Glossary
Latency
For an AI phone assistant, latency is the time between the end of something the caller says and the start of the audible reply.
Latency is made up of several steps: detecting that the caller has finished, converting speech to text, generating the reply, synthesising speech and carrying the audio over the telephone network. Every step takes time. Tool calls, such as a calendar query, are added to that.
In conversations between people, pauses are short. If the pause becomes noticeably longer, the conversation feels halting and callers start speaking again. Vendor figures often refer to a single component only. The only meaningful figure is a measurement across the whole chain, on a real phone and over several calls. Shorter latency is not always better: if the system answers too early, it cuts the caller off.
In practice
You can measure latency yourself: call the assistant from an ordinary phone, record the call and measure the pauses between your last word and its first syllable. Repeat this with short questions, with long sentences and with questions that trigger a look at the calendar. The pauses will differ in length. The pause can be shortened through briefer answers, leaner instructions and fewer tool calls per turn. Where one step clearly takes longer, a short filler such as “One moment, I’ll check” helps.
An example
A hairdresser tests her assistant. “Are you open on Saturdays?” is answered almost at once. After “Do you have anything free on Saturday for a cut and colour?” the line stays noticeably silent, because the assistant queries the calendar first. Two test callers say “Hello?” into that silence. Since the assistant says “I’ll just look at the calendar” before the query, the very same waiting time bothers nobody.
A common misunderstanding
Latency is often understood as a fixed property of a provider. Yet it varies from call to call, depending on the question, the load and the connection. A single average hides the outliers, and those are exactly what callers remember. So watch the long pauses and not only the mean. And never compare a measurement in the browser with one on the phone; the path through the telephone network always comes on top.
Read more
As of: 12 October 2026