PolyAI

PolyAI sells to enterprise contact centres that want the phone queue answered before a human picks up. The buyer already runs a CCaaS platform or telephony provider, and calls reach the assistant over a SIP or PSTN link into it, so nothing is ripped out first. Named verticals run from banks to hotels and utilities.

The stack is PolyAI's own: the Raven LLM, an in-house ASR called Owl, and a patented dialogue policy trained on what it reports as a billion real conversations. Billing runs on conversation minutes, with no seat anywhere in the model. Every plan carries a 99.9% uptime SLA on your phone lines.

What it does not do is publish the number. The pricing page is a form wanting your employer, phone number and annual call volume; its whole disclosure is one sentence saying the minute rate covers maintenance and support. No plan names, no minimum, no allowance, no trial. Whether telephony sits inside that minute or arrives as a separate carrier charge is unstated, as is whether a minute rounds up. Language support reads 45 on one page and 75 on another, and every export path is undocumented.

Retell AI meters the same axis in the open, $0.055 a minute for voice infrastructure plus $0.015 for telephony and a per-model LLM line, so its bill is calculable in advance. Sierra is equally silent on rate but bills a resolved outcome, moving the risk from call length to that definition. All three sit in AI voice agents.