Setting up an AI receptionist properly takes an afternoon. Setting it up badly takes twenty minutes, and the difference shows up as bookings that do not fit your day.
Here is the order that works, and the two steps people skip.
1. Write down what you actually sell
The step everybody underestimates, and the one that determines whether any of it works.
For each service: what it is called, how long it really takes, what it costs, and who can do it.
Two things go wrong here. The durations are the ones in your booking system rather than the ones that are true — if a fill takes fifty minutes and the system says thirty, the phone will build a day that collapses by two o'clock. And the qualifications are assumed rather than recorded, so the system offers colour correction with someone who does not do it.
If this information lives only in one person's head, it has to come out before anything else happens. That is genuinely worth doing regardless of what you do about the phone.
2. Set three kinds of hours, not one
These are different and people set them all to the same thing:
- When you are open — what the public is told
- When someone can actually reach a phone — narrower, because your team is with clients
- When a transfer to a person can succeed — narrower still
Then put your holidays and closures in. A system that does not know you are shut on a public holiday will take bookings for it.
More on the trade-offs in setting answering hours.
3. Decide what it must never do
Write the list before you go live, not after an incident.
For most businesses: no allergy or medical questions, no prices that depend on inspection, no complaints, no large or unusual bookings. In each case the behaviour should be to say a person will call back — and then someone actually has to.
A system that improvises here is a liability regardless of how well it books.
4. Point only your missed calls at it
Do not hand over the line.
Conditional call forwarding sends the calls nobody answers — and only those — wherever you choose. Your number does not change, your existing handsets keep working, and nothing that currently works stops working. Start at four or five rings.
The codes differ by carrier and are in the call forwarding guide. This is the step that makes the whole thing low-risk: you are testing on calls that were already going to voicemail.
5. Call it yourself, five times
Before any customer does. Five specific calls, each testing a different failure:
- Ask for a vague time — "tomorrow afternoon" — and listen to whether the options are genuinely spread apart
- Make a booking, ring back, and ask to move it — listen for whether it names the booking before changing anything
- Ask for a time it should question, like 2 AM
- Interrupt it mid-sentence
- Ask it something it cannot know, like a price for something not on your menu
The last one is the one that matters. If it invents an answer for you, it will invent one for a customer. The full version, with what to listen for in each, is five calls to make before you trust an AI receptionist — every one is a failure we found in our own system first.
6. Read every transcript for two weeks
This is the step people skip, and it is the one that makes the difference between a system that works and one that quietly underperforms.
You are looking for three things: words it consistently mishears (usually your service names and staff names — most systems let you add a vocabulary list), questions your callers ask that you never anticipated, and anything it answered that it should have handed over.
Almost everyone finds that their callers ask something they had not thought about. That hour of reading is the highest-value part of the whole setup.
7. Widen it, slowly
Once the transcripts look right, extend coverage. Fewer rings before it picks up, or all calls during a rush, or everything with a transfer path.
The mistake is doing this on day one. The right sequence is: prove it on the calls you were already losing, fix what the transcripts reveal, then give it more.
The two things to have ready before you start
Your real service list — names, true durations, prices, who does what.
Someone who will read the transcripts. Not much time; an hour in the first week. A system nobody looks at drifts, and six months later it is quoting a price you no longer charge.
If you have those two, the rest is an afternoon. If you do not, get them first — the setup is not the hard part.