Approvals

Approvals let an authorized reviewer inspect a proposed data mutation before execution. Open AI → Approvals or the proposal linked from Chat or a run.

Review a proposal

1. Confirm the project, automation or conversation, and requested operation.
2. Inspect the reason, affected entities, before/after values, and available operation details.
3. Check the affected count. A preview can be truncated and is not necessarily the full dataset.
4. Approve only the intended change, or reject it if the target or effect is wrong.
5. Inspect the resulting run and data after the decision. Approval and completed execution are separate states.

Pending proposals await a decision. Historical proposals can be approved, rejected, expired, or cancelled. All includes history; an empty Pending view does not mean there have never been proposals.

Mutation policy

Require approval pauses covered mutations for review. Auto-approve permits the covered data mutations without that manual decision. Use the latter only after testing the automation’s inputs, scope, and failure behavior.

Custom tools execute application code and can have effects outside GraphQL. Do not assume a GraphQL mutation approval policy gates every email, external API call, or other tool side effect. Implement authorization and any required confirmation in the tool’s design. See tool security.

If a proposal expires or execution fails after approval, inspect the error and current data before initiating another run.