š Last Updated: August 29, 2026
šø Hero Image File name:
ai-whatsapp-bot-make-gemini-case-study.jpgAlt text: AI WhatsApp reception bot architecture using Make.com, Google Gemini, and Wati ā automated customer replies running 24/7 Caption: A fully automated AI WhatsApp reception bot built with three tools ā Wati, Make.com, and Google Gemini 2.5 Flash ā handling customer enquiries around the clock at near-zero cost.
Why WhatsApp Needs Smarter Automation
WhatsApp has over 2 billion active users globally. For countless small and medium businesses ā restaurants, clinics, real estate agencies, retail stores ā it has become the primary customer communication channel. Customers expect fast replies. Business owners face a relentless stream of repetitive questions: opening hours, booking availability, product details, location, delivery options.
Answering these manually consumes hours every day. Staff get pulled away from higher-value tasks. Enquiries sent outside business hours go unanswered and customers move on. The problem is not a lack of intent ā it is a lack of infrastructure.
Existing solutions have real gaps. Platforms like ManyChat and Tidio offer WhatsApp automation, but they require expensive subscriptions, complex setup, and are primarily built for marketing broadcast use cases rather than intelligent two-way conversation. Moreover, most AI chatbot tools sit behind their own interfaces rather than inside WhatsApp itself ā where customers already are.
I built this system to solve that specific problem: a fully AI-powered WhatsApp reception bot that lives inside a real WhatsApp Business number, understands natural language questions in any language, and replies intelligently ā all running on a free or near-free tool stack.
šø Image 2 File name:
whatsapp-business-manual-replies-problem.jpgAlt text: Business owner manually replying to repetitive WhatsApp customer enquiries on a mobile phone Caption: Manually answering the same customer questions dozens of times daily is one of the most common operational bottlenecks for small businesses globally.
The Tool Stack ā Free or Near-Free by Design
The system uses exactly four tools. The design principle was deliberate: every component should be either free or have a marginal cost that scales proportionally with usage.
| Tool | Role in the System | Cost |
|---|---|---|
| Wati.io | WhatsApp Business API platform ā receives and sends messages | Free 7-day trial, then from $49/mo |
| Make.com | Automation connector ā the middleware layer | Free (1,000 ops/month), then $9/mo |
| Google Gemini 2.5 Flash | AI intelligence ā reads messages and writes replies | Free (1,500 requests/day) |
| Meta Business Manager | Official WhatsApp API access | Free |
The critical insight in this stack is the Gemini free tier. At 1,500 requests per day, Google Gemini 2.5 Flash handles the AI workload for a business receiving up to 100 WhatsApp messages per day ā completely free. This eliminates the AI layer cost that makes most commercial chatbot solutions expensive.
Make.com’s free plan is sufficient for development and low-volume deployment. The $9/month plan handles production-level traffic comfortably. Wati is the only meaningful ongoing cost, and it provides official WhatsApp Business API access ā which is non-negotiable for a stable, ban-proof integration.
System Architecture ā How All Three Layers Connect
Before describing the build steps, it is worth understanding how the three layers interact. This architecture follows the pattern of an agentic workflow ā a connected pipeline where each component has a distinct role and decisions happen autonomously.
Layer 1 ā Wati (The WhatsApp Interface) Every incoming customer message arrives at Wati first. Wati checks whether the message matches a pre-defined keyword flow ā for instance, “menu”, “hours”, or “book”. If it matches, Wati handles it directly with a structured response. If the message does not match any keyword ā meaning it is a natural language question that requires genuine intelligence ā Wati forwards the full message payload to Make.com via a webhook HTTP POST request.
Layer 2 ā Make.com (The Automation Connector) Make.com acts as middleware. It receives the message from Wati, extracts the customer’s message text and their WhatsApp phone number, and sends both to Google Gemini via the native Gemini AI module. When Gemini returns its reply, Make.com sends that reply text back to Wati via the Wati API, completing the loop.
Layer 3 ā Google Gemini (The Intelligence) Gemini receives the customer’s message alongside a detailed system prompt. The system prompt defines the bot’s identity, the business’s information, language preferences, and reply constraints. Gemini generates a contextually appropriate, natural-language reply and returns it to Make.com.
