Questions AI should never answer in customer support

Write this list before you write the greeting

Picture the failure this list exists to prevent. Your chat widget tells a customer that a return is fine. Your policy says thirty days; the order is eleven weeks old. The customer has a screenshot of the conversation. You pay the refund, because the alternative is arguing with your own words in public.

Nothing was broken. The AI was helpful, fluent and wrong, and it was wrong in the one category where being wrong costs real money.

Searchers phrase this as "questions AI should not answer customer support", and the honest version of the list is short — five categories, plus two catch-all rules. They are in this order deliberately: ranked by how much damage a single wrong answer does, which is also the order to write them into your escalation rules.

1. Anything past your written policy

The rule. If the answer requires making an exception, a person makes it.

Refunds after the window. A replacement on something that looks worn. A goodwill discount because a delivery was late.

Why it is first. These are the only support answers that move money directly, and they are the ones a customer will screenshot. An AI exception is binding in practice, whatever your terms say, because you will not win the argument that followed your own chat widget saying yes.

What goes wrong. Models are agreeable by default. Asked twice, politely, for something just outside the rules, the likely failure is not refusal — it is accommodation.

How to write it. Give the AI your policy as text it can quote and nothing else. Any question whose answer is a number or a date not in that text escalates. "Our policy is thirty days, and I am not able to make exceptions — let me get someone who can" is a good answer. It is also an honest one.

2. Anything legal, medical, financial or about safety

The rule. Route it, every time, with no attempt at a partial answer.

Is this safe during pregnancy. Does it contain nuts. Is this electrical rating OK for my country. Will this interact with a medication. Is this tax deductible. What does your warranty legally entitle me to. Is this fabric flame retardant.

Why it is this high. This is the only category where a wrong answer can hurt somebody. Everything else on this list costs money or goodwill.

What goes wrong. A model will reason from the product description to a conclusion the description does not support. "Made with natural ingredients" becomes "suitable for sensitive skin" in one plausible step, and that sentence is now your store's position on a medical question.

How to write it. List the words that trigger an automatic handoff: allergy, allergen, pregnant, medication, child, safety, voltage, legal, warranty, tax. Crude keyword rules are the right tool here, because the cost of escalating unnecessarily is a mildly inconvenienced customer and the cost of a miss is not.

3. An angry customer

The rule. Detected frustration escalates before it answers anything else.

Why. An angry customer is not asking a question. They are asking whether anybody at your company cares. There is no sentence a bot can produce that answers that, and a fluent, calm, perfectly reasonable reply reads as contempt.

What goes wrong. The AI addresses the literal question and ignores the emotion. Then the customer repeats themselves louder, and your transcript becomes the thing they post.

How to write it. Escalate on profanity, on all caps, on the words refund, complaint, lawyer, review, Trustpilot, chargeback, scam, and on any message that is the third in a row from the same person on the same issue. Escalate early and apologize for the interruption later.

4. A mistake you made

The rule. If the store is at fault, a person responds.

A late order. The wrong item. A charge that went through twice. A product that arrived broken. A price that was wrong on the page.

Why. An apology that nobody chose to make is not an apology. The repair is not information, it is accountability, and a model cannot be accountable — it has no authority to promise anything and the customer knows it.

What goes wrong. The AI apologizes in the voice of the brand and offers the standard remedy. The customer, correctly, hears a form letter. What would have fixed this is one sentence from a named human saying "that was our error, here is what I am doing about it."

How to write it. Any conversation that contains an acknowledgment of store error routes immediately — including when the AI itself is about to acknowledge it.

5. A promise the AI cannot verify

The rule. If the answer is a commitment about the future that your own data does not contain, it does not get made.

Will it arrive before Friday. Will you restock this. Is this really handmade. Can you hold this for me until payday.

Why. These are the most innocent-looking entries on the list and they generate the most disappointment per sentence, because the customer heard a yes and planned around it.

What goes wrong. Carrier estimates get restated as guarantees. "Usually two to three days" becomes "it will be there Thursday", and Thursday is now a promise your store made.

How to write it. Let the AI state facts with their source attached — "the carrier's estimate is two to three business days" — and never let it convert an estimate into a date. Anything about restocks, holds, custom requests or special handling escalates.

The two rules that catch everything else

Asked for a person once, transferred once. No "let me see if I can help first." The customer has already decided and every extra turn is a cost.

Asked twice, escalate. If someone has rephrased the same question, the AI has lost, whatever its confidence score says. Rephrasing is the cheapest and most reliable failure signal you have.

What the AI should say when it stops

A handoff done badly is nearly as damaging as a wrong answer, because the customer experiences it as a refusal.

Three properties make the difference: it says why it is stopping, it names a time, and the customer does not have to repeat themselves.

How to test the list

Message your own store — and call it, if you are on a plan with a phone number — with the five worst questions you can invent. One per category. Then do it again next month, because a model's behavior changes when its instructions or version do.

If any of the five gets answered instead of routed, you have found a rule to write today, before a customer finds it for you.

What the list costs you, honestly

Some customers will be routed to a human who could have been helped by the AI, and they will wait longer than they needed to. That is the trade, and it is a good one. Over-escalation costs you a slower reply. Under-escalation costs you a refund you did not owe, a public review, or a safety answer you had no business giving.

In mAutomate the underlying principle is that the AI drafts and recommends and nothing goes live until you approve it, which is the same instinct this list applies to support: the model does the volume, you keep the judgment. The AI Call Center answers on your website on Pro, $149/month, and on a real phone number on Scale, $349/month. Every plan comes with a 14-day free trial, no card required — see pricing or get started.

For the decision one level up — which questions to automate at all, and when answering personally is still your advantage — see this guide. For what happens on a voice call specifically, this one.