How to automate customer email replies with AIstart here
Set up an AI assistant that drafts your repetitive email replies in Gmail, with you approving every send. A full, no-code walkthrough for small businesses.
If you run a small business, you probably answer the same handful of emails over and over. “What are your hours?” “Can you send a quote?” “Is this still available?” None of them are hard. Together they quietly eat an hour or more of your day.
This guide shows you how to set up an AI assistant that drafts those replies for you, right inside Gmail, so your morning job changes from writing twenty emails to skimming and approving twenty drafts. The rule we never break: the AI writes, a human approves. Nothing sends on its own until you decide it is safe.
No code. You will use Gmail, a tool called Zapier to connect things, and an AI model (ChatGPT or Claude). Plan on about 45 minutes the first time.
Heads up:some links below are affiliate links. If you start a paid plan through one, we may earn a commission at no extra cost to you. We only mention tools we would set up ourselves.Full disclosure.
What you are building
Here is the whole flow in plain English:
- A customer email arrives and gets a Gmail label, “AI-draft”, automatically.
- That label triggers an automation.
- The automation sends the email’s text to an AI model, along with a short playbook that tells the AI your facts, your tone, and your rules.
- The AI writes a reply.
- The automation saves that reply as a draft in the same email thread.
- You open Gmail, read the draft, tweak if needed, and hit send.
The magic is not the AI. The magic is the playbook plus the “draft, never auto-send” safety rule. That combination is what makes the replies sound like you and keeps you in control.
Step 1: Create the “AI-draft” label and filter
First, decide which emails the assistant should handle. Start narrow. Pick one type of email you answer constantly, for example quote requests or availability questions.
In Gmail, create a label called AI-draft. Then create a filter that applies this label
automatically. You can filter by words in the subject or body (“quote”, “availability”,
“booking”), or by the address your contact form sends to.
Step 2: Write your one-page playbook
This is the part that matters most, and the part most people skip. Your playbook is a single page that tells the AI everything it needs to answer like you would. It has three parts.
Facts. Your hours, your prices or price ranges, your service area, your turnaround times, links to your booking page. The things you repeat all day.
Tone. How you sound. Warm but brief? Formal? First names? Do you sign off with your name or the business name? Give one or two real examples of replies you are proud of.
Rules. The guardrails. “Never quote a firm price over email, give a range and invite a call.” “Never promise same-day service.” “If the customer sounds upset, draft a short, human apology and flag it for me.” “If you are not sure, say you are not sure and leave it for me.”
Keep it to one page. A tight playbook beats a long one, because the AI follows clear rules better than vague essays.
Step 3: Connect Gmail to your AI with Zapier
Zapier is the glue. It watches for the label and runs the steps. In Zapier, create a new automation (a “Zap”).
Trigger. Choose Gmail, then the “New Labeled Email” trigger, and pick your AI-draft
label. Now the automation runs every time a new email gets that label.
AI step. Add a step that calls your AI model. Both ChatGPT (OpenAI) and Claude (Anthropic) work well here; pick whichever you already pay for. You will paste your playbook into this step and feed in the incoming email’s body.
Step 4: Write the prompt
The prompt is where your playbook meets the incoming email. Here is one you can copy and adapt. Replace the bracketed parts with your own playbook text.
You are the email assistant for [BUSINESS NAME]. You draft replies to customer emails. You never send anything. A human reviews every draft before it goes out.
Here is your playbook. Follow it exactly.
FACTS: [paste your facts: hours, price ranges, service area, turnaround, booking link]
TONE: [describe how you sound, and paste one or two example replies you are proud of]
RULES:
- Keep replies short. Lead with the answer.
- Never quote a firm price. Give a range and offer a quick call.
- If the customer sounds upset or the question is unusual, write a brief, human reply and add this line at the very top: “NEEDS REVIEW:”.
- If you are missing a fact you need, do not guess. Write the reply with a clear [BLANK] where the fact should go.
- Sign off as [your sign-off].
Now write a reply to the email below. Return only the reply text, nothing else.
EMAIL: [the incoming email body is inserted here by the automation]
Notice the rules do real work: they stop the AI from inventing prices, flag anything emotional for your eyes, and leave obvious blanks instead of guessing. That is what keeps the drafts safe to trust.
Step 5: Save the reply as a Gmail draft
Add a final Zapier step: Gmail, “Create Draft Reply”. Point it at the original thread and pass in the AI’s reply as the body. Now the draft lands in the right conversation, ready for you.
Crucially, this step creates a draft, not a sent message. Do not use a “Send Email” action yet. We want you reading every reply for now.
Step 6: Test with ten real emails
Turn the Zap on and test it against ten real emails from your sent folder or inbox. Label
them AI-draft one at a time and watch the drafts appear.
For each draft, ask three questions: Is the information correct? Does it sound like me? Would I send this with no edits, small edits, or a rewrite? Track the answers. If most need a rewrite, your playbook needs sharpening, not the AI. Tighten the rules and the tone examples and test again.
Within a batch or two, most drafts should need only small edits. That is the finish line for phase one.
Common mistakes (and the fixes)
| What goes wrong | Why it happens | The fix |
|---|---|---|
| Replies sound generic or robotic | Playbook has no tone examples | Paste two real replies you wrote and liked |
| AI invents a price or a policy | Rules do not forbid guessing | Add “never guess, leave a [BLANK]” rules |
| Drafts miss the customer’s real question | Whole inbox pointed at one prompt | Narrow the filter to one email type first |
| Angry emails get a cheerful auto-reply | No escalation rule | Add the “NEEDS REVIEW:” flag rule |
| It sends before you review | A “Send Email” action was used | Use “Create Draft Reply” only |
A note on privacy
This is also why we exclude finance, health, and legal topics across this whole site. Those carry real responsibility, and a drafting assistant is the wrong tool for them.
When to move from draft to auto-send
After a few weeks, you will notice a pattern: some reply types you approve every single time with zero edits. A simple “yes, we are open until 6pm” confirmation, for example.
Those, and only those, are candidates for full auto-send. Move one category at a time. Keep a weekly check where you read a sample of what went out. Anything with a price, a promise, or a judgment call stays in draft mode for a human. There is no prize for removing yourself from decisions that deserve a person.
FAQ
FAQ
Do I need to know how to code?
No. Everything here uses Gmail settings and Zapier's point-and-click steps. The only writing you do is your plain-English playbook and prompt.
How much does this cost?
You need a Zapier plan that supports multi-step Zaps and an AI provider subscription. Exact prices change often, so check each tool's current pricing page. Start on the smallest plan that runs one multi-step Zap.
Will customers know an AI wrote the reply?
You are reviewing and sending every reply yourself in phase one, so it is genuinely your message. If you later move specific categories to auto-send, consider a light touch of transparency where it matters to your customers.
What if the AI writes something wrong?
That is exactly why nothing sends without your approval at first. Wrong drafts are feedback: tighten your playbook rules and the drafts improve. The human-approves rule is your safety net.
Can I use Make instead of Zapier?
Yes. The same trigger-plus-AI-plus-draft pattern works in Make. We compare the two in our Zapier vs Make guide so you can pick the one that fits your budget and comfort level.
Your next step
You now have the full picture: a label, a one-page playbook, a three-step automation, and a safety rule that keeps you in control. Set it up against one email type this week and test it with ten real messages.
When that is humming, the natural next move is picking the right automation tool for the jobs you will add next.