A customer asks for a person, and you are the one who answers.
OpusDone hands over the moment a conversation needs you: it stops, tells the customer a person is taking over, and puts the chat in front of you. If a customer asks for a person and the model does not escalate, the runtime escalates anyway. Two taps and you are typing.
Reviewed

Most handovers break in one of three places: the model does not notice, the alert is held back, or you get so many alerts that you stop reading them.
A polite customer asks twice for a person, the model keeps answering helpfully, and nobody tells you. Behaviour this important should not live in a prompt.
Quiet hours exist for a reason, and then they swallow the message from the customer who was about to walk away.
A visitor becomes a lead, books an appointment, then asks for a human. Three events, one person, and a phone that has now trained you to ignore it.
Five reasons escalate a conversation: low_confidence, sensitive, explicit_request, tool_failure and other. sensitive is locked on and cannot be switched off.
An EN/ES keyword guardrail catches “speak to a human” and “quiero hablar con una persona”, and if the model does not escalate, the runtime escalates anyway.
The in-app bell always rings, email is always sent and is never held by quiet hours, and Telegram arrives too if you connected it.
Customer, channel, reason and the last five messages, with [Take over], [Show more] for fifteen more and [Open in dashboard].
The assistant pauses, the conversation is assigned to you, the card is rewritten in place, and colleagues’ cards update to say who took it.
Thirty minutes of silence hands the conversation back to your assistant and tells you it has done so.
That reason is locked on, and both the email and the card withhold the content entirely. All you are told is that a sensitive topic needs you, open the dashboard.
Who it is, which channel they came in on, why it escalated, a three-bullet AI summary, the recent transcript, and one button that opens the conversation. It is sent whatever the hour.

Customer name, channel, reason and the last five messages, with [Take over], [Show more] and [Open in dashboard]. Show more reveals 15 more. The full transcript is never dumped into a chat automatically.
Telegram alerts are free and optional.

Take over, assign to a teammate, snooze, tag, reply with attachments, insert a saved reply by typing “/”, or leave an internal note the customer never sees. “Help me write” and translation produce drafts that never send themselves.
A website visitor who closed the tab gets your reply by email.

Taking over is a small feature with a lot of ways to embarrass you. These are the ones we closed.
You reply from the card, and the message is delivered on the customer’s own channel. Every receipt names the destination, in the form “Sent to Maria on the website chat.”
Handling two conversations at once means quoting the right card. Ambiguity is refused by name rather than resolved with a guess.
The runtime re-checks the pause flag immediately before committing a reply, and drops the assistant’s answer if you stepped in mid-turn.
A takeover started in the dashboard ends the Telegram binding, and one started from Telegram ends the dashboard one.
A delivery ledger records one row per recipient per channel, including alerts that were deliberately suppressed and the reason they were.
Any tool that fails escalates to you and the chat carries on. The customer sees a sentence, not a dead end.
Replies you type in Telegram are delivered on the customer’s own channel, and every receipt names the destination; the customer is never in your Telegram. For a sensitive escalation, both the email and the Telegram card withhold the content entirely. Free to build and test, no card; 14 days from going live. Starter €39/mo, Pro €99/mo, 15% off yearly.
Solicitar acceso