Voice is not chat with a microphone: what changes when the AI agent speaks

A voice agent has to answer inside the rhythm of speech, where people expect a reply within about 200 milliseconds, cannot show a list, cannot be re-read and can be interrupted at any moment. Everything built for text needs three changes: shorter turns, spoken confirmation of anything that matters and a plan for silence.

Teams that already run a chat agent often assume voice is the same agent with a speech layer on top. The knowledge carries over. The writing does not.

How fast do people answer each other?

Faster than a typing indicator would allow. A study of turn-taking across languages found the same pattern in all of them:

all languages show a similar distribution of response offsets (unimodal peak of response within 200 ms of the end of the question).

The paper is about people, not machines, and it is the reason a pause of a second on a phone call feels long. In chat nobody notices two seconds of typing. On a call, two seconds is where the caller says hello to check the line is alive.

What does an AI agent do with a gap it cannot close?

It chooses what the gap sounds like. There are three honest options, and one dishonest one.

• A short acknowledgement. "Let me check that" tells the caller the line is alive and that something is happening.
• A streamed first phrase. The agent starts speaking before it has the whole answer, as long as the first words do not depend on the rest.
• A deliberate pause, used rarely, before something the caller should hear carefully.

The dishonest option is silence with no signal. The caller cannot tell whether the agent is working, stuck or disconnected, and some will hang up or talk over it.

Why can a voice agent not read a list?

Because the caller cannot see it, scroll back or copy a code. A chat answer can offer five options. A spoken answer can offer two or three, in the order the caller is most likely to want them, and then stop and ask.

Numbers are the sharper case. A chat message with a booking reference can be read twice. A spoken one has to be chunked, repeated back and confirmed. Treat anything the caller will write down, such as an amount, a date, a reference or an address, as something the agent says slowly and confirms aloud.

What happens when the caller interrupts?

It will happen on most calls, and a good agent stops talking. People interrupt to correct, to answer early or to say they have heard enough. An agent that finishes its sentence over the top of the caller teaches them it is not listening.

The design consequence is that turns must be short enough to be cut off at any point without losing the thread. An agent that delivers its important fact in the last clause of a long sentence has designed for a caller who never interrupts.

What about noise, accents and a bad line?

Treat every input as uncertain. On chat the text is exact. On a call the agent hears a transcript that may be wrong, and a wrong name or number is expensive. The rule that follows is simple: the more a mistake costs, the more the agent confirms. A greeting needs no confirmation. A payment amount needs one.

The phone number is a hint about who is calling, not proof of it. Anything that moves money or changes an account should ask for something the caller knows, as it would on any other channel.

Does the agent have to say it is an AI?

In some countries, yes, and a call is where it is hardest to hide. In the US the FCC ruled in 2024 that AI-generated voices are an artificial voice under the TCPA, which brings consent and opt-out rules with it. In the EU, Article 50 of the AI Act requires people to be told they are interacting with AI from 2 August 2026, unless it is obvious. We summarise the rules by country in another post, and the question of when to say it is a design decision we cover in a tool.

What happens when a person has to take over?

A handoff on a call is a transfer, and the caller is waiting while it happens. The person who picks up should already have the reason for the call and what the agent did, so the first thing they say is not "how can I help you". We cover what has to travel in the post on handoffs.

Where does StudioChat sit on voice?

Our published channel list includes a phone-call agent with live transcription and tool execution, next to web chat, WhatsApp Business and email. The same knowledge and the same actions serve all of them, and the voice-specific work is in the turn design described above. For what orchestration across channels means, see our post on platforms.