Your AI assistant needs a way to reach a person
87% of customers say a company using AI for support must still offer a human. What that means in practice: the three endings a stuck conversation can have, and what each one costs you.
Every assistant will eventually be asked something it cannot answer. That moment is not a failure — the failure is what most of them do next, which is nothing. The difference between a dead end and a conversation that continues is worth more than any amount of model quality.
The research is unusually blunt on this. Gartner (2026) found that 87% of customers say a company using generative AI for service must still offer a way to reach a person. And when it goes wrong, it goes wrong permanently: only 27% would try a chatbot again (2026) after one bad experience. You get one.
Three endings to the same conversation
Here is a question an assistant genuinely cannot answer, because the answer is in an order system it has no business reading. What it says next is a product decision, not a model decision.
Where is my order? It said delivered but nothing arrived.
I’m sorry, I don’t have information about that.
The middle option is the one most people skip, and it is the one that matters most for a small team. You do not need somebody sitting at a desk for an assistant to be useful when it is stuck — you need it to know your email address and your opening hours, and to say them without being asked twice.
An assistant that says “I don’t know” has answered the question. An assistant that says “I don’t know, and here is who does” has kept the customer.
The arithmetic of a dead end
The share of conversations an assistant cannot finish is the number nobody guesses correctly. Gartner's 2024 work on self-service found only 14% of issues fully resolved — and in 43% of the failures, the customer could not find relevant content at all. Even a very good assistant on a well-documented site will be stuck on a tenth of what it is asked.
What a good handoff actually does
- It does not restart the conversation. Whoever picks it up reads what came before. Making a customer repeat themselves is the single most reliable way to annoy them — Zendesk's 2026 trends work (2025) found 74% are frustrated by exactly that.
- It is honest about who is typing. The visitor should be able to tell a person from the assistant without asking. A reply from a human should carry a human's name.
- It does not promise a time it cannot keep. “Someone will reply here” is a promise you can keep. “Someone will be with you in two minutes” is one you cannot, at 11pm on a Sunday.
- It costs nothing while a person is in the chat. If the meter runs while a human types, the pricing is working against the behaviour you want.
- The visitor can ask for it. Not everyone will wait for the assistant to give up. A button beats a magic phrase.
If you are the whole support team
Most of the advice written about handoff assumes a rota. If you are one person — and plenty of good products are, including PixGuard, which is run solo — the shape is different but the principle holds. The assistant covers the dull repeated questions; it hands out your contact details when it is stuck; and when you are at a desk, you can step into a live conversation rather than answering the same thing by email tomorrow.
What makes that workable is being told. A conversation waiting for a person is worth nothing if you find out about it a day later, which is why the notification matters more than the inbox — Zinx OS makes the same bet for a small team's work, and for the same reason: the thing that needs you should come and find you.
In Zinx Chat, handoff is off until you turn it on — an assistant that promises a person nobody is going to send is worse than one that is honest about being alone. With it on, the visitor gets a button, the assistant can call someone in when it is stuck, and everyone in the workspace is notified on their phone. Without it, the assistant hands out whatever contact details you have set. Both are on every plan, including the free 25 replies.
Common questions
- Does an AI chatbot need a human handoff?
- Yes. Gartner found in 2026 that 87% of customers say a company using generative AI for customer service must still provide access to a human agent, and only 27% would try a chatbot again after one bad experience. An assistant with no exit converts a stuck conversation into a lost customer.
- What should a chatbot do when it cannot answer a question?
- Three things, in order: say plainly that it cannot answer, offer a way to reach the team, and — if someone is available — bring them into the same conversation rather than starting a new one. The visitor should never have to explain themselves twice.
- Do I need live agents for AI chat to work?
- Not necessarily. A small team can publish contact details and let the assistant hand them out, which still rescues the conversation. Live takeover matters most when somebody is usually at a desk; contact details matter always.