> ## Documentation Index
> Fetch the complete documentation index at: https://docs.callivox.com/llms.txt
> Use this file to discover all available pages before exploring further.
# AI in the inbox
> What your team sees when AI is involved: AI-written replies, drafted suggestions, the AI panels, handoffs, automatic tags, and turning AI off.
AI does two different jobs in the inbox. It answers customers on its own, and it helps your team answer faster. This page is about what that looks like from an agent's seat — the configuration lives under [AI agents](/ai-agents/overview).
## The AI switch
Every conversation header carries an **AI** pill. It reads the agent's name when an agent is answering here, or **On** or **Off**.
Tap it to turn AI replies on or off for this one conversation. It does not change anything anywhere else. When you turn AI off, a small panel lets you decide whether it should come back automatically.
### A temporary human takeover
**Automatically turn AI back on** is enabled by default, with **30m** selected. Choose **15m**, **30m** or **60m**, or choose **Custom** and enter any duration from 1 minute to 7 days. Callivox remembers your latest choice for the next conversation where you turn AI off.
Turn the option off when the conversation should stay with a person until someone deliberately turns AI on again.
Auto-enable applies only when a person turns AI off with this control. If the AI hands the conversation to a person, a routing rule disables it, a flow transfers it, or the customer asks for a human, no timer is started. Those handoffs stay with your team until someone chooses to bring AI back.
Turning AI on yourself before the timer expires cancels the pending timer. A later handoff or automation also cancels it, so an old manual timer can never switch AI back on over a newer decision.
The pill shows the state that actually applies, taking your channel setting, your agent bindings and anything you have set on this conversation into account. If AI has been paused — because an agent handed the conversation over, or because someone switched it off — the pill reads as off, and one tap brings it back.
Every change writes a marker into the thread, so you can always see when AI was turned on, off, paused or resumed, and follow what happened next.
## AI replies in the thread
An AI reply is an ordinary message in the thread, sent to the customer on their channel like any other. The only difference is the name above it: instead of a teammate's name, it carries the **AI agent's name**.
That means the history reads honestly. You can see at a glance which parts of the conversation the AI handled and where a person took over.
### When the AI does not reply
Sometimes a customer's message arrives and the AI says nothing. Rather than leave you guessing, Callivox puts a short line under that message:
> AI didn't reply — \
The reason is spelled out — AI is off for the conversation, the sender is blocked, the agent decided a person should take it, and so on. The **AI details** panel keeps a fuller record of the same thing under **Diagnostics**.
## Agent Assist — drafted suggestions
An assist agent never messages a customer. It works for your team, inside the **Agent Assist** panel on the right of the conversation.
| Section | What it gives you |
| --------------- | -------------------------------------------------------------------------------------------------------------------------- |
| **Summary** | What the conversation is about, so you can pick up someone else's thread without reading it all. |
| **Mood** | Whether the customer sounds happy, neutral or upset, with a short explanation. |
| **Suggestions** | Drafted replies, each labelled by style. **Insert** puts one into your composer; **Regenerate** asks for a different take. |
| **Knowledge** | An answer drawn from your knowledge base, with chips naming the documents it came from. |
Each section has its own refresh control, so you ask for what you need rather than waiting for everything. **Summary** and **Mood** also refresh themselves shortly after a customer's message, so they are current when you open the conversation.
**Insert** appends the draft to your composer. Nothing is sent — read it, edit it, then send it yourself.
The panel header has an **On/Off** switch for this conversation. When it is off you will see "Agent Assist is off for this conversation.", along with **Turn off by default for me** if you would rather not have it running at all. If no assist agent is set up yet the panel says "Agent Assist not configured." See [Assist agents](/ai-agents/assist-agents).
## The AI detail panels
Two more panels on the rail show what the AI has been doing.
**Objective** — the goal the agent was given for this conversation, with a status pill reading *In progress*, *Completed* or *No objective*.
**Collected variables** — the facts the agent has gathered, such as an order number or a preferred date. Each row shows the value, or "not collected" when the agent has not got it yet, with a small red marker on anything required and still missing. Hover a row to **Edit** or **Delete** it, and use **Add variable** to record something the customer told you outside the conversation. Values you type by hand are treated exactly like the ones the AI collected — dashboards and automations see no difference.
**Diagnostics** — a per-message record of why the AI replied or did not. When there is nothing to report it reads "No AI reply issues on the loaded messages."
Every action the agent took — looking up an order, creating a ticket, writing to your CRM. Each row expands to show what went in and what came back, and how long it took.
This is the panel to open when a customer says the agent told them something odd. Empty conversations read "No tool calls in this conversation yet."
## Handoffs
There are two kinds of handoff, and they look different in the thread.
**AI to a person.** When an agent decides a conversation needs a human, it stops replying. The **AI** pill no longer reads as on, and the thread carries an **AI paused** marker at the point where it stepped aside. From there the conversation is yours: everything the agent learned is still in the thread, and in the **Collected variables** list. What triggers a handoff — a phrase, a sentiment, a failed attempt — is set on the agent under [Behavior and handoff](/ai-agents/behavior-and-handoff).
**AI to another AI.** When agents work as a team, one can pass the conversation to a colleague better suited to it. That shows as a highlighted line across the thread:
> ⇄ \ transferred to \ — "reason"
The reason is the transferring agent's own words. Transfers are never collapsed together, so a chain of them stays legible. See [Teams](/ai-agents/teams).
## Automatic tagging
Callivox reads conversations and tags them by topic on its own — "Refund request", "Delivery", "Pricing" — so you can see what people are actually contacting you about without anyone filling in a form.
* Tags appear as coloured chips on the conversation row in the list, up to two, with a **+N** chip when there are more.
* Tagging happens early, while a conversation is still live, and again when it closes — the second pass reads the whole conversation and settles on the final set.
* Automatically created tags carry an **AI** badge on the [Tags page](/dashboards/overview).
* Tags you add by hand are kept separate and are never overwritten by the automatic pass.
* **Re-evaluate tags** in the **More actions** menu re-runs it on demand.
Tags share one vocabulary with your contacts, so the same tag means the same thing in both places. Rename, recolour, merge and delete them on the Tags page.
A tag can also drive routing: a [rule](/settings/folders-and-rules#routing-on-ai-assigned-tags) can match on **Conversation tag (AI-assigned)** and file the conversation into a folder as soon as it is tagged.
## When AI is switched off
Turning off **AI auto-reply** on a channel stops your agents answering customers there. It does not switch off AI altogether:
* **Agent Assist keeps working.** Your team still gets summaries, mood and drafted suggestions.
* **Automatic tagging keeps working.**
* Nothing is sent to a customer without a person pressing send.
That combination is a good way to introduce an agent: leave auto-reply off, read the drafts it produces for a week, and turn it on once you trust them. See [Managing channels](/channels/managing-channels).
Blocking a sender has a narrower effect: AI stops answering that one person across every conversation on that channel, while your team can still reply by hand. The conversation header shows a red **Blocked** pill.
## Next steps
How agents are built, and how they get attached to a channel.
Set up the agent behind the Agent Assist panel.
Decide when the AI should stop and bring in a person.
See what your conversations are about.