Human-in-the-loop
The AI can research the itinerary and flights, but it must ask me before paying or issuing a ticket.
The AI can research the itinerary and flights, but it must ask me before paying or issuing a ticket.
Search and comparison are read-only, so the agent can complete them automatically.
Human-in-the-loop keeps low-risk research automatic and places a human decision before the high-risk action.
Pause before the action: Search and comparison can run automatically. Charging, ticketing, publishing, or deletion must be blocked before the high-risk action occurs.
Show enough information to decide: An approval cannot be only “continue?” The interface should show the object, scope, cost, rules, and expected consequence.
Return the decision to the same run: Approval resumes from the pause, rejection stops it, and corrections return with human feedback without restarting the whole task.
You may search and compare weekend flights from Shanghai to Tokyo automatically, but pause before payment and ticketing. Show the round-trip dates, passenger, total with taxes, baggage, and change rules. Continue only after my explicit approval, and stop if I reject or the information changes.