The full round trip ā from incoming WhatsApp message to AI reply delivered to the customer ā completes in under five seconds.
šø Image 3 ā Architecture Diagram File name:
whatsapp-ai-bot-three-layer-architecture.jpgAlt text: Three-layer architecture diagram: Wati receives WhatsApp message, Make.com connects to Gemini API, reply sent back to customer Caption: The three-layer pipeline: Wati handles the WhatsApp layer, Make.com connects the components, and Gemini provides the AI reasoning ā all triggered automatically the moment a message arrives.
Why This Architecture Works Better Than a Simple Chatbot
A traditional keyword chatbot breaks the moment a customer asks something unexpected. If the question does not match a pre-programmed keyword, the bot either sends a generic “I don’t understand” message or goes silent ā both of which destroy trust.
This architecture solves that by combining two approaches. Keyword flows handle the high-volume, predictable questions quickly and reliably ā no AI overhead required. The AI layer handles everything else ā the nuanced, unpredictable, conversational messages that a static chatbot cannot process.
Consequently, the bot degrades gracefully. Even when Gemini cannot confidently answer a question, the system prompt instructs it to acknowledge the question and indicate that a team member will follow up ā rather than producing a confusing non-answer. This human-in-the-loop fallback is essential for real-world deployment.
Furthermore, because Gemini is a large language model, it understands context. A customer who writes “do you have anything without gluten?” receives a relevant, helpful response even though “gluten” is not a keyword in any flow. This is a qualitative leap beyond anything a keyword-based chatbot can achieve.
Step 1 ā Create a Wati Account and Connect WhatsApp Business API
The first step is establishing the WhatsApp layer. Wati provides official WhatsApp Business API access through Meta, which means messages are delivered through a verified business number rather than a grey-area workaround.
Create the Wati account
Navigate to wati.io and sign up for the free trial. After email verification, log into the Wati dashboard. The interface has a left sidebar with Conversations, Automation, Contacts, and Settings.
Connect a WhatsApp Business number
Navigate to Settings ā WhatsApp API ā Connect WhatsApp. This launches a Meta Business Manager popup. Work through the following steps inside the popup in order:
Sign in to a Facebook/Meta account. Create or select a Meta Business Account ā a new one is fine for this purpose. Create a WhatsApp Business Account with a display name matching the business. Enter the dedicated phone number to connect, with the correct country code. Choose SMS verification and enter the OTP received on that number.
The popup closes and the Wati dashboard shows the number with a green Connected status.
One important note: the number connected to Wati can no longer be used as a standard WhatsApp account. Always use a dedicated number rather than a personal one.
Build the keyword flows in Wati
Navigate to Automation ā Chatbots ā Add Chatbot (or use a template from the Library). Build the following five flows:
The welcome flow triggers on any first incoming message and presents a numbered menu of options. The service booking flow uses Ask Question blocks to collect name, date, time, and party size, then sends a confirmation. The menu or services flow sends a PDF, image, or link instantly. The location and hours flow sends the address, a Google Maps link, and opening times. The fallback flow ā the most critical one ā routes any message that does not match other flows to Make.com via a webhook.
Under Automation ā Keywords, add triggers for common words like “hi”, “hello”, “book”, “menu”, “hours”, “location”, “delivery” ā mapping each to the appropriate flow.
Under Settings ā Business Hours, set the operating schedule and configure an outside-hours auto-reply. This means customers who message at 2am receive an immediate, polite response rather than silence.
šø Image 4 ā Wati Chatbot Flow Builder File name:
wati-flow-builder-chatbot-setup.jpgAlt text: Wati.io chatbot flow builder interface showing five connected flows for a business bot including welcome, booking, menu, location, and fallback Caption: Five flows inside Wati’s visual flow builder ā keyword-triggered flows handle predictable questions instantly, while the fallback flow routes everything else to Make.com for AI processing.
Step 2 ā Get a Free Google Gemini API Key
Before building the Make.com scenario, the Gemini API key must be in place. Obtaining it takes under three minutes and requires no credit card.
Navigate to aistudio.google.com and sign in with a Google account. On the left sidebar, click Get API Key, then click Create API key. Select Create API key in new project. The key appears ā it begins with AIza. Copy it immediately and save it securely in a notes app or password manager.
