Most organizations adding an AI assistant to their website spend their effort on what the assistant can answer. Far less goes into what happens when it can't. Yet that moment decides whether someone leaves with their problem solved or leaves angry.
People tend to forgive a chatbot that says it doesn't know. What they don't forgive is going in circles, repeating themselves to a human, or not knowing whether anyone is coming. Here are five rules we use when designing the hand-off.
1. Make the way out visible from the start
Don't hide the option to reach a person until the third failed answer. Show it from the first message, in plain words: "Talk to a person." People who know they can leave are more willing to give the assistant a try.
2. Recognize when to step aside
Some signals should trigger a hand-off offer automatically: the same question asked twice, words like "complaint", "wrong" or "agent", anything about money already taken, health, safety or a deadline, and visible frustration. The assistant doesn't need to be certain. Offering a person costs little; keeping someone trapped costs a lot.
3. Never make people repeat themselves
Pass the whole conversation, and anything the person already entered, to the human who picks it up. The first thing the person should hear is a summary of their problem, not "How can I help you today?"
4. Be honest about the wait
Say what happens next and when. "A person will reply by email within one business day" is better than a spinner that never stops. If no one is available right now, say so, and offer the next best thing: a callback, an email, or office hours.
5. Say what the AI is, and isn't
Tell people they are talking to an automated assistant, what it can help with, and that it can make mistakes. Clear limits build more trust than a confident tone. For government and regulated services, also say where the information comes from and how to check it.
Test the hard paths, not just the happy path
Most chatbot testing checks whether common questions get good answers. Add sessions where people bring a messy, emotional or unusual problem, and watch what happens when the assistant reaches its limit. That is where the real design work is.
We help teams design and test conversational services, including the hand-offs, with the people who will actually use them.