Build an Auto-Reply System — A 2-Hour Workshop from Form Submission to AI-Drafted Email

OUTCOME — What you have after these two hours

Following this workshop end-to-end leaves you with four things.

What gets built

• A live Google Form attached to your actual service (email + one core question + one context question, 3 fields total)
• A connected spreadsheet where responses pile up automatically
• A visual automation scenario — Sheets trigger → AI node → Gmail send, three nodes wired and toggled ON
• Proof of operation: when a friend submits your form, an AI-drafted reply lands in their inbox within 15 minutes, while you do anything else

📖 Term: Trigger and Action A trigger is the "when something happens" start signal. "When a new row is added," "when a new email arrives," "at a scheduled time" — those are all triggers. An action is the "then do this" execution. "Send an email," "notify Slack," "record in a sheet" — those are actions. Automation = trigger + action. The scenario you'll build in this workshop is structured as new-row detection (trigger) → AI response + Gmail send (actions).

Workshop metadata

Item	Value
Time	2 hours (fast: 90 min, unhurried: 3 hr)
Level	Intermediate (basic Forms/Sheets/Email experience assumed)
Tools	Web browser + Google account + one AI API key (Claude or GPT)
Coding knowledge	None (everything via GUI)
Payment	Zapier/Make free tiers work; $5 prepaid to AI API recommended for testing
Deliverable	A running auto-reply system (not a demo)

Why build this — customer abandonment happens in minutes

"I don't get that many inquiries; why bother automating?" Fair question. The numbers tell a different story. For early-stage services, a first response within 10 minutes correlates with conversion rates in the 10% range, while responses after an hour drop to 1% or below, a pattern that shows up across industries. Of three inquiries arriving during your 8 hours of sleep, two are already mentally gone by breakfast. The AI doesn't have to be perfect. Even "We've received your inquiry, here's our understanding of it, a team member will follow up with specifics" — auto-sent — prevents the drop-off. Immediacy beats accuracy.

Where you are now: you understand what this workshop produces. Next we verify the tools and accounts you need are in place.

---

PREREQUISITES — What you need

Four things, and you'll finish in 2 hours.

Required

• One Google account: Forms, Sheets, and Gmail all connect through the same one. A dedicated business account is even better
• Visual automation tool account: Zapier (zapier.com) or Make (make.com), either one. Both free tiers cover this workshop. Personally I recommend Make — its node limits are more generous, and as of 2026 it remains the best free-tier option
• One AI API key: Anthropic Console (console.anthropic.com) or OpenAI Platform (platform.openai.com). After adding a card and prepaying about $5, you can test all workshop long and still not spend $1
• Your service URL: a one-person shop, a course landing page, a portfolio site — anywhere inquiries could arrive. No URL yet? A temporary Notion page or Google Site works too

Not required

• Server administration, domain purchase, programming experience
• Paid CRM or customer management platforms
• A polished FAQ (one AI prompt replaces it)

📖 Term: API Key An API key is a "delegation password" that lets an external service talk to an AI on your behalf. The password you type into ChatGPT's web login is for humans; the API key is a password for programs. When Zapier or Make says "I'll send a question to Claude on this user's behalf," the API key proves the identity. Warning: this key has the same sensitivity as a password. If leaked, someone else can run AI workloads on your account's dime. Never post it to GitHub, blogs, or screenshots — only drop it into the automation tool's Credentials store.

Environment check (3 min)

Type each of these into your browser's address bar. All four should load while you're logged in.

• forms.google.com — Google Forms dashboard
• sheets.google.com — Google Sheets dashboard
• make.com or zapier.com — your automation dashboard
• console.anthropic.com or platform.openai.com — your AI API console

All four loaded, and an API key copied? You're ready. The key on the console screen will start with sk-ant- or sk-; copy it into a scratch note temporarily, and delete the scratch note when the workshop ends.

Where you are now: all tools are in hand. Before clicking anything, we decide what to receive. Skipping the design step means redoing two hours of work two hours later.

---

WHY — Why this build order

Lots of people try similar automations and get tangled mid-build. The common pattern: start with the AI prompt. Or start with the Gmail template. That reverses the data flow and forces rework at every step.