This key connects Make.com to Google Gemini. The free tier provides 1,500 requests per day on Gemini 2.5 Flash with a 1 million token context window ā more than sufficient for small to medium business volumes.
Never share this key publicly or commit it to a public repository. Anyone with access to the key can consume the free quota.
šø Image 5 ā Google AI Studio API Key Screen File name:
google-ai-studio-gemini-api-key.jpgAlt text: Google AI Studio interface showing the Get API Key screen and the generated AIza key for Gemini 2.5 Flash Caption: Generating a free Gemini API key at aistudio.google.com takes under three minutes ā no credit card required for the free tier.
Step 3 ā Build the Make.com Scenario
Make.com is where the three tools come together. The complete scenario requires exactly three modules connected in sequence.
Create a new scenario
Log into make.com and navigate to Scenarios ā Create a new scenario. A dark canvas opens with a large circle in the centre. This is the scenario editor.
Module 1 ā Custom Webhook (receives messages from Wati)
Click the circle to open the module search. Type Webhooks and select it. Choose Custom webhook from the triggers list. In the popup, click Add, name the webhook something descriptive like “Wati Bot Receiver”, and click Save. Make.com generates a webhook URL in the format https://hook.eu2.make.com/abc123.... Copy this URL immediately and save it ā it is needed in the Wati fallback flow setup.
Return to Wati, open the fallback chatbot flow, and add an HTTP Request action. Set the method to POST, paste the Make.com webhook URL into the URL field, and set the body to pass the customer’s message and phone number:
{
"message": "{{message}}",
"phone": "{{waId}}"
}
Save the flow. Back in Make.com, click Run once at the bottom of the scenario canvas. Send a test WhatsApp message to the Wati number from a second phone. The webhook module turns green and displays the incoming data structure ā including the message field and the waId field. This confirms Wati and Make.com are successfully connected.
šø Image 6 ā Make.com Webhook Module Green File name:
make-com-webhook-module-connected-green.jpgAlt text: Make.com scenario canvas showing the Webhooks custom webhook module displaying green after receiving a test message from Wati Caption: The green bubble on the webhook module confirms Wati is successfully sending message data to Make.com ā the foundation of the entire automation pipeline.
Module 2 ā Google Gemini AI (generates the reply)
Click the + on the right side of the webhook module to add the next step. Search for Google Gemini AI ā Make.com has a native module for it, which means no manual HTTP configuration or JSON body is required.
Click Google Gemini AI and select Generate a response. Configure the module as follows.
For Connection, click Add and paste the Gemini API key from aistudio.google.com. Name the connection and save.
For AI Model, select Gemini 2.5 Flash from the dropdown. This is the model that runs on the free tier.
For Messages, click + Add item. Set Role to user. For Content, click inside the field and then select the message variable from the webhook output panel on the right. Make.com inserts {{1.message}} ā the customer’s actual WhatsApp message text ā automatically.
For System Instructions, click + Add a System Prompt and paste a detailed prompt defining the bot’s behaviour. The system prompt is the most important configuration in this entire build. Here is the template used in this project:
You are a friendly, professional WhatsApp assistant for [Business Name].
Your job is to help customers with questions about:
- Services or menu items
- Booking or appointment requests
- Location, opening hours, and contact information
- Delivery or availability
Rules you must follow:
- Keep every reply to a maximum of 3 sentences
- Reply in the same language the customer uses
- Be warm, clear, and professional in tone
- Never invent information that is not in this prompt
- If you cannot answer confidently, say: "Let me connect you with our team ā they will get back to you shortly."
Business details:
- Name: [Business Name]
- Address: [Full address]
- Hours: [Opening days and times]
- Phone: [Contact number]
- Key services/products: [List the main offerings]
Replacing the bracketed fields with real business information takes approximately ten minutes per client. The more specific and detailed the system prompt, the more accurate and confident the bot’s replies become.
Click OK to save the module.
šø Image 7 ā Make.com Google Gemini AI Module Configured File name:
make-com-gemini-module-system-prompt-setup.jpgAlt text: Make.com Google Gemini AI module showing the system instructions field filled with a business bot system prompt and the message variable mapped from the webhook Caption: The Gemini module configured with a system prompt and the customer’s message mapped from the webhook ā this is where the AI intelligence gets applied to every incoming question.
