Build an AI Customer Service Bot with Make.com and Claude (No Code, 2026)

Most “AI chatbot” tutorials build decorated FAQs — rule-based flows with an AI facade. This post builds a real conversational bot: one that actually reasons about what the customer is asking, looks up real information, and responds with the right answer in natural language.

We’ll use Make.com for orchestration, Claude for reasoning, and a Google Sheet as the knowledge base. You can adapt the input channel (email, WhatsApp, website chat) to whatever your business uses.

What We’re Actually Building

Input: a customer message (email, form submission, or chat) asking a question about your business — pricing, availability, services, hours, wherever.

Processing:

1. Claude reads the message and figures out the intent (what they’re actually asking)
2. If the question needs specific info (prices, hours, etc.), fetch from a knowledge base
3. Claude drafts a response using the retrieved info
4. Reply goes back to the customer

Output: a response that sounds human, uses your actual business info, and covers the scenarios your FAQ doesn’t.

This is roughly the pattern behind the booking agent on ospipo.com.au — a premium chauffeur service where customers chat with an AI that handles enquiries end-to-end.

Step 1: Build Your Knowledge Base

Before touching Make.com, create a Google Sheet called Business Knowledge with these columns:

Topic	Question variations	Answer
Pricing	What are your prices, how much does it cost, pricing	List your current prices, specific and complete
Hours	When are you open, business hours, operating times	Your actual operating hours
Services	What do you offer, services, what do you do	Your service list
Location	Where are you, address, location	Your address and areas served
Booking	How to book, book, reserve	How a customer can book

Fill in the Answer column with actual info about your business. Be specific: “Tuesday–Sunday, 8am–10pm AEST” not “regular hours.”

Add as many rows as you want. The AI will figure out which row matches each question.

Step 2: Choose Your Input Channel

For this tutorial I’ll use email because it’s universal. Later you can swap in WhatsApp Business API, a website form, or a chat widget — the core flow is identical.

In Make.com:

1. New scenario → Gmail → Watch emails
2. Folder: Inbox
3. Criteria: (empty — we want to consider all incoming messages)
4. Maximum results: 1

Connect your business Gmail.

Step 3: Claude — Classify and Extract Intent

Add an Anthropic Claude module → Create a Prompt.

• Model: claude-haiku-4-5
• Max Tokens: 200
• System Prompt:

You are an intent classifier for a customer service bot.

Given a customer email, return ONLY a JSON object with these fields:

{
  "intent": "one of: pricing | hours | services | location | booking | other",
  "customer_name": "extracted name or null",
  "summary": "brief 1-sentence summary of what they want"
}

Return ONLY valid JSON. No explanation, no code fences.

• Messages → Add item → Role: user → Content: map {{1.Text content}} from Gmail

Step 4: Parse the Classification

Add a JSON → Parse JSON module:

• JSON string: map {{2.content[].text}}
• Data structure: create one with sample:

{"intent": "pricing", "customer_name": "Alice", "summary": "wants pricing info"}

Step 5: Look Up the Knowledge Base

Add Google Sheets → Search Rows:

• Connection: your Google account
• Spreadsheet: Business Knowledge
• Sheet: Sheet1
• Filter: Topic equals {{3.intent}} (from Parse JSON)
• Maximum results: 1

This pulls the relevant row based on Claude’s classification.

Step 6: Claude — Draft the Reply

Add another Claude module. This one has a richer job: it has the customer’s message, the retrieved business info, and now needs to compose a natural response.

• Model: claude-sonnet-4-6 (worth the upgrade for customer-facing writing)
• Max Tokens: 500
• System Prompt:

You are a friendly customer service representative for [YOUR BUSINESS NAME].

You've been given a customer's enquiry and the relevant business information. Write a polite, clear email reply that:

1. Addresses the customer by name if known
2. Uses the business info provided to answer their specific question
3. Keeps the tone warm but professional
4. Ends with a clear next step (how to book, how to reach you, etc.)
5. Signs off as "[YOUR SIGNATURE]"

Only use the business info provided. Do not invent details. If the info doesn't fully answer their question, acknowledge what you know and offer to follow up directly.

Return ONLY the email body. No subject line, no greeting formatting like "Dear Customer:", just the message text.

• Messages → Add item → Role: user → Content:

Customer email:
{{1.Text content}}

Customer name: {{3.customer_name}}

Relevant business info:
{{5.Answer}}

Please draft a reply.

(Variable numbers will match your actual scenario — check the outputs of each module to confirm.)

Step 7: Send the Reply

Add Gmail → Send an Email:

• To: map {{1.Sender email address}}
• Subject: Re: {{1.Subject}}
• Content: map {{6.content[].text}} from the second Claude module
• Content type: Text

Step 8: Test

Turn on Run once. Send yourself an email asking something like:

Hi there, I’m Sam and I wanted to check what your operating hours are. Thanks!

After about 10 seconds, check your inbox. You should have a reply from your business email that:

• Addresses Sam by name
• States your actual business hours (from the sheet)
• Sounds natural

If it looks good, turn the scenario ON.

Making It Better

The base flow works. These upgrades take it to production quality:

Human Approval Step

For a solo business, you might not want the bot auto-replying to everyone. Add a Telegram module between step 6 and 7: send the draft reply to yourself via Telegram with “approve” / “edit” buttons. Only send if you approve.

(This is easier to set up than it sounds — Telegram bot approval is a single module in Make once you have a bot token.)

Handling “other” Intents

When Claude classifies as other, the Google Sheets lookup returns nothing, and the second Claude call has no info to work with. Handle this with a Router:

• If intent = other: send a generic “Thanks, I’ll get back to you personally within a few hours” reply
• Forward the original email to yourself for manual handling

Memory Across Conversations

Basic version treats each email as standalone. For follow-up threads, add an email-thread lookup that retrieves prior messages in the same thread and includes them in the prompt. Claude will understand context.

Cost Control

Haiku for classification + Sonnet for replies costs pennies per conversation. But track it: the Anthropic Usage dashboard shows spending per model. Set a hard monthly limit in Anthropic Billing so a broken loop can’t drain your account.

When This Pattern Fits

• Services businesses where questions are predictable (pricing, hours, booking)
• Lead qualification for sales
• First-line tech support for software products
• Appointment-based businesses (salons, clinics, tutors)

When It Doesn’t

• Complex technical support needing deep product knowledge
• Regulated industries where every customer message needs human review
• Businesses where hallucination risk is unacceptable (medical, legal, financial advice)

For those, use AI only as a first-pass classifier, and route everything to humans for the actual response.

A Real Example

The booking agent at ospipo.com.au is a variant of this same pattern, adapted for a web chat interface instead of email:

• Website chat messages trigger a webhook instead of a Gmail poll
• Intent extraction also pulls pickup/dropoff locations
• Google Places API replaces the Google Sheet for live pricing
• Response goes back through the webhook to the website chat bubble

Same core architecture. Same 6-module structure. Live and taking real customer enquiries.

Next Steps

The full customer service bot build — including Telegram approval, thread memory, WhatsApp integration, and the router patterns — is in the Implementation Blueprint course ($29). Walks through production patterns step by step.

If you want to start simpler, the free Quick Start builds a single-scenario email automation that you can extend into this pattern over time.

---

Last updated: 20 April 2026.