Article hero background

AI & Technology

The One Metric That Decides If Callers Stay on the Line: Latency

AI receptionist latency decides if callers stay or hang up. See the targets that matter, what causes slow replies, and how to test any system yourself.

by Vinod Jethwani October 26, 2026

Callers rarely hang up on an AI receptionist because the voice sounds a little synthetic. They hang up because it goes quiet. AI receptionist latency, the gap between the moment a caller stops talking and the moment the agent starts replying, decides whether a phone call feels like a conversation or a dead line. A polished demo hides that gap. A real phone line exposes it within the first two turns.

What AI Receptionist Latency Means

AI receptionist latency is the silence between the end of a caller’s sentence and the first sound of the agent’s reply. It’s measured in milliseconds on a real phone line, and it includes every step from hearing the caller to speaking back.

Engineers call it voice-to-voice latency or response latency. Callers don’t call it anything. They just notice the pause, and they decide what it means within a second.

Answer Speed vs Latency

Answer speed is how fast the phone gets picked up. Latency is how fast the agent replies once the conversation starts. Atty’s AI Receptionist answers in under a second, but pickup happens once per call, and latency happens on every turn after it.

That difference adds up fast. A service booking collects at least 4 details, such as a name, an address, the problem, and a time slot. That’s at least 8 turns of question and answer. An agent that takes 1.5 seconds per reply adds 12 seconds of dead air to a call that’s only meant to last a couple of minutes.

How Fast People Expect a Reply on the Phone

People expect a reply in about 200 milliseconds. A study of natural conversation in 10 languages found a mean gap of 208 ms between a question and its answer, with most replies starting within 0 to 200 ms.

The research comes from Stivers and colleagues, published in PNAS in 2009. The team recorded everyday conversations on 5 continents, from English and Japanese to Lao and Danish. Japanese speakers were fastest, at a mean of 7 ms. Danish speakers were slowest, at 469 ms. Every language in the sample kept its average gap under half a second.

That’s the timing every caller brings to the phone. It isn’t learned from technology. It’s how people talk.

Why Silence Sounds Like Bad News

Silence sounds like bad news because in human conversation, it usually is. The same PNAS study found that answers arrive faster than non-answers like “I don’t know,” and that confirmations arrive 100 to 500 ms faster than disconfirmations.

People learn that pattern without thinking about it. A quick reply means “yes, I’ve got this.” A slow reply means trouble: the speaker doesn’t know, doesn’t agree, or isn’t listening. When an AI receptionist pauses for 1.5 seconds after a caller says “my water heater is leaking,” the caller doesn’t hear a computer processing. They hear hesitation, and hesitation doesn’t sound like a company that’s about to fix their problem.

Why an AI Can’t Reply in 200 ms Like a Person

People hit 200 ms by cheating the clock. Levinson and Torreira’s 2015 review of turn-taking research shows that planning even a single word takes over 600 ms, which is 3 times longer than the typical gap. People manage it by planning their reply while the other person is still talking.

An AI receptionist works under 2 handicaps a person doesn’t have. It has to decide the caller has finished before it commits to a reply, and it has to send audio across the phone network in both directions. The ITU-T G.114 standard, which telephone networks are designed around, treats one-way delay under 150 ms as essentially unnoticeable and delay above 400 ms as unacceptable for network planning. The phone line uses part of the budget before the AI has heard a single word.

The Right Latency Target for an AI Receptionist

A well-built AI receptionist holds median latency under 600 ms and keeps 95 of every 100 replies under 800 ms. Above 1,200 ms, callers start repeating themselves, talking over the agent, or hanging up.

Latency per replyWhat the caller hearsWhat the caller does
Under 300 msNo gap at allTalks as they would with a person
300 to 600 msA short, natural pauseStays engaged and keeps going
600 to 800 msA noticeable beatWaits, but the rhythm slows down
800 to 1,200 msAn awkward silenceSays “hello?”, repeats, or talks over the agent
Over 1,200 msA frozen or dropped lineHangs up or asks for a person

These bands come from teams running voice agents on live phone traffic, and they line up with the human research above. They aren’t the output of a controlled trial. You’ll also see claims online that callers hang up 40% more often past one second. No published study stands behind that number, so don’t build a business case on it. Your own call recordings are a better source, and they’re free.

The 95th Percentile Matters More Than the Average

The 95th percentile matters more because callers don’t experience averages. They experience the slowest reply on their call.

