---
title: First message and conversation
description: "Who speaks first on a call and what the agent says when it does, set in the Conversation card on a voice agent, with the wording that works."
---

<Badge variant="accent">Voice agents</Badge>

The **Conversation** card on **Settings** decides how a call opens. **Who talks
first** chooses between the agent greeting immediately and the agent waiting for
the caller; **First message** is the exact line the agent says when it opens.

:::note[Text agents]
Text agents have no Conversation card. The person always writes first, and the
agent's first reply comes from the prompt.
:::

![The Conversation card.](/media/build/conversation-card.webp)

## Configuring it

| Field | Options | Default |
|---|---|---|
| **Who talks first** | **Agent speaks first**: the agent greets the caller immediately. **Human speaks first**: the agent waits for the caller to speak. | **Agent speaks first** |
| **First message** | Free text. The opening line, shown only while **Who talks first** is **Agent speaks first**. Dynamic variables such as `{{customer_name}}` are filled in. | Empty |

![The two Who talks first options.](/media/build/who-talks-first-options.webp)

## Choosing who talks first

- **Outbound calls**: the agent. The person answered a call they did not expect,
  and silence makes them hang up. Say who is calling and why in the first sentence.
- **Inbound calls to a business line**: the agent. Callers expect a greeting.
- **Transfers from a person or an IVR**: usually the caller, who was told to state
  their request. Pair it with the [reminder](/build/speech-pipeline/silence-reminders-and-call-duration)
  settings so the agent prompts if nothing is said.
- **Answering machine detection** on outbound calls: with **Agent speaks first**
  the message starts as soon as the line opens. See
  [Answering machine detection](/build/speech-pipeline/answering-machine-detection).

## Writing the first message

- Name, company, reason, question. "Hi, this is Nora at Riverside Dental. How can I
  help you today?"
- One breath. A long opener is talked over, and the turn detector then treats the
  interruption as the caller's first turn.
- Match the prompt. The identity in **Agent Identity** and the name in the first
  message must agree; the model notices when they do not.
- With **First message** empty and **Agent speaks first**, the model improvises an
  opener from the prompt. Fine for tests; write one before going live.

## Verify

Open **Call** under **Testing** and start a call. With **Agent speaks first**, the
first message is spoken before you say anything and appears as the first transcript
line.

## Troubleshooting

- **The agent does not speak first.** **Who talks first** is on **Human speaks
  first**. Switch it and **Save**.
- **The opener differs from what I wrote.** The change is unsaved, or a dynamic
  variable in it has no value. Check the unsaved bar and the **Variables** dialog.

## Related

- [Prompting guide](/build/prompting/prompting-guide)
- [Dynamic variables](/build/dynamic-variables)
- [Turn detection](/build/speech-pipeline/turn-detection)
- [Silence, reminders and call duration](/build/speech-pipeline/silence-reminders-and-call-duration)
