How fast must an AI phone assistant answer? Latency explained
On average, people answer each other after around 200 milliseconds. An AI phone assistant takes longer because several processing steps follow one another: a Twilio guide gives around 1.1 seconds as the target for the whole chain, 1.4 seconds as the upper limit. This article explains the chain — and how to test it on a call.
Why every tenth of a second counts on the phone
People take turns in conversation remarkably fast: in a study across ten languages, the gap between a question and its answer averaged around 200 milliseconds, and in all ten languages the average stayed under half a second (Stivers et al., PNAS 2009). Longer pauses get noticed: in a listening experiment with German listeners, a yes to a request sounded less willing the longer the pause before it — markedly so from about 600 milliseconds (Kohtz and Niebuhr, Interspeech 2017). On the phone there is no eye contact and no nodding to bridge a pause either. If it stays silent too long, uncertainty sets in immediately: is the line dead? For an AI assistant, response speed is therefore not a question of comfort; it helps decide whether callers stay on the line — more on that in the article "How natural does an AI phone assistant really sound today?".
The latency budget: five items, one chain
Between your last word and the first sound of the answer, a whole chain of processing steps takes place — each one costs time, and they add up. There is no standard for it. Useful reference values are the targets that the communications platform provider Twilio gives in a guide of November 2025 for launching a voice agent (median values):
| Step | What happens there | Reference value per Twilio (target / upper limit) |
|---|---|---|
| Phone network | Your audio travels through the network to the assistant, and the answer travels back | around 230 ms (difference between the total and the value within the platform) |
| Endpointing | The system recognises: "the caller has finished speaking" | included in the speech recognition value; fixed silence thresholds mostly around 500 ms |
| Speech recognition | Speech becomes text | 350 / 500 ms |
| Language model | The AI formulates the start of its answer | 375 / 750 ms to the first word |
| Speech synthesis | Text becomes the first audible sound | 100 / 250 ms to the first sound |
For the whole chain — from the end of your sentence until the answer reaches your ear, phone network included — Twilio gives around 1.1 seconds as the target (1,115 ms) and 1.4 seconds as the upper limit. That is a multiple of the gap between two people: an AI chain answers more slowly than a person, and every tenth of a second saved counts. The values are reference values from a platform provider, not measurements of individual assistants: real values vary with the system, the load and the time of day.
Endpointing: the hidden item
The least known item is the second one — and with a fixed silence threshold one of the largest. Before the assistant is even allowed to start answering, it has to be sure that you have finished speaking. That is harder than it sounds: a pause for breath in the middle of a sentence ("I'd like an appointment … on Thursday") must not count as the end of the sentence, or the AI will talk over you. But if the system waits too long to avoid that, you get exactly the awkward silence it is trying to prevent. Systems with a fixed threshold wait for a set amount of silence — according to Twilio mostly around half a second; a shorter threshold is risky because natural pauses then count as the end of a sentence. This waiting time, in which technically "nothing" happens, is easily overlooked in advertised figures. When two assistants feel differently fast, the difference often lies here too.
How to test the latency yourself
No measuring equipment needed — one test call and three observations:
- Count the pause: Ask a question and then silently count "one thousand and one" — that takes about a second. If the answer comes at about that point, the system is in the range of the reference values above; if you get to "one thousand and two" or further, the pause is considerably longer.
- Pause in the middle of a sentence: Speak with a distinct thinking pause ("I need an appointment … for next week"). If the assistant cuts into your pause, the endpointing is set too impatiently.
- Interrupt: Start talking while the assistant is still speaking. Good systems stop and listen; bad ones stubbornly talk to the end.
Honestly: where the figures mislead
Advertised latency figures should be read with care. Common patterns: only one item of the chain is named (the speech synthesis, say, with "under 100 ms") and misread as the total; a best case is stated instead of a typical value; or it is measured under laboratory conditions while latency fluctuates over the course of the day. A single number with no statement of what was measured from where to where says little. Your own test call — several times, at different times of day — is more informative than any data sheet. What an AI phone assistant fundamentally is and has to be able to do is summed up in "What is an AI phone assistant? Technology, limits, law, cost".
That is why we deliberately give no figure for scoco.ai here either, but the yardstick that counts: the assistant picks up and answers before the caller hangs up — and that is exactly what you can check for yourself.
Frequently asked questions
What response time is normal for an AI phone assistant?
There is no standard, and providers measure differently. As a reference value, Twilio (guide of November 2025) gives around 1.1 seconds as the target and 1.4 seconds as the upper limit for the whole chain of network, endpointing, speech recognition, language model and speech synthesis. For comparison: people answer after around 200 milliseconds on average.
Why does the AI answer faster sometimes and slower at others?
Because several items in the chain depend on load — the systems behind them are used by many conversations at once, and the length of the answer plays into it as well. Variation in the range of tenths of a second is normal.
What is endpointing?
The detection that a caller has finished speaking. Systems with a fixed silence threshold mostly wait for around half a second of silence (Twilio) — an often overlooked part of the response time. Set too impatiently, the AI interrupts; set too patiently, awkward silence follows.
Can I check a provider's latency before signing a contract?
Yes, if it offers a test number or trial access. Call several times, at different times of day, pay deliberate attention to the pauses and do the interruption test. That tells you more than any data-sheet discussion.
Status & sources
As of 2026-11-15.
- Stivers, T. et al.: Universals and cultural variation in turn-taking in conversation. PNAS 106 (26), 10587–10592, June 2009 — ten languages; mean response offset +208 ms, per-language means from +7 ms (Japanese) to +469 ms (Danish)
- Kohtz, L. S.; Niebuhr, O.: How Long is Too Long? How Pause Features After Requests Affect the Perceived Willingness of Affirmative Answers. Interspeech 2017 — perception experiment with 28 German listeners; perceived willingness falls with pause length, markedly beyond about 600 ms
- Twilio, "A Guide to Core Latency in AI Voice Agents (Cascaded Edition)", 17 November 2025: median targets for launching a voice agent — mouth-to-ear turn gap 1,115 ms (upper limit 1,400), platform turn gap 885 (1,100), speech-to-text 350 (500), LLM time to first token 375 (750), text-to-speech time to first byte 100 (250); fixed endpointing thresholds "mostly around 500ms"
- Note: the time values are reference values from a platform provider, not an assurance from any individual provider and not a measurement of scoco.ai; real values vary with the system, the load and the time of day
- scoco.ai product information: as of 15 November 2026