Approval Gate
An approval gate is a control point where an AI agent pauses before a consequential action and a named human decides whether it proceeds. The reviewer sees what the agent proposes, why, and the evidence behind it — then approves, rejects, or returns the case with a reason.
What an approval gate means
Approval gates are the human-in-the-loop mechanism that makes agent autonomy acceptable in regulated work. The design questions are concrete: which actions get gated (a usable rule — anything that changes a record someone else relies on, leaves the team, or cannot be cleanly reversed; reading, drafting and comparing stay ungated), what the reviewer sees (the proposal, the evidence, the conflicting values, in order), what happens on timeout (a default that is safe, not silent), and how the decision is recorded (proposed action exactly as shown, decision, reason, identity of the decider). Gates also guard against rubber-stamping: decisions are sampled and reviewed, and gate health — approval rates, time-to-decision, rejection reasons — is measured. Decisions with regulatory weight should be unreachable by the agent entirely, which is stronger than gating them.
Frequently asked questions
Which AI agent actions need an approval gate?
Anything that changes a record someone else relies on, anything that leaves the team, and anything that cannot be cleanly reversed. Leave reading, comparing and drafting ungated so the agent stays useful, and limit what the agent can reach so ungated actions stay harmless.
What should the human reviewer see at an approval gate?
The proposed action exactly as it will execute, the evidence the agent relied on, any conflicting values between sources, the policy rule that requires the gate, and the consequence of approving. Enough to decide — not so much that review becomes rubber-stamping.