Open Inbox and pick a conversation. Everything the visitor and the AI said is there, with the AI's citations intact.
Reply
Type in the composer and send. The moment you send a reply, the AI stops answering that conversation. The visitor is told a person has joined, and nothing is sent over the top of you.
You don't have to do anything special to "take over" — sending a reply is the takeover. The AI sees your message land and stops queuing further responses for that thread. If the AI was mid-draft when you sent, that draft is discarded.
If you want to quote something the visitor or the AI said, click the small Reply icon on any message bubble. The quoted text appears in your composer with a grey border, so the visitor knows exactly what you're addressing. This is handy when a conversation has gone back and forth a few times and you need to anchor your answer to a specific point.
Taking over mid-thread
You can intervene at any point — there's no need to wait for the AI to reach a stopping point. Open the conversation, write your reply, and send. The handoff happens the moment your message is delivered.
If you'd rather stop the AI from responding without immediately sending a reply yourself, use the Take over button in the conversation header. This mutes the AI for the thread without putting words in your mouth. The visitor won't see anything change; they're just waiting for a response from whoever picks it up. Use this when you need a minute to look something up or check a policy before writing.
What "taking over" means in practice: the conversation moves from AI-auto mode to Human mode. In AI-auto, the AI replies on its own to every incoming visitor message. In Human mode, nothing goes out until you type it. The AI can see the thread, but it won't act on it.
Note
Switch the composer to Note to leave an internal note instead. Notes are never sent to the visitor.
Notes are useful for flagging things for teammates, logging a call summary, or leaving a reminder before you hand back to the AI. If you're passing a conversation to a colleague, a note telling them where things stand is far more useful than making them read the whole thread from scratch.
Handing it back
When you are done, choose Let the AI reply on the conversation to release it.
Once you do, the AI resumes normal behaviour and will reply to the visitor's next message using your trained content. If the visitor hasn't sent anything new yet, the AI waits — it doesn't send a follow-up on its own just because control was returned.
Timing matters here. If you sent a closing message like "Let me know if there's anything else I can help with," hand back to the AI right after so it's ready if the visitor follows up. If you leave the conversation in Human mode, no reply will go out until someone on your team writes one.
Assist
The Assist tab drafts a reply from your own help content and tells you honestly whether it found grounding for it. A grounded draft comes with citations; an ungrounded one says so.
Assist is not the same as letting the AI reply autonomously. It drafts text for you to review and edit before you send — you're still the one hitting send, so the conversation stays in Human mode. Think of it as a search shortcut that pulls the most relevant passage from your docs and turns it into a reply draft.
When to use each mode
Most support conversations don't need human intervention at all. The AI handles repetitive questions well as long as the answer is somewhere in your training content.
Step in yourself when a visitor is upset, when the request is too complicated or multi-part for the AI to handle cleanly, or when the issue needs a judgment call — a refund decision, an exception to policy, anything where the "right" answer depends on context the AI doesn't have.
Use Assist when you're in Human mode and want a head start from your docs. It's especially useful for team members who are still learning the product and might not know where to look.
Leave a Note when you need to hand off context to a teammate or record something for your own reference. Notes don't interrupt the visitor and don't affect which mode the conversation is in.
Troubleshooting
The "Take over" button is greyed out. Someone else on your team has probably already claimed the conversation. Check the Assignee field in the right panel — if it shows a teammate's name, the thread is theirs. Re-assign it to yourself from that same panel, and the Take over option will activate.
You sent a reply but the AI replied again straight after. If the AI had a response queued before your message landed, it may have slipped through. This shouldn't happen twice. To be safe, hit Take over after your reply to make sure the conversation is locked in Human mode. Also check that no automation rule is reassigning the conversation back to AI after each human reply — look under Rules in your workspace settings.
The visitor didn't receive your reply. Check that the composer is set to Reply and not Note before you send. If you're sure it was set to Reply, open the Events log in the right panel. Each message shows its delivery status and the channel it was sent on, which makes it straightforward to spot where something went wrong.