AI Escalation Rules: When an AI Assistant Should Hand a Customer to a Person

A customer asks an AI assistant for a refund on a broken item. The assistant, trained to be helpful, explains the refund policy in four cheerful paragraphs and asks whether there is anything else it can help with. The customer wanted a person five messages ago. A badly handled moment like this does more damage than having no AI at all. AI escalation rules decide when the assistant steps aside, and they deserve as much design as its answers.

Why handover is the critical feature

An assistant that answers well ninety percent of the time is valuable only if the other ten percent reaches a person quickly. Without clear rules, an AI tends to keep trying. It apologises, rephrases and loops, because producing another reply is what it does. The customer experiences this as being trapped.

A good handover turns the AI's limit into a service: the customer feels that the business noticed their problem needed attention.

Eight triggers

Write these as explicit rules.

1. The customer asks for a person

Any form of "can I talk to someone", "admin please", "is this a bot". Hand over at once. Do not try one more answer first.

2. Complaint or strong emotion

Anger, repeated punctuation, words such as disappointed, cheated, terrible. Acknowledge briefly and hand over. A person should own the recovery.

3. Money decisions

Refunds, discounts, compensation, payment disputes, special prices. The AI may collect details. It should not grant or refuse.

4. Exceptions to policy

Requests outside the standard rules: delivery outside the usual area, a return after the deadline, a custom order.

5. Missing information

The answer is not in the knowledge base. The AI should say it does not know and pass the question on, never guess.

6. A loop

The same question has been asked, or the same answer given, more than twice. Something is not being understood.

7. High-value or sensitive customers

A large order, a business buyer, a customer flagged as important, or a sensitive subject such as health, safety or legal matters.

8. The AI is unsure

Low confidence about intent, ambiguous messages, or content it cannot read, such as an unclear image or voice note.

It is better to hand over slightly too often at the start. You can relax the rules once you see which handovers were unnecessary.

How to hand over well

A handover has three parts, and each can fail.

Tell the customer

Plain, short and honest:

"I will pass this to our team so they can help you directly. They usually reply within 15 minutes during working hours."

Do not say "I have resolved your issue". Do not promise a time you cannot keep. Outside working hours, say when the team returns.

Give the person the context

The team member should not have to read fifty messages or ask the customer to repeat anything. Pass along:

• A two-line summary of what the customer wants.
• What has already been said or tried.
• Key details collected: order number, product, photos.
• Why it was escalated.
• The customer's mood, if relevant.

Make sure someone owns it

A handover that lands in a shared inbox where nobody is responsible is a silent failure.

• Assign the conversation to a named person or a rota.
• Notify them.
• Flag it as waiting for a human.
• Alert a supervisor if it is not picked up within a set time.

After the handover

• The AI stops replying in that conversation unless the person hands it back. Two voices answering at once is confusing.
• The person replies as a person, ideally with their name.
• The conversation is closed deliberately, and the AI resumes only for a new topic.
• The reason is recorded, so that you can analyse it later.

Outside working hours

When no one is available:

• The AI says so and states when the team will respond.
• It collects the details needed, so the person can act at once in the morning.
• It records a ticket or flags the conversation.
• It does not pretend a person is on the way.

A customer who is told honestly "our team is back at 8 am and will see this first" is usually satisfied. A customer left wondering is not.

Enforce it in the system

Instructions in a prompt are a start. Reliable escalation needs rules the model cannot argue with:

• Keyword and pattern triggers for "talk to a person" in the languages your customers use.
• A hard limit on consecutive AI replies without resolution.
• No tool for refunds or discounts, so the AI cannot grant them even if it tries.
• A handover function that assigns, notifies and pauses the AI in one step.
• A check that the AI does not claim an action it did not perform, such as "I have forwarded this", unless the forwarding really happened.

Testing the rules

Add handover cases to your test set and run them before every change:

• A direct request for a person, phrased ten different ways.
• An angry complaint.
• A refund request.
• A question outside the knowledge base.
• The same question repeated three times.
• A request for a discount.
• A message at midnight.

Each should end in a handover with a correct summary. Any case in which the AI keeps answering is a failure to fix before launch.

What to measure

• Handover rate. The share of conversations passed to a person. Very low may mean the AI is holding on to things it should release.
• Time to pick up. How long the customer waits for the person.
• Reasons. Which triggers fire most. A frequent "missing information" trigger means the knowledge base needs work.
• Unnecessary handovers. Cases the person resolved with a standard answer. Candidates for the AI to learn.
• Missed handovers. Conversations where the customer was unhappy and no handover happened. The most important figure.

Review a sample of conversations each week. Numbers show the pattern. Reading the conversations shows the cause.

Improving over time

Each handover is a lesson. If the same question is escalated repeatedly because the answer was missing, add the answer. If handovers for a simple topic are always resolved the same way, let the AI handle it. The aim is not zero handovers. It is that every handover is one that a person should handle.

In Polynusa, a customer who asks to speak to a person becomes a ticket for your team, and anything involving money or identity always asks a person first, on every plan. See the pricing page.

Summary

Define explicit triggers: a request for a person, complaints, money, exceptions, missing information, loops, sensitive customers and uncertainty. Tell the customer plainly, pass a summary to a named person, and stop the AI in that conversation. Enforce the rules in the system, test them before every change, and learn from each handover.

FAQ

When should an AI assistant hand over to a human?

When the customer asks for a person, when the message is a complaint or shows strong emotion, when money, contracts or exceptions are involved, when the AI lacks the information, and when the same question has gone round more than twice. These should be explicit rules, not left to the model's judgement alone.

What should the AI say when handing over?

It should say plainly that it is passing the conversation to a team member, what will happen next and roughly when. It should not pretend that the handover has resolved the issue, and it should not promise a time that the team cannot keep.

What if nobody is available to take the handover?

The AI should say so honestly, give the hours when the team is available, and confirm that the message has been recorded for follow-up. A ticket or a flagged conversation must really exist, so that the customer does not have to explain everything again.