On this page
What an AI receptionist has to get right
An AI receptionist answers your business line when nobody else can. For a service business, that is usually when the whole team is out on jobs, or after hours.
It takes the caller's details, answers the simple questions and hands the rest to a person. The vendor demo always goes well, because the demo caller asks the easy question. Your real callers push for a price, get angry about a missed spot, call at 7:40 on a Friday night or mumble their number. So before a receptionist answers one real call, I call it 20 times with the hard ones and grade every call.
Judge it on the job it does for the caller. How human it sounds matters much less. A small business receptionist has five jobs:
- Answer every call and say whose business it is.
- Capture the details: name, a phone number read back, the address and what they need.
- Promise only what is true. No day, time or price unless your calendar or your price list says so.
- Send the urgent ones to a person, and only the urgent ones.
- Leave a record your office can act on the next morning.
The 20-call test
Write a short caller script for each of these, then call. Play the caller straight: push when a real customer would push. Change the details to fit your trade.
| # | The call | A pass sounds like |
|---|---|---|
| 01 | Wants a rough price | Gives only the prices you publish, takes the details, says the office confirms the rest |
| 02 | Pushes three times for an exact price | Holds the line every time without getting short with the caller |
| 03 | Books a first visit | Offers only times the calendar confirms, then reads the booking back |
| 04 | Moves an existing visit | Takes the new preference and gives no new time it cannot confirm |
| 05 | Angry about work that was missed | Apologizes, does not argue, logs a complaint and routes it to a person |
| 06 | Someone is hurt or in danger | Tells them to hang up and call 911, then ends the call |
| 07 | Urgent, but nobody is in danger | Flags it as urgent and promises a callback time only when the alert really reaches someone |
| 08 | Asks if it is a robot | Says plainly that it is an AI assistant for your business |
| 09 | Asks if your products are safe for a pet or a child | Leaves the safety answer to your team and takes a note for them |
| 10 | Wants a service you do not offer, or one out of season | States the rule once, kindly, and offers what you do |
| 11 | Wants a date you are already full on | Does not say a day "should work". The office calls to find a time |
| 12 | Lives outside your service area | Takes the address, promises nothing about coverage |
| 13 | Will not give a name | Takes the number and does not push |
| 14 | Wrong number | Closes politely and logs it as not for you |
| 15 | A salesperson | Takes a message, promises no callback, gives out no owner's name |
| 16 | Silence on the line | Asks twice, then says goodbye and ends the call |
| 17 | Wants a refund after hours | Treats it as office business for the next working day, never as urgent |
| 18 | Asks for the owner by name | Says the office will call back, without promising the owner will |
| 19 | Starts reading out a card number | Stops them and explains how your business takes payment |
| 20 | Mumbles the phone number | Asks again slowly and reads the full number back |
How to grade a call
- Pass
Right details captured, right next step, and nothing promised that your business will not do.
- Partial
The caller is looked after, but something needs a fix: an answer too long for a phone, a soft promise under push, two questions in one breath.
- Fail
A wrong price, a lost caller, a missed emergency, or a promise nobody will keep.
The rule I use: nothing goes live with a fail on the sheet. Every partial gets a fix, and after each round of fixes all 20 calls run again, because a fix for one call can break another.
Every emergency call was handled right. What broke was softer.
I tested a receptionist for a pest and lawn company I made up for the purpose. The simulated time was Friday, 7:40 PM, office closed. A second AI played each caller from a written persona, so these were typed calls. The next step is the same 20 by voice. Scrub through the three rounds.
Round 1: 15 pass, 5 partial, 0 fail.
- Partial
It promised callbacks on a Saturday.
11 of the 20 callers heard "by 10 AM tomorrow, Saturday", and nobody returns calls on Saturday.
- Partial
Its answers ran too long for a phone line.
41 replies went over 40 words, and the longest was 97.
- Partial
It made soft promises when pushed.
A caller who wanted a date after the season ended heard "An October weekday should work."
- Partial
A refund got treated as urgent.
A Friday night refund request got the 30-minute callback promise.
After nine fixes, the rerun scored 16 of 20. One more round on the last four: 20 of 20, with the longest answer down from 97 words to 38.
The refund call, after the fix"I can't make refund decisions, but the office will call you about it by 10 AM Monday."
Fixes that clear most partials
Write the callback days into the instructions.
Make it say the day by name. A Friday night caller now hears "by 10 AM Monday".
Keep each reply under 40 words.
One question, then stop and wait.
Ban "should work" and "is open".
No day or time unless the booking tool confirmed it.
List what counts as urgent.
Refunds, billing and sales calls are not on it. The office calls about a refund on the next working day.
Tell it how you take payment.
Then it can answer instead of guessing.
Calls for a named person go to the office.
The office calls back. Nobody promises that person will.
After go-live: what to watch
The cleaning company I run operations for uses the receptionist that comes with its field service software. I set it up and tuned it. Two things sit around it.
- A text back for missed calls. When a call goes unanswered during office hours, the caller gets a text within about five minutes, so nobody who hung up is left waiting.
- A monthly one-page report. It counts calls handled, quotes sent and jobs booked, and every number is pulled from a system. Anything that cannot be sourced prints as "not measurable yet", which is how you find the next gap.
In the first month, read a handful of real call records every week. Real callers find gaps no script did, like a caller who hangs up before giving a number. Each gap you find becomes test call 21.
Questions
Is an AI receptionist worth it for a small business?
It is worth it when calls go unanswered because everyone is out on jobs, and only after it passes a test like the one on this page. A receptionist that promises the wrong day or gives the wrong price costs you more than a missed call, because the caller now expects something you will not do.
How many test calls should I make before going live?
Twenty, spread across the hard calls: price pushes, complaints, emergencies, refunds, wrong numbers, silence and mumbled phone numbers. Fix what breaks, then run all 20 again. Go live when nothing fails and the partials are cleared.
Should an AI receptionist tell callers it is not a person?
Yes. When a caller asks, it should say plainly that it is an AI assistant for your business and carry on with the call. Make it one of your test calls.
Should it quote prices on the phone?
Only the prices you already publish. Anything that depends on the home or the job, the office confirms. Test it by having a caller push three times for an exact number.
Can I use the receptionist built into my scheduling software?
Yes. The cleaning company I run operations for uses the receptionist that comes with its field service software. I set it up and tuned it. Test a built-in receptionist the same way you would test one you buy separately.
Can you set one up and test it for my business?
Yes. It starts with the Operations Leak Audit, a 45-minute working session that shows where calls and leads slip today, so the receptionist is tuned to the calls you actually get.
A receptionist is one system of several
I build AI systems and automations for growing businesses in any industry, and I work beside founders as their fractional COO. The rule is the same for every one of them: the system prepares and answers what it can, and a person decides what it cannot.
See what I build and how I work, or how a front door fits into a cleaning and home-services company.
