How it works2 min read
What happens when it does not know
Every assistant meets a question it cannot answer. What it does next is the most important thing about it, and the least demonstrated in any sales call.
Published 13 May 2026 · Updated 15 August 2026
Demonstrations show the happy path. Your customers will not stay on it. They will ask about the thing you do once a year, describe a situation that fits no category, or want a decision only you can make.
The behaviour at that boundary is what separates a product you can put in front of customers from one you cannot.
The right behaviour
Say plainly that it does not know. Offer to take the details and have somebody ring back, or transfer to a person if one is available. Then actually do that, immediately and visibly.
That sequence costs the business nothing. The caller has been heard, the enquiry is captured and a person is involved. Very few people are annoyed by an honest I will find out.
The wrong behaviour, and why it is worse than silence
Producing a confident, plausible, incorrect answer. This is the failure that matters, because your customer will believe it, act on it, and hold you to it.
A wrong price quoted on the phone becomes a price you either honour or argue about. A wrong statement about whether you can treat something becomes a complaint. This is genuinely more damaging than an unanswered call, and it is the one thing worth testing hard before buying.
Easier to hear than to read about
Callwise answers your phone the way this article describes. Ring it, ask it something awkward, and see whether it holds up.
Talk to our AI assistantHow to test for it
Ring a demo line and ask about something adjacent but outside its knowledge. Ask for a specific price for an unusual variant. Ask whether it can do something you suspect it cannot.
Listen for hedging that is actually a guess. Probably around and I believe it is are both worse than I do not know, because they carry authority without accuracy.
Configuring the boundaries
Be explicit about what it must never answer. For clinical, legal and financial businesses this is the first thing to write down and the least negotiable.
Give it permission to fail. A line telling it that saying it does not know is always acceptable, and guessing never is, changes behaviour more than most people expect.
Escalation paths that work
Have more than one. Transfer to a person if it is working hours and someone is free. Take a message and alert immediately if not. For anything urgent, both.
Then check the escalations weekly for the first month. They tell you exactly what is missing from the assistant's knowledge, and each one you fix is a question it can answer next time.
Questions people ask
How often does it happen?
Early on, more than you would like, and it drops quickly as you feed the gaps back in. The escalation log is the single most useful thing to read in the first month.
Can we make it always transfer instead?
You can, during hours when somebody is available. Out of hours there is nobody to transfer to, which is where taking a proper message matters.
Read next
Do they actually sound human?
The voice is rarely the problem. What gives most of them away is timing.
Setting one up, step by step
The technical part takes about ten minutes. The part that decides whether it works is writing down what your business says to callers, which most businesses have never done.
How to choose an AI receptionist
Almost every product in this category demos well on a scripted call. The differences show up on the calls nobody scripted.
Hear it answer a call
Ring the assistant and ask it the awkward question your customers always ask. If you like what you hear, it will book you in there and then.
Talk to our AI assistantNo form, no sales call. Thirty seconds on the phone with the thing itself.