Approvals keep supervised automation under human control. They appear when an agent proposes an external change or when a workflow reaches an explicit review step.
When approval is required
Connection tools are grouped into three levels:
- Reads retrieve information without changing the connected service.
- Actions create, send, or update information and require approval.
- Destructive tools can remove or irreversibly change data and require both explicit access and approval.
If a tool cannot be classified safely, AI Agent treats it as an action rather than silently allowing it as a read. To configure which accounts and tools an agent can use, see Connection tool access.
How a proposal moves
Approval applies to the displayed proposal. Approving continues that exact action; rejecting stops it before execution.
flowchart LR
Proposal["Proposed action"] --> Review["Review account and change"]
Review --> Decision{"Approve?"}
Decision -- "Yes" --> Execute["Execute displayed action"]
Execute --> Outcome["Success or failure"]
Decision -- "No" --> Rejected["Rejected before execution"]
Review an external action
An approval card can appear in Chat, a ticket, the Inbox, or the Workflows area.
Confirm the account
Verify the provider and exact connected account shown on the proposal.
Inspect the change
Review the action summary and every displayed field, recipient, record, or piece of content.
Compare it with the request
Check that the proposed change matches the original task and intended scope.
Approve or reject
Choose Approve to execute the displayed proposal or Reject to stop it.
Approval and execution are connected: after approval, the action moves into execution and then reports success or failure. Rejecting it prevents execution.
Review a workflow step
A workflow can pause at an approval block. Open the run detail, inspect the completed steps and proposed output, then approve or reject the waiting step. The workflow resumes from that point with your decision.
Approval states
- Pending — waiting for a decision
- Executing — approved and currently running
- Executed — completed successfully
- Rejected — stopped by a reviewer
- Failed — execution returned an error