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

# Autopilot settings

> Choose per channel whether the AI replies on its own or drafts for your team, and set prohibited topics, the might-solve rule and blocklists.

**Settings → Autopilot** decides, channel by channel, whether the AI answers customers on its own, drafts replies for your team to review, or stays out. The same page holds the guardrails: topics the AI must not handle, when web chat asks whether a reply helped, and who the AI ignores.

> [!NOTE]
>
> **Settings → Autopilot** in your [OpenCX dashboard](https://platform.open.cx/settings/autopilot). Viewing it needs the **Settings** and **Autopilot** permissions. Editing channel settings, the might-solve rule, the blocked visitor message and the blocklists needs edit access to **Autopilot**. Adding or editing prohibited topics needs edit access to **AI knowledge base**.

| Section | What it controls |
| - | - |
| **AI Autopilot** | Per channel: whether the AI replies, drafts, or stays out, and what happens on handoff |
| **Prohibited topics** | Topics the AI must not handle on its own |
| **Might solve user issue** | When web chat asks the visitor whether a reply helped |
| **Blocked visitor message** | What a blocked web chat visitor sees instead of a reply |
| **Blocked contacts**, **Blocked domains**, **Blocked IPs** | Who the AI ignores |

## Turn the AI on or off for a channel

The **AI Autopilot** section lists Phone, Web, Email, Slack, SMS, WhatsApp, Instagram, Messenger, X (Twitter), X Mentions, App Reviews and Internal. When a channel's switch is on, the AI answers customers and runs actions on that channel.

![Turning on the SMS switch under AI Autopilot and selecting Save changes](https://mintcdn.com/openchat/hiyaCL2S28Hya6Q7/images/settings/autopilot-channel-on.gif?s=e3155240622bb1c5e260688e538aa7a8)

Switch SMS on, save, and the AI starts answering customers on SMS.

**1. Open Autopilot settings**

Go to **[Settings → Autopilot](https://platform.open.cx/settings/autopilot)**.

**2. Flip the channel switch**

Turn the switch on or off. To change more than on and off, select the settings icon (**Configure autopilot settings**) on the same row.

**3. Save**

Select **Save changes** at the bottom of the section. Nothing applies until you save; **Reset** discards your edits.

When the switch is off and your team works in OpenCX Inbox, a new customer message on that channel is handed straight to your team with no AI reply. Handoff assignment and handoff workflows run as usual.

> [!NOTE]
>
> The Slack switch is unavailable: Slack is currently only for your internal team members, not shared channels. WhatsApp also has a per-number AI switch; both must be on. See [Connect WhatsApp](https://docs.open.cx/whatsapp/connect-whatsapp.md).

## Let the AI draft instead of reply

Turn on **Assist Mode** in a channel's settings to have the AI suggest replies for your team to review instead of sending them. The dialog shows **Where you'll receive AI suggestions**: your active ticketing system. In OpenCX Inbox, suggestions appear in the session for an agent to review; with a connected helpdesk they go to that helpdesk. See [Inbox](https://docs.open.cx/inbox/intro.md) and your helpdesk's integration guide.

Assist Mode is not available for Phone. While it is on, **AI Autopilot Coverage** is disabled, and two more option groups appear:

| Setting | What it does | Default |
| - | - | - |
| **Respond Only with Potential Answers** | Only suggests replies for messages that need an answer, skipping greetings and other non-support messages | Off |
| **Should Take Actions** | Lets the AI run your [actions](https://docs.open.cx/actions/introduction.md) while drafting. Not recommended for time-sensitive actions | Off |
| **Post Replies as Email Drafts** | Email channel with Salesforce only: suggests replies as editable drafts on the case instead of case comments | Off |

**Allow Autopilot on First Message Only**, **AI Auto Learning Language**, **Do Not Assist On Topics** and **Allow Autopilot On Topics** fine-tune assist mode for sessions that come from a connected helpdesk such as Zendesk or HubSpot.

## Channel settings reference

Each channel's settings dialog (**Autopilot settings – *channel***) contains:

| Setting | What it does | Default |
| - | - | - |
| **AI Autopilot Coverage** | Share of sessions the AI tries to handle (0%, 25%, 50%, 75% or 100%); the rest are left to your team. It currently applies to sessions from HubSpot and Genesys Cloud | 100% |
| **Assist Mode** | The AI suggests replies for review instead of sending them | Off |
| **AI Agent Version** | Which AI agent answers the channel: **V1 (Classic)** or **V2 (Improved)** | — |
| **Auto-Assign on Handoff** | On handoff, assigns the session to an available agent. Off, the session stays unassigned in the inbox | On |
| **Silent Handoff** | The AI sends no reply on the turn it hands off. Assignment and workflows still run. Not applied in assist mode | Off |
| **Prohibited Topic Handoff Behavior** | What happens when a prohibited topic is detected. Applies when **AI Agent Version** is **V2 (Improved)** | Always hand off |

## Set different rules for one email inbox

Under the **Email** row, each email inbox connected to OpenCX has its own row: **Override AI Autopilot rules for this inbox**. Use it when one address needs different handling, for example assist mode on `billing@` while the rest of your email runs on Autopilot.

