What your AI agent should not answer

Most advice about AI agents is a list of what they can do. This is a list of what they shouldn't, which we think matters more, because the questions an agent handles badly do far more damage than the ones it handles well do good. You can answer "do you ship to Ireland?" perfectly a thousand times and it's worth less than one botched refund reply is expensive.

The instinct, once you've got an agent that answers well, is to let it answer everything. Resist it. The skill isn't breadth. It's knowing where to stop, and building an agent that stops gracefully instead of guessing confidently. A good "let me get a person for that" is not a failure of the agent. It's the agent working correctly.

Here's where we draw the line, and why.

Anything about money moving the wrong way

Refunds. Disputed charges. "I've been billed twice." "I cancelled and you charged me anyway." These should almost never get an autonomous answer, and it's not because the agent can't find the policy. It's because these are decisions, not lookups, and the customer is usually already unhappy when they ask.

An agent that reads your refund policy and recites it to someone who's just been double-charged is technically correct and completely tone-deaf. The person doesn't want the policy. They want their money back and they want to feel heard. Even when the answer is "no," a human saying it lands differently, and a human can make the small exceptions that keep a customer — the ones no policy page will ever contain. Let the agent recognize the topic, express that it's on it, and hand straight to a person with the full context attached. Don't let it adjudicate.

Legal, security, and anything about someone's data

"Can I get a copy of everything you hold on me?" "Are you GDPR compliant?" "What happens to my data if I delete my account?" "Has there been a breach?"

These deserve a careful, correct, on-the-record answer, and an AI agent free-associating from your privacy page is the opposite of that. The stakes are asymmetric in the worst way: get it right and nobody notices, get it slightly wrong and you've made a commitment you can't keep or admitted to something that isn't true. Security and legal questions are exactly the category where "roughly right, phrased confidently" is more dangerous than silence.

Our rule is that anything touching law, security, or a specific person's data goes to a human every time, even when the agent "knows" the answer. Especially when it knows the answer, actually, because that's when it's most tempted to just say it.

The angry customer

This one's less about the topic and more about the temperature. Someone who's furious does not want to talk to a machine, and every extra AI reply they get while angry raises the heat. It reads as a company hiding behind automation to avoid dealing with them, which is often exactly what it is.

An agent should be able to notice frustration — the all-caps, the "this is ridiculous," the third message in a row — and treat that as its own signal to step back and bring in a person, regardless of whether it technically has the answer. Handing an angry customer to a human quickly is one of the highest-value things automation can do. Making them argue with a bot first is one of the worst.

The honest heart of it

A confident wrong answer is worse than no answer, because the customer acts on it. "I don't know, let me find someone who does" costs you nothing and buys you trust. Guessing costs you the trust and sometimes the customer.

This is the whole argument, really. The failure mode of AI agents isn't that they don't know things. It's that they don't know that they don't know, and they say the wrong thing in the same fluent, assured voice they use for the right things. The customer can't tell the difference. They act on it. And then you're not just fixing a wrong answer, you're rebuilding trust in every answer.

Which is why we built Engine64 to answer only from the content you actually give it, to show its source on every reply, and — the part that matters most for this piece — to say plainly when something isn't covered and pass it to your team instead of improvising. Not because refusing is impressive. Because the alternative, an agent that would rather be confidently wrong than admit a gap, is a liability wearing the mask of helpfulness.

Drawing your own line

You don't need our exact list. You need a line, drawn on purpose, before you turn the thing on. A quick way to find yours: look at your recent support conversations and ask, for each one, "would I be comfortable with this being answered by software while I was asleep?" The ones where you flinch are your no-go list. Usually it's money, law, data, and anyone who's angry — but your business will have its own, and you'll know them by the flinch.

Then build the hand-off to be genuinely good. Not a dead-end "please email support," but a warm pass that carries the whole conversation across so the customer never repeats themselves and your team picks up mid-stride. The measure of a good agent isn't how rarely it hands off. It's how well it hands off when it should — and how reliably it knows when that is.