Human Approval
Human approval is a control for actions with meaningful impact. It should be based on action sensitivity, not applied indiscriminately to every AI interaction.
Default Approval Matrix
| Action | Default | Reviewer should verify |
|---|---|---|
| Search or read permitted data | Automatic | Access scope and source |
| Create an internal draft | Automatic or confirmation | Sensitive content and intended audience |
| Create an internal ticket or task | Confirmation for broad agents | Project, owner, priority, duplicates |
| Send or publish externally | Confirmation | Recipient, content, attachments, classification |
| Update a business record | Confirmation | Target ID, changed fields, downstream effects |
| Export data | Confirmation | Purpose, destination, minimum necessary data |
| Change permissions, budgets, production, or legal state | Confirmation plus provider-side authorization | Identity, scope, rollback, owner approval |
| Delete or perform an irreversible action | Disabled or exceptional approval | Necessity, backup, exact target, recovery path |
Effective Approval Requests
An approver needs enough context to make a real decision. Show:
- the action and target system,
- the exact recipient, record, environment, or resource,
- the arguments that matter,
- a concise reason for the action,
- the expected effect and whether it is reversible,
- which agent, conversation, workflow, and connection requested it.
Do not hide a batch of materially different changes behind one generic confirmation.
Review and Tuning
If approvals become noisy, separate read, draft, and write functions or narrow the workflow. Do not remove confirmation from high-impact actions only to reduce friction. Review approvals and outcomes in Tool Executions.