OpenAI has launched a Data agent for ChatGPT Work, giving teams a conversational way to investigate company data, explain changes, and create interactive dashboards. The underlying Data plugin is also available in Codex.
This is more ambitious than placing a chat box in front of a database. The agent can connect analysis to an organization’s definitions, permissions, documents, and existing business-intelligence tools. That surrounding context is what determines whether an answer is merely plausible or useful enough to inform a decision.
One question can cross several governed sources
The Data agent can work with approved sources including Amazon Redshift, ClickHouse, Databricks, Google BigQuery, MongoDB, and Snowflake. It can also bring documents and files from services such as Google Drive and SharePoint into an investigation when those connections are available.
A user can ask why retention changed, which accounts carry renewal risk, or where spending is rising. The agent can examine the relevant data, explain its reasoning, show the evidence behind a finding, and refine the analysis through follow-up questions. The workflow is designed to keep the investigation in one conversation instead of forcing the user to move between query tools, documents, spreadsheets, and presentation software.
Business definitions matter as much as database access
Enterprise data is full of terms that look simple until two teams calculate them differently. Revenue, active customer, qualified lead, churn, and gross margin can each depend on internal rules. OpenAI says the agent can use a company’s semantic layer: authoritative metric definitions, custom calculations, dataset relationships, documentation, and trusted dashboards.
That is an important product decision. Natural-language analysis becomes more credible when the system knows which definition the business has approved and can identify the source, filters, time period, and calculation behind a result. Users still need to review those details before acting, especially when a finding conflicts with an existing report.
Permissions travel with the connected account
The Data agent does not create a new path around existing access controls. Queries use the connected account’s permissions, including applicable table-, row-, and column-level restrictions. Workspace administrators control which plugins and connections are available and who can use them.
This makes governance part of the workflow rather than a separate promise. A sales manager, finance analyst, and executive may ask similar questions yet receive access to different underlying records. The system must preserve those boundaries not only while querying data, but also when creating, publishing, refreshing, or sharing an artifact.
The output can become a working dashboard
After an investigation, users can ask the agent to turn its findings into an interactive dashboard, then adjust the metrics, breakdowns, charts, layout, or visual styling through conversation. OpenAI’s documentation says the plugin can also work with connected BI tools including Power BI, Tableau, Sigma, ThoughtSpot, Omni, and Oracle BI.
That shifts the value proposition from answering an isolated question to producing something a team can revisit and share. The agent can help explain what supports a chart and refresh the analysis when updated information is needed. Available actions still depend on the connected tool, workspace policy, user permissions, and approval requirements.
Why it matters for builders
The launch shows where enterprise agents are heading: systems that inherit the organization around the data, not simply access the data itself. Useful agents need governed connections, shared definitions, traceable evidence, familiar outputs, and explicit approval points.
It is the enterprise-data counterpart to the infrastructure shift behind OpenAI’s Agents API and managed Codex harness. The Agents API packages the machinery required to keep an agent working; the Data agent packages a specific, high-value business workflow on top of governed tools and context.
Product teams should also treat the launch as a warning against oversimplifying “chat with your data.” A polished answer can still be wrong because of a stale source, an ambiguous metric, a hidden filter, or an incomplete connection. Strong implementations must make evidence easy to inspect, surface the definitions used, preserve least-privilege access, and require review before consequential actions or broad sharing.
The opportunity is not to replace every analyst or dashboard. It is to shorten the distance between a business question and a reviewable piece of analysis—then let people continue the work in the systems they already trust.