With behavior configured, decide how the agent should operate and where it should run.
Choose a mode
Modes are set under Settings > AI > Setup.
Supervised Mode. The agent drafts replies. A teammate reviews each one and clicks send. Recommended for every launch and for any time you change instructions or knowledge sources.
Autonomous Mode. The agent sends replies automatically. Use it once you have run Supervised Mode long enough to trust the suggestions, typically a few weeks of active review. Do not switch to Autonomous Mode until reviewers are sending most suggestions without edits, unassignment rules have been tested on real conversations, and a monitoring owner is in place.
Assign the agent to inboxes
Open Settings > Inboxes and add the AI Agent as a member of any inbox where it should reply.
In the inbox settings, enable Assign all new conversations to AI Agents to route incoming chats to the agent automatically. You can leave this off and assign the agent manually for a more cautious rollout.
How the agent appears in conversations
In an open chat, the AI Agent appears in the assignee dropdown alongside human teammates. You can hand off to and from the agent at any time.
A few details worth knowing:
- Color-coded bubbles. AI Agent messages render in purple. Human teammate messages render in blue. The visual difference makes it easy to scan a transcript and see who said what.
- Auto-response Automatic AI Agent replies only happen when the AI Agent is assigned to the conversation and the latest message is inbound from the contact.
- Typing indicator. While a suggestion is being generated, the chat shows "Agent is typing..." just like with a human teammate.
- Manual generation. Even when the agent is not assigned, you can click Generate Suggested Reply to get an on-demand draft.
- Humans can reply at any time. Sending as a human does not interrupt the agent's state. Outbound human messages do not trigger a follow-up suggestion from the agent. To fully take over, reassign the chat away from the AI Agent. Otherwise, the agent can respond again when the contact sends the next message.
- Only incoming messages invoke the agent. New outbound messages from your team will not generate a suggestion.
Rollout strategy
Start narrow:
- One inbox, your lowest-stakes one
- Supervised Mode
- A small group of reviewers
- A daily review of suggestions for the first week
Expand only after you have data showing the agent is producing useful replies.
Next steps
Continue to Article 6: Test Your AI Agent Before Going Live.