Human conversational turn-taking runs on gaps of roughly a couple of hundred milliseconds. That is not a target anyone chose; it is what people are used to, and departures from it carry meaning.
A longer gap is not neutral. In human conversation, a pause before answering signals hesitation, reluctance or disagreement. Callers apply the same reading to a machine, which is why a slow voice agent is perceived as unsure rather than merely slow.
Three thresholds
- Around the natural turn gap: the exchange feels like a conversation.
- Around half a second to a second: noticeably machine-like, still workable.
- Beyond about a second and a half: callers start talking again, assuming they were not heard - which produces a collision the agent then has to recover from.
These are orders of magnitude from observing calls, not measured thresholds, and they vary with the caller and the context. The shape is robust; the boundaries are soft.
A consistent 700 ms is easier to talk to than a system that alternates between 200 ms and 1,400 ms. Callers calibrate to a rhythm; an unpredictable one prevents that and produces exactly the collisions above. Optimising the tail matters more than optimising the mean.
The silence problem
The special case is a tool call. Waiting on a calendar or CRM lookup produces dead air, and dead air on a telephone line reads as a dropped connection. Callers hang up on it. Saying "let me check the calendar" before the wait - rather than after - costs nothing and removes the failure.
This is the clearest example of latency being a design problem before it is a performance problem: the wait did not get shorter, and the complaint disappeared.