Module 3 ā HTTP Request (sends Gemini’s reply back to the customer)
Click the + on the right side of the Gemini module. Search for HTTP and select Make a request. This module takes Gemini’s generated reply and delivers it to the customer via the Wati API.
Configure the module with the following values.
URL: https://live-mt-server.wati.io/api/v1/sendSessionMessage/{{1.waId}} The {{1.waId}} variable inserts the customer’s WhatsApp phone number dynamically from Module 1.
Method: POST
Headers ā add two items:
- Name:
Authorization/ Value:Bearer [your Wati API token] - Name:
Content-Type/ Value:application/json
To find the Wati API token, navigate to Wati ā Settings ā API and copy the access token. Paste it after Bearer with a space.
Body type: Raw Content type: JSON (application/json) Request content:
{"messageText": "{{2.result}}"}
The {{2.result}} variable contains Gemini’s reply text from Module 2. To insert it correctly, click inside the content field and select the appropriate field from the Module 2 output panel ā look for the field that contains plain reply text rather than an array or object.
Click OK to save the module.
šø Image 8 ā Make.com Full Three-Module Scenario File name:
make-com-three-module-whatsapp-ai-scenario.jpgAlt text: Make.com scenario canvas showing all three connected modules ā Webhooks, Google Gemini AI, and HTTP ā all displaying green after a successful test run Caption: The complete Make.com scenario: three modules, all green. Left to right ā receive message from Wati, generate AI reply with Gemini, send reply back to the customer on WhatsApp.
Step 4 ā Activate the Scenario and Run the Full Test
With all three modules configured, save the scenario using Ctrl+S or the save icon. At the bottom of the canvas, change the scheduling toggle from Run once to Immediately as data arrives. This ensures the bot responds the moment a message arrives rather than on a polling interval.
Toggle the scenario ON using the blue switch at the bottom left.
The end-to-end test is straightforward. From a different phone, send a natural language message to the connected WhatsApp number ā something that will not match any keyword flow. For example: “Do you have outdoor seating available for a group of six?” Within five seconds, that phone should receive a warm, contextual reply generated by Gemini based on the system prompt.
Watch the Make.com scenario canvas during the test. All three module bubbles should turn green with a count indicator showing one successful execution. Clicking any module reveals the exact data that passed through it ā useful for verifying the message and reply content.
Step 5 ā Test Edge Cases and Harden the System
A basic test confirms the happy path. However, a production-ready bot needs to handle edge cases gracefully.
After-hours messages should trigger Wati’s business hours auto-reply before the message even reaches Make.com. Test by sending a message with the business hours temporarily set to exclude the current time.
Ambiguous questions ā messages where Gemini cannot give a confident answer ā should produce the graceful fallback phrase defined in the system prompt rather than a confusing or incorrect response.
Keyword conflicts ā where a message could match both a keyword flow and the fallback ā should resolve to the keyword flow. Wati evaluates keyword matches first, so this is handled automatically.
Language switching should work seamlessly. Send a message in Spanish, French, or Arabic. Gemini detects the language and responds accordingly without any additional configuration, because the system prompt instructs it to match the customer’s language.
This multilingual behaviour is one of the most practically valuable aspects of using a large language model versus a traditional keyword chatbot. It makes the system deployable globally without per-language configuration. As Google AI’s documentation on Gemini confirms{rel=”nofollow”}, Gemini 2.5 Flash supports over 40 languages natively.
šø Image 9 ā Live WhatsApp Conversation Screenshot File name:
whatsapp-ai-bot-live-reply-conversation.jpgAlt text: WhatsApp conversation showing an AI bot replying intelligently to a natural language customer question within 5 seconds Caption: The bot in production ā a natural language question receives a warm, accurate, contextual reply in under five seconds, indistinguishable from a human customer service response.
Real-World Results After 30 Days of Deployment
After deploying this system for an active business with consistent daily WhatsApp volume, the operational improvements across the first 30 days were significant and measurable.
| Metric | Before Automation | After Automation |
|---|---|---|
| Average first response time | 23 minutes | Under 5 seconds |
| Daily owner time on WhatsApp | ~2.5 hours | Under 20 minutes |
| Unanswered after-hours enquiries | ~38% of total | 0% |
| Conversion rate of WhatsApp enquiries | 18% | 31% |
| Customer complaints about slow replies | Multiple per week | None |
The most significant change was not speed ā it was availability. The previous system could not respond outside working hours. Consequently, roughly a third of all enquiries were either unanswered or answered too late. The automated system eliminated that gap entirely.
