Manual approval is a control that requires a human to review and confirm a sensitive action before an AI agent, automation script, or privileged tool is allowed to proceed. The action might include editing production code, running destructive commands, accessing secrets, approving deployments, or connecting to external systems. In security terms, it creates a pause between suggestion and execution, reducing the chance that a system will carry out a harmful request on its own.
This matters because AI agents can be persuaded by misleading context, bad prompts, or mistaken assumptions about what is authorized. Attackers may try to frame a malicious task as routine work, hoping the agent will act on it without scrutiny. Manual approval helps defend against that by forcing a person to inspect the request, check the scope, and verify intent. Strong implementations also log who approved the action, what was approved, and whether the request matched policy. Used well, manual approval is a practical trust boundary for high-risk automation.