What an AI Receptionist Call Sounds Like: An Annotated Walkthrough
Feature lists do not tell you whether an AI receptionist is any good. Transcripts do. Here are three composite calls for a locksmith shop, annotated line by line so you can see exactly what gets captured, what gets decided, where the bot hands off to a human, and what the message on your phone looks like at the end of it.

What an AI Receptionist Call Sounds Like: An Annotated Walkthrough
Most evaluations of AI receptionists never get past the feature list. Twenty-four seven coverage, bilingual, books appointments, integrates with your calendar. Every vendor says the same six things, which means the list tells you nothing about whether the thing is any good on a real phone call at 11 PM with a stressed caller and traffic noise in the background.
Transcripts tell you. So this post is three calls, written out turn by turn, with annotation after each exchange explaining what is happening underneath — what got captured, what decision the system just made, and what it deliberately did not do.
These three calls are composite illustrations, not recordings of real customers. They are written to show the mechanics accurately, using the intake fields and behaviors an automotive locksmith setup actually uses. No customer name, address, phone number, or outcome below belongs to a real person. If you want a genuinely real call, the fastest route is the one at the end of this post: make the thing call you.
As of August 2026, this is the format we would want if we were the buyer, so it is the format we are publishing.
Call one: the 11:04 PM lockout that books
This is the call that pays for the subscription. A car lockout, after hours, caller is outside standing next to the vehicle.
KeyBot: Thanks for calling Sunrise Locksmith, this is Mike. What can I help you with tonight?
Caller: Yeah, hi — I locked my keys in my car. I'm at the Kroger on North Collins.
What just happened. The greeting names the business and gives the assistant a name, because callers who reach an unnamed system hang up more often than callers who reach a named one. The caller volunteered both the problem and a rough location in the first sentence, which is normal and which a rigid script would have stepped on by asking for a name first. The intake is ordered by what the caller is likely to say, not by what the CRM form wants.
KeyBot: Sorry, that's a rough one — I can get somebody out to you. Are you somewhere safe right now?
Caller: Yeah, I'm fine, I'm right by the entrance.
KeyBot: Good. Is anyone else in the car — a child or a pet?
Caller: No, just me out here.
What just happened. Two safety checks before anything commercial. A child or an animal locked in a vehicle changes the call completely: it becomes an emergency, and the correct behavior is to tell the caller to call 911 immediately rather than to continue collecting a service address. Asking this early rather than late is the whole point — an intake that discovers it on turn nine has already wasted the minutes that mattered. The escalation boundaries are covered in more depth in what an AI receptionist cannot do.
KeyBot: Perfect. What's the year, make and model?
Caller: It's a 2019 Honda Accord.
KeyBot: Got it, 2019 Accord. And do you have a spare key anywhere — at home, with a family member?
Caller: No, that's the thing, this is the only one and it's sitting on the passenger seat.
What just happened. Year, make and model is the single highest-value field in an automotive locksmith intake, because it determines whether this is a straightforward entry job or a key that has to be cut and programmed. The spare-key question is doing quiet triage: keys visible inside the car means this is a lockout, not an all-keys-lost. Those two jobs have completely different prices, different time on site, and different tools on the truck. Getting that distinction on the phone is the difference between a tech arriving equipped and a tech arriving twice.
KeyBot: Okay — keys are in the car and it's a straight lockout, that's the good news. Let me get you an exact price. One moment.
KeyBot: For a 2019 Accord lockout at this hour, your all-in total is one hundred and ninety-five dollars. That includes the after-hours rate, and there are no other fees on top of it.
Caller: Okay. How long?
What just happened. Two things worth pulling apart.
First, the price came from the shop's own uploaded price book, not from the model's imagination. A receptionist that improvises numbers on a recorded line creates disputes the owner inherits the next morning; the correct architecture is a lookup against confirmed prices, which we described in running a price book through phone quoting.
