Most AI support tools focus on the chat widget — the AI agent that sits on your website and handles inbound questions in real time. That's useful. But a large portion of support volume doesn't come through a chat widget. It arrives as email tickets, submitted through a contact form, processed by a helpdesk like FreeScout, Help Scout, or Zendesk.
AI Inbox Reply is built for that channel. It reads every incoming ticket, searches your knowledge base for a relevant answer, and drafts a reply — ready for your agent to review and send in one click.
Here's how to configure it so the drafts are actually good.
How it works under the hood
When a new ticket arrives in your connected helpdesk:
- SupportBot reads the full conversation thread — not just the last message
- It searches your knowledge base for the most relevant content
- It drafts a reply using that content, formatted as a professional support response
- The draft appears in your helpdesk as a draft reply, attributed to your AI agent
Your agent opens the ticket, reads the draft, adjusts if needed, and sends. Instead of writing from scratch, they're editing — which is dramatically faster.
If the AI agent can't find a relevant answer (confidence below your threshold), it leaves the ticket undrafted and flags it for manual handling. You never send a hallucinated response because the AI agent never sends anything — it only drafts.
Setting up your first inbox connection
Go to the Integrations tab in your SupportBot dashboard, select your helpdesk, and connect via the API key or OAuth flow.
Once connected, configure:
Which mailboxes to monitor. Don't enable Inbox Reply on every mailbox immediately. Start with your highest-volume, most predictable mailbox — usually a general support address. Leave billing disputes and escalations for later.
Draft vs. auto-send. Start in draft mode. Every reply goes through an agent before sending. Once you've reviewed 200–300 drafts and trust the quality, you can enable auto-send for replies above a high confidence threshold (90%+).
Confidence threshold. This is the most important setting. Set it too low and the AI agent drafts low-quality replies. Set it too high and it barely drafts anything. Start at 70% — the AI agent drafts when it has a solid answer, skips when it doesn't.
What makes a good draft
The difference between a useful draft and a generic one comes down to your knowledge base quality.
A generic draft looks like this:
"Thank you for contacting us. Please refer to our returns policy for information about your request."
That's useless. It's what you get when the AI agent found a returns policy page but the page doesn't have specific, actionable information.
A good draft looks like this:
"Thanks for reaching out about your return. Orders placed within 30 days are eligible for a full refund. To start the process, visit our returns portal at [link] and enter your order number. Refunds are processed within 3–5 business days once we receive the item."
That's what you get when your knowledge base has a dedicated, well-written returns page with specific timelines, links, and instructions.
The AI agent can only be as specific as your documentation. If your docs say "contact us for returns," the draft will say "contact us for returns."
Building a citation trail
Every draft SupportBot generates includes source citations — which knowledge base articles it used to construct the reply. This does two things:
- Your agents can verify the answer before sending, without digging through docs themselves
- You can spot documentation gaps when the cited sources are thin or off-topic
Review the citations on your first 50 drafts. You'll quickly see which pages are being referenced heavily (keep them updated) and which tickets are getting weak citations (those docs need to be written or improved).
Handling the edge cases
Some ticket types should never be auto-drafted, regardless of confidence:
- Billing disputes and refund requests above a certain value
- Tickets from VIP customers flagged in your helpdesk
- Anything that contains words like "legal", "lawyer", "media", "escalate to manager"
- Second-contact tickets where the customer has already replied once without resolution
Configure these as exclusion rules in the Inbox Reply settings. The AI agent will see these tickets but will not draft a response — it flags them for immediate human review instead.
Measuring the impact
After 30 days, look at:
Average handle time. The time from ticket creation to resolution should drop. Editing a good draft takes 20–30 seconds; writing from scratch takes 2–5 minutes.
First-reply time. With drafts ready when the agent opens the ticket, first-reply time drops significantly — especially during high-volume periods.
Draft acceptance rate. What percentage of drafts are sent with minimal edits? This is your quality signal. A rate above 70% is good. Below 50% means your docs need work.
Escalation rate on drafted tickets. If drafts are triggering more escalations than manual replies, the AI agent is drafting for the wrong tickets. Tighten the confidence threshold or add exclusion rules.
The right mental model
AI Inbox Reply is not an auto-responder. It doesn't replace your agents — it removes the blank-page problem from their workflow. Every ticket that arrives already has a draft answer waiting. Your agents spend their time on judgment, tone, and edge cases — not on typing the same answers they've typed a thousand times.
That shift — from writing to editing — is where the time savings come from.
AI Inbox Reply is available on Pro plan and above. Connect your FreeScout, Help Scout, or Zendesk mailbox in under 5 minutes. See pricing →
