Human Handoff
How conversations escalate to humans, how a person takes over mid-thread, and how approval mode keeps a human in the loop.
The agent is not supposed to handle everything. It is supposed to handle what it can answer well, and get out of the way — cleanly, with context — for everything else.
Three mechanisms cover that: escalation (the agent asks for a human), takeover (a human steps in), and approval mode (a human signs off before anything sends).
Escalation
The agent hands a conversation to a human when it should not be the one answering.
When it escalates
| Trigger | Example |
|---|---|
| No confident source | The question is about your business and nothing in your knowledge supports an answer. The agent sends a templated "let me get someone" rather than inventing one — see knowledge |
| A configured trigger | Rules you set: pricing disputes, refund requests, cancellations, legal questions, angry tone |
| The customer asks | "Can I talk to a person" is honoured |
| A guardrail fires | The runaway message cap, or an action the agent is not permitted to take alone |
What happens on escalation
The conversation flips to human control. The agent goes quiet — it does not keep replying underneath a human. A task is created for your team, the conversation is flagged as needing attention, and the notification bell fires.
The human gets the full context, not a ticket stub. The thread timeline carries everything: the messages, transcripts of voice notes, descriptions of images the customer sent, the tool calls the agent made, and — when the agent chose not to reply — an inline explanation of why. Alongside it: the contact card, the deal and its qualification so far, and the contact's history on your other channels as read-only context.
Silence is explained, not mysterious
When the agent stays silent — paused, out of hours, low confidence, contact excluded — the thread says so inline. You should never have to guess why your agent did not answer.
Live takeover
Any human on your team can take over a conversation at any point, mid-thread, without the agent fighting them for the reply.
Take over
Control of the thread flips from AI to human. The agent stops generating replies on that thread immediately.
Reply as yourself
You type. The customer sees a normal message on the channel they are already using. The agent does not comment, correct, or double-send.
Hand back
Return the thread to AI when you are done. The agent picks up with the full history, including everything you said, so it does not repeat or contradict you.
Two details that matter in practice:
- Takeover is scoped to one thread. Taking over a customer's WhatsApp conversation does not pause the agent on that same customer's Instagram DMs. Actions apply where you took them.
- A manual reply is a soft pause. If a human replies without formally taking over, the agent backs off for a period rather than talking over them. This includes replies you send from your own phone on a connected channel — CoPhrase notices the message it did not send, records it as a human reply, and steps back.
On Instagram and Messenger, human replies get a longer send window than bot replies do under Meta's messaging rules, and takeover uses it. See channels.
Approval mode
Approval mode is the middle setting between "AI replies to my customers unsupervised" and "I answer everything myself".
With approval mode on, the agent does all the work — retrieves knowledge, decides on tool calls, writes the reply — and then stops. The draft is held for a human.
You can:
- Approve it and send as-is
- Edit it and send your version
- Reject it and write your own reply
Drafts collect in a queue so you can work through them in a batch rather than hunting for them thread by thread.
Drafts expire on some channels
WhatsApp, Instagram, and Messenger only allow free-form replies inside a limited window after the customer's last message. A draft waiting for approval as that window closes will show its expiry — and once the window shuts, it is regenerated or discarded rather than sent into a channel that will reject it. Approval mode works best when someone is actually watching the queue.
Training wheels
New agents can start in approval mode automatically for their first stretch of real conversations, then prompt you to graduate to autonomous once the drafts have been approved consistently. It is the honest answer to "I am not ready to let an AI talk to my customers": do not. Watch it work first, then decide.
Turning the agent off
Beyond per-conversation control:
- Per contact — exclude a specific contact from AI entirely, or mark them do-not-disturb.
- Per agent — pause AI on the whole agent (this is also what a failed BYOK key does, see persona).
- Per org — a daily budget cap that pauses AI rather than spending past it, see billing.
Learning from your team
Roadmap: when a human answers a question the agent could not, that answer is proposed as a knowledge candidate. You approve it, and it becomes a first-class FAQ the agent can use next time. Escalations stop being pure cost and start being training data — with a human deciding what gets learned.