Database permissions decide which rows an agent can read. Business semantics decide what those rows mean. OpenAI's Data agent puts both in the path from a question to an approved action.

What OpenAI Released

On September 10, 2026, OpenAI introduced a Data agent in ChatGPT Work. Administrators can install it from the Plugins directory and choose which data connections and roles are available.

The announcement lists connections to warehouses and databases including Redshift, BigQuery, ClickHouse, Databricks, MongoDB, and Snowflake, plus documents from Google Drive and SharePoint. Queries enforce the connected account's table, row, and column restrictions.

The Semantic Layer Is An Execution Input

The agent can also use business terms, metric definitions, calculations, and data relationships from semantic layers and trusted sources such as dbt, Databricks Genie, Snowflake Horizon, GitHub, and existing BI dashboards.

A result that changes pricing, staffing, outreach, inventory, or cash needs more than a valid query. The agent can still make the wrong decision when it uses an expired definition, joins entities at the wrong grain, or compares a stock measure with a flow. Permission checks do not catch those failures.

Evidence And Action Stay Visible

OpenAI says users can inspect evidence behind a finding, refine the analysis, create and refresh interactive dashboards, and work inside tools including Power BI, Tableau, Sigma, Omni, Oracle BI, and ThoughtSpot. The agent can recommend next steps and carry out connected Slack or email actions after approval.

The release does not publish an independent accuracy study or error rate across those connectors. Its customer examples and internal usage figures are vendor-provided evidence. Treat availability as confirmed and performance as something each operator must measure.

A Minimum Semantic Control Record

  • Metric name, definition, owner, version, and effective date.
  • Source tables, joins, grain, exclusions, and freshness expectation.
  • Identity and row-level policy used for the query.
  • Query, result snapshot, evidence links, and known data-quality limits.
  • Decision threshold and the person or policy allowed to approve an action.
  • Downstream action ID so the analysis can be tied to what changed.

Operator move: Pick one recurring operating decision and publish its metric contract before automating it. Run the agent against a known historical period and require it to reproduce both the number and the evidence trail before enabling an action.