

OpenAI is positioning the Data agent as a layer across existing systems rather than a replacement for them. According to the company’s launch materials and data solutions overview, it can work with data warehouses and analytics platforms such as Snowflake, Google BigQuery, and Databricks, along with business intelligence dashboards, reports, spreadsheets, presentations, documents, and Slack conversations.

The agent can also incorporate operational context from CRM systems, customer support platforms, product analytics services, feedback repositories, and internal product tools. That combination matters because many business questions cannot be answered from structured event data alone. A drop in usage may be visible in a warehouse, while the explanation sits in support tickets, launch notes, or customer interviews.

This approach reduces the need for a large migration before a company can experiment with AI-assisted analysis. It also makes the quality of the connections critically important. If two dashboards define active users differently, or sales and finance maintain conflicting revenue figures, the Data agent still needs a reliable way to identify the authoritative definition.

OpenAI says the system accounts for those conflicts by examining context and confidence rather than treating every record as equally trustworthy. For identity resolution, for example, direct identifiers receive more weight than inferred matches. Ambiguous evidence should produce a lower-confidence conclusion instead of a definitive claim. That is the intended behavior, but organizations will need to test how consistently it works with their own data.

A Polished Dashboard Must Still Show Its Work

Generative AI can produce confident explanations and professional-looking charts even when the underlying reasoning is weak. For an enterprise data product, provenance is therefore more important than presentation.

The Data agent exposes its sources and analytical methods, including generated SQL, Python code, and intermediate work. A data analyst can inspect the query, confirm that filters were applied correctly, and check whether the agent joined the right records. A business user can see which dashboards, documents, or conversations contributed to the answer.

OpenAI reports that expert analysts preferred Data agent outputs by a ratio of more than three to one when comparing them with existing internal analytics tools. The evaluation reportedly covered work in product, finance, research, revenue, and operations.

That result should be treated as an internal product claim rather than an independent benchmark. OpenAI’s announcement does not provide a public task set, sample size, scoring rubric, or detailed breakdown of the comparisons. A preference test can indicate that analysts found an output more useful, but it does not by itself establish factual accuracy, query reliability, or consistent performance across unfamiliar data environments.

The traceability features make the claim easier for customers to evaluate themselves. If users can inspect the evidence and reproduce an analysis, they do not have to rely exclusively on the agent’s prose or OpenAI’s benchmark.

Actions Stay Behind a Human Approval Step

The Data agent can move beyond analysis once it reaches a conclusion. OpenAI gives examples such as drafting a Slack message, preparing a launch brief, generating recommendations, creating a CRM audience for a campaign, or writing code and opening a pull request.

These capabilities turn the agent into a workflow coordinator. A retention analysis could identify an at-risk segment and then prepare a campaign for that group. A product investigation could find a technical issue, generate a proposed fix, and create a pull request with the relevant context attached.

OpenAI says the user must confirm workflows that could produce downstream effects before the agent executes them. The system can prepare the action and preserve the context, but it should not silently message customers, alter a campaign, or submit code.

Confirmation is a sensible boundary, although it does not remove every operational risk. Companies still need narrowly scoped permissions, visible action logs, approval rules for sensitive systems, and a recovery process when an approved action produces an unintended result. A human approval dialog is useful only when the reviewer can understand exactly what will happen.

OpenAI's Internal Results Show the Intended Use Case

OpenAI has already deployed the Data agent across much of its own organization. The company says nearly all of its Product team and more than two-thirds of its go-to-market organization have adopted it, with hundreds of employees using it each day.

The reported time savings are substantial:

OpenAI also describes a product manager using the agent to answer more than 70 launch-related questions within hours, work that the company says could otherwise have taken weeks.

These examples help explain the target workload. The agent is not limited to executive dashboards or occasional company-wide reports. OpenAI expects teams to use it for recurring questions, launch reviews, operational checks, customer research, and the many small investigations that are often delayed because an analyst has more important work in the queue.