Sort and Draft Support Emails With AI in Make, and Keep a Person on the Send Button
Quick verdict: for a small team sharing one support inbox, the safest useful automation is not auto-reply. It is auto-sort plus auto-draft. Make reads each new email, an AI step labels it and writes a reply, and the reply is saved as a draft in the original thread. A person reads it and presses send.
This guide is part of our AI Automation for Small Business series.
Why Draft Instead of Auto-Reply
An AI reply that goes straight to a customer can be wrong in ways that cost you: a refund promised that policy does not allow, a delivery date invented, a complaint answered with a cheerful template. A draft has none of that risk, because nothing leaves until a person approves it. You still save most of the time, since sorting and first-draft writing are the slow parts.
Saving the reply as a draft is also the simplest approval step available. It needs no extra app, no extra plan, and no training. The reviewer opens the thread, reads, edits if needed, and sends.
What You Need
- A Make account. The free plan includes 1,000 credits a month, two active scenarios, and a 15-minute minimum interval, which is enough to build and test. Core is $9 a month for 10,000 credits and a 1-minute interval. See Make pricing.
- A Gmail or Google Workspace support mailbox. Outlook works with the equivalent modules.
- An AI connection. Either Make’s built-in AI tools or your own OpenAI or Anthropic API key connected in Make.
- A short reply guide: your refund policy, opening hours, delivery times, and the five questions you answer most. One page is enough.
- About 90 minutes for the first build and test.
The Scenario at a Glance
- Gmail: watch emails in the support inbox.
- Filter out newsletters, auto-replies, and your own outgoing mail.
- AI step returns a category, an urgency level, and a draft reply.
- Router sends each category down its own branch.
- Gmail: create a draft in the same thread, and add a label.
- Notify the team in Slack or by email for urgent items.
- Log the email, category, and timestamp to a Google Sheet.
Step 1: Watch the Inbox
Create a new scenario and add the Gmail module that watches for new emails. Point it at the inbox folder and set it to process a small batch per run, for example 5, so one busy hour does not use your whole credit allowance. On the free plan the scenario can run every 15 minutes at most. On paid plans you can go down to every minute.
Step 2: Filter Out What Should Never Be Answered
Add a filter on the line after the trigger. Let an email continue only when all of these are true:
- the sender is not your own domain,
- the sender address does not contain “no-reply” or “noreply”,
- the subject does not start with “Automatic reply” or “Out of office”.
Every email the filter stops saves the credits the later modules would have used.
Step 3: Classify and Draft in One AI Step
Add an AI module and ask for structured output, so later modules can read each field separately. A prompt to start from:
You are helping a small business support team. Read the customer email below. Return JSON with exactly these fields: “category” (one of: order_status, refund_request, technical_problem, billing_question, complaint, sales_enquiry, other), “urgency” (low, normal, high), “summary” (one sentence), and “draft_reply”.
Rules for draft_reply: use only the facts in the reply guide below and in the customer’s email. If the guide does not cover the question, write a short reply that acknowledges the question and says a team member will follow up, and set category to “other”. Never promise a refund, a date, or compensation. Keep it under 120 words. Sign off as “The [Business Name] Support Team”.
Reply guide: [paste your one-page guide]
Customer email: [map subject and body from the Gmail module]
Follow the AI module with a JSON parsing step if your AI module does not return separate fields on its own.
Two rules protect you here. First, the reply guide is the only source of facts. Second, the fallback for anything outside the guide is a holding reply, not a guess. Before you paste customer emails into any AI service, check its data terms. Our AI compliance guide lists what to look for.
Step 4: Route by Category
Add a Router after the AI step with one branch per group of categories:
| Branch | Categories | What happens |
|---|---|---|
| Routine | order_status, billing_question, sales_enquiry | Draft saved in thread, label “AI draft” |
| Sensitive | refund_request, complaint | Draft saved, label “Needs owner”, team notified |
| Technical | technical_problem | Draft saved, label “Tech”, assigned to the right person |
| Fallback | other, or anything the AI could not classify | No draft, label “Manual”, team notified |
Set the fallback as the router’s default route, so an unexpected value never disappears silently.
Step 5: Save the Draft in the Thread
On each drafting branch, add the Gmail module that creates a draft. Reply to the original sender, reuse the original subject, attach the draft to the original thread, and put the AI’s draft_reply in the body. Then add a Gmail module that applies the label for that branch. Labels are how the team sees at a glance what is waiting.
Step 6: Notify for Urgent Items
On the Sensitive branch, and on any item where urgency is “high”, add a Slack or email module that posts the one-sentence summary and a link to the thread. Keep routine items out of notifications, or people will mute the channel within a week.
Step 7: Log Every Email
Finish each branch with a Google Sheets module that adds a row: date, sender, category, urgency, branch. After a month this tells you which questions are most common, which is the best input for improving the reply guide.
What About a Formal Approval Step?
Make lists a Human in the Loop app that pauses a scenario until a reviewer responds. Its listing is labelled Enterprise, so check whether your plan includes it before designing around it. See Make’s Human in the Loop listing. For most small teams, the draft-in-thread approach above does the same job at no extra cost.
Add an Error Handler Before You Switch It On
If the AI service is briefly unavailable, the scenario should not lose the email. Make offers five error handlers: Break, Commit, Ignore, Resume, and Rollback. Add a Break handler to the AI module so a failed run is stored and retried, and add the “Manual” label on error so a person picks the email up. See Make’s error handler documentation.
Test With Real Emails
- Send the support inbox six test emails from a personal address, one per category.
- Run the scenario once by hand and check each landed on the right branch.
- Read every draft against your reply guide. Look for invented facts.
- Send one question the guide does not cover and confirm you get the holding reply and the “Manual” label.
- Send an auto-reply and confirm the filter stops it.
What It Costs to Run
Make counts each module action as one credit. An email that goes through the full path uses about six to eight credits: watch, AI, parse, draft, label, log, and sometimes notify. At 300 support emails a month that is roughly 2,000 to 2,500 credits, inside the 10,000 on Core. AI modules can use additional credits or your own API budget depending on how you connect the model, so check Make’s pricing page for how AI usage is billed on your plan.
Common Mistakes
1. Letting the AI answer from general knowledge
Without a reply guide and a “use only these facts” rule, drafts will sound confident and be wrong about your policies.
2. No fallback route
If the router has no default branch, an email with an unexpected category is processed by nothing and nobody notices.
3. Turning on auto-send after one good week
A good week proves the common cases work. The damage comes from the rare ones.
4. Notifying on everything
Notifications are for sensitive and urgent items only.
FAQ
Can I build this on Make’s free plan?
Yes, for testing and for very low volume. The 15-minute interval and 1,000 credits a month are the limits you will reach first.
Does this work with Outlook or a help desk tool?
Yes. Swap the Gmail modules for the Outlook or help desk equivalents. The filter, AI step, router, and log stay the same.
Should any category be fully automatic?
Possibly simple acknowledgements, such as “we have received your message”, after a month of clean results. Refunds and complaints should stay with a person.
Why Make instead of Zapier for this?
The router makes several branches easy to see and maintain on one canvas, and credits are inexpensive at this volume. Our Zapier vs Make vs n8n comparison covers the trade-offs.
Final Recommendation
Build the sort-and-draft version first and run it for a month. Use the log to improve the reply guide. Only then consider letting one low-risk category send on its own.
