An AI receptionist speaks to your customers before you do, in your name, with your reputation. These are the questions we would ask a vendor if we were buying rather than building — and the answers that should end the conversation.
Every demo you will be shown is a call that goes well. That is not dishonest; it is just not informative. A receptionist — human or otherwise — is judged on the calls that go badly, and those are the ones a demo never includes.
So here are nine questions that get at the awkward parts. We have answered them for ZoiVox where the answer is a straight one, and left the answer out where it would be a promise rather than a fact.
This is the first question and the most important one, because the default behaviour of a language model under pressure is to produce something plausible. Not something true. Something plausible.
We have watched this happen. During development one of our agents hit a scheduling conflict it could not resolve and told the caller that the slot "just got taken" — a tidy, reasonable-sounding explanation that was simply invented. The real conflict was somewhere else entirely. No customer was harmed because it was our own test call, but that is the failure mode, and it is not rare.
What a good answer sounds like: the vendor can describe specific mechanisms — rules that forbid guessing, a defined fallback like taking a message or transferring to a person, and error messages designed so the model cannot paper over them. Walk away if they tell you the model is accurate enough that this does not come up.
A great many "AI receptionists" are a transcription service with a nice voice. They listen, they summarise, they email you, and then a human — you — does the actual work. That can still be worth paying for, but it is a different product from one that writes into your calendar while the caller is still on the line.
Ask to see a booking land in a real calendar during the demo. Ask what happens when the requested time is already taken, when the caller wants to move an appointment, and when they want to cancel. Rescheduling and cancelling are much harder than booking, and plenty of products only do the easy one.
ZoiVox: books, reschedules and cancels in Google Calendar during the call, and texts a confirmation.
When you connect a calendar you are handing over an access token, and most people click through the Google consent screen without reading which boxes it covers. Read it. The difference between scopes is the difference between "can manage appointments on the calendar I chose" and "can read everything in my Google Calendar".
ZoiVox: we request permission to read and write events, plus permission to list your calendars so you can pick one. We deliberately do not request Google's broader read-only calendar scope — our verification submission with Google says exactly that. A practical consequence you can verify: we check availability by listing events on the calendar you selected rather than using Google's free/busy endpoint, because free/busy requires the broader scope we chose not to ask for.
Some calls should not be handled by software. A dental emergency, an angry customer, a lawyer's client in distress. The question is not whether the AI can take a message for you — it is whether it can get a live human onto that call while the caller is still there.
Ask whether the transfer is warm: does the person picking up hear who is calling and why, or do they get a cold handoff and have to start over? Ask what happens if nobody answers the transfer.
ZoiVox: warm transfer to a number you set, with context on who is calling.
If a caller asks "am I talking to a real person?", there is one acceptable answer. Beyond the ethics, this is moving quickly in regulation and in platform policy, and a vendor whose selling point is that nobody can tell is selling you a liability.
Test it yourself on the demo. Ask the question three different ways — "are you a robot?", "is this a recording?", "are you a real person?" — because these are often handled by a rule that matches one phrasing and not the others.
Not if. A voice AI sits on top of a telephony provider, a speech-recognition service, a language model and a text-to-speech service, any of which can have a bad afternoon.
The question to ask is what the caller hears when that happens. The good answer is that the call falls back to something — forwarding to your real number, a voicemail, a human — rather than dying silently or leaving the caller listening to nothing. Ask specifically: "if your service is down at 9 AM tomorrow, what does my customer experience?" A vendor who has thought about this will answer immediately. One who has not will talk about uptime percentages, which is not the same question.
This is where we want to be especially pointed, because it is the easiest claim in the industry to make and the least often verified.
"Supports 30 languages" usually means the underlying model has those languages in it. It does not mean anyone has called the thing in Punjabi and listened to what came back. Those are very different claims, and the gap between them produces specific, ugly failures. Non-Latin scripts cost roughly three times as many tokens as English for the same sentence, so a response limit that is generous in English can cut a Hindi or Punjabi reply off mid-word. Speech recognition has to be explicitly configured for a language or it simply will not hear it. Both of those look completely fine in an English demo.
Ask: "has a human made a real phone call to this agent in that language and listened to the whole thing?" If the answer is vague, treat the language as unsupported no matter what the website says — and that standard should apply to us as readily as to anyone else. If a language matters to your business, make the test call yourself before you pay.
An oddly specific question that tells you a great deal about whether a vendor has run real calls.
When someone reads out a phone number, speech recognition does not hand over ten clean digits. It inserts punctuation it thinks it heard. A real transcript of a Canadian number can arrive looking like 6:07, 4:10, 4:16 — the recogniser has decided those pauses were timestamps. If nobody has dealt with that, the number written into your booking is wrong, and you find out when the confirmation text fails to arrive.
Names are the same problem in a different coat. Industry words get mangled — in our case a "court" was repeatedly heard as a "coach", which broke bookings until we gave each business a custom vocabulary list.
Ask: can I add my own terminology — treatment names, product names, the street my shop is on? And read a phone number out during the demo at normal speed, then check what it repeats back.
Headline pricing is for an average month. Ask about the other kind.
Specifically: what is the per-minute rate once the included minutes run out, is there a setup fee, what is the contract length, and — the one people forget — does the meter run on spam and wrong numbers? It does, for everyone. A robocall that your AI answers politely for forty seconds is forty seconds of billed minutes.
ZoiVox: plans run $69 to $329 a month with 150 to 1,000 included minutes, and per-minute overage from $0.45 down to $0.35 depending on plan. Setup is $200, currently waived for our first five Canadian customers. Month to month, no contract. It is all on the pricing section rather than behind a sales call.
Can I change what it says without filing a support ticket? Your hours change, you add a service, a price goes up. If every edit is a request to the vendor, you will stop making them and the agent will slowly start lying to your customers on your behalf. You want a screen you can edit yourself, and ideally a version history so a bad edit can be undone.
Where does my call data live, and what is kept? This one is long enough that we gave it its own post — under Canadian privacy law the customer data flowing through your phone line is your responsibility, not your vendor's, and a surprising number of vendors cannot tell you which country their database is in.
Phone the vendor's own demo number and try to break it. Interrupt it mid-sentence. Change your mind about the time twice. Give a name it will not expect. Ask a question that is not on their website. Five minutes of being an awkward caller will tell you more than any feature list — ours included, which is why our demo is on the front page with no signup in front of it.
Talk to ZoiVox in your browser, or have it ring your phone. No signup, no credit card.