Agent Assist in the inbox

Agent Assist is Aura AI working next to you rather than instead of you. It sits in the details rail beside a ticket, under a tab named after the agent — Aura AI, alongside Details. The same panel appears on the standalone ticket page.

The one fact to hold on to: the panel never sends anything to a customer. Everything it produces goes into your composer or your clipboard, and sending stays a thing a person does.

Drafting

The panel does nothing when you open a ticket. A draft is a click, because a draft costs a call to the agent and most tickets do not need one.

Draft a reply reads the ticket and the knowledge base and writes a reply for you to edit. It takes about twenty seconds and reports its progress while it works. Afterwards the button reads Draft again.

Above the button is a Draft language pill. It offers ten languages plus Match customer, which is the default and means the agent picks up the language of the conversation.

Ask about this ticket is the other half of the panel: a question you type, answered for you rather than for the customer. The answer is labelled Answer for you and carries a note saying it was not written for the customer, and when you insert it the panel defaults to putting it in an internal note rather than a reply. That default is the safety rail — you can override it, but you have to mean to.

Reading the result

Under every draft is a Sources block listing the knowledge-base articles it used.

Read it. When it is empty, the panel says so in as many words: nothing in the draft is grounded in an article, Aura wrote it from the conversation alone, and you should check every fact in it before you send. A confident, fluent, ungrounded draft is exactly the one that needs checking, and this is how you spot it.

Rewrite draft reveals controls for regenerating in a different tone or against particular sources. Regenerating from a previous draft does not search again, and the panel says as much: the citations you are looking at are still the previous draft's.

Three more things the panel tells you and means:

• It read only the most recent messages when a conversation is long, and it says how many.
• The draft is saved nowhere. Leaving the ticket discards it.
• A draft eventually expires, at which point the button becomes Start a new draft.

Getting it into the reply

Three destinations: Insert (into the reply composer), Insert as note (into an internal note), and Copy. Once there is text in the composer already, the first two read Add to reply and Add to note instead, so you know it will append rather than replace.

A destination you are not allowed to use, or that the ticket's channel cannot accept, is absent rather than disabled. On a source with no working reply channel, an insert aimed at the reply lands in an internal note instead.

The thumbs under a draft are worth using. A thumbs-down opens a short comment box asking what was wrong with it, and that is the most direct route from a bad answer to a better index.

When the panel cannot help

If Aura AI is not enabled the panel says so and gives an admin a link to enable it. If the stored connection has stopped working it says Aura is not answering right now and points at the Overview. If your role is below the floor for drafting, the button is not there at all and the panel names the role you have and the one it needs. None of these are silent.

Finding the AI's work in the inbox

The inbox sidebar has an Aura AI section with four views:

• AI answered — the agent replied.
• AI resolved — the agent finished the job.
• AI unresolved — it did not.
• AI web chats — conversations the agent handled on the web widget. These open solved, so they do not appear in All open until a human is handed the conversation.

These are mutually exclusive with the ordinary status views: you are looking at one or the other, never a mixture.

How AI messages are marked on a ticket

An AI-written message on the timeline is visibly not a colleague's. It carries the Aura mark and a badge reading Auto-resolved by Aura, or Action completed by Aura when it is an approved action reporting back rather than a resolution.

Under it, Auto-resolve details expands to three rows: the model, the confidence score, and the citations. Anything that was not recorded says so rather than showing a blank. That disclosure is what makes an automatic reply auditable months later, and it is the reason auto-resolve is safe to turn on at all.