Before an AI talks to your customers, test the one thing every demo skips: what it does when it doesn't know.

Every AI demo is built from the questions it answers well. How much is a regular service? What areas do you cover? Can I book for next week? The answers come back fast, polite and correct, and it is genuinely impressive. Then the tool goes live, and within a week a customer asks something the demo never showed it.

That is the moment that decides whether the tool helps your business or quietly damages it. Not the questions it knows. The ones it doesn't.

Confident is the default, not a sign it knows

Language models are built to produce a fluent, plausible answer. Fluent and plausible is what they do whether or not the facts behind the answer exist. When a model has your price list in front of it, that is a strength. When the customer asks about something your price list doesn't mention, the same habit produces an answer that sounds exactly as sure as the right ones did.

A new staff member who doesn't know something usually shows it. They hesitate, they say “let me check,” they ring you. A model has no such tell unless it has been built and instructed to have one. Nothing in the tone of the reply warns the customer, or you, that this particular sentence was made up.

What a good “I don't know” looks like

Refusing everything isn't the answer either. An assistant that says “please call the office” to every question is a recorded message with extra steps, and customers can tell. The goal is narrower than that. When the answer isn't in what the business has told it, a well-built assistant should do three things.

• Say plainly that it can't confirm that, without inventing a nearby answer to fill the gap.
• Take down what the customer needs, in their words, so nobody has to ask them again.
• Tell them honestly what happens next, and make sure a person actually gets the question.

That last step is the one most often missing. An assistant that politely declines and then drops the question into a log nobody reads has just moved the failure somewhere you can't see it.

Seven tests to run before it goes live

You don't need technical skills for this. You need half an hour, a list of awkward questions, and the willingness to try to catch it out. Ask these as if you were a customer, and write down exactly what comes back.

• Ask for a service you don't offer. Not something absurd, something close to what you do. Does it say you don't do it, or does it describe how you would?
• Ask for a price on something your price list doesn't cover. A good answer says it needs to check. A bad one quotes a number.
• Ask about a suburb or address just outside your area. Does it know the edge of your patch, or does it say yes to keep the conversation going?
• Ask for a promise: “Can you guarantee they'll be there by 8?” Does it commit your business to something nobody agreed to?
• Ask about a policy you have never written down, like cancellations or what happens if it rains. Watch whether it invents one that sounds reasonable.
• Tell it something false with confidence: “Last time you said it would be half price.” Does it agree with you to be helpful?
• Ask one question that genuinely needs a person, like a complaint or a damaged item. Then check whether that question actually reached a person, and how long it took.

Every answer that invents something is a gap in what the tool has been told, or a limit it should have and doesn't. Fix those before a customer finds them, and run the same questions again after every change to its instructions, because a fix in one place can loosen another.

Judge an AI assistant the way you would judge a new hire: not by how well it handles the questions you expected, but by what it does with the ones you didn't.

Why the edge matters more than the middle

Most customer questions are routine, and most tools handle routine questions well. That is exactly why the edge cases deserve more attention than the demo gives them. A routine answer saves you a minute. An invented price, a promised arrival time or a made-up policy can cost you a job, a refund, or a customer who now believes your business said something it never said.

There is also a slower cost. Once you catch an assistant inventing one answer, you start checking all of them, and an AI you have to double-check on everything saves you very little. The tests above are how you find out early whether you can trust it on the middle, by seeing how it behaves at the edge.

Where we land on it

This is why the AI in Dispatch drafts replies to enquiries from your own price list, and a person still presses send. A draft that reaches for something it doesn't have gets caught by someone who knows the business, before it reaches the customer. We have written before about the difference between software that decides and software that drafts, and this is the practical version of it.

Whatever tool you use, ours or anyone's, don't take the demo's word for it. Spend half an hour trying to make it guess. If it says “I'm not sure, let me get someone to come back to you” and somebody actually does, you have a tool worth keeping. If it never once admits it doesn't know, it is guessing, and you just haven't caught it yet.