Every AI receptionist sounds good in a demo. The vendor calls it, asks for a haircut at two, and it books one. That tells you almost nothing, because the demo is the one call it was built to pass.
What matters is what happens when a real caller is vague, interrupts, mumbles a service name, or has three appointments and wants to move one. We know where those calls go wrong because we found almost every failure below in our own system over the past few weeks — on test calls, before a customer did. They're fixed. They're also exactly the things worth testing on anyone else's.
Make these five calls — to your own shop if you're on a trial, or to the vendor's demo line — and listen for the specific failure.
1. Ask for "the pedicure one"
Book two or three appointments for yourself, then call and ask to move or cancel "the pedicure one".
Speech-to-text turns the word "one" into the digit 1 more often than you'd expect. On one of our test calls, "the pedicure one" arrived as "The Pedicure 1", the system treated the 1 as an appointment number, moved a different booking entirely — and confirmed it as the pedicure.
Listen for: does it tell you which appointment it found — the service and the time — before it changes anything? If it has more than one and isn't sure, does it ask? A system that confidently changes the wrong booking is worse than voicemail, because nobody finds out until the client arrives.
2. Ask for a time with an AM/PM trap
Ask for "2pm", ideally on a line with a little background noise.
We watched our transcription return "2 AM" for "2 PM" — with high confidence. Most systems will work out that a salon isn't open at 2am. What matters is what they do next.
Listen for: the good answer is "we're not open at 2am — did you mean 2pm?" The poor answer recites your opening hours and makes you pick a time all over again. The bad answer books the nearest slot without mentioning that it changed your time.
3. Ask for "tomorrow afternoon"
A vague window is a completely normal request, and a surprisingly good test.
Listen for: does it offer times that are actually in the afternoon, spread far enough apart to be a real choice — something like 1:00, 2:30 or 4:00? We caught ours offering 12:00, 12:15 and 12:30: technically afternoon, practically one option said three times. How many appointment times to offer a caller explains why that matters.
4. Interrupt it
Ask what services it offers and, while it's still listing them, cut in with "actually, wait".
Listen for: does it stop within about half a second and let you talk? And then — the part people miss — does it do something sensible with what you said?
On our system, "actually, wait" was once transcribed as "Ashley Way", which sounds exactly like a name, and the caller's profile was renamed. The interruption worked. What it did with the interruption didn't. A good system treats a name that turns up mid-interruption as something to confirm, never something to save.
5. Ask for something it can't do
Ask a question only a person should answer: a price for something that isn't on the menu, whether a technician can fix a nail done at another salon, or — for a restaurant — whether a dish can be made without nuts.
Listen for: does it say it doesn't know and offer to have someone call you back, or does it make up an answer? An assistant that invents a price or reassures a caller about an allergen is the most dangerous kind, because it sounds right.
What you're really testing
All five come down to one question: does it know when it isn't sure?
Every system mishears. Ours did, on four of these five, until we fixed it. What separates a usable receptionist from a liability is whether a mishearing turns into a question — "did you mean the pedicure at 2?" — or into a confident change to your book.
All five take about ten minutes. If you want to try them on ours, the demo line is on the demo page. And if you're weighing it against a human answering service instead, the cost difference is laid out here.