Human in the loop

• The system does the routine work and stops at set approval points.
• The approval points are usually steps that cannot be undone.
• The person sees the actual action before saying yes.

Human in the loop describes an automated process in which a person is still part of the decision at certain points. The machine does the preparation, and a human approves, edits or rejects before something important happens. The opposite is a fully automatic process with no checkpoint.

Example: an agent waiting for approval before sending

Say you ask an agent to reply to a customer. The agent reads the thread, drafts a reply, and stops. It shows you the exact message and offers three choices: send, edit or deny.

1. You read the draft and see it promises a refund you did not mean to offer.
2. You choose edit, change the line, and approve.
3. Only then is the message sent.

Without the approval point, the wrong promise would have gone straight to the customer. With it, the cost of the agent's mistake was a few seconds of your attention. DuckyHelper works like this: it's built to ask first before it sends a message or pays for anything, and its agent work is in early access.

Why it matters

Not every step needs a person. Asking about everything makes an assistant slow and teaches people to click yes without reading. The useful rule is to place approval points where an action cannot be taken back, such as sending, paying or deleting, and let the rest run. The person also sees the real action, not a summary, so the approval means something.

This is a safeguard for any system that can act on its own, including computer use. For a longer look at the reasoning, read why an AI agent should ask before it sends or pays.