Create a WhatsApp AI Agent: Step-by-Step Tutorial 2026

A WhatsApp AI agent answers, qualifies, and converts your prospects 24/7. Here's how to set it up step by step, along with the tools and limitations you need to know.

Why a WhatsApp AI Agent in 2026

WhatsApp now has over 2.7 billion active users worldwide, including more than 30 million in France. It's no longer just a personal messaging app — it has become consumers' preferred communication channel for interacting with businesses.

The numbers speak for themselves:

73% of consumers prefer sending a message over calling customer service (Meta Business, 2025).
The open rate for WhatsApp messages reaches 98%, compared to an average of 20% for emails.
The response rate on WhatsApp sits between 45 and 60%, while a cold email struggles to exceed 5%.
Conversations initiated via WhatsApp convert 3 to 5 times more than those started by email or web form.

For an SMB, a tradesperson, or an e-commerce merchant, this means every prospect who contacts you on WhatsApp represents a warm opportunity. But if no one responds within 15 minutes, that opportunity goes cold. That's exactly where a WhatsApp AI agent changes the game: it responds instantly, around the clock, without tiring, without forgetting, and qualifies each contact before passing it along to you.

WhatsApp Agent vs WhatsApp Chatbot — the difference

These two terms are often confused, but the difference is fundamental. A WhatsApp chatbot follows a rigid decision tree: the user chooses from predefined buttons, and the bot returns a static response. If the prospect steps off the expected path, the chatbot gets stuck or sends back an "I didn't understand."

A WhatsApp AI agent, on the other hand, understands natural language. It interprets intent, rephrases, asks contextual questions, and adapts in real time thanks to a language model (LLM). It doesn't just respond — it converses.

Criterion	WhatsApp Chatbot	WhatsApp AI Agent
Logic	Fixed decision tree	LLM + dynamic scenarios
Understanding	Exact keywords	Natural language, intent
Adaptability	Low	High
Personalization	Static templates	Contextual, real-time
Qualification	Predefined buttons	Tailored open-ended questions
Maintenance cost	Low initially, high at scale	Moderate, but scalable

For an in-depth analysis of this distinction, check our article AI Agent vs Chatbot.

Technical architecture of a WhatsApp AI agent

The components

A functional WhatsApp AI agent relies on four technical building blocks:

WhatsApp Business API — This is the official gateway. Without it, it's impossible to receive or send messages in an automated way. The API is accessible through a Business Solution Provider (BSP) like Twilio, 360dialog, or directly via Meta. The unofficial API exists but violates WhatsApp's terms of service and risks having the number banned.
LLM (Large Language Model) — This is the agent's brain. OpenAI (GPT-4o, GPT-4.1), Anthropic (Claude), or Mistral provide models capable of understanding and generating natural language in French. The LLM receives the user's message, the conversation context, and the system instructions (role, tone, objectives), then generates an appropriate response.
Workflow engine — This is the orchestrator. It connects WhatsApp to the LLM, manages conditions, loops, external API calls, and transitions between scenarios. Our n8n vs Make vs Zapier comparison details the options, but in short: n8n is the most suitable tool for this use case thanks to its flexibility, possible local hosting, and native webhooks.
CRM — This is the memory. HubSpot, Pipedrive, Airtable, or even a Google Sheet — the agent must be able to write the collected data (name, budget, need, status) somewhere. Without a CRM, every conversation starts from scratch.

The conversation flow

Here's how data flows between these components:

[User sends a message on WhatsApp]
        |
        v
[WhatsApp Business API -> Outgoing Webhook]
        |
        v
[n8n receives the webhook -> extracts message + number + timestamp]
        |
        v
[n8n sends to the LLM (OpenAI/Anthropic) with system prompt + history]
        |
        v
[LLM generates the response -> n8n receives it]
        |
        v
[n8n sends the response via WhatsApp Business API]
        |
        v
[n8n writes collected data to the CRM]
        |
        v
[User receives the response on WhatsApp]

Each cycle — from receiving the message to sending the response — must execute in under 5 seconds to maintain a natural conversation.

Step-by-step tutorial

Step 1 — Get WhatsApp Business API access

Create a Meta Business account on business.facebook.com.
Go to WhatsApp Manager and create a WhatsApp Business profile.
Verify your phone number (it must not already be registered on the WhatsApp Messenger app).
Choose a Business Solution Provider. 360dialog is a popular option in Europe for its simple integration with n8n. Twilio also works but is more expensive per conversation.
Retrieve your API Key and Phone Number ID — you'll need them to configure the webhook.

Important: signing up for the Business API requires Meta's approval. The process typically takes 24 to 72 hours. Your business must legally exist.

Step 2 — Configure the webhook

In n8n (self-hosted or cloud version):

Create a new workflow.
Add a Webhook node as a trigger. Configure it with the POST method.
Copy the webhook URL generated by n8n.
In your BSP (360dialog, Twilio), paste this URL as the receiving webhook.
Add a Respond to Webhook node configured as 200 OK — WhatsApp requires a quick HTTP response to confirm receipt.

