> ## Documentation Index
> Fetch the complete documentation index at: https://docs.conduit.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Proposed replies

> How the agent drafts replies in copilot, what changes on autopilot, and what to do with a draft you do not like.

When a guest message arrives, the agent drafts an answer. What happens to that draft depends on the mode your workspace is in.

## Copilot and autopilot

| Mode | What happens to the draft |
| - | - |
| **Copilot** | The draft waits in the conversation as a **proposed reply**. Nothing goes out until a person sends it |
| **Autopilot** | The draft is sent to the guest automatically, unless the agent flagged it |

The difference matters when you read a conversation. In copilot, a proposed reply is a suggestion nobody has seen but you. It has not reached the guest.

See [Road to Autopilot](/build-your-agent/road-to-autopilot/overview) for how teams typically move from one to the other.

## Message types

Four kinds of AI message appear in a conversation:

| Type | What it means |
| - | - |
| **Sent message** | The agent was confident and sent the reply automatically. Autopilot only |
| **Proposed reply** | The agent drafted a response and is waiting for you to approve it |
| **Held reply** | The agent drafted a response and did not send it, because someone had already replied to the contact. The card says who: a teammate, a workflow, or a reply sent outside Conduit (in your PMS or on the OTA). This happens on autopilot too, and does not mean autopilot is switched off |
| **Urgent message** | The conversation was escalated based on a conversation tag. See [Escalations and AI tags](/inbox/agent/escalations) |

## Working a proposed reply

Expand the proposed reply to read the full draft, then take one of three paths:

* **Send** it as written, if it is right
* **Edit** it in the composer first, if it is close
* **Teach agent** to correct the underlying gap, if it is wrong for a reason that will recur

**Teach** (`Command + K`) opens from the conversation and lets you create or edit Knowledge entries and add or change Guardrails without leaving the inbox. Updates apply to future replies straight away.

**Draft** (`Command + J`) is how you write with the AI rather than from scratch. Outline what you want to say in brief notes, and the agent turns them into a properly worded reply in the composer.

<Tip>
  Editing fixes this one message. Teaching fixes every future message like it. If you find yourself making the same edit twice, that is the signal to teach instead. See [Teaching the agent from the inbox](/build-your-agent/road-to-autopilot/teach#teaching-from-real-conversations).
</Tip>

## Why the agent drafted what it did

Every AI message carries tabs that open up its working:

| Tab | What it shows |
| - | - |
| **Messages** | The draft itself |
| **Reasoning** | How the agent read the conversation and why it landed on this answer. Press `i` to open it |
| **Attribution** | The knowledge articles, guardrails, and other sources it drew on. Sources can be edited directly from this view |
| **Tools** | The tools it called while composing the reply |

These are how you tell a knowledge gap apart from an instruction gap. If Attribution shows the agent found nothing relevant, the answer is missing from your knowledge base. If it found the right material and still answered badly, the problem is in how it was told to behave.

## Checking your changes worked

After teaching the agent, do not assume the fix landed. Test it.

Use **Teach agent** or the operator sidebar and ask it to draft again, for example:

> Draft a new reply to this contact using the updated knowledge.

1. Review the generated message card
2. Choose **Open in chat input** to move it into the composer, where you can edit it before sending

If the message was already sent, there is no draft left to regenerate. Recreate the scenario in Sandbox instead:

1. Open the agent in Sandbox
2. Send the same message the guest sent, with the same context where possible
3. Check that the new reply reflects what you taught

Sandbox is also the better choice when you want to test the exact scenario rather than a fresh draft in the live conversation.

See [Teaching the agent from the inbox](/build-your-agent/road-to-autopilot/teach#teaching-from-real-conversations).

## Frequently asked questions

<AccordionGroup>
  <Accordion title="Has the guest seen a proposed reply?">
    No. On copilot, a proposed reply exists only in your inbox until somebody sends it.
  </Accordion>

  <Accordion title="Would this reply have gone out on autopilot?">
    A plain proposed reply, yes. A message tagged AI is missing information or flagged as conflicting would not. See [Escalations and AI tags](/inbox/agent/escalations).
  </Accordion>
</AccordionGroup>


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.