AI writes. You decide if it sends.
Topics it already handles well can go out on their own. Anything that touches a discount, money, or an angry customer waits for you. You change your mind with one click.
A Conext reply does not go out just because the model feels sure. The agent has to be on, the day open, and you inside hours. If one is missing, it stays a draft.
The model feeling sure is not enough.
For a reply to go out on its own, several things have to be true at once: the message was classified as automatic, the model marked it safe to send, the agent has auto-send on, a work day is open, and you are inside business hours. If one is missing, it stays a draft.
That draft lives on the thread. You see it, you edit it, you send it. If the send to Meta fails, it stays there too: the text is not lost and the CRM does not pretend it went out.
What you review is what the customer would see.
WhatsApp is not Markdown. Bold is *this*, with one asterisk. Models tend to write **this**; the CRM translates before showing the draft, not only when sending. Translate too late and you would be approving extra asterisks.
Draft mode is the way to start for the first weeks. When an agent already handles a topic well, you turn auto-send on for that one only and leave the rest waiting.
In practice
- 1
The message comes in
It is classified and a reply is drafted in the same model call.
- 2
The CRM decides if it goes out
Agent, category, hours, and an open day. If something does not fit, it does not send.
- 3
You see it, or the customer does
A draft on the thread, or a WhatsApp already sent. Never a limbo.
Control is not a hidden prompt. It is a switch you see on every agent.
Want to see it with your own threads?
Essential includes one channel, 500 AI conversations per month, and one active agent.