This workshop's order matches the direction data actually flows.

Customer (form) → Sheet (store) → Automation (detect) → AI (generate) → Gmail (send) → Customer (receive)

In this chain, each stage's output is the next stage's input. Form questions must be locked before sheet columns; sheet columns before AI prompt variables; AI output shape before Gmail body content. Reverse the order, and every new stage forces you to reopen the previous one.

Five tools, but each tool has exactly one role. Form collects, sheet stores, automation detects and orchestrates, AI judges, Gmail sends. When roles don't overlap, the system stays simple.

Where you are now: you know why this order. Now we execute it, one stage at a time, within two hours.

---

STEPS — Five-stage build

Step 1. Design form + connect spreadsheet (20 min)

First, decide what to receive. Skipping this and jumping to the form builder means changing questions mid-build — and every question change cascades into sheet columns and automation variables.

1-1. Pick one input type (3 min)

Three broad categories. Pick one today.

Type	Example question	AI response flavor
Inquiry	What are you curious about?	Explanatory, referral to resources
Signup	What problem do you want to solve?	Welcome, next-step guidance
Feedback	How was your experience?	Thanks, improvement commitment

This workshop uses "Inquiry" as the example, but the structure is identical for the other two. Later, when you want to add signups or feedback, just duplicate this scenario and swap the prompt.

1-2. Lock three questions (5 min)

On paper or in a note, fill in this template. Exactly three questions. More than that crushes submission rates.

Q1 (required): Email address → short answer + email-format validation ON
Q2 (required): [your service's core question] → long answer
Q3 (optional): [additional context] → short answer OR multiple choice

Example — if your service is a fictional calendar SaaS called "SmartDay":

Q1: Email
Q2: How many meetings do you have in a day on average? When do you need focus time?
Q3: Which calendar do you currently use? (Google / Outlook / Other)

1-3. Create the Google Form (7 min)

Go to forms.google.com and click "Blank form."

1. Title: [your service] — Inquiry
2. Description: a one-liner like "Submit below, an AI will draft a first reply within 24 hours"
3. Q1: short answer → three-dot menu → "Response validation" → "Text" → "Email" → required ON
4. Q2: long answer → required ON
5. Q3: your preferred type → required OFF

1-4. Connect the spreadsheet (3 min)

Top tab "Responses" → green Sheets icon → "Create a new spreadsheet" → done. The sheet should be empty with headers only — that's correct.

1-5. Test submission (2 min)

Top-right preview (eye icon) → new tab opens the form → submit a response to your own email ("Test #1") → back to the sheet, refresh → row 1 should now exist. Step 1 complete.

Step 1 checkpoint: at least one test response row must appear in the sheet. If not, the form-sheet connection is broken — reconfigure from the "Responses" tab.

Step 2. Create a scenario in the automation tool (25 min)

I'll use Make for the walkthrough, but Zapier is the same concept with different labels ("Zap" = "Scenario," trigger and action terminology matches).

2-1. New scenario (3 min)

make.com → sign in → left menu "Scenarios" → top-right "Create a new scenario" → blank canvas. The large gray circle in the middle is where the first node (trigger) goes.

2-2. Google Sheets trigger node (10 min)

1. Click the gray circle → search "Google Sheets" → select it
2. Submenu: "Watch Rows" or "Watch New Rows"
3. "Create a connection" → Google sign-in → permission grant
4. Spreadsheet field: select the sheet from Step 1
5. Sheet Name: "Form Responses 1" (or similar)
6. "Table contains headers" → Yes
7. "Row with headers" → 1
8. Limit → 10 (max 10 rows per run)
9. Save

2-3. Test the trigger (5 min)

Right-click the node → "Run this module only" → your Step 1 test response should appear as output. Email address, Q2 text, Q3 text should each be captured as separate fields. If no output, re-check sheet permissions and tab name.

2-4. Scheduling (5 min)

Clock icon at the bottom of the screen → "Run scenario" → "At regular intervals" → select 15 minutes. Once toggled ON later, it will check every 15 minutes. Don't turn it ON yet.

2-5. Save (2 min)

Bottom Save button (floppy disk). Name it something like ${your service} — Inquiry auto-reply.