To verify that the webhook works, send a test message to your WhatsApp Business number. The message should appear in the execution data of the n8n Webhook node.

Step 3 — Connect the LLM (n8n + OpenAI/Anthropic)

Add an OpenAI or HTTP Request node (for Anthropic) in your n8n workflow:

Recommended model: GPT-4o-mini for simple conversations (low cost, short latency) or GPT-4.1 for complex scenarios requiring deep reasoning.
System prompt: This is the centerpiece. Here's an example:

You are the virtual agent of [Company Name], a real estate agency in Lyon.
Your role: qualify each prospect in 3-5 messages max.
Information to collect: budget, type of property sought, timeline, geographic area.
Rule: if the budget is below 150,000 EUR, politely indicate that we do not cover that range.
Always offer to schedule an appointment with an advisor at the end of qualification.
Tone: professional, warm, concise. No technical jargon.

History: For the agent to remember previous messages in the same conversation, pass the last 5 to 10 messages as context in the API call. n8n can store this history in a Redis node or in a database (PostgreSQL, Supabase).

Check out AI Automation Complete Guide for advanced context management patterns.

Step 4 — Create conversation scenarios

A good agent doesn't just respond — it follows structured scenarios. In n8n, use Switch or IF nodes to create branches:

Welcome scenario: "Hello! How can I help you?" -> Intent analysis.
Qualification scenario: Series of contextual questions tailored to the industry.
Appointment scheduling scenario: Propose time slots, confirm, send a Calendly link.
Support scenario: Answer frequently asked questions, escalate to a human if necessary.
Follow-up scenario: Automatic follow-up after an unanswered quote (via approved WhatsApp template).

Each scenario should have a clear exit condition: the prospect is qualified, the appointment is booked, or the agent must transfer to a human.

Step 5 — Integrate the CRM

At each key step in the scenario, the agent must write to the CRM:

After qualification: Create a contact with the collected information (name, budget, need).
After appointment scheduling: Create an opportunity/deal with the appointment date.
After escalation: Tag the contact as "to call back" and assign to the right advisor.

In n8n, add a HubSpot, Pipedrive, or HTTP Request node (for Airtable, Supabase, etc.) after each qualification action. Map the fields extracted by the LLM to the CRM properties.

For SMBs without a formal CRM, a Google Sheets node can serve as a temporary database — but this solution quickly reaches its limits in terms of collaboration and automation.

Step 6 — Test and deploy

Before putting the agent into production:

Test scenarios in a closed loop: Send messages from a personal number and verify each branch of the workflow.
Test edge cases: Empty messages, emojis only, voice messages, foreign languages, off-topic questions.
Measure latency: The time between the user sending the message and receiving the response must not exceed 8 seconds. Beyond that, the user thinks no one is responding.
Enable logging: Every conversation should be recorded (without sensitive personal data) for later analysis.
Set up a human fallback: If the agent can't answer after 2 attempts, transfer the conversation to a human operator with a notification (Slack, email, or CRM notification).

Then deploy gradually: start with off-hours (evenings and weekends), then extend to business hours once confidence is established.

5 concrete use cases

Real estate lead qualification

A Lyon-based real estate agency serves as our case study. Before the agent, WhatsApp prospects received an automated acknowledgment and sometimes waited 24 hours for a callback. The response rate was 12% — most leads had already contacted another agency in the meantime.

After deploying a WhatsApp AI agent that asked 3 qualification questions (budget, property type, area) and directly offered a viewing slot, the response rate climbed to 47%. The agent qualified the lead, scheduled the appointment in the CRM, and sent a confirmation — all in under 2 minutes.

This anecdote best illustrates the power of an agent compared to a simple auto-responder: the ability to act, not just react.

Automated appointment scheduling

For independent professionals (osteopaths, coaches, consultants), the WhatsApp AI agent replaces the front desk. The patient sends a message, the agent proposes available slots via the Calendly API or directly through Google Calendar, confirms the appointment, and sends a reminder 24 hours before.

The average gain observed: 3 to 5 hours per week less phone management, and a no-show rate reduced by 30% thanks to WhatsApp reminders.

E-commerce customer support

An e-commerce merchant handled 200 requests per day on WhatsApp — returns, order tracking, product questions. The AI agent now manages 70% of these requests in full autonomy: querying the Shopify API for order status, sending return links, answering questions about sizing and delivery times.

The remaining 30% (complex complaints, exceptional refund requests) are transferred to a human with the full conversation context.

Order tracking

The agent queries the carrier's API (Colissimo, UPS, DHL) using the order number and responds in natural language: "Your package is currently at the Lyon sorting center. Delivery is expected tomorrow between 2 PM and 5 PM."

It's simple, but it eliminates 90% of inbound order tracking calls.

Quote follow-up

After sending a quote, the WhatsApp AI agent can automatically follow up with the prospect on D+2 and D+5 with personalized messages. The first message asks if the client has any questions. The second offers a call slot.

Note: these follow-up messages must use approved WhatsApp templates if 24 hours have elapsed (see the section on limitations below).

Limitations to know

WhatsApp restrictions (24-hour window, templates)

