Live Inbox
The takeover console — the queue, the conversation and the reasoning rail. How to review a draft properly, when to take a conversation over, and how to hand it back.
/admin/ai/support/inbox. This is the only screen in the addon with a customer
waiting on the other end of it.
It is a full-screen console: the site header and footer are suppressed so the conversation gets the viewport rather than four rows of chrome. The back arrow in the top bar is the way out.
Three panes
| Pane | What it is for |
|---|---|
| Queue | Every conversation the assistant is handling or has escalated |
| Conversation | The thread, exactly as the customer sees it |
| Reasoning rail | What the assistant retrieved, and how confident it was |
The rail is the point of this screen. An agent deciding whether to send a drafted answer needs to see what it found and how sure it was, not only the sentence it produced. Without that, the review is theatre — you are reading fluent prose and guessing.
Queue order
The queue opens on Priority: conversations waiting for a person come first, then the ones a colleague is already holding, then the ones the assistant has been switched off for. Within each, higher importance and the longest wait come first. Those are the ones where a customer has already been told somebody is coming.
Change it from the sort control under the search box — last activity, ticket created, importance, AI answers, cost, subject or customer — and use the arrow beside it to reverse the direction. The choice is saved in your browser.
The escalation reason is written in plain language on the row and again on the card it opens — one rule, so the queue and the case cannot describe the same conversation differently.
Filters, search and paging
The tabs above the search box choose what the queue holds: Waiting, Mine, AI handling, Open (the default — everything the desk still owns) and Closed. The number on each tab is the number of conversations in it, not the number on the current page.
Search matches the subject, the customer's name or email address, the escalation reason, and a ticket id pasted whole.
The foot of the queue says which rows you are looking at and how many there are in total. Rows per page — 25, 50 or 100 — is in the sort menu. All of it runs in the database, so what you see is the top of the whole queue rather than the top of an arbitrary slice of it.
Reading a draft properly
-
Read the rail before the answer. If the top source is only loosely related, the answer is fluent and wrong, and reading it first will make it sound right.
-
Check the numbers against the citation, not against memory. The one thing an assistant does that a person does not is state a figure with total confidence.
-
Send, edit, or take over. Editing is fine and it is measured — see Autonomy, where the edit distance on your own tickets is what decides whether it is ready to answer unsupervised.
Asking for a draft
Generate a draft asks the assistant for a suggested reply on a conversation that has none — one it escalated, one you have taken over, one where the customer has just added something.
The button is draft-only regardless of your autonomy rung. On AUTO_TICKET or
AUTO_ALL it used to take the send decision from the install's configured
autonomy, so pressing it appended the answer to the ticket and emailed it to
the customer — and then handed the same text back to the composer with a Send
button under it. An agent who edited and sent delivered a second, contradictory
reply; an agent who discarded it had already sent the first.
Now nothing is appended, no email goes out, and the session is not moved on. The conversation is handed to the assistant for the length of the generation and returned to you whatever happens, including on a failure.
A draft that cannot be produced says why rather than failing generically:
| What you are told | What it means |
|---|---|
| The knowledge base has nothing relevant… | Retrieval found nothing above the floor. Write the article — it is already on Gaps |
| No AI provider is configured. | Provider screen, then restart the backend |
| The AI spend cap has been reached. | A daily or monthly ceiling is hit |
| A spend cap is set but this provider reports no cost… | Price the endpoint or set the caps to 0 — see Error codes |
Some drafts cost nothing — on an install where direct answers are switched
on. They ship off and no screen in this build renders the switch, so most
installs never see this; the key and the way to set it are in
Settings reference. Where they are on and one of your
own published articles answers the question well enough, the assistant serves
that article with no model call at all — and in COPILOT that arrives here as a
draft like any other rather than going straight to the customer. It still counts
as a turn against the escalation cap, because the customer had an exchange.
Taking over and handing back
Take over moves the conversation to you. The assistant stops drafting and the customer is talking to a person from that message on.
Release hands it back. Use it when the thing that needed a human is done — a refund approved, an account unlocked — and the rest of the conversation is ordinary.
A half-corrected AI answer reads worse than either a clean AI answer or a clean human one, and the edit distance it records makes the promotion gate read your install as less ready than it is.
What a customer sees
They see your agent's name once you take over. They do not see the rail, the scores, the retrieved passages, or that a draft existed.
Citations do appear in their thread — the title and the passage — so they can check an answer themselves. The links are stripped: a citation never sends your customer to somebody else's documentation site.