What an "AI receptionist" really is
They cost wildly different amounts and fail in different ways. Work out which one you are being quoted for before you agree a price.
"AI receptionist" is four products wearing one name. The gap between the cheapest and the most expensive version is enormous, and most of the disappointment I hear about comes from someone buying one and expecting another.
A chat on your website. Answers questions from your own material, takes a booking or an enquiry, hands off when it does not know. This works today, it is the cheapest to run, and it is the one I recommend first almost every time.
A chat on WhatsApp or Instagram. Same brain, much harder plumbing. Every platform has its own approval process and rules about who may message whom first. Doable, but it is a project with a waiting period, not a switch you flick.
Something that answers the phone. A different animal entirely. Voice adds delay, accents, background noise and a caller who cannot see what is happening. It works for short repetitive calls and it fails loudly, in public, in the caller's ear.
Something that "handles everything". This does not exist for a business your size, whatever the demo showed. The clearest signal in this year's write-ups is that generalist agents promising to replace several roles at once are the ones being cancelled.
Ask which of the four you are being quoted. If the answer wanders, that is the answer.
The three things that decide whether it is useful or embarrassing
It answers only from your material. Not from the internet, not from what it knows about businesses like yours. If it has not been given your prices, services and process, it will produce something plausible and wrong, and your customer cannot tell the difference. You find out when they arrive expecting a price you have never charged.
It can say it does not know. The most valuable sentence such a system can produce is "I am not sure, let me have someone call you." A system that always has an answer is a system that is sometimes confidently wrong, and confidence is exactly what makes a wrong answer expensive.
What it captures lands somewhere real. A conversation that ends with no record is worse than no conversation, because now the customer thinks they have told you. The name, the number and what they wanted belong in the customer record automatically, not in a transcript nobody opens.
The part nobody quotes for: the first month
Every one of these I have set up has answered two or three things badly in its first month. Not catastrophically, just wrong enough to matter. Every single one was fixed by editing a document, never by changing the technology.
So budget an hour a week for a month to read what it actually said. If your supplier is not offering that, they are selling you the demo, not the system.
What it is genuinely good at
The questions that eat a front desk and never change. Where are you, what time do you open on a Saturday, do you do this service, roughly what does it cost, do I need a referral, what should I bring, can I get in this week.
In most businesses that is the bulk of contacts, and every one it handles is an interruption that never reaches your team. That is the whole promise, and it is enough.
What it should not touch
Anything clinical, legal or financial where being wrong costs money or safety. Anything needing judgement about one person's particular situation. And complaints, always. The one guaranteed way to make an unhappy customer angrier is a machine being cheerful at them.
On cost, briefly
Running one of these is cheap if it is built carefully and expensive if it is not, and the difference is not the model. It is how much material gets sent with every single question.
I had an assistant that was sending its entire library to answer anything, which worked fine for one team and would have cost about five thousand US dollars a month at the scale its owners were planning. Rebuilding it to send only the handful of passages relevant to the question took the cost to roughly a third of a US cent per question, about a tenth of what it was, with better answers because there was less noise around the right passage.
Ask your supplier what a question costs and what happens to that number at ten times the usage. If they do not know, they have not built it to scale, and you will find that out on a bill.
The order I would do it in
Website chat first, in text, answering from your own documents, with a clean handoff. Read the transcripts for a month and fix the material. Then consider messaging platforms. Treat voice as a completely separate decision, with its own budget and its own patience.
That order costs the least, fails the most quietly, and tells you honestly whether the loud version is worth it for your business.