> ## Documentation Index
> Fetch the complete documentation index at: https://docs.emanate.ai/llms.txt
> Use this file to discover all available pages before exploring further.
# Channels
> Configure the AI email agent: how it drafts, when it replies, when it escalates, and how it learns your style
**Channels** is where you set how the AI email agent behaves on each of your connected inboxes. The agent is what
reads incoming mail and writes the drafts you see in [Demand](/demand/managing-email) and your
[Home](/home/agent-inbox) feed. This page is the dial that controls how much it does on its own.
## How involved should AI be?
Set an auto-response mode per inbox:
| Mode | Behavior |
| -------------- | ------------------------------------------------ |
| **Disabled** | No AI: you write every reply yourself |
| **Auto-Draft** | The agent prepares a draft; you review and send |
| **Auto-Reply** | The agent sends replies automatically, no review |
Start on **Auto-Draft** to build confidence in how the agent responds. Move an inbox to **Auto-Reply** only once
you trust its drafts for that kind of mail.
## One agent per inbox
Each inbox can be assigned its own AI agent, so your `support@` inbox can run a support-trained agent while `sales@`
runs a sales-focused one, each with its own tone, knowledge, and instructions. Assign the agent from the inbox's
**AI Settings**.
### Choosing the model
Each inbox can pick the AI model that powers its drafting. The current options are strong general-purpose models;
newer, high-reasoning models (for example, ones tuned for long, nuanced procurement and RFQ threads) can be selected
for the inboxes that need them. Model choice is **per-inbox**, so a procurement inbox and a support inbox can each
use what fits best without affecting one another.
### Long-range memory
The agent has a memory that reaches back across your relationship with a contact, so a reply can reference prior
commitments, past objections, and earlier orders instead of just the current thread. The history window is
configurable (from one year up to many), scoped per agent and per inbox.
## Escalation and handoff
Not every email should be answered by AI. Set escalation rules (keywords or patterns like "urgent" or "legal")
that hand a thread off to a human instead of drafting a reply. When the agent hands off, it can attach a short brief
summarizing who the customer is, what they want, and what's already been said, so the person picking it up has
instant context.
## Fine-tuning per thread, lead, or account
The inbox setting is the **default**. You can override it at three finer levels when one inbox handles a mix of
audiences (inbound sales, enterprise renewals, procurement RFQs) that each deserve different behavior.
Emanate resolves which agent runs in this order, first match wins:
| Priority | Level | Where to set it |
| -------- | ----------- | ----------------------------------------------------- |
| 1 | **Thread** | Inside the thread view, in the AI Auto-Response panel |
| 2 | **Lead** | Lead drawer → Interactions tab |
| 3 | **Account** | Account drawer → Interactions tab |
| 4 | **Inbox** | The default on the inbox itself |
So a thread override on a VIP beats their account-level setting, which beats the inbox default. Each override can
change the agent, the mode (draft vs auto-reply), whether it uses your knowledge base, and whether it reads prior
threads for context. Overrides are **sticky**. To stop using one, set its agent to **None**.
## How the agent writes a draft
When the agent drafts a reply, it pulls together everything a good rep would:
* **The thread:** the subject and recent back-and-forth.
* **Your knowledge base:** the product docs, specs, and pricing most relevant to the question.
* **Past emails:** how you and your team have handled similar conversations.
* **Attachments:** it reads incoming RFQs, spreadsheets, spec sheets, and even scanned drawings, so its reply reflects what the customer actually sent.
* **Your brand:** the tone and signature from your [Business DNA](/business-dna/overview).
## It learns your style
When you edit an AI draft before sending, Emanate quietly captures the difference (shorter, warmer, more direct,
different sign-off) and folds it into how it drafts for you next time. Over weeks, drafts drift toward sounding like
you wrote them. You can review what it's learned and turn the learning on or off per agent.
## Next steps
Where the agent's drafts show up
The context the agent draws on
Route mail to the right agent automatically
Give the agent tools like inventory lookups