Confidence & escalations
The AI reply queue
Every held draft in one place, ready for a verdict. This is the ten minutes a day.
What each card tells you
Customer, channel, the suggested reply, and its confidence tier. Context chips show what the AI used: the matched order, the knowledge base, Shopify data, or general knowledge. Risk flags surface anything sensitive. A "wait" chip explains why it held.
Three verdicts
Approve and send. Edit before sending, and your edit becomes training data. Or discard. Select many and approve or discard in bulk when the queue is obviously fine.
Nothing is ever marked sent until it actually is.
It refreshes itself
The queue updates every 30 seconds, pausing politely while you type. Come back after lunch to a queue that is already current.