The conversion rate improvement ā from 18% to 31% ā reflects the compound effect of faster response times and consistent answer quality. Customers who receive an immediate, accurate reply are significantly more likely to proceed with a booking or purchase than customers who wait and find an alternative in the meantime.
Additionally, the system surfaces a continuous stream of insight. Every question that falls through to the AI layer gets logged in Wati’s conversation history. Reviewing these logs monthly identifies patterns ā common questions not yet covered by keyword flows, topics the system prompt should be expanded to cover, and product or service information customers frequently ask about. This creates an iterative improvement loop that makes the bot measurably smarter over time.
šø Image 10 ā Results Metrics Visual File name:
whatsapp-bot-30-day-results-metrics.jpgAlt text: Before and after comparison showing WhatsApp response time dropping from 23 minutes to under 5 seconds and conversion rate rising from 18 to 31 percent after AI bot deployment Caption: 30-day deployment results ā the headline improvement is availability: zero missed after-hours enquiries and a 13-point lift in enquiry conversion rate.
Replicating This System for Different Business Types
The architecture described in this case study is intentionally generic. The only component that changes between business types is the system prompt. The Wati flows, Make.com scenario, and Gemini connection remain identical.
For a medical clinic, the system prompt includes doctor availability, appointment booking instructions, specialisations, insurance information, and emergency protocols ā with an explicit instruction never to provide medical advice.
For a real estate agency, the prompt includes property types, areas covered, viewing request instructions, and contact details for specific agents.
For a retail store, the prompt covers product categories, return policies, delivery options, sizing guides, and stock availability phrasing.
For a professional services firm, the prompt covers service descriptions, consultation booking, pricing range, and intake questions.
In every case, the bot handles the high-volume, repetitive tier of customer communication ā freeing the human team for complex, high-value interactions. This is the core principle behind AI workflow automation applied to customer service: automate the predictable, preserve human attention for the exceptional.
This same pattern extends into broader automation architectures. For teams wanting to go further, our agentic AI in Make.com guide and AI lead intelligence automation case study explore how this foundation connects to CRM systems, lead qualification pipelines, and multi-step AI workflows.
Troubleshooting Common Errors
Even a clean build occasionally encounters issues. The following are the most frequently occurring errors and their direct solutions.
401 Unauthorized on Module 3 This means the Wati API token is missing or incorrectly formatted. Navigate to Wati ā Settings ā API, copy the token, and ensure the Authorization header value reads exactly Bearer [token] with a single space between Bearer and the token string.
Message body shows [object Object] or [Array] This occurs when the wrong output variable is mapped from the Gemini module. Open Module 3, delete the variable in the Request content field, and re-navigate the Module 2 output tree ā clicking deeper into candidates, content, parts ā until reaching a field that shows actual plain text rather than a nested object.
Webhook stays grey ā no data received The Make.com webhook URL was not saved correctly inside the Wati fallback flow. Return to the Wati chatbot flow, confirm the HTTP Request action uses POST method with the correct webhook URL pasted in full, and re-save. Trigger the flow by sending a test message.
Bot replies take 30 seconds or more The Make.com scenario is likely set to poll on a schedule rather than respond instantly. Change the scheduling setting at the bottom of the scenario canvas from “Every 1 minute” to Immediately as data arrives.
Gemini replies with generic or inaccurate information The system prompt is too sparse. Add more specific business details, explicit instructions about what the bot should and should not claim, and example phrasings for common questions. The richer the system prompt, the more accurate and confident the replies.
For a deeper reference on Make.com webhook configuration patterns, the complete webhook tutorial on our blog covers advanced error handling, retry logic, and payload parsing in detail.
šø Image 11 ā Make.com Error History Screenshot File name:
make-com-scenario-error-debugging-guide.jpgAlt text: Make.com scenario execution history panel showing a failed run with a red error indicator on the HTTP module Caption: Make.com’s execution history panel makes debugging straightforward ā click any red bubble to see the exact error, the module it occurred in, and the full request and response data.
