Yes on the calls that follow a pattern, and in a restaurant that is most of them: Skai Solutions handles bookings, hours, location and takeaway orders after 5 days of setup, then hands the odd, upset or unclear call to a person.
Four pieces, and understanding them tells you exactly where it breaks.
Our fuller walkthrough of the category sits on the AI receptionist guide, which is the page to read if you are still working out whether the category is for you at all. This page is narrower: it is about whether the thing works.
Four failure modes, in the order they will actually bite you.
Accents and noise. Speech recognition is the floor under everything else. If the transcription is wrong the language model reasons perfectly about the wrong sentence. A busy dining room behind the caller is worse than a strong accent.
The unusual request. A caller asking whether you can do a gluten free birthday cake for 14 on Thursday with a high chair and street parking is not a pattern. A good agent recognises that it is out of depth and takes a message. A bad one confidently books a table for 14 and tells the caller the cake is fine.
Emotion. Complaints, a caller in distress, a supplier chasing money. These calls need a person and a competent system should route them to one fast.
Silent failure. The worst one, because you never see it. The booking is taken, the caller hangs up happy, and nothing reaches your diary because the integration quietly stopped working three weeks ago. Ask any vendor how you would find out.
Most of the statistics in this industry are folklore. The number you will see quoted everywhere is that 60 percent of callers hang up after 45 seconds on hold, attributed to an unnamed study. We traced it and could not find a study behind it at all, so we do not use it.
What does exist is a peer reviewed census rather than a survey. Brown, Gans, Mandelbaum and colleagues analysed more than 1,200,000 calls to a call centre over a full year and published it in the Journal of the American Statistical Association in 2005. Their findings: about 20 percent of callers requesting service abandoned before being served, mean caller patience was 446 seconds, and there were two distinct spikes in hanging up, one in the first few seconds as the hold message starts and a second at about 60 seconds when the message repeats.
Read that carefully, because it cuts both ways. Callers are more patient than the industry claims, roughly seven minutes on average rather than 45 seconds. But one in five still leaves, and there is a real cliff at the one minute mark. In a restaurant the caller has an easier alternative than waiting: the venue two doors down. So the honest case for answering every call is not that people hang up instantly. It is that a fifth of them leave, and you never learn which fifth.
A scripted demo tells you nothing. These take twenty minutes and tell you everything.
The same tests apply to any option in this category, including a human service. Our comparison of what these services actually are sits on what is a telephone answering service, and the mechanics of the older human staffed model are covered in how a virtual receptionist works.
Three groups, and we would rather say so now.
If your phone rings a handful of times a day and someone is always within reach of it, the maths does not work. Buy nothing.
If almost every call is an enquiry that ends in a message rather than a booking or an order, a basic message-taking plan does the job for a fraction of the price and it is not us. Those plans advertise from about NZ$19 a month as at August 2026. That is the right buy for a business whose phone carries messages rather than revenue.
And if your bookings all arrive through an online platform and the phone is genuinely dead, fix the platform, not the phone.
Where it does work is the venue in the middle: a restaurant, cafe or takeaway where the phone rings through service, where the person nearest it is carrying plates, and where a missed call is a lost booking or a lost order rather than a lost message. That is the case set out in full on our AI phone answering service for restaurants page.
Book a demo, then ring the number yourself with the worst call you get all week and see what happens.
Book a Free DemoIt depends entirely on the speech model behind it. Most are trained on North American English and degrade on a broad New Zealand vowel or a fast Australian speaker with background noise. This is the single most important thing to test, and the easiest, because you can simply ring it.
A properly configured agent should fail over to a person or take a message rather than guess. Guessing on a booking is worse than not answering, because a wrong table at 7pm on a Saturday costs you the cover and the reputation.
Yes, and this is where it earns its keep in a takeaway, because those calls are short, repetitive and follow a fixed menu. Long custom orders with substitutions are where accuracy drops.
Five days at Skai Solutions, which covers the menu, hours, the booking rules and testing against your own phone number.
Most will work it out within a sentence or two, and that is fine. What annoys people is not the machine, it is a machine that loops, mishears and will not let them reach a person.