The pause is what tells a caller nobody is there.
Every vendor in this category says instant, and almost none of them says a number. This page is the other half of that conversation: what the gap before a reply does to a garage door call, where the time goes inside a voice pipeline, and how to time any vendor yourself in ten minutes with a phone and a stopwatch.
A caller who has to repeat themselves assumes nobody is listening.
Conversation runs on a rhythm people can feel to the fraction of a second. Break it and a caller does not think the system is thinking. They think the line is dead, and they behave accordingly, in this order.
They say hello again
A gap long enough to notice reads as a dropped call, not as thinking. The caller says hello into it, which lands on top of the reply and starts a collision the rest of the call never quite recovers from.
They start the sentence over
Nobody repeats themselves in a conversation that is working. The moment a caller re-tells you the door will not close, they have decided they are talking to a machine, and the tone of the whole call changes.
They dial the next company
Standing in a garage at seven in the morning with a car shut in, the search results are still open on the screen. A line that feels unattended is not a bad experience. It is the reason the job goes to the next number down.
Half a second between responses is the pace this trade quotes at each other. It is a convention rather than a measurement, it carries no primary source we could trace, and it is a target to measure against, not a pass mark.
It is still the yardstick worth carrying into a vendor call, because it is about the gap two people leave in an easy conversation. Everything past it is a matter of measurement, not of adjectives, which is why the rest of this page is a method rather than a claim.
Five things happen between your last word and the reply.
A voice pipeline is not one thing being fast or slow. It is five legs in series, each with its own failure mode, and on a garage door call three of them are made harder by where the caller is standing.
Hearing the words
Speech recognition turns audio into text. This is the leg the trade punishes: a phone on speaker, bouncing off bare block walls, with an engine running and the door still trying to cycle. Clean audio is fast. A garage at 7am is not clean audio.
Deciding they finished
Something has to judge the end of a turn. Tuned tight, the receptionist interrupts a caller mid address. Tuned loose, it sits there. This single judgement is made on every turn of every call, and it is where most vendors' pauses actually live.
Composing the reply
The model works out what to say: which triage question comes next, whether this is a photo eye or a spring, which published range applies, whether the answer just changed the urgency of the job.
Starting to speak
Speech synthesis has to put the first audio on the wire. What matters is when speech starts, not when the sentence finishes generating, because the caller stops waiting the instant they hear a voice.
The line itself
The public telephone network adds its own delay, and a browser demo skips it entirely. Any number measured in a browser tab is measuring a shorter journey than the one your customers will make.
This is why a single average tells you so little. A line can post a respectable median and still stall on exactly the turn where a caller reads out a long address, which is the turn that has to work.
Time it yourself, on any vendor, in ten minutes.
You do not need telemetry to compare two receptionists. You need a phone, a stopwatch, and the discipline to run the same seven steps on each of them.
Call it from a phone
Not the browser demo on the vendor's website. A browser demo runs over your office wifi and misses the telephone leg completely, which is the one leg you cannot tune later.
Stand where your callers stand
On speaker, in a garage, with the door open. If it can hold a conversation there it can hold one anywhere, and that is where the calls you are worried about are actually made from.
Start the clock on your last syllable
Not on the end of your thought. The gap a caller feels begins the moment they stop making sound and ends the moment they hear a voice, so that is what the stopwatch has to cover.
Time ten turns, not one
The greeting is pre-loaded and tells you nothing. It is turns four through ten, once the conversation has real content in it, that show you what the line is like on a busy morning.
Include the four awkward turns
A long street address. A callback number read at speed. An interruption in the middle of its sentence. A correction: no, sorry, it is a torsion spring, not extension. Those are the turns that separate systems.
Write down the slowest one
Averages hide the turn that loses the job. One long pause in a three-minute call is what the caller remembers, and it is the only figure that predicts whether they stay on the line.
Then ask for their numbers
Median, 90th percentile, how many turns, measured when, over the phone or in a browser. A vendor that measures its own line can answer in one sentence.
Five questions, and what the answers tell you.
A vendor that measures its own line answers these in a sentence each. A vendor that has never measured reaches for an adjective, every time, and the pattern is easy to hear once you are listening for it.
| Ask | A real answer | A dodge |
|---|---|---|
| What is your median response time? | A number, with a unit and a date it was measured. | Instant. Real time. Near-human. |
| What is the 90th percentile? | A second, larger number, offered without flinching. | We do not track that. |
| Measured over how many turns, and when? | A count, a window, and how calls were selected. | Thousands of calls. Constantly. |
| Phone network or browser? | Phone, and they know why the difference matters. | The demo on our site is the same thing. |
| What makes your slow turns slow? | Named causes: noisy audio, long tool lookups, a mid-call save. | That does not really happen. |
We publish measurements, not adjectives, and we have not taken ours yet.
There is no GarageLine AI latency number on this page, because the garage line is not running calls yet and we do not publish figures we have not measured. When it is, this page gets the median and the 90th percentile, the number of turns behind them, the dates they were taken, and the slow turns explained rather than trimmed out of the sample. That is the same standard we are asking you to hold every vendor to, applied first to ourselves. Hold us to it: the method above works on our line exactly as it works on anyone else's.
The questions worth asking about speed.
What counts as good latency for an AI receptionist?
The pace the voice-AI trade quotes is roughly half a second between responses, which is a convention rather than a traced measurement, and it is about the gap two people leave in an easy conversation. Treat it as the target to measure against rather than a pass mark, and weigh consistency alongside it: a line that answers evenly on every turn sounds more alive than one that is quick most of the time and then disappears for four seconds.
Why does this page not publish GarageLine AI's own latency figures?
Because we have not measured them. GarageLine AI publishes numbers it has taken off live calls, with the method, the sample size and the slow turns included, and it publishes nothing before that. The day the line is running, the measurements land on this page in that form.
Does latency matter more on a garage door call than on other calls?
It matters at a specific hour. The trapped-car call comes in before work from somebody who is already late, standing next to a door that will not lift, with the search results still open. Every second of dead air on that call is a second they spend deciding to try the next number.
Can a demo be made to sound faster than the real thing?
Yes, and it is worth knowing how: a browser demo skips the telephone network, a warmed-up session skips the first-call delay, and short scripted replies start speaking before any real work has been done. Time it over the phone, with a real address and a real callback number, and the difference shows up immediately.
Keep going.
Bring the stopwatch to ours.
We build the receptionist around your company first, then you talk to it and time it the way this page describes. If the gaps are wrong, you will hear it before your customers ever do.
Put it on your line