Step 2 checkpoint: "Run this module only" should return the most recent form response as structured output, with email, Q2, and Q3 each in its own mappable field.

Step 3. Add AI node + write the prompt (30 min)

This is the heart of the workshop. Prompt quality = reply quality.

3-1. Add the AI node (7 min)

Click + to the right of the first node → search "Anthropic" (for Claude) or "OpenAI" (for GPT) → pick "Create a Message" (Claude) or "Create a Chat Completion" (OpenAI).

"Create a connection" → paste the API key from Prerequisites → Save.

3-2. Write the prompt (15 min)

Copy this template into the AI node's "Messages" or "Prompt" field and replace bracketed parts with your service's specifics.

You are a customer service representative for [service name].

[service name] provides [one-line description].
Core value: [one line]
Primary audience: [one line]

A customer just submitted the following inquiry:

- Email: {{1.Email}}
- Inquiry: {{1.Q2 answer}}
- Additional context: {{1.Q3 answer}}

Write a friendly, helpful reply in English (or the customer's language).

Guidelines:
1. Address the customer as "Hi there" or similar — no assumptions about name
2. First paragraph: one-line summary of their inquiry to show you understood
3. Second paragraph: 3–4 concrete sentences answering or guiding
4. Third paragraph: close with "reply to this email if you have follow-ups"
5. Overall 3–4 paragraphs, ready to paste into an email body
6. If uncertain, don't guess — say "A team member will follow up with specifics"

Variable insertion: inside the prompt, replace the {{1.Email}} placeholder by deleting the curly braces and clicking the matching field in Make's variable picker (blue bubble on the right panel). The tool auto-inserts the correct binding. Zapier uses a similar variable picker.

📖 Term: Variable mapping Variable mapping is the visual way of specifying "which output of the previous node connects to which input of this node." Instead of writing email = row.email in code, you click fields in a GUI and link them with arrows. Make uses drag-and-drop, Zapier uses dropdowns, but the idea is the same. Bad mapping means the wrong customer's email gets the wrong reply. Always test with your own email first.

3-3. AI node settings (5 min)

• Model: claude-sonnet-4-6 (Claude) — cost-efficient and well-calibrated for auto-reply workloads, or gpt-5.4 (OpenAI)
• Max Tokens: 1024 (length cap, fits 3–4 paragraphs)
• Temperature: 0.7 (balanced — not robotic, not wildly creative)

3-4. Test the AI node (3 min)

Right-click the node → "Run this module only" → the output pane should show the AI-generated reply. Natural, well-formed text with the customer's inquiry reflected — success. Robotic or translation-flavored text — add "write in natural, fluent English" to the prompt.

Step 3 checkpoint: output shows a 3–4 paragraph reply, opens by summarizing the inquiry, and closes with "reply if you have follow-ups." Miss any of these? Make the prompt rules more explicit and retry.

Step 4. Gmail send node + end-to-end test (25 min)

The AI wrote a reply. Time to actually send it to the customer.

4-1. Add the Gmail node (5 min)

+ to the right of the AI node → search "Gmail" → "Send an Email" → "Create a connection" → Google sign-in → permission grant.

4-2. Field mapping (10 min)

Field	Value
To	{{1.Email}} (customer email from first node)
Subject	[${your service}] Thanks for reaching out as fixed text
Content Type	HTML
Content	{{2.content[0].text}} (AI node output)
From Name	${your service} Support

The most common mapping mistake is the Content field. The exact field holding "the actual reply text" differs per AI service. Claude: content[0].text. OpenAI: choices[0].message.content. Expand the variable picker and find the field that contains real text (not metadata).

4-3. Send test (10 min)

This is the real test. Do these steps in order.

1. Open the form URL in a new tab
2. Submit a fresh test response to your own email (label it "Test #2")
3. Refresh the spreadsheet — confirm the new row exists
4. Back in Make, click "Run once" (green play button at the bottom)
5. Watch all three nodes turn green in sequence
6. Check your Gmail inbox — the AI reply should have arrived

If the email landed with clean subject and body, success. If not, click the red exclamation icon on the failing node in Make → read the error. 90% of the time it's a mapping typo or Gmail permission not granted.

