AI-to-human handoff

An AI-to-human handoff is the moment an AI system stops handling something and a person takes over. In customer support it usually means escalating from a chatbot to a human agent. In tools and agents it often means handing control back to the user, for example when a travel-booking agent can't complete a payment step. Either way, a good handoff is triggered at the right time, is easy to request, and carries the context with it.

Why it matters

Every AI system has limits, and users get stuck when there is no way past them. The US Consumer Financial Protection Bureau's 2023 report on bank chatbots describes "continuous loops of repetitive, unhelpful jargon or legalese without an offramp to a human customer service representative", which it calls "doom loops", often caused when an issue "falls outside the chatbot's limited capabilities" (CFPB). The report also notes that scripted systems may only recognize a dispute if the customer uses specific words or syntax.

The handoff also carries responsibility. When Air Canada's chatbot gave a passenger wrong fare information, a Canadian tribunal ordered the airline to pay, rejecting the idea that the chatbot was responsible for its own actions (Daily Hive). For questions about money, eligibility or policy, routing to a person can be cheaper than a confident wrong answer.

How to apply it

• Do make "talk to a person" always available and easy to find, not unlocked only after several failed attempts.
• Do trigger handoff automatically on clear signals: repeated rephrasing, frustration, complaints, disputes, and high-stakes or out-of-scope topics.
• Do pass the context. PAIR says the person taking over needs "awareness of the situation, what they need to do next, and how to do it" (PAIR). Send the transcript, the user's details and what the AI already tried, so the customer doesn't repeat themselves.
• Do set expectations about wait times and channels, and offer a callback or email if no one is available now.
• Do tell users when they are now talking to a person, and when they are back with the AI. See AI disclosure.
• Do hand partial work back to users when an agent stops, such as a filled-in form they can finish themselves.
• Don't make the AI argue the user out of escalating.
• Don't route to a human who can't see the conversation or doesn't have the authority to resolve it.

Common mistakes

• Measuring the AI by containment rate alone. Keeping people away from humans looks efficient until it drives complaints and churn.
• Requiring magic words such as "agent" before escalation works.
• Losing the conversation history at the handoff, forcing users to start again.
• Treating handoff as only a support-bot feature. Any human-in-the-loop workflow needs a clean way to pass control both ways.