Take an agent with a fast average but a 95th percentile of 1,500 ms. One reply in 20 lands in the hang-up band. Across an 8-turn booking call, the chance of hitting at least one of those slow replies is 34%. One caller in three meets a long, unexplained silence, even though the dashboard average looks healthy.

How Slow Replies Break Phone Conversations

Slow replies break calls through a chain reaction. The caller thinks the line dropped, speaks into the silence, the agent cuts off its own answer to listen, and both sides talk over each other until the caller gives up.

Here’s how a single slow reply turns into a lost booking:

  1. The caller assumes the line dropped. After about a second of silence, most callers say “hello?” or “are you there?” to check.
  2. The caller talks into the gap. The agent’s reply finally starts, right as the caller starts speaking again.
  3. The agent stops to listen. Systems built to allow interruptions hear the caller and cut their own answer short.
  4. The conversation loops. The agent answers the “hello?” instead of the original question, the caller repeats the address, and a 2-minute call drifts toward 4.

On a no-heat call in January, an AI answering service for HVAC companies that falls into this loop loses the caller before it ever reaches the booking question.

This is why a slightly synthetic voice that replies on time beats a beautiful voice that lags. Voice texture doesn’t carry meaning in human conversation. Timing does.

What Causes High Latency in an AI Receptionist?

Latency builds up across 6 stages: the phone network, end-of-speech detection, speech-to-text, the language model, text-to-speech, and the hops between separately hosted services. Running those stages one after another, instead of overlapping them, is the most common cause.

StageWhat happensTypical time added
Phone networkAudio travels from the caller to the system and back50 to 150 ms
End-of-speech detectionThe system decides the caller has finished talking150 to 300 ms
Speech-to-textThe caller’s words become a final transcript50 to 300 ms
Language modelThe system works out what to say and produces its first words150 to 800 ms
Text-to-speechThe reply becomes audio100 to 400 ms
Service hopsAudio and text move between separately hosted vendorsOne extra round trip per hop

Add the best case for each stage and you land between roughly 500 and 900 ms. Run the same stages one after another, waiting for each to finish completely, and published engineering budgets put the total between 800 and 2,000 ms before the phone network is counted.

Streaming Makes the Biggest Difference

Streaming lets every stage start on partial input instead of waiting for the stage before it to finish. The transcript builds while the caller is still talking, the language model starts its reply from the first finished words, and speech starts playing before the full sentence exists.

It’s the same trick people use. The fastest human conversations overlap listening and planning, and the fastest AI receptionists overlap their stages the same way.

End-of-Speech Detection Is a Trade-Off

End-of-speech detection is a trade-off because the same setting controls speed and interruptions. A short silence threshold makes the agent reply quickly, but it also jumps in whenever a caller pauses mid-sentence. A long threshold stops the interruptions and adds delay to every single turn.

Trade callers pause in predictable places. They stop to read a model number off a furnace label, to remember a street number, or to ask someone in the room. An agent tuned to treat every half-second pause as the end of a turn will talk over a homeowner who’s reading “Rheem, model… X-R-90…” off the side of a water heater. An AI answering service for plumbing companies has to wait out that pause without losing pace on the next question.

Phone Audio Slows Things Down

Phone audio is harder to process than studio audio. Standard phone calls carry a narrow band of sound, roughly 300 to 3,400 Hz, while most speech systems are trained on fuller, cleaner recordings. Add a running tap, a barking dog, or a truck engine, and recognition takes longer and makes more mistakes.

Systems that perform best on real calls are tuned for real phone audio. That means background noise reduction and spell-back confirmation of names, numbers, and addresses, so a fast answer is also a correct one.

How to Test an AI Receptionist’s Latency Yourself

Call the live phone number, record the call, and measure the silence after each of your sentences in a free audio editor. Make at least 10 calls and judge the system by its slowest reply, not its average.

Skip the website demo widget. Browser demos skip the phone network and use cleaner audio, so they run faster than any real call your customers will make. Then work through these steps:

  1. Call the real business number from a mobile phone.
  2. Record the call on a second device or with a call recording app.
  3. Say a short sentence, then stop cleanly.
  4. Measure the gap from your last word to the agent’s first sound, using the waveform in a free editor like Audacity.
  5. Repeat with a long description, a mid-sentence pause, background noise, and an interruption.
  6. Call again during a busy hour and after business hours.
  7. Rank every reply and judge the slowest one.

Use this scorecard while you listen. The emergency row matters most for an AI answering service for electricians, where a burning smell at a breaker panel has to reach a person without a single long pause.