Second, and this one is a hard-won lesson: the after-hours premium is inside the first number the caller hears. Not added at the end, not mentioned when the tech arrives. The bait-and-switch version of this call — quote $95, then add $100 on arrival — loses the job and earns the review. The structure is explained in after-hours emergency pricing.
Notice also what the caller did not ask: they did not push back on price. A confident all-in number delivered immediately converts far better than a hedged range delivered slowly.
KeyBot: I've got a tech who can be there between eleven thirty and midnight. Does that work?
Caller: Yeah, that works.
KeyBot: Great. Can I get your name and the best number to reach you? I've got you calling from a number ending five-five-eight-one — is that the best one?
Caller: Yeah that's my cell. It's Daniel.
KeyBot: Thanks Daniel. And what color is the Accord, and roughly where in the lot are you parked?
Caller: Silver. I'm on the side by the pharmacy entrance.
What just happened. Four details, each of which prevents a specific failure.
The arrival window is a concrete one-hour window, not "sometime tonight" and not "between eight and five". Vague windows lose bookings on the call and generate where-is-my-tech calls afterward.
The callback number was confirmed against caller ID rather than dictated digit by digit. Reading ten digits back over a phone connection is where automated systems most reliably embarrass themselves — text-to-speech garbles long digit strings, and background noise makes the correction loop worse. Confirming the last four is faster and more accurate. If the caller needs a different number, a keypad fallback lets them type it, which removes speech recognition from the equation entirely. The general problem is covered in audio quality, accents and background noise.
Color and position in the lot are not CRM fields anybody designs on day one. They come from techs asking why they spent eleven minutes walking a parking lot at midnight.
KeyBot: You're all set, Daniel — silver Accord at the Kroger on North Collins, tech between eleven thirty and midnight, one ninety-five all in. You'll get a text confirmation right now. Anything else?
Caller: Nope, thanks.
What just happened. A read-back of every material fact before the call ends. If the address is wrong, this is the cheapest possible moment to find out. Then a text confirmation, which gives the caller something to look at and gives the shop a timestamped record of what was promised.
The message that lands on the owner's phone, within seconds of the call ending, looks roughly like this:
Name Daniel, callback ending 5581, 2019 Honda Accord, silver, service is a car lockout with keys visible inside, location Kroger on North Collins by the pharmacy entrance, quoted $195 all-in including after-hours, window 11:30 PM to midnight, source of call is the Google ad, recording link attached.
That is a dispatchable job, not a note to call someone back.
Call two: the price shopper who does not book
Publishing only the call that books would be dishonest. Here is the far more common one.
KeyBot: Thanks for calling Sunrise Locksmith, this is Mike. How can I help?
Caller: How much for a key for a 2014 Camry?
KeyBot: Happy to get you a number. Is this a spare key while you still have a working one, or are all your keys lost?
Caller: All of them, I lost the only one.
What just happened. The caller asked a price question with no information attached, which is the most common opening in the trade. The receptionist did not answer it — it qualified first, because "a key for a Camry" spans a very wide price range and the spare-versus-all-keys-lost split is most of that range. Answering the literal question with a number would have been worse service and worse business.
KeyBot: Understood. For a 2014 Camry with all keys lost, that's three hundred and forty-five dollars, all in, and that covers cutting and programming the key on site.
Caller: Oof. The dealer told me two eighty.
KeyBot: That can happen — though the dealer price usually doesn't include the tow, since the car has to get to them. We come to you and it's done in about an hour. Would first thing tomorrow work?
Caller: Let me think about it. I'll call you back.
KeyBot: No problem. I'll have that quote saved under this number, so if you call back it'll be right here at the same price. Have a good one.
What just happened. Three things, and the third is the important one.
The objection got exactly one response — a real, specific counterpoint about the tow, not a wall of justification — followed immediately by a close attempt. Long defensive wind-ups after a price objection are where callers hang up mid-sentence.
The close was a concrete alternative slot, not "would you like to book". Give a caller a thing to say yes to.
