You can cut in mid sentence
When the customer starts speaking, the agent stops and listens. It does not finish its sentence out of stubbornness or ask for a tone.
The customer clicks and speaks. No app to download, no account to create, no queue to wait in. The agent listens, stops when interrupted, and answers from the same sources as in chat.
Four behaviours that separate a conversation from reading out announcements.
When the customer starts speaking, the agent stops and listens. It does not finish its sentence out of stubbornness or ask for a tone.
A short pause for thinking is not treated as the end of a sentence, so the agent does not talk over someone who is still deciding.
Feedback suppression keeps the agent's own voice from returning as customer speech. Without it, a speakerphone conversation falls apart after two sentences.
A separate watchdog ends a session that stopped responding, instead of leaving the customer with silence.
Worth knowing which parts run on our servers and which at a provider, because it maps directly onto the subprocessor register.
| Component | Where it runs |
|---|---|
| Audio transport | Our server in Warsaw |
| Speech recognition | The model provider (OpenAI), for the duration of the call |
| Understanding and content selection | Language model provider |
| The agent's voice | Speech synthesis provider |
| Knowledge base and citation gate | Our server |
| Transcript and audit log | Our server |
Voice in the browser and voice on a call are the same conversation over a different transport. The second half has been working since 1 September 2026.
The fastest way to judge a voice channel is a conversation, not a description of one. Use the button in the avatar on the home page.