Who speaks first

When someone answers the phone, they usually say hello before anything else. On outbound calls the agent lets the person say hello first. It listens from the moment the call is answered and speaks once the person has finished, so its opening does not land on top of their greeting and they do not have to say hello twice.

Configure the opening

Open AI agents → your AI agent → Identity. Beside First message, choose:

• Wait for the person (default): listen for their greeting. If nobody speaks, wait two seconds by default; choose a silent wait between one and five seconds.
• AI agent speaks first: request the opening as soon as the outbound call is answered. The saved silent wait is retained for switching back.

These settings apply to Standard, GPT-Live 1 and Gemini Live. Voice and engine selection stay on Voice. Keep your opening words in First message and behavioral instructions in System prompt; there is no duplicate editor. Publish changes to use the new opening on subsequent calls. A call already admitted keeps the opening policy in its publication snapshot.

The wait applies to outbound telephone calls. An inbound call or browser test still opens normally, so a browser test alone does not verify pickup timing.

API clients save the same policy with PATCH /v1/agents/{agent_id}/voice-stack:

{
  "opening": {
    "first_speaker": "callee",
    "silence_wait_seconds": 2
  }
}

Use "agent" for an AI-agent-first opening. An empty opening object resets both settings to their defaults; omitting opening preserves the saved policy. The API rejects values outside one to five seconds and unknown fields.

Before the call is answered

On every outbound call, whatever the conversation engine, the agent neither hears the line nor speaks until the person answers. Ringback, carrier announcements and any other sound the line carries while it is still ringing never reach the agent, so it cannot mistake them for the person or talk over them.

What the agent waits for

• The person speaks. The agent waits until they have finished, then replies. There is no time limit on a person: a long greeting such as "Good afternoon, Harbour Street Pharmacy, Sam speaking, how can I help?" is heard to the end. A short pause between sentences does not end their turn; a pause of about 0.6 seconds or more (0.8 seconds in a language without a turn model; see Turn detection by language) starts the response decision. The agent also allows a brief window for the person to resume before releasing a reply.
• Nobody speaks. If the line stays silent for your configured wait after the call is answered (two seconds by default), the agent requests its opening.
• The line plays sound that is not a voice. Hold music, a tone or other continuous sound is not silence, so the two-second rule does not apply to it. After six seconds of it, the agent requests its opening once. Recognising an answering machine or a looping recording is a separate matter, covered in Answer detection and Voicemail detection.
• The person carries on. If the person pauses and then keeps talking before the agent has started its reply, the agent waits for them to finish again.

What the agent says first

• If the person only greets the agent, it uses your First message. Standard reads a literal opening as written; live models follow the opening instruction and may phrase it differently.
• If they said something else, such as "wrong number", "who is calling?", "can you call me back later?" or another question, the agent answers that first and then carries on toward the purpose of the call. It does not read its opening over what it has just heard. On Standard, that reply can use the agent's tools like any later one: a call-back request made on picking up is recorded with set_call_disposition as it is answered. This works in Danish, Dutch, English, Estonian, Finnish, French, German, Italian, Norwegian, Polish, Portuguese, Russian, Spanish, Swedish and Ukrainian; in other languages a greeting cannot be told from a request, so the reply to a pickup uses no tools. If that reply has to be asked for a second time, the second try uses no tools either. On GPT-Live 1 and Gemini Live, the reply Skysay asks for in place of the model's own uses no tools yet, so a request made only on picking up is not always recorded; the same request made later in the call is. An opening said before the person speaks, and the reply to a greeting alone, use no tools.
• If your agent's opening is generated rather than a fixed line, it is written with what the person said in view.

On GPT-Live 1 and Gemini Live, the model hears the pickup while Skysay waits to release its first response. If the person resumes during that wait, Skysay discards the pending response to the incomplete pickup. The model still decides the words: a timing safeguard does not guarantee that it understands every request or repeats a configured opening exactly. Test your agent's languages and instructions with representative calls.

The two-second silent-line rule is when the opening is requested. Model and audio processing add time before the person hears it. If caller audio briefly stops arriving, Skysay waits for a fresh observed quiet interval; it does not count the missing audio as silence.

The opening is an ordinary reply. The person can interrupt it, and the agent stops and listens, as it would at any other point in the call.

Which calls this applies to

The outbound rows describe the default Wait for the person setting. Choosing AI agent speaks first changes those outbound rows only.

For a workflow, pickup speech can trigger a configured global interrupt, such as a callback or stop request, before the entry step. If no interrupt matches, the workflow speaks its entry step. A greeting or other pickup never counts as an answer to a question that has not been spoken. Which interrupts exist and where they lead remain part of the workflow. Live-engine workflow binding is not added by this opening setting.

Listening during speech and protecting a pending hangup are automatic engine behavior. They do not introduce additional Identity switches. Question-attempt limits are a separate workflow capability and are not part of this change.

Time on the call

The time the agent spends listening is part of the call. Outbound time is billed from the moment the call is placed, as described in Billing. The rules that end a call nobody speaks on are unchanged: on a silent line the AI agent requests its opening after the configured wait, and the call then ends as described in Answer detection if they never reply.