* Until you change an inbox row, it follows the **Email** row.
* Your first change copies the Email channel's current settings to that inbox. From then on the inbox keeps its own settings, and later changes to the **Email** row do not reach it.
* The rules match the address the customer wrote to, ignoring letter case.
* An inbox override is honored even when the rest of your sessions go to a connected helpdesk.

## Keep the AI away from sensitive topics

**Prohibited topics** are topics the AI must not handle on its own, such as refunds, legal advice or account deletion. For choosing good topics, see [Handoff topics](https://docs.open.cx/handoff/topics.md).

![Adding the prohibited topic Refunds with a description and seeing it join the list](https://mintcdn.com/openchat/hiyaCL2S28Hya6Q7/images/settings/prohibited-topic-add.gif?s=f5293acf6c3bfbf9fd3bcf7c5a5d3484)

Add Refunds as a prohibited topic, with a description of when it applies.

**1. Add a topic**

Under **Prohibited topics**, select **Add topic**.

**2. Name it**

Enter the **Topic**, for example *Refunds*. Up to 255 characters, and each name can be used once.

**3. Describe when it applies (optional)**

Add a **Description** of up to 500 characters, for example *Only when the customer asks us to take legal action, not when they merely mention it.* Conditions you state here are followed strictly.

**4. Save**

Select **Add topic**. Select a topic's name later to edit it, or its **×** to remove it.

When a message matches a topic and the channel runs **V2 (Improved)**, the AI does not answer that message. The session timeline records the detection, and the **Prohibited Topic Detected** workflow trigger fires. What happens next follows the channel's **Prohibited Topic Handoff Behavior**:

| Option | Result |
| - | - |
| **Always hand off** | The session is handed off to your team |
| **Within office hours only** | Handed off during office hours: the hours of the session's team if it has its own, otherwise your default office hours. Outside them, behaves like **Never hand off** |
| **Never hand off** | The AI stays on the session. It skips the matching message and keeps answering messages that do not match |

With **V1 (Classic)**, the AI hands the session off when a topic matches.

## Control when web chat asks whether a reply helped

When the AI judges that a reply might solve the visitor's issue, the chat widget asks the visitor whether it helped, and the widget API returns that verdict with the reply ([custom clients](https://docs.open.cx/widget/custom-components-headless.md#ask-whether-a-reply-helped) can show their own prompt).

Leave **Might solve user issue** empty to use the built-in rule. It does not count replies where the AI says it cannot help, sends the customer to another team, asks for something only a teammate can act on, or answers an automated message, and it stops once the customer confirms the issue is solved.

To use your own rule, say when to return `true` or `false`, for example: *If your answer might solve the issue, return true, otherwise false.* Your text replaces the built-in rule entirely, so include any exceptions you still want.

## Block contacts, domains and IP addresses

| List | What it blocks | Where it applies |
| - | - | - |
| **Blocked contacts** | One existing contact, found by email address | Every text-based channel |
| **Blocked domains** | Everyone whose email address is on the domain. Matching is exact, so `mail.example.com` is separate from `example.com` | Email, and any other channel where the contact's email address is known |
| **Blocked IPs** | One address or a range, IPv4 or IPv6 | Web chat. Once a session is seen from a blocked address, the block stays on that session on any channel |

* **Block contact** asks for the contact's **Email** and an optional **Reason**. You can also block from the **Block Contact** card in an Inbox session's sidebar, or from the contact's page.
* **Block domain** takes one domain or several separated by commas. The results show how many were blocked, skipped (already blocked or invalid) or failed.
* **Block IP** takes addresses or ranges separated by commas, and an optional **Expires after (hours)** from 1 to 8,760. Leave it empty to block permanently. One address can be shared by a whole office or mobile network, so an expiry limits the damage of an over-broad block. You can also block from the **IP Address** card in an Inbox session's sidebar.

To unblock, use the delete icon on the entry's row.

### What a blocked person sees

On web chat, a blocked visitor gets the **Blocked visitor message** in place of a reply. Leave it empty to use the default, *At the moment you can't contact this company.*, which does not say why. On email, messages from a blocked sender are ignored. On other channels, the AI does not reply.

To manage the lists from your own systems, use the [Blocklist API](https://docs.open.cx/api-reference/blocklist/index.md).

## Troubleshooting

| Symptom | Likely cause | Fix |
| - | - | - |
| **No contact found with that email** when blocking | The address does not belong to a contact yet | Block from the session sidebar once they write in, or block their domain |
| A sender on a blocked domain still gets replies | They write from a subdomain, which is matched separately | Add the subdomain as its own entry |
| A topic does not save | Another topic already has that name, in any letter case | Edit the existing topic instead |
| An email inbox ignores a change to the Email row | That inbox has its own override | Change the inbox's own row |

***

## Related Documentation

> [card] [Handoff topics](https://docs.open.cx/handoff/topics.md)
>
> Choosing topics the AI should never handle.

> [card] [Handoff settings](https://docs.open.cx/handoff/settings.md)
>
> Transfers, office hours and Team in the loop.

> [card] [Blocklist API](https://docs.open.cx/api-reference/blocklist/index.md)
>
> Manage blocked contacts, domains and IPs by API.

> [card] [AI safety](https://docs.open.cx/safety/introduction.md)
>
> Every guardrail in one place.


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