OpenAI introduced the Data agent for ChatGPT Work on September 10, 2026. It is not positioned as another chatbot that only answers questions about data. It combines enterprise data, business semantics, interactive dashboards, and follow-up actions into one workflow. Users can ask questions in natural language, request an answer or dashboard, and—when connected tools and approvals allow it—hand the result to the next step.
The data layer is the foundation of the release. OpenAI lists connectors for Amazon Redshift, Datadog, Google BigQuery, ClickHouse, Databricks, MongoDB, and Snowflake, along with access to files and documents from Google Drive and SharePoint. That gives the agent a way to combine queryable systems with the written context around them. The actual availability of each connector still depends on the organization’s administrators and account configuration.
The more important design choice is the semantic layer. OpenAI says the Data agent can use an organization’s business terms, metric definitions, custom calculations, and relationships from semantic layers or trusted sources such as Databricks Genie Ontology, dbt, GitHub, Snowflake Horizon, and existing BI dashboards. This addresses a common enterprise failure mode: two teams can use the same word, such as revenue or active customer, for different definitions. A model can query the right table and still answer the wrong business question without that context.
Governance is part of the operating model. Admins choose which data connections are available and which roles can use the Data agent. Queries inherit the permissions of the connected account, including table-, row-, and column-level restrictions. The agent’s view of data should therefore come from the organization’s access policy, not from a user claiming access in a prompt. That boundary matters most when analysis touches customer, financial, or personnel data.
The Data agent also separates finding an answer from delivering it. Users can ask follow-up questions, inspect the evidence behind a finding, edit and share dashboards, and refresh them into tools such as Omni, Oracle BI, Power BI, Sigma, Tableau, and ThoughtSpot. OpenAI says the agent can recommend next steps, identify people who should be involved, and share findings through Slack, email, or other connected tools. External changes and actions still depend on approved connections and organization settings; the announcement should not be read as unrestricted autonomous writing.
OpenAI says nearly all of its product team and more than two-thirds of its go-to-market organization now use data agents internally. The post also cites early Alpha participants including NTT DATA, Thermo Fisher, and ServiceTitan. Those are vendor-reported early-use examples, not an independent cross-company effectiveness study. The reusable lesson is the operating discipline underneath them: shared business definitions, explicit data permissions, and human judgment before and after an agent acts.
For an enterprise, the adoption sequence should start with data governance rather than with a demo of how polished a report can look. Define trusted sources, metric owners, eligible roles, sensitive fields, and actions that require approval. Then choose one repetitive workflow with clear inputs and outputs. Preserve the source evidence, calculation path, permission context, and approval record so the team can reproduce and audit the result instead of accepting a plausible paragraph on faith.
The larger shift is that the Data agent’s endpoint is no longer the query result. It is a governed chain from data to decision to action. Model quality matters, but confidence in the workflow depends on consistent semantics, inherited permissions, visible evidence, and explicit approval points. OpenAI is putting those pieces together in one product surface, raising the enterprise question from whether AI can answer quickly to whether it can deliver the right answer to the right person and tool inside a controllable boundary.
