What voice AI does when it doesn’t know the answer
Any sales demo of a voice agent shows it succeeding. That is what a demo is for, and it is also why demos are close to worthless for choosing one. The information you need is in the other cases: what it does with a question it was never told the answer to, a name it did not catch, or a request outside what it is allowed to handle. A system that handles those badly is not slightly worse than one that handles them well — it is a different category of thing to put in front of your customers.
The two ways of being wrong
A system that does not know something can fail loudly or quietly, and the difference matters more than the frequency.
Loudly means saying so. "I don’t have that — let me take your number and have someone call you back." The caller is mildly disappointed, the business gets a lead with a note attached, and nobody has been misled.
Quietly means answering anyway, fluently, with something plausible and untrue. A price that is not your price. An opening time you do not open at. A service you stopped offering last year. The caller believes it, acts on it, and turns up — and the first time you find out is when somebody is standing at your counter quoting a number nobody at your business has ever said.
Language models are good at fluency, which makes quiet failure their natural mode unless it is designed against. This is the single most important thing to check and the least likely to come up in a sales conversation.
What designing against it looks like
Three properties, in roughly this order of importance.
- A bounded set of facts. Prices, hours, services and policies come from what the business actually told the system, not from what the model finds plausible. Anything outside that set is not something to answer.
- Confirming rather than assuming. Phone numbers, names, dates and amounts get read back to the caller. This costs three seconds and prevents the failure that is most expensive to discover later — a booking attached to a number nobody can reach.
- An escalation path that is used. Anything requiring judgement — clinical, legal, financial, or an angry customer — should stop and hand over with a written summary, rather than being attempted well.
The five-minute test
Run this against any product you are evaluating, including ours. It is faster than a sales call and considerably more informative.
Ask something they cannot have configured — an obscure question about a service, a policy nobody would have thought to specify. Listen for whether you get an admission or an answer. An answer is a red flag, however good it sounds.
Give a phone number quickly, in the middle of a sentence, and see whether it reads the number back. Then give a name that is not common in the language you are speaking, and check the same.
Ask something that clearly needs a person, and see whether it hands over or has a go.
Finally, interrupt it mid-reply. A system that keeps talking over you has not been built for phone calls, whatever else it does well.
Why nobody advertises this
Because the honest version sounds worse in a pitch. "It tells callers when it doesn’t know" is a weaker sentence than "it handles any question", and the second one is what wins the meeting.
It is also genuinely harder to build. Getting a model to answer is easy; getting it to reliably decline, in the middle of a conversation, at speed, in five languages, is the part that takes the time — and the caller never sees the work, because a correct refusal sounds like nothing happening.
The bar we would suggest holding vendors to is not accuracy. Every system in this category will occasionally mishear a name in a noisy shop. The bar is what happens next: whether it asks, or whether it commits.