For a decade, “voice AI” on a business line meant a tree: press 1, wait, press 2, wait, leave a message that nobody transcribed. The demos at conferences were always a calm speaker in a quiet room. Miami shops do not run quiet rooms. Someone is yelling an address over a compressor, and they will talk over the agent the moment they hear a pause.
What changed in the last two years is barge-in. If the system cannot stop talking when the caller starts, it still feels like a tree. Latency under a few hundred milliseconds is the other half. Anything slower and people hang up, even if the voice is pretty.
Why the lab versions failed here
South Florida inbound is bilingual, impatient, and often from a car. Models trained on headset audio in English-only offices miss the first turn. The shops that tested early versions in 2024 told us the same story: the agent sounded fine until the caller said “no, wait — Hialeah,” and then it looped.
What to listen for on a test call
Call your own number from a noisy parking lot. Interrupt the greeting. Give a partial address. If the system apologizes and restarts the script, it has not left the lab. If it keeps the thread and asks one clarifying question, it might actually survive a Thursday in August.
This is a product story, not a payroll story. The labs got the voice. The field still punishes anything that cannot be talked over.