TestWhat to listen forPass mark
Short answer, like “yes”Gap before the next questionUnder 800 ms
Long problem descriptionGap after you stop talkingUnder 800 ms
Pause while reading a model numberDoes the agent cut in?No interruption
Interrupt the agent mid-sentenceDoes it stop and listen?Stops within a beat
TV or running tap in the backgroundAre your details right?Correct spell-back
“This is an emergency”Does it route you fast?Transferred without asking twice

A system that consistently runs past 800 ms is creating friction on real calls, even if every transcript reads perfectly.

Why Latency Matters More for Plumbers, HVAC Companies, and Electricians

Trade calls are urgent, short, and easy to lose. A homeowner with a burst pipe or a dead furnace wants a booked time in the first minute, and every silent pause in that minute pushes them back toward the next number on Google.

Emergencies aren’t rare events for homeowners. HomeServe’s 2025 State of the Home survey of 2,333 U.S. homeowners found that 81% had a home repair emergency in the previous year. Heating and cooling failures topped the list at 30%, followed by appliance breakdowns at 29% and clogged or overflowing toilets at 26%.

Those callers are standing in water or sweating through a broken AC unit while they talk. They have less patience for silence than anyone else who’ll ever call your business, and they have a list of other contractors one tap away.

Is a Slow AI Worse Than Voicemail?

A slow AI breaks a promise that voicemail never makes. Voicemail sets a clear expectation: leave a message and wait. A slow AI starts a live conversation and then fails to hold up its end, so the caller spends 30 seconds unsure whether anyone is there. Many hang up with nothing recorded at all, which is worse than a message left in a voicemail box.

What a Trade Business Needs From Its AI Receptionist

A trade business needs 4 things from its AI receptionist on every call:

  • Speed under 600 ms median and 800 ms at the 95th percentile, measured on a real phone line.
  • Interruption handling that lets a caller cut in without breaking the flow of the call.
  • Emergency triage in the first 2 questions, so a burst pipe reaches a person immediately.
  • Spell-back confirmation of the address and callback number, so speed never costs accuracy.

That first minute is where an AI receptionist earns its keep. Atty screens every caller for service type, location, and urgency, and routes the emergencies you define straight to a person on your team.

How Latency Shows Up in Your Booking Numbers

Latency shows up as short calls that end before a booking, callers saying “hello?” mid-call, overlapping speech, and repeated answers. Search your call recordings for these 4 signals before you blame the script or the voice.

  • Short calls under 30 seconds that end before the booking question.
  • Check-ins like “hello?” or “are you still there?” in the middle of transcripts.
  • Overlap where the caller and the agent speak at the same moment.
  • Repeats where the caller gives the same address or phone number twice.

Every call through Atty’s AI answering service for home service businesses produces a recording, a transcript, and a summary, so these signals take minutes to find. Pull 20 calls from your busiest morning and count them. That count is a better measure of caller experience than any number on a vendor’s pricing page.

Frequently Asked Questions

What Is a Good Latency for an AI Receptionist?

Under 600 ms at the median and under 800 ms at the 95th percentile, measured from the end of the caller’s speech to the start of the agent’s reply on a real phone line.

Is One Second of Latency Too Slow?

For booking calls, yes. At around one second, callers start saying “hello?” and repeating themselves, which stretches the call and raises the risk of a hang-up.

Does a Filler Phrase Like “Let Me Check That” Fix Latency?

It hides one real wait, such as a calendar lookup. Used on every turn, it becomes its own annoyance and doesn’t shorten anything.

Do Callers Care More About Speed or Accuracy?

Both, in that order on every turn. A fast agent that writes down the wrong address still loses the job, which is why spell-back confirmation is worth the extra second it takes.

Can an AI Receptionist Reply as Fast as a Person?

Not on a standard phone line today. The phone network alone uses a meaningful share of the 200 ms that people expect, before any processing starts.

Latency Is a Conversion Metric, Not a Technical Detail

Latency decides whether a caller stays long enough to book. Every reply either keeps the rhythm of a normal conversation or breaks it, and a broken rhythm sends the caller to the next contractor on the list.

The voice, the script, and the integrations all matter, but none of them gets a chance to work if the caller has already hung up. Measure the silence on your own line before you judge anything else.

Call the number your business uses today, record 10 calls, and find your slowest reply. Then book a demo with Atty and run the exact same test on a live line, with your own services, your own service area, and an emergency scenario from your trade. You’ll know within 10 calls which system your customers will stay on the line for.

“Atty has transformed how we handle after-hours calls. Our client satisfaction scores have increased by 40% since implementation.”
Adam Waknine Managing Partner, Chen & Associates