> ## Documentation Index
> Fetch the complete documentation index at: https://sitegpt.ai/docs/llms.txt
> Use this file to discover all available pages before exploring further.

# Reply to visitors and take over conversations

> Answer visitors from Chat History, take a conversation from the AI, give it back, and close it when you finish.

Your team answers visitors from **Chat History** in the dashboard. You can reply, take the conversation from the AI, give it back to the AI, and resolve it. For how these states work together, see [Conversations and handoff](/docs/concepts/conversations-and-handoff.md). You can also reply from ChatGPT, Claude, or another AI assistant. See [MCP server](/docs/developers/mcp.md#remote-server-tools).

## Before you start

* **Plan:** any plan. See [Plans and limits](/docs/reference/plans-and-limits.md).
* **Role:** any team role, including Agent. See [Team members](/docs/guides/account/team-members.md).
* To get alerts for new escalations, set up [notifications](/docs/guides/human-support/notifications.md).

## Find conversations that need you

1. In your chatbot's sidebar, select **Chat History**.
2. Select the **Attention** view. It lists open conversations that are escalated, have an unread visitor message, or have negative feedback.
3. Select a conversation to open it.

Escalated conversations show the **Escalated** badge. For all views and filters, see [Review conversations](/docs/guides/monitor/review-conversations.md).

## Reply to the visitor

**1. Write your reply**

At the bottom of the conversation, type your message in **Reply as an agent...**.

**2. Send**

Select the send button.

When you reply, the conversation switches to human mode. The AI stops answering in this conversation. The visitor sees your reply in the chat.

Replies from **Chat History** reach visitors who chat in the website widget, on WhatsApp, on Messenger, or by email. On WhatsApp, a reply is delivered only within 24 hours of the visitor's last message, and only when the WhatsApp add-on covers the number. See [WhatsApp](/docs/integrations/whatsapp.md#reply-from-the-dashboard). On Messenger, a reply is delivered within 24 hours of the visitor's last message, or up to 7 days later if your Meta app has Meta's Human Agent approval. See [Messenger](/docs/integrations/messenger.md#reply-from-the-dashboard). In an email conversation, the reply goes to the customer as an email in the same thread. See [Email](/docs/integrations/email.md#reply-from-chat-history). Replies do not reach visitors on Slack and other connected channels. Reply to those visitors in that tool. See [Integrations](/docs/integrations/overview.md).

A reply alone does not lock the conversation. The visitor can still select **Continue with AI** to go back to the AI. To stop this, take over the conversation before you reply. The dashboard shows **Take over the chat** only while the AI answers the conversation, or while an escalated conversation waits for your team to answer the visitor's last message. In human mode it shows **Release to AI** instead. To lock that conversation, use **Lock conversation**. See [Lock a conversation that your team holds](#lock-a-conversation-that-your-team-holds). With API v2, `POST /api/v2/chatbots/{chatbotId}/conversations/{threadId}/take-over` locks it.

A reply from an AI assistant (the SiteGPT MCP server) or from API v2 works differently. If the AI still answers the conversation, or an escalated conversation waits for your team, the reply takes the conversation over first. The visitor sees "A human agent has joined the conversation" and cannot select **Continue with AI**. See `reply_to_conversation` in the [MCP tool reference](/docs/developers/mcp-tools.md#chatbot-and-conversation-tools).

## Take over the chat

Take over when you want the AI fully out of the conversation.

1. Open the conversation in **Chat History**.
2. At the top of the conversation, select **Take over the chat**. It shows while the AI answers the conversation, or while an escalated conversation waits for your team to answer the visitor's last message.

While you hold the conversation:

* The AI does not reply. If an AI answer is being written when you take over, it stops.
* The visitor sees "A human agent has joined the conversation".
* The visitor cannot switch back to the AI. The **Continue with AI** button does not show.
* The visitor's messages still arrive in the conversation.

Taking over does not send escalation emails and does not mark the conversation as escalated.

### Lock a conversation that your team holds

After a team member replies, or in a conversation that started in human mode, the conversation is in human mode without a lock. The top of the conversation shows **Release to AI**, and the visitor can still select **Continue with AI**. To stop this:

1. Open the conversation in **Chat History**.
2. At the top of the conversation, open the **…** menu.
3. Select **Lock conversation**. The menu says "The visitor cannot switch to the AI."

**Lock conversation** does the same as **Take over the chat**. The visitor sees "A human agent has joined the conversation", and the **Continue with AI** button does not show. **Release to AI** still gives the conversation back to the AI in one click.

The menu shows **Lock conversation** only while your team holds the conversation without a lock. It does not show after a take over or in a resolved conversation. It also does not show while the AI answers, or while an escalated conversation waits for your team to answer the visitor's last message. In these two cases, use **Take over the chat**.

## Release to AI

When you finish helping and the visitor may have more questions, give the conversation back to the AI.

1. Open the conversation.
2. Select **Release to AI**.

**Release to AI** shows at the top of the conversation while your team holds it: after you take over, or while the conversation is in human mode, for example after a team member replied or because it started in human mode. One exception: an escalated conversation that no team member answered since the visitor's last message shows **Take over the chat** instead.

The visitor sees "Continuing with AI". The AI answers the visitor's next message in the same conversation.

## Mark as Resolved

Resolve a conversation when it is finished.

1. Open the conversation.
2. Select **Mark as Resolved**.

A resolved conversation is closed:

* It leaves the **Attention** view and moves to **Closed**. It shows the **Resolved** badge.
* If an AI answer is being written, it stops.
* The reply box is hidden, and you cannot take over the conversation.
* Lead forms and feedback buttons in that conversation stop working.

