November 2026
How we design approval gates for AI agents
An approval gate is a named person and a specific action. Silence is not a yes.
The gate sits on the action
A prompt that says “ask a human if you are unsure” is not a control. The gate is a named person, a listed action, and a refusal when that person has not acted. Conversation can be free. Changing a system cannot.
Write the list before the pilot
Before a sprint we write the action list. Each row has a permission (read, propose, or act), whether approval is required, who may give it, and what is logged. If a row cannot be filled, that action is not available to the agent.
Silence does nothing
An expired approval is not a yes. An unread message is not a yes. The incident demo uses the same rule: the agent may draft a runbook step. The timeout does not change until a technician approves it.
What we will not call a gate
A confidence score alone is not a gate. A second model reviewing the first is not a gate. They can inform the person. They do not replace the person on an action that spends money, changes access, or touches production.