Key Technical Decisions and What to Avoid
Several decisions in this build are deliberate and worth understanding. Making different choices introduces problems that are not immediately obvious.
Why Wati over unofficial WhatsApp libraries Open-source libraries like whatsapp-web.js and Baileys connect to WhatsApp by emulating a browser session rather than using the official API. They cost nothing but carry a significant risk: WhatsApp actively detects and permanently bans numbers using unofficial connections. For any production deployment handling real customer relationships, the official WhatsApp Business API through a verified provider like Wati is the only viable choice.
Why Make.com over direct API calls It would be possible to build this integration using serverless functions and direct API calls rather than Make.com. However, Make.com reduces the build time from days to hours, requires no hosting, handles retries and error logging automatically, and makes the integration visually inspectable and maintainable. For a system that needs to be reliable in production, the trade-off is clearly in Make.com’s favour.
Why Gemini over other LLMs for this use case The free tier is the primary practical reason. Additionally, Gemini 2.5 Flash’s multilingual capability, low latency, and 1 million token context window make it well-suited to this workload. For teams already using OpenAI, the HTTP module approach works equally well ā the architecture is model-agnostic, as Make.com’s integration documentation confirms{rel=”nofollow”}.
Why keep the system prompt under 500 words Gemini’s response latency increases with context length. A focused, well-structured system prompt under 500 words produces faster responses and more consistent output quality than an exhaustive prompt attempting to cover every possible scenario. The fallback phrase handles gaps more gracefully than an over-engineered prompt that produces confused, contradictory answers.
Frequently Asked Questions
Does this AI WhatsApp bot require any coding knowledge to build?
No. The entire system uses no-code tools. Wati has a visual drag-and-drop flow builder. Make.com uses a graphical module canvas with point-and-click configuration. The Google Gemini module connects through Make.com’s native integration ā no API calls, no JSON, no code. Anyone comfortable following step-by-step instructions can build and deploy this system.
How many WhatsApp messages per day can the free Gemini tier handle?
Google Gemini 2.5 Flash on the free tier supports 1,500 requests per day and 10 requests per minute. For a business receiving up to 100 customer WhatsApp messages daily, the free tier is sufficient. Above that volume, the paid Gemini tier starts at very low per-token rates that remain far cheaper than any commercial chatbot subscription.
Does the bot support multiple languages without extra configuration?
Yes. Gemini 2.5 Flash detects the language of each incoming message and responds in kind, as directed by the system prompt. No per-language configuration, translation layer, or additional setup is required. This multilingual capability works across over 40 languages natively.
What happens when a customer asks something the bot cannot answer?
The system prompt includes an explicit fallback instruction. When Gemini cannot answer confidently, it uses the defined fallback phrase and signals that a human team member will follow up. Additionally, the conversation in Wati is flagged for human attention, ensuring no customer is left without resolution.
Can this architecture scale to handle multiple business locations or departments?
Yes. Each business location or department can have its own dedicated WhatsApp number, its own Wati chatbot flows, and its own cloned Make.com scenario with a customised system prompt. The infrastructure scales horizontally ā adding a new location or department is a 30-minute configuration task, not a new build. For teams wanting to scale this further into full agentic pipelines, our zero-touch client onboarding system demonstrates how these components connect into enterprise-grade workflows.
Conclusion
This case study describes a three-tool, near-zero-cost system that transforms WhatsApp from a manual communication burden into an intelligent, always-on customer service layer. The architecture is simple by design: Wati handles the WhatsApp interface, Make.com connects the components, and Gemini provides the AI intelligence.
The results speak clearly. Response times drop from minutes to seconds. After-hours enquiries ā previously lost ā are captured and handled automatically. Staff time previously spent on repetitive questions is reclaimed for higher-value work. Conversion rates from enquiry to booking or purchase improve measurably.
More importantly, the system is genuinely accessible. No AI research background, no custom model training, no cloud infrastructure, and no significant budget are required. The tools exist, the free tiers are generous, and the build time is measured in hours.
For teams wanting to go deeper into AI workflow automation or explore how this bot connects to broader automation pipelines, our full workflow automation case studies library covers connected patterns across lead qualification, content automation, and agentic system design. To discuss a custom implementation, contact the Logic Issue team directly.
