Connector And Governance Guide

OpenAI Data Agent Connectors And Permissions

Map each system to the job it performs: source data, business meaning, dashboard delivery, or an approved follow-up action.

Official list checked Sep 29, 2026Four connection layersPermission-first rollout
The Short Answer

Connectors Are Only One Part Of A Reliable Data Agent

A warehouse supplies records. A semantic layer supplies the meaning of metrics and relationships. A BI destination turns analysis into a working dashboard. An action channel distributes or carries out an approved next step. A production pilot needs an owner and permission test across all four layers.

OpenAI Data Agent Connector Map

LayerOfficial ExamplesPrimary Job
Data sourcesAmazon Redshift, Datadog, Google BigQuery, ClickHouse, Databricks, MongoDB, SnowflakeSupply governed operational and analytical data
Files and documentsGoogle Drive, SharePointAdd documents and unstructured business context
Semantic and trusted contextDatabricks Genie Ontology, dbt, GitHub, Snowflake Horizon, BI dashboardsDefine terms, calculations, and data relationships
BI destinationsOmni, Oracle BI, Power BI, Sigma, Tableau, ThoughtSpotBuild or interact with dashboards in existing tools
Action channelsSlack, email, other connected toolsShare findings or carry out user-approved actions

This is a list of examples OpenAI named in its announcement, not a promise that every connector is available to every account, region, or plan. Confirm the current plugin listing and provider requirements inside your workspace.

Choose A Source For The Decision

Start with the system that contains the evidence needed for one recurring decision. Do not connect every available source just because the connector exists.

Structured

Warehouses And Databases

Use for governed events, transactions, product usage, financial records, and operational measurements. Begin with read-only access and a narrow schema.

Validate totals against a known report
Unstructured

Files And Documents

Use for plans, definitions, research, feedback, and policy context. Confirm which folders and files the connected identity can actually retrieve.

Test with allowed and restricted files

Give The Agent Trusted Business Context

Two teams can use the same word—such as “customer,” “active,” or “revenue”—and mean different things. Semantic layers, metric definitions, custom calculations, and documented data relationships reduce that ambiguity.

Minimum context package: define the decision, the primary metric, its formula, its owner, the source of truth, the comparison period, and any exclusions before testing the first prompt.
  • Name one source of truth when multiple systems contain similar fields
  • Record the time zone, currency, fiscal calendar, and aggregation level
  • Explain how deleted, late, duplicate, and test records are handled
  • Define which dimensions are safe to combine
  • Attach a human owner to every metric used for a decision

Decide Where The Result Should Live

OpenAI says Data Agent can create interactive dashboards directly and can build or interact with dashboards in named BI tools. Pick the destination based on who must use, edit, refresh, and govern the result—not on novelty.

If The Result IsPreferReason
A quick investigationConversation plus evidence reviewKeep iteration fast before publishing
A shared operating viewInteractive dashboardMake refreshes and follow-up analysis repeatable
An established executive metricExisting governed BI toolPreserve ownership, definitions, and access patterns
A recommendation requiring actionApproval plus connected channelKeep a person responsible for the consequential step

Run A Real Permission Test

  1. Create representative test users for the roles included in the pilot.
  2. Verify each user can see an allowed table, row, column, dashboard, and document.
  3. Verify the same user cannot retrieve one deliberately restricted item in each system.
  4. Test a question that combines sources and confirm the most restrictive rule still holds.
  5. Review what is shared when a dashboard or finding moves to another person or channel.
  6. Record the owner, test date, evidence, and next review date.
A connector being available does not mean it should be enabled: begin with the smallest source and permission scope that can answer the pilot question.

Turn The Map Into A Pilot Plan

Select your sources, governance maturity, risk level, and economics. The readiness planner will identify blockers and produce a copyable 30-day checklist.

Build My Readiness Plan

Primary source: OpenAI’s Data Agent announcement, last checked Sep 29, 2026. Product availability and connector support can change. AmpysApps is independent and is not affiliated with OpenAI.