Handover and guardrails: what the AI may do on its own
How replies differ from actions, the three autonomy levels, every way a conversation passes to a person, and how to take over or switch the AI off for one conversation.
On this page
The AI agent does two different things: it writes replies, and it can take actions. Replies go out on their own. Actions are limited to the ones you allow, and you decide which of those wait for a person first. Separately, a set of rules decides when the agent stops and hands the conversation to your team.
Replies and actions are different
A reply is the message the customer reads. Once the agent is on, replies go out without anyone approving them. The limits on what it says are:
- the knowledge base — what it knows (more);
- Guardrails — things it must never do, on the Behaviour tab. One short sentence each, such as "Never promise a refund";
- Response guidelines, also on Behaviour. Each is a single standing rule, such as "Always confirm the order number before quoting a delivery date".
An action changes something: it sends a payment link, labels the conversation, assigns it to someone. The agent may take only the actions ticked under Actions → Allowed actions. They are split into two groups:
- Reaches the customer — Send payment link, Capture order draft. Ticking one of these asks you to confirm first, because it affects live conversations straight away.
- Tidies the workspace only — hand off to the team, update the conversation status, assign, set labels, remember customer facts, update contact attributes, record appointment requests.
Customer confirms first is a separate setting on the same tab. It makes the agent ask the customer for a yes before doing a particular action. It has nothing to do with staff approval.
How much it may do on its own
How much it may do on its own, on the Actions tab, decides whether an allowed action runs straight away or waits for a person:
| Level | What happens |
|---|---|
| Shadow | Every action waits for a person, except handing over to the team, which always happens at once. What a team runs for the first week. |
| Review (recommended) | Internal, undoable work happens straight away. Anything that leaves Juno waits. |
| Full autonomy | Every allowed action runs the moment the agent decides on it, including sending a payment link. |
The agent keeps replying at every level. A held action shows up in the conversation's Copilot tab under Waiting for you, with Approve and Reject. A private note also appears in the thread so nobody misses it.
Handing over to your team never waits, at any level. When the agent decides a person is needed, it stops replying and passes the conversation on straight away, even in Shadow. Only business actions, such as labelling or sending a payment link, wait for approval.
When the conversation goes to a person
The agent hands over when any of these happens:
- The customer uses a handover keyword. Set under Channels & handover → Hand over immediately when the customer says. The keyword is matched anywhere in the message, ignoring case, before the AI even reads it. This is the one trigger that needs nothing else switched on.
- The agent decides a person is needed. That covers a customer asking for a person, a message touching one of your Topics the AI should hand over instead of answering, a question the knowledge base cannot answer, or a request it cannot carry out. When the AI is unsure sets how quickly it gives up: after one clarifying question, or by offering first and handing over once the customer agrees. All of these need Hand off to the team ticked on the Actions tab. Without it, the agent can decide a person is needed but has no way to act on it. While it is off, the Channels & handover tab says so above the handover settings, with a Turn on button that ticks it for you.
- It cannot read the message. A voice note, or a photo with no caption.
- It hits a limit. Max turns (under Behaviour → Advanced parameters) is reached, the plan's AI usage for the period runs out (plans and usage), or generating a reply fails.
- A person steps in. Someone presses Take over, replies, or assigns the conversation to a person (by hand, or with a macro). An automation rule that assigns a conversation does not count: the AI keeps answering. To have a rule hand conversations to people, add Stop the AI agent on this conversation to it (automation rules).
Hand over to decides where the conversation goes: a team (a member is picked round-robin), a named person, or Unassigned (whoever picks it up). Under Handover wording and auto-close you can change what the customer is told: one message for when the team is online, one for outside business hours, and one for when the AI cannot continue. Leave them empty to use the defaults.
What your team sees
- In the conversation list, a handed-over conversation is marked Needs you.
- The Copilot tab shows Where this conversation is, a summary written at the moment of handover. Whoever picks it up can start there instead of scrolling the whole thread.
- On AI agent → Performance, Why conversations left the agent counts handovers by reason. Missing knowledge is the reason you can fix directly.
Take over, hand back, switch off
While the agent is handling a conversation, a banner says AI agent is handling this conversation. The reply box is replaced with a note, so nobody can type over the AI by accident.
- Take over (on the banner) stops the AI on that conversation. The reply box comes back and you write the next message yourself.
- To hand it back, open the Conversation tab in the right-hand panel and turn AI Agent auto-reply back on. The setting is shared, so every teammate sees the same state.
- For one channel or certain hours, change Channels or Schedule on Channels & handover.
- For everyone, right now, flip the switch at the top of the AI agent page to Not answering customers.
Anyone who can reply to conversations can take over and hand back, including the built-in Front Desk role (members and roles).
Reviewing what it did
- Messages the agent sent are marked AI agent in the thread.
- What the AI changed, in the Copilot tab, lists the actions it took on the conversation. Most can be reversed with Undo. The ones that can't, such as a payment link already sent, say Can't be undone.
- Performance also shows actions planned against actions executed, which guidelines the agent says it followed, and Where the AI did not take over (outside schedule, or on a channel it is not deployed to).
Approving, rejecting and undoing in a conversation need permission to update conversations: the built-in Administrator and Agent roles have it, Front Desk can see what is waiting but not decide. Changing the AI agent's own settings stays with Administrator.