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

# AI reply drafts

> How Guhan drafts responses to replies — and when it hands off to you.

When a prospect replies to your sequence, Guhan reads what they said, classifies the intent, and drafts a response. You approve, edit, or take over. Here's how the AI reply flow works.

## The classification step

Every reply gets classified into one of these buckets:

| Classification       | What it means                                                   | What Guhan does                                                         |
| -------------------- | --------------------------------------------------------------- | ----------------------------------------------------------------------- |
| **interested**       | "Yes, tell me more" / "Sounds interesting"                      | Drafts a follow-up + option to auto-send OR hand off                    |
| **meeting\_request** | "Sure, let's talk" / "How about 3pm?"                           | Creates a Schedule task with booking link                               |
| **not\_now**         | "Not the right time" / "Ask me in Q3"                           | Pauses sequence + creates a Notify task; snoozes prospect for N months  |
| **not\_interested**  | "No thanks" / "Not a fit"                                       | Ends sequence + adds prospect to Do Not Contact                         |
| **unsubscribe**      | Explicit unsubscribe request                                    | Adds prospect (+ their email/LinkedIn) to Do Not Contact + acknowledges |
| **question**         | "What does the pricing look like?" / "Do you integrate with X?" | Drafts an answer OR hands off to you (configurable)                     |
| **out\_of\_office**  | Auto-responder                                                  | Ignores, sequence continues (Guhan retries in 3 days)                   |
| **spam\_filter**     | Filter response, not a real reply                               | Ignores                                                                 |
| **wrong\_person**    | "Not the right contact" / referral                              | Notify task with the referral note                                      |

Classification runs in \~2 seconds per reply. You'll see the classification badge on every task in the Inbox.

## The drafting step

If the classification triggered a draft (interested / question / not\_now / etc.), Guhan writes a response using:

* The prospect's original message (what they said)
* The full conversation history
* Your Brand & Voice settings
* The product being pitched
* The prospect's research brief

Every draft:

* Is 2-4 sentences typical
* Matches your tone
* Answers what they asked
* Suggests the next step (meeting, next email, more info)

## Where to configure this

**On the Outreach Agent's Settings tab** — you can set the workspace-wide default AND override per Outreach Agent:

<CardGroup cols={3}>
  <Card title="AI drafts + auto-send" icon="wand-magic-sparkles">
    Guhan writes the response + sends it without asking. **Safest** for confirmed follow-ups (they said yes, you're just replying "great, let me send you a link"). Risky for anything requiring judgment.
  </Card>

  <Card title="AI drafts + hand off" icon="hand-holding">
    Guhan writes the response + creates an Approve task. You read + approve OR edit + approve. **Recommended default** — you stay in control without doing the writing.
  </Card>

  <Card title="Always hand off" icon="user">
    Guhan classifies the reply but writes nothing. Creates a Take-over task with a "Reply here" button. **Best for high-touch sales** or high-stakes accounts.
  </Card>
</CardGroup>

## AI reply nodes on the sequence canvas

You can also set this per-step by dropping an **AI reply node** on the sequence canvas.

Configure:

* **Send for approval** — hand off vs auto-send (per this node)
* **Max consecutive replies** — how many rounds Guhan handles before mandatory handoff (default: 1)
* **Loop wait between replies** — 48h / 72h / etc. (how long to wait for their next reply)
* **Outgoing edges** — where the flow goes after N rounds

Common shape:

```
Wait for event: reply
     ↓
AI reply (auto-send, max 1 round)
     ↓
    │
    ├──→ conversation_done (they closed the loop)
    ├──→ meeting_requested (they asked for a call)
    └──→ not_interested (they said no)
```

## Approving a draft

Task detail → the draft is shown in an editor. You can:

* **Approve as-is** → sends immediately
* **Edit → Approve** → your changes get sent
* **Discard + write your own** → free-form reply

Your edits go back to Guhan as implicit feedback — over time it learns your patterns.

## When Guhan won't draft

Some replies always go to Take-over (regardless of your settings):

* **Long / complex replies** — over 500 characters, Guhan pauses
* **Technical questions** — pricing, integrations, security — high-stakes, judgment-heavy
* **Angry / rude replies** — human handling required
* **Multiple people on the thread** — group replies need context

You'll see the reason in the task detail — "handoff reason: complex\_question".

## Cost

Each AI reply draft = 1 credit. Doesn't matter if you approve or discard — the credit is spent on generation. Classification is free.

Rough budget: 1 credit per reply your sequence generates. If your sequence gets a 5% reply rate on 500 sends, that's 25 drafts = 25 credits.

## Related

<CardGroup cols={2}>
  <Card title="Inbox tasks" icon="inbox" href="/inbox/overview">
    All task modes explained.
  </Card>

  <Card title="Sequences" icon="diagram-project" href="/agents/outreach-agents/sequences">
    AI reply node on the canvas.
  </Card>
</CardGroup>