How to build an AI-first support team

An AI-first support team isn't one where AI answers customers. It's one where every ticket starts with a draft, every agent works from the same written voice, and the team measures how much of the AI's work survives review. Built that way in FreeScout with ReplyRabbit, the team spends its time on judgement (is this right, is this what we promised, is this customer about to leave) instead of on typing. Here's how to build it in stages, with the numbers to watch.

Stage one: a draft on every ticket

Install ReplyRabbit, connect a provider, and turn on Auto Draft (New Email) so a reply is waiting before the agent opens the conversation. Spam, newsletters and receipts are filtered out first, so drafts only appear on real customer mail. This alone changes the shape of the day: agents review and send instead of starting from blank.

Do this on one mailbox first. The free plan covers one mailbox on OpenAI with 10 drafts a month. Pro removes the monthly draft limit when you're ready for real volume.

Stage two: one voice, written down

Most inconsistency in a support team is unwritten. Put the voice in the mailbox's Response Style as rules a stranger could apply, put business facts in Company Context, and add two or three real replies as Examples. Every draft now starts from the same place, and new hires inherit the voice on day one. On Pro, ReplyRabbit learns from the edits agents make, so the drafts converge on the team's style rather than drifting.

Stage three: measure edits, not impressions

The number that tells you whether AI-first is working is how much agents change drafts before sending. Three buckets are enough: sent with light edits, substantially rewritten, contained an invented fact. Track them for a fortnight. Team analytics on the Team plan record drafts used, edits made and time saved across the team, so the numbers don't depend on a spreadsheet.

If the rewritten bucket is large, the settings are thin. If the invented-fact bucket isn't zero, the knowledge sources are missing something. Fix the input, then measure again.

Stage four: ground the drafts

Connect the FreeScout Knowledge Base and your docs as knowledge sources on Pro, so drafts cite the article that answers the question instead of paraphrasing. Add a store connector if you sell products, so "where's my order" drafts know the order. On Team, connect an Outline wiki for internal knowledge that shouldn't be quoted to customers. Each source you add moves tickets from the rewritten bucket to the light-edit bucket.

Stage five: route on signals

Turn on AI Signals and Workflow Tags. Critical conversations get rr_priority_critical and go to whoever's on call; rr_needs_human gets a note; rr_no_reply_needed leaves the queue. Bug reports become Linear, GitHub or Jira issues, automatically on Team, and Slack hears about it. The team's attention now follows the signals rather than the inbox order.

What stays human, permanently

• Sending. No plan sends a customer reply. That's the line that makes the rest safe.
• Commitments. Refund amounts, delivery promises, exceptions to policy: an agent decides.
• Escalation judgement. Whether a report is a bug, and how urgent it is, gets a human check even when the AI has an opinion.
• The voice. The AI applies it; the team owns it and updates it.

Roles in an AI-first team

Someone owns the settings (context, style, knowledge sources) and reviews the edit numbers monthly. Agents fix drafts rather than replacing them, so the learning has something to learn from. A lead reads ten sent replies a month for voice drift. That's the whole management overhead.

Start with Getting started with ReplyRabbit, then How to install the ReplyRabbit module and How to set up ReplyRabbit. The FAQ answers the questions that come up in the first week.