Step 4 checkpoint: your Gmail shows an inbox email with subject "[your service] Thanks for reaching out," body contains the AI's 3–4 paragraph reply, and sender name is set correctly.

Step 5. Toggle ON + live monitoring (20 min)

Manual execution isn't automation. Real automation runs without you pressing a button.

5-1. Activate the scenario (3 min)

Make scenario, bottom-left toggle: OFF → ON. Zapier: top-right On/Off switch. Now it checks the sheet every 15 minutes on its own.

5-2. Ask a friend to test (5 min)

Send the form URL to one friend via messenger. "I built an auto-reply system — mind submitting once with your real email?" When they submit, the next run cycle (within 15 minutes) will send them an AI-drafted reply automatically.

5-3. Collect friend feedback (7 min)

Ask your friend "How was the reply?" One of three patterns usually comes back.

Feedback	Cause	Where to fix
"Too stiff"	Prompt lacks tone guidance	Add "warm, friendly tone" to the AI prompt
"Didn't seem to understand me"	Q2 variable not mapped into prompt	Re-check {{1.Q2}} mapping
"Took 15 minutes"	Schedule interval too long	Shrink Make clock to 5 min (mind free-tier ops limit)

5-4. Check execution logs (5 min)

Make left menu "Scenarios" → your scenario → "History" tab. A table of every run in the last 24 hours. Click any row to see exactly what data came in and what reply went out. In live operations, skim this log once a day to catch odd replies before they pile up.

Step 5 checkpoint: scenario toggle is ON, your friend's test submission got an AI reply via email, and History shows at least two successful runs. Workshop complete.

Where you are now: the auto-reply system is actually running. Next we do a final 4-point verification that you built it correctly.

---

VERIFICATION — What "done" looks like

All four should be "yes" before you call this done.

1. [ ] The form link is attached to your service URL or landing page. One button click should open the form
2. [ ] Form responses auto-record into the spreadsheet. At least 2 test-response rows should be stacked up
3. [ ] The Make/Zapier scenario is ON. Toggle is green, or "Active" label is visible
4. [ ] A third party (your friend) submitted a form, and the AI reply auto-sent to their email. Your friend confirmed receipt verbally

If any is "no," trace backwards through the stages. The most common failure spots are Step 4 variable mapping (empty Content field, wrong field linked) and Step 5 scenario OFF (saved but not activated). The Make scenario's History tab surfaces the precise failure reason in English — and if the error jargon confuses you, paste it into Claude with "explain this error in plain language" and you'll have a translation in seconds.

Why the friend test is mandatory: when you test with your own Google account sending to itself, Google does special handling that differs from how external recipients receive mail. A friend's inbox (different domain) is the closest proxy to a real customer.

---

VARIATIONS — When the situation differs

Variation 1 — AI auto-classifies inquiry type

Add this single line to the end of your prompt and the AI will classify too.

Before the reply, output a line: [Category: urgent / general / spam]

Then add a Router (branch module) in Make to handle categories differently: urgent → instant Slack/SMS to you; general → AI reply auto-sends; spam → do nothing. This single variation cuts your operational load in half.

Variation 2 — Slack mirror alerts

Add a Slack node next to the Gmail node, mirroring the same data to your personal Slack channel. The instant a reply goes out to a customer, your phone pings. Far faster than discovering in the morning that the AI sent something strange at 3 a.m.

Variation 3 — Notion database log

Add a Notion node that writes each inquiry to a database row. Three months in, when you ask "which inquiries come up most?", this DB becomes the seed of your FAQ. Pulling the top 3 recurring questions onto your landing page typically cuts inquiry volume by 30%.

Variation 4 — Multi-language auto-handling

Add "Detect the customer's language and respond in the same language" to the prompt. English, Japanese, Korean inquiries all get handled by the same AI. The biggest wall for a solo global service is "the overhead of customer service in English" — one prompt line removes that wall.

---

TROUBLE — Common stuck points

❓ Make shows "Module execution failed" Usually API key expiration or credit exhaustion. Anthropic/OpenAI console → Usage → check balance. Under $1? Top up. Or you may have burned through Make's free-tier "operations" quota (1,000/month). Upgrade plan, or stretch schedule from 15 min to 30 min to conserve.

