Definition
A checkpoint in an AI agent workflow that requires a human — or a separate authorisation system — to review and confirm before the agent takes a consequential action such as deleting files, sending messages, or executing commands. Approval gates are the primary safety brake on autonomous AI behaviour.
Why it matters
When approval gates are disabled or hardcoded to auto-approve — as found in real production systems — organisations lose the only human check on potentially destructive AI actions, exposing them to accidental or maliciously triggered harm at machine speed.