Somewhere between 2023 and 2025, talking to a machine stopped feeling like leaving a voicemail and started feeling like dialogue. The change was latency, mostly: real-time speech models got fast enough to interrupt, and interruption is the entire difference between narration and conversation. The voice wave followed, and it changed customer AI's social contract.
Why latency was the whole ballgame
Earlier voice assistants failed on turn-taking, not comprehension: the pause between your sentence and theirs was long enough to break the rhythm humans call conversation, so every exchange became a command and a wait. When the round-trip collapsed, the interaction inherited conversation's native physics, you could cut in mid-sentence, redirect, talk over, trail off, and the machine kept up. Tone survived the trip too, and with it the things tone carries: urgency, confusion, satisfaction. Mid-sentence language switching, the way real multilingual committees actually talk, became followable. None of this was a feature; it was a threshold, and crossing it re-opened every customer-facing use case voice had previously flunked.
What voice actually changed
Bandwidth first: people explain situations several times faster by talking than typing, and the interactions where customer AI earns its keep, the stuck user, the evaluating buyer, the learner mid-question, are exactly the high-back-and-forth kind. Presence second, and more subtly: a voice is a participant. It can join a kickoff call, hold a demo's attention, sit on the video call where the deal actually happens, places a chat window structurally couldn't go. The wave produced AI SDR calls, AI phone support, avatar-fronted everything, some of it excellent, some of it uncanny, all of it proof that the interface had genuinely changed.
The blindness the wave inherited
Most voice agents were chat's architecture wearing a headset: they could hear the user and still couldn't see the screen or touch the product. That reproduced the oldest support failure at higher fidelity, a fluent voice asking "can you describe what you're seeing?" is still an interview, just a warmer one. A voice that cannot see is a phone call from someone who has never used your app, and users felt the gap even when they couldn't name it: the conversation was natural; the help was still secondhand.
Voice's real place in the story
Voice was never the destination; it was the stage that made the destination social. Perception and action, the next wave, gave agents something to say worth hearing ("I can see it, let's fix it together"); voice is what lets them say it like a colleague instead of a form. In the synthesis, voice is the steering wheel: the channel through which users direct eyes and hands they trust. (Disclosure: Skippr's agents speak ten languages and switch mid-conversation, but always as one organ of four, the wave taught everyone that a beautiful voice alone is half an agent.)
See what a live agent actually does
The category is easier to watch than to define. Fifteen minutes is enough to see where the mechanism differs from everything it gets confused with.