❓ AI reply comes back in the wrong language Add to the top of the prompt: "Respond only in [target language]. Avoid foreign phrases unless strictly necessary." Claude fixes almost immediately; GPT complies under the same instruction.

❓ Variable picker doesn't show the AI reply text field The AI node's output is nested. In Make's variable picker, click the triangle next to a field to expand nested sub-fields. Claude path: content → [1] → text. OpenAI path: choices → [1] → message → content. After running the scenario once with "Run once," reopen the picker — you'll see the actual returned structure, making mapping much easier.

❓ Replies only go to my email, not the customer's The To field has your email hard-coded. Click the To field → variable picker → connect it to the Google Sheets node's "Email" field (Q1 mapping). Must be a variable, not a static value.

❓ 15 minutes pass and no auto-run Double-check the toggle is ON. Make: circle toggle at scenario card's bottom-left. Zapier: On/Off switch top-right of the Zap page. People most often save but forget to activate.

❓ AI replies are consistently too long or too short Adjust the AI node's Max Tokens. Too long? Try 800. Too short? Try 1500. Or add explicit length constraints to the prompt: "Respond in exactly 3 paragraphs, each 2–4 sentences."

---

NEXT — What comes next

Two hours ago you may have been the person who manually typed replies to every inquiry. Now you're the person running a system that replies on your behalf while you sleep, train, or travel. This isn't just convenience — it's the lever that makes a solo-scale service survivable as volume grows.

This workshop is the operational-phase hands-on of the SPIN series. The path continues:

1. Auto-reply system ✓ (done)
2. Content auto-generation pipelines — blog and newsletter automation (in prep)
3. Segment-based auto-response — VIP / standard / free-tier branching (in prep)
4. Dashboard build — real-time inquiry / revenue / churn tracking (in prep)

The best way to keep this system alive is a daily 5-minute sweep of Make's History tab. Any weird AI replies? Any unusually slow responses? Any new inquiry types? Those 5 minutes are your service's feedback loop and the source of every future prompt improvement. What you built is a beginning, not an end. The first wheel is rolling; you now just oil it daily so it keeps rolling.

고객(폼 입력) → 시트(저장) → 자동화(감지) → AI(생성) → Gmail(발송) → 고객(수신)

Q1 (필수): 이메일 주소 → 단답형 + 이메일 형식 검증 ON
Q2 (필수): [당신 서비스의 핵심 질문] → 장문형
Q3 (선택): [추가 맥락] → 단답형 OR 객관식

Q1: 이메일 주소
Q2: 하루에 회의가 평균 몇 개 있으신가요? 집중 업무 시간이 언제 필요하신가요?
Q3: 현재 쓰시는 캘린더는 무엇인가요? (구글/아웃룩/네이버/기타)

당신은 [서비스명]의 고객 서비스 담당자입니다.

[서비스명]은 [한 문장 서비스 설명]을 제공합니다.
핵심 가치: [한 줄]
주요 타겟: [한 줄]

방금 고객 한 분이 다음 문의를 보내셨습니다.

- 이메일: {{1.이메일 주소}}
- 문의 내용: {{1.Q2 답변}}
- 추가 맥락: {{1.Q3 답변}}

이 문의에 대해 친절하고 도움이 되는 답변을 한국어로 작성해주세요.

답변 작성 지침:
1. 호칭은 "고객님"으로 통일할 것
2. 첫 문단에서 문의 내용을 한 줄로 요약해 "이해했음"을 보여줄 것
3. 두 번째 문단에서 구체적 답변·안내를 3~4문장 제공할 것
4. 세 번째 문단에서 "추가 질문이 있으시면 이 메일에 회신 주세요"로 마무리할 것
5. 전체 3~4단락, 이메일 본문으로 바로 쓸 수 있는 깔끔한 형태로
6. 확실하지 않은 정보는 추측하지 말고 "담당자가 확인 후 추가 안내드리겠습니다"로 처리할 것

답변 시작 전 맨 위에 [카테고리: 긴급/일반/스팸] 형식으로 분류를 붙여주세요.