Human-in-the-Loop AI Agent

The pattern matters because an AI agent's two failure modes are guessing (inventing a policy or a price) and overreach (making a change nobody sanctioned). A human-in-the-loop design closes both: questions go to a person instead of being guessed, and sensitive actions are proposed rather than executed. The quality of the implementation shows in what happens while the human is busy — whether the customer is kept informed, whether the request finds the next person, and what happens when nobody answers.

In Lodgestory, the AI Agent step in a Bot Journey can ask the team, hold changes for approval, and escalate turn by turn — to the next person, then to a phone call on your own number when nobody is online, then to a fallback you chose. Held changes are rejected when nobody answers, never made silently. Your team answers from the Kaylee console, the conversation, team chat, the mobile app, the phone, or your own systems through the API and MCP.