The 700ms wall: what callers actually do at each latency band

Everyone benchmarks voice-AI latency. Almost nobody talks about the thing that actually matters: what callers DO at each delay band.

Latency isn't a number on a slide. It's a behavior trigger. Cross the wrong threshold and the caller stops cooperating with the conversation.

The three bands

Under 500ms — human

The reply lands in the rhythm a person expects. Callers don't notice the technology; they just have a conversation. This is the band you're aiming for on every turn.

Around 700ms — "robotic"

It still works, but something feels off. The caller can't name it, but they downgrade their trust: they speak more slowly, repeat themselves, start "talking to a machine" instead of talking to an agent. You haven't lost the call — but you've lost the illusion.

900ms and up — they take over

This is the wall. The silence is long enough that the caller assumes it's their turn: they talk over the agent, repeat the question, or hang up. No model quality saves you here — the answer arrives after the caller has already moved.

Why this is a conversion line, not a benchmark

A demo can hide latency. A real call can't. The difference between a 480ms agent and a 720ms agent doesn't show up in a spec sheet — it shows up in completed bookings, in callers who stay on the line, in the share of calls that resolve without a human.

That's why we treat the sub-500ms target as a product requirement, not a brag.

Measure the gap the caller feels

Per-component latency is for tuning. The number that predicts behavior is end-of-speech → first-audio: the moment the caller stops talking to the first sound of the reply. Optimize that, and the bands above take care of themselves. For the engineering picture, see how we get voice AI under 500ms — and for the nuance the infra community pushed back with, latency, corrected: P95, P99, and jitter.