What the visitor sees depends on where they chat:

| Where the visitor chats | After you resolve |
| - | - |
| Website widget | The chat shows "This conversation has been closed." and a button to start a new conversation. The resolved conversation accepts no new messages. |
| API, with the resolved conversation ID | Message requests return `409 CONVERSATION_CLOSED`. |
| Slack, Messenger, WhatsApp, Google Chat, Crisp, Freshdesk Messaging, Zoho SalesIQ, or Zendesk Messaging | The visitor's next ordinary text message starts a new SiteGPT conversation. The resolved conversation stays closed. |
| Email | The customer's next email, also a reply in the same email thread, starts a new SiteGPT conversation. The resolved conversation stays closed. |

A new conversation uses your chatbot's [chat mode](/docs/guides/human-support/set-up-human-support.md#choose-who-answers-new-conversations). Resolving does not guarantee that the AI answers next. To continue the same conversation with the AI, use **Release to AI** instead.

To act on many conversations, select them in the list and select **Mark as Resolved** in the bar that appears.

## Mark as Unresolved

Only your team can reopen a resolved conversation. Open it and select **Mark as Unresolved**. A new message from the visitor never reopens it.

## When SiteGPT emails your reply to the visitor

Visitors often leave before you answer. SiteGPT can email your reply to them. This also applies to replies that you send from ChatGPT, Claude, or another AI assistant through the SiteGPT MCP server, and to replies sent with API v2.

This section does not apply to WhatsApp, Messenger, and email conversations. There, your reply goes to the person on WhatsApp or Messenger, or is itself the email, so SiteGPT sends no separate email. See [WhatsApp](/docs/integrations/whatsapp.md#reply-from-the-dashboard), [Messenger](/docs/integrations/messenger.md#reply-from-the-dashboard), and [See what happened to each reply](/docs/integrations/email.md#see-what-happened-to-each-reply).

SiteGPT sends the email only if all of these are true:

* SiteGPT has the visitor's email, for example from a lead form.
* The visitor was not active in the chat after your reply.
* The conversation is not resolved.
* The visitor did not mark an earlier SiteGPT email as spam.
* The chatbot owner has an email address.

SiteGPT waits about 2 minutes after your first reply. Replies you send in that time go out in one email. The subject is "New reply from" and your chatbot name. The email contains your reply and an **Open the conversation** link that opens the chat.

### When the visitor answers the email

If the email says "You can reply to this email.", the visitor's answer comes back into the conversation. This is the case for every chatbot that is not in a HIPAA workspace, also when the chatbot has no [email address](/docs/integrations/email.md). If it has an email address that answers, the answer goes back through that address.

* The answer shows in the conversation as the visitor's message, without the quoted earlier messages and the signature. Attachments show by name, type, and size. The files are not kept. The AI does not answer it.
* Your team gets an email with the subject "New email reply -" and your chatbot name. It goes to the same people as your escalation emails. The iPhone app also sends a push under **Handover requested**. See [Notifications](/docs/guides/human-support/notifications.md).
* SiteGPT takes only answers from the visitor's own email address. It ignores other senders and automatic replies, with no alert.
* SiteGPT adds the answer only if the sender's email provider vouches for the visitor's address, and the conversation is not resolved. Otherwise, the alert email contains the answer, and the answer is not added to the conversation.

In a HIPAA workspace, the email does not say "You can reply to this email.", and the answer goes to the first of these addresses that is set:

1. The chatbot's **Support Email** in **Settings** > **General**.
2. The first address in your escalation email list, if **Enable Escalation Notifications** is on.
3. The chatbot owner's email.

SiteGPT skips an address that is not a valid email address.

If the first address in the escalation email list is not valid, the answer goes to the chatbot owner's email.

On HIPAA workspaces, the email does not include your reply. It only tells the visitor that there is a new response.

### See whether the visitor got the email

The conversation shows a line under your last reply of each email batch. WhatsApp, Messenger, and email conversations show no such line, also for older replies.

| Line | Meaning |
| - | - |
| "Email to the visitor scheduled, unless they reply in the chat first." | The email waits for the end of the 2-minute window. |
| "Emailed to the visitor at" and the time | The email was sent. When it holds more than one reply, the line also shows the number of replies. |
| "Not emailed: the visitor has no email address." | SiteGPT does not have the visitor's email. |
| "Not emailed: the visitor was in the chat." | The visitor was active in the chat after your reply. |
| "Not emailed: the conversation was closed." | The conversation was resolved. |
| "Not emailed: the visitor marked an earlier email as spam." | SiteGPT does not email a visitor who reported its email as spam. |
| "Not emailed: there was no reply to send." | No reply was left to send. |
| "Not emailed: the chatbot owner has no email address." | SiteGPT needs the owner's email to send this email. |
| "The email to the visitor could not be sent." | Sending failed after the last retry. |

The conversation keeps the 20 most recent results. The line shows only in the dashboard, not in the iPhone app.

## Check that it works

1. In your chatbot preview, ask for a person, then send a message.
2. In **Chat History**, open the conversation from **Attention** and select **Take over the chat**.
3. Send a reply. The preview shows "A human agent has joined the conversation" and your reply. The AI does not answer.
4. Select **Release to AI**. Send another message in the preview. The AI answers.
5. Select **Mark as Resolved**. The preview shows "This conversation has been closed."


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.