When should an AI support chatbot hand off to a human?

An AI support chatbot should hand a conversation to a person when it cannot answer from the available business information, when the customer asks for a person, or when the request needs human judgement. Keep the conversation history, explain that the issue needs review and give your team responsibility for the next reply.

Start with the boundary between an answer and an action. A documented returns policy can help answer a general question. Approving an exception or promising a refund needs the appropriate authority. Write down which topics your team must handle, and test those rules with realistic questions before putting the chatbot in front of customers.

In Chat Support's website chat flow, the agent checks the available business context before generating an answer. If there is no usable context, it hands the conversation over. The flow also checks for requests for a person, frustration, legal or safety concerns, and configured handover topics. An uncertain generated reply or an AI error can trigger escalation too. These checks help limit unsupported answers; they do not make every AI response correct.

A handoff needs more than a message saying someone will help. Your teammate needs the customer's question, the conversation so far and the reason for escalation. Chat Support records the conversation as awaiting human attention with a handover reason and a summary when one is available. The team can use that record to understand what remains unresolved before replying.

Decide who will check the queue and when. In the website chat flow, Chat Support attempts an email alert if a forwarding inbox is configured. An attempted alert is not proof that a teammate has read the conversation. Check the queue as part of your support routine, verify the forwarding destination and test notification delivery. Your business supplies the people who respond.

Set expectations around your team's actual availability. When designing a handoff message, avoid promising an immediate reply unless your team can provide one. For an unanswered delivery-policy question, a suitable message might be: 'I don't have enough information to answer that. Your question needs a team member to review it.' Treat this as suggested wording, and check what your own widget displays. A handoff is not a promise of a response time.

Before launch, test a question with a documented answer, a question missing from your knowledge base, an explicit request for a person and a topic reserved for your team. Check what the customer sees, whether the conversation reaches the queue, what context your teammate receives and whether the configured alert arrives. Use a test scenario without real customer data. Do not stop at checking the chatbot's first reply.

After a handoff, review the cause. If a routine question lacked an answer, update the relevant source and test that question again. If the request needed approval or judgement, keep the human boundary. Compare your plan's help desk and handoff options before choosing how to run this process, then create an agent to test with your own business information.