How AI phone calls work: the technology, the delay and what callers notice

Two years ago an AI phone call was a novelty that mostly embarrassed whoever tried it. In 2026 it is ordinary: businesses use AI voice agents to ring new enquiries back, answer out of hours, confirm appointments and work through lists. The technology underneath has changed a great deal in that time, and knowing roughly how it works makes you a much better buyer.

This is the non-engineer's version: what happens during a call, why the pause before the AI answers matters so much, and what still goes wrong.

What happens during an AI call

From the caller's side it is a phone call. Their phone rings, or they dial your number, and a voice speaks. Behind it, four things are connected:

1. A phone carrier connects the call over the normal telephone network and turns it into a stream of audio.
2. A voice model listens to that audio and produces speech in reply.
3. Your instructions and knowledge tell the model who it is, what it is trying to achieve and what it may say.
4. Tools let it do things during the call: check a diary, book a slot, transfer to a person.

The interesting part is the second. There are two ways to build it.

Two designs: cascade and speech-to-speech

The older design is a cascade of three models. One transcribes the caller's speech into text. A language model reads the text and writes a reply. A third turns the reply into speech. Each step is mature and controllable, but each adds delay, and everything about how the caller spoke, such as hesitation, irritation or a half-finished sentence, is lost when their words become text.

The newer design is speech-to-speech. One model takes audio in and gives audio out, hearing tone and timing directly. It can listen while it talks, which is what lets a caller interrupt naturally.

Neither wins outright on getting the task done. In ServiceNow's EVA-Bench, the most accurate system was a cascade and the best conversational experience came from a speech-to-speech model; no system scored above half on both at once. The direction in late 2026 is a hybrid: a fast speech model that holds the conversation and passes harder reasoning and tool use to a larger model in the background. We compare the current models in voice AI models compared.

The pause: how fast is fast enough?

Human conversation runs on remarkably tight timing. Across languages, the gap between one person finishing and the next starting averages about 200 milliseconds, a fifth of a second. That is faster than a person can plan a sentence, which means we start preparing our reply before the other person has finished.

We also read meaning into gaps. In one study, listeners judged a speaker as less willing to agree once the silence before their answer reached about 600 to 700 milliseconds. A pause isn't neutral. On a phone call, with no face to watch, it reads as doubt, confusion or a dropped line.

Now compare the machines. Measured at the model, the leading systems in November 2026 take between about 0.6 and 1.4 seconds to begin speaking, according to Artificial Analysis. That is before the phone network, the carrier and any tool calls are added.

On real phone lines the picture is slower. The testing firm Cekura called nine vendors' public phone agents and measured median response delays from 2.4 to 3.8 seconds. Marketing pages for the same category commonly claim around half a second.

Interruptions and turn-taking

Speed is half of timing. The other half is knowing when it is your turn. Real callers pause mid-sentence, say "um", change their minds and talk over the agent. An agent has to decide, many times a minute, whether a silence means "I've finished" or "I'm thinking".

Models differ widely here. In Sierra's τ-Voice benchmark, in realistic conditions, one leading model interrupted callers on 14% of turns and another on 84%. In Full-Duplex-Bench, which uses real hesitant human speech, the hardest case was a caller correcting themselves mid-sentence: the best speech-to-speech model handled 59% of those, the cascade 18%.

When you test an agent, this is what to listen for. Interrupt it. Start a sentence, stop, and start again. Say "sorry, no, I meant Thursday".

What still goes wrong

The benchmarks of 2026 agree on where the weak points are.

• Names, emails and codes. τ-Voice called authentication the dominant bottleneck. EVA-Bench found failures cluster on alphanumeric codes. "Was that F for Freddie or S for sugar?" is still hard for a machine on a compressed line.
• The phone line itself. Ordinary calls carry a narrow band of sound, far less than a video call. Transcription that is nearly perfect on clean audio is noticeably worse on real phone conversations.
• Accents. In τ-Voice, accented speech cost about ten percentage points of task success on average, and one model lost eighteen. We found no published benchmark for UK regional accents, so test with your own customers.
• Background noise. A caller in a van or a café mostly damages turn-taking: the agent hears noise as speech and stops, or misses the end of a sentence.
• Consistency. On recordings of real callers, the best model in Cekura's test handled 43% of scenarios correctly on all three attempts. Passing once is much easier than passing every time.
• Voicemail. Working out whether a person or an answering machine has picked up is a guess made in under a second. Twilio's own documentation says its detection will not always return the right answer.

None of these is a reason not to use AI calls. They are reasons to design around them: confirm anything exact in writing, keep a person within reach, and let the system, not the model, enforce the rules. The same themes run through what AI agents can't do yet.

What an AI call is good at

Within those limits the case is strong.

• It is immediate. A new enquiry can be rung back in seconds, at any hour you are allowed to call.
• It is patient. It will ask the same six qualifying questions the thousandth time as carefully as the first.
• It is available. Evenings, weekends and the lunch hour, when small teams miss calls.
• It is consistent about the rules. Calling hours, opt-outs and limits can be enforced on every call.
• It leaves a record. Every call is transcribed and saved against the contact.

The calls it suits are the ones with a clear purpose and a small number of likely paths: qualify and book, confirm or rearrange, answer and take a message. Open-ended, emotional or high-stakes calls still belong to people, ideally with the AI's notes in front of them.

How to test an AI voice agent

You do not need a lab. You need twenty minutes and a mobile phone.

1. Call it from a mobile on an ordinary signal, not from a laptop on office wifi.
2. Time the pauses. Count "one thousand" after you stop speaking. If you get to two, callers will notice.
3. Interrupt it in the middle of a long answer. Does it stop, and does it pick up sensibly?
4. Give it something exact: spell a surname, read out an email address and a postcode. Check what it saved.
5. Go off script. Ask something it shouldn't know. A good agent says so. A bad one invents an answer.
6. Add noise. Try again with the radio on or from the street.
7. Ask for a person. See what actually happens next, and what your team would see.
8. Let it go to voicemail. Hear what it leaves, if anything.

Then ask someone with a different accent to do the same.

Where this leaves you

AI calling works, on the right calls, when the design accepts what the technology is like today: quick but not instant, fluent but unreliable with exact details, and best with a person one step away. In Unibound an AI call is one step in a flow, so what happens after "no answer", "wants a person" or "booked" is drawn out in advance, and you design the voice and hear it on your own phone first.

Next: what an AI voice agent costs in the UK, and the UK rules for AI calls and texts.

Questions people ask