If you are going to put an AI in front of your customers, there is one question you should ask any vendor before all the others: what happens when it doesn't know the answer?
It is not a technical question. It is the business question. Because a language model's natural response to a gap in its information is not silence. It is to fill the gap.
Why AI completes what it doesn't know
A language model works, at bottom, by continuing text in the most plausible way it can. That is what makes it so good at conversation: it always has something coherent to say.
But note the word: plausible. Not true. When a customer asks "do you do this model in white?" and that information is nowhere the model can reach, the most plausible continuation of that conversation is "Yes, it's available in white." It sounds like what a shop would say. It fits. And it may be false.
The model is not lying in the human sense. It does not know that it does not know. It is simply completing the pattern. Your customer does not care about the distinction: you have just confirmed a white model that does not exist, and someone on your team will have to unpick it — possibly after the order has gone in.
"Behave yourself" is not a guarantee
The first fix everyone reaches for is to write it into the instructions: "Do not invent prices. If you don't know something, say so."
It helps. It genuinely helps — it makes the problem rarer. But it is worth being clear about what it is: a request, not a limit. The model weighs it alongside everything else — the tone of the conversation, a pushy customer, the thousand other instructions piling up in the prompt — and most of the time it complies. Most of the time.
In a sales conversation, "most of the time" does not cut it. Nobody would hire a salesperson who quotes correct prices 97% of the time and invents the other 3% with total confidence. The problem with that salesperson is not the 97%.
There is a second line of defence that sounds better: a detector that reviews the replies and catches the fabrications. Better, yes. But a detector built from the same kind of technology is best-effort: it catches a lot, it does not catch everything. It is still probability, not guarantee.
What an actual guarantee looks like: the hard floor
The distinction we care about is this one: between asking the system to behave, and building it so it cannot misbehave.
A deterministic hard floor is a check that does not depend on the model. It works like a circuit, not a judgement call: if the reply contains a price, that price must exist in the real catalogue, character for character. If it names a product, that product must be on your list. If the check fails, the reply does not go out — it gets replaced, reworded, or turned into a "let me confirm that for you". There is no accuracy percentage to argue about, because it is not a prediction. It is a gate.
That is how Zatio is built. The model does the conversing, but prices, stock and products pass through floors the model cannot talk its way past, no matter who is asking.
Look it up, don't remember it
There is a second principle, simpler and arguably more important: the AI should not remember your catalogue. It should look it up.
If the catalogue lives inside the model — because someone pasted it into the instructions three months ago — your bot is selling at three-month-old prices. You change a rate, discontinue a line, run out of a finish, and the AI keeps answering from the old snapshot, with complete confidence.
The alternative is that on every conversation the agent consults the live source — your catalogue, your stock, your current terms — and answers with what is true today, not what was true when someone wrote the prompt. It is the difference between an employee who read the manual once and one who has the database open in front of them.
The commercial value of "I don't know"
Which leaves the unavoidable case: the question whose answer is nowhere. Do you also cover Northern Ireland? Can you do a bespoke size?
There, the right answer is the one a good new hire would give: "I don't actually know that one — let me put you through to the team." Zatio says exactly that, and hands the conversation to a human with the full context — the whole thread, not a bare notification.
It can look like a weak answer. It is the opposite. An honest "I don't know" is what makes every preceding "yes" believable. The customer who watches the bot admit a limit trusts the rest of what it said. The one who catches it fabricating never believes it again — nor, while we are at it, your business.
We ask exactly this of people: do not assert what you do not know. There is no reason to ask less of a system. The difference is that with a person, you can only ask. With a system, you can guarantee it — if it is built for that. It is also, incidentally, the spirit of the transparency rules now applying to chatbots in the EU: say what you are, and don't pretend to know what you don't.
If you want to see hard floors working over a real catalogue, get in touch. Ask our instance about a product that doesn't exist, and see what it says.