And then the receptionist let go cleanly and saved the quote against the phone number. This is the part that matters for the business. When Daniel-the-Camry-owner calls back on Thursday, the system recognizes the number and honors $345. It does not re-interview him from scratch and it does not quote him $395 because the second conversation happened to run differently. Silently re-quoting a returning caller higher is one of the fastest ways to lose a customer who had already decided to buy.
This call produced no revenue today. It produced a record, a saved price, and a caller who was not annoyed. That is the correct outcome for a call that was never going to book on the first attempt.
Call three: the one you should never see
KeyBot: Thanks for calling Sunrise Locksmith, this is Mike. How can I help?
Caller: Hi, I'm calling regarding your business's eligibility for a working capital advance of up to two hundred and fifty thousand dollars —
KeyBot: We're not interested, thanks. Take us off your list.
What just happened. No intake, no message, no notification. The owner's phone stays silent.
The value here is subtractive and it is easy to underrate. A shop that forwards every call to an owner's cell trains that owner to stop looking at notifications, and the moment that happens the real messages start getting missed too. Screening is what keeps the notification channel worth watching. This category is bigger than most owners think — the identification patterns are in spam calls and robocalls at locksmith shops.
One caution: this is also where an over-aggressive filter does real damage, because a genuine commercial account calling from an unknown number can superficially resemble a solicitation. The screening should be based on what the caller actually says, not on the number they called from.
What each turn bought you
Reading the three calls back against the alternative makes the difference concrete. The left column is what a voicemail box would have captured from the same three callers.
| What happened | Voicemail | AI receptionist |
|---|---|---|
| 11:04 PM lockout | Nothing — most callers hang up rather than leave a message | Booked job, $195 quoted all-in, one-hour window, tech dispatched |
| Vehicle year/make/model | Only if the caller thinks to say it | Always captured, and used to route lockout vs all-keys-lost |
| Child or pet in the vehicle | Discovered on callback, if at all | Checked on turn three, before anything commercial |
| Exact service location | Often partial and mis-heard | Confirmed, with lot position and vehicle color |
| Callback number | Whatever the caller recites | Confirmed against caller ID, keypad fallback available |
| Price shopper at 2 PM | No record | Quote saved to the number and honored on callback |
| Robocall at 9 AM | Fills the message box | Screened, no notification |
| Time before you see it | Next time you check the box | Seconds, as a structured message |
The honest limits
Four things the three calls above should not be read as claiming.
This was a Core-plan call, not a Lite call. Quoting from a price book and booking a slot are full-platform behaviors, starting at $500 a month on Core with a 14-day free trial. KeyBot Lite at $149 a month — 100 calls included, 50 cents per minute after, first 5 answered calls free on the 7-day trial — is a message-taker. On Lite, call one would have run identically through the safety checks and the vehicle and location capture, and then ended with a promise that the team will call straight back, with the message in Telegram within seconds. That is a genuinely different product and it is the right one for a shop whose bottleneck is capture rather than conversion. The comparison is in when to upgrade from Lite to Core and the Lite mechanics in the KeyBot Lite walkthrough.
Real calls are messier than written ones. People interrupt, restart, take a second call, walk into a parking garage and lose signal, and describe a Highlander as a 4Runner. A transcript on a page is always cleaner than the audio it came from. Judge the audio.
A booked job is not a completed job. Call one ends with a dispatchable message. Somebody still has to roll. The gap between a booking alert and a truck actually moving is an operations problem that no receptionist solves, and it is where more revenue leaks than most owners want to believe.
Transfers are one-way. When a call does need a human, the handoff behavior matters a great deal — a blind transfer to a line nobody picks up is worse than no transfer, because the system has already let go. That trade-off is set out in warm versus blind transfer, and it is the single question we would press hardest on with any vendor.
Hear a real one instead
Everything above is an illustration. The real version takes about thirty seconds to arrange: go to the instant demo, enter your business information, and the AI calls your phone answering as your own company — no signup, first demo free. Ask it the awkward questions. Interrupt it. Give it a vehicle it has no price for and see whether it invents a number or says a tech will confirm.
If you would rather listen cold without entering anything, call the live demo line at +1 (716) 350-6391, which answers for a demo shop named Sunrise Locksmith. And if you want the intake fields laid out as a checklist rather than as dialogue, they are in the first-90-seconds call handling checklist. The mechanics behind the call layer are on the AI call handling page, and plan pricing is at pricing.
The bottom line
A feature list cannot tell you whether an AI receptionist is worth buying, because every vendor's list is identical. A transcript can, because it exposes ordering, judgment and restraint — whether safety questions come before commercial ones, whether prices come from a confirmed book or from the model's imagination, whether an objection gets one crisp answer or a defensive paragraph, whether a returning caller gets their old quote honored, and whether the thing knows when to stop talking and hand off.
Read the three calls above as a specification rather than as marketing: those are the behaviors to test. Then go get a real call placed to your own phone and check them one at a time. Thirty seconds of setup and one honest listen will tell you more than any comparison page, including ours.
Frequently asked questions
Are these real customer call transcripts?
No — all three calls are composite illustrations written to show the mechanics accurately, and no name, address, phone number, vehicle or outcome in them belongs to a real customer. They use the intake fields and behaviors an automotive locksmith configuration actually uses, so they are representative of structure rather than transcripts of recorded calls. To hear a genuinely real call, use the instant demo at https://www.thekeybot.com/try and the AI will call your own phone.
What does an AI receptionist actually capture on a locksmith lockout call?
On a lockout it captures whether anyone is in danger, whether a child or pet is locked in the vehicle, the year make and model, whether keys are visible inside or all keys are lost, the exact service location including position in a lot, the vehicle color, a callback number confirmed against caller ID, the quoted all-in price, the agreed arrival window, and where the caller heard about you. That set is what turns a message into a dispatchable job rather than a note to call somebody back.
Does the AI make up prices if it does not have one?
No, and this is the behavior worth testing hardest during a demo. Prices come from the price sheet the owner uploads and confirms on screen, and the receptionist states those exact figures rather than negotiating or improvising. When a caller asks about a vehicle or service that is not on the confirmed sheet, the correct response is to capture the request and tell the caller a technician will confirm the exact figure, because invented numbers on a recorded line create disputes the shop inherits later.
Would KeyBot Lite handle the lockout call the same way?
Partly. Lite runs the same safety checks and the same vehicle and location capture, but it is a message-taking receptionist, so it ends by telling the caller the team will call straight back rather than quoting a price and booking a slot. Lite is $149 per month with 100 calls included and 50 cents per minute after, and the first 5 answered calls are free. Live quoting, booking and dispatch start at $500 per month on Core with a 14-day free trial — see https://www.thekeybot.com/pricing.
What happens when a robocall or a solicitor gets through?
It gets screened and produces no notification at all, which is a genuinely valuable outcome rather than a missing feature. A shop that forwards every call to the owner's cell quickly trains that owner to ignore notifications, and once that happens the real messages start getting missed too. Screening should key on what the caller actually says rather than on the number they called from, so a legitimate commercial account calling from an unknown number still gets a full intake.
How do I test an AI receptionist properly before buying?
Place a real call and deliberately make it difficult. Interrupt mid-sentence, give a vehicle or service that is not on your price sheet and check whether it invents a number, describe a situation that should escalate to a human and see whether the handoff is warm or blind, and give it a different callback number than the one you are calling from to test digit capture. Go to https://www.thekeybot.com/try to have the AI call your own phone in about 30 seconds, free and with no signup.
About the Author
TheKeyBot Team is dedicated to helping locksmiths grow their businesses through AI automation and smart technology. With years of experience in the locksmith industry, our team provides actionable insights and proven strategies.
