Ana içeriğe geç

Put an assistant on a live call

POST 
/api/calls/:callId/ai

Runs the whole chain: resolve which assistant should answer, take one of its concurrent slots, deliver the EU AI Act Art. 50 disclosure, ask the session gate, build the provider's dial plan, and hand it to the media bridge.

A refusal is a 409 with a reason, not an error​

Nothing about the request is wrong when an assistant may not open — the answer is "not on this call, not right now". The reason is a machine-readable key from VoiceAgentRefusalReason, and every one of them has a translated sentence in the panel catalogue.

The disclosure is delivered BEFORE the model is dialled, never after and never instead: a session whose disclosure could not be played is refused with disclosure_unavailable and no provider socket is opened.

Today this works on PHONE LINES ONLY​

No AI provider surveyed takes Opus at 48 kHz, and the media plane refuses to resample 48 -> 8 kHz rather than do it badly. A WhatsApp or Messenger call therefore cannot carry an assistant: it is refused with ai_leg_wire_unsupported, and VoiceCapabilities.control.ai_leg is false on those channels so the control is not drawn at all.

If the provider socket drops, we do NOT re-dial​

The bridge parks the customer on hold music and reports ai_stream_dropped; the leg moves to escalating and a PERSON is the expected next actor. Re-dialling would be a NEW session with no history (OpenAI's realtime WebSocket has no resumption), so the assistant would greet the caller again mid-conversation — and would owe a second Art. 50 disclosure. Calling this endpoint again re-runs the whole chain deliberately, disclosure included.

Request​

Responses​

The assistant is on the call.