Maaz Zindani

Loading the index…

↑↓ move↵ openesc close4072 pages

Help for owners

My AI chatbot keeps going when it should pass the customer to a person

Short answer

A handoff is a rule, not a feeling: the bot passes the conversation to a person when a trigger fires, such as an angry customer, a refund, a legal or safety topic, repeated misunderstanding or a direct request. Write the triggers down, give the person the full chat, and make sure someone is actually there to take it.

  • Written for business owners
  • Updated
  • By Maaz Zindani, Senior Software Engineer

Does this sound like your situation?

  • Customers type 'agent' or 'human' and the bot carries on.
  • People complain they are stuck going round in circles.
  • A refund, complaint or safety question was answered by the bot alone.
  • When it does hand over, my staff get no context and ask everything again.
  • The handoff happens at night and nobody sees it until morning.

Why does this happen?

  1. No written list of when to stop

    A language model does not know your escalation policy unless it is given one. Without explicit triggers it tries to answer everything, because answering is what it is built to do.

  2. The bot cannot tell it is failing

    A model usually sounds equally sure when it is right and when it is wrong. Repeated rephrasing by the customer, rising frustration or a topic outside its sources have to be detected by rules around the model, not by the model itself.

  3. The handoff is a dead end on the human side

    The conversation is passed to an inbox, a shared email or a chat tool that nobody watches out of hours, or it arrives without the history. The customer is then worse off than before.

  4. Some topics should never be answered by a bot

    Refunds, cancellations, complaints, medical, legal and safety questions, and anything involving money moving, carry a cost if answered wrongly. They need a person or an approved fixed answer.

What can I do about it today?

  1. Write down your stop topicsOn one page list the questions a person must answer: refunds, cancellations, complaints, anything about health, safety, legal or payments.
  2. Put a visible way to a person on the chatA plain line such as an email address or phone number, shown in the chat window, costs nothing and does not depend on the bot working.
  3. Decide who receives handoffs and whenName a person or a shared inbox, the hours they watch it, and what the customer is told outside those hours.
  4. Read the last conversationsOpen a sample of real chats, especially the long ones. The long ones are usually where a customer was stuck.
  5. Turn off the risky topics until the rules existIf the bot is handling refunds or complaints with no safeguard, switch it off for those topics. Do not delete the bot or its history.

When should I bring someone in?

  • A customer has already been harmed or lost money because the bot answered when it should not have.
  • The chat tool you use has no way to pass the conversation with its history.
  • You need the handoff to work across chat, email and phone.

How do I handle it?

  1. Turn your policy into triggersEach stop topic and each frustration signal becomes a rule: explicit request for a person, repeated failed answers, a sensitive keyword or intent, a missing source.
  2. Put the rules outside the modelTriggers are checked by ordinary code around the model, so they fire the same way every time and can be tested.
  3. Hand over the whole conversationThe person receives the chat, the customer's details and a short summary, so the customer does not repeat themselves.
  4. Handle the empty roomWhen nobody is available, the customer is told honestly when to expect a reply and the case is queued, not dropped.
  5. Test with real failed chatsOld conversations where customers got stuck are replayed to check that the handoff now fires.

The price depends on what I find, so I quote in writing after a short discovery call. Tell me what is happening and I reply within 24 hours.

What do owners ask about this?

Should the bot always offer a person?

A visible route to a person should always exist. Whether the bot offers it after one message or after three failed answers is a choice for your business.

Can the AI decide by itself when to escalate?

It can help flag a conversation, but the rule that actually transfers it should be explicit and tested, not left to the model's mood.

What if my team is too small to take handoffs?

Then reduce what the bot promises, publish a clear reply time, and use a queue. An honest wait is better than a bot that pretends to help.

How was this page made?

I drafted this page with AI assistance. It is general guidance for business owners, not advice for your own system. Found a mistake? Tell me.

Next step

Want this off your desk?

Send me a few lines on where things stand. I reply within 24 hours.