Realtime API

The Realtime API powers speech to speech agents that listen, think and reply at the pace of a normal conversation, either over the phone or inside your own product. The difference from older voice systems is the timing: there is no long pause while it transcribes, thinks and then speaks.

Yes, because we tell them. Disclosure is good practice and callers respond well when the agent is fast, useful and honest about what it is.

Immediately, and it should. Escalation paths are part of the design, with context passed along so the caller never repeats themselves.

The models handle a wide range of accents and languages well, and we test against your actual caller base during the pilot rather than assuming.

Usage based audio pricing, driven by call volume and length. We model it against your current cost per call, which is usually the more confronting number.

After hours enquiries, booking and rescheduling, order status, triage before a human picks up, and outbound reminders. The good use cases are high volume and repetitive, where the alternative is a caller waiting on hold or leaving a message.

Handle a distressed caller, make a decision about someone’s money or care, or pretend to be a person. We design the handover points before we design the conversation, because the exception path is what people judge you on.

Narrow scope, approved content and a hard limit on what it will discuss. A voice agent that answers anything will eventually answer something badly, and unlike a chat window there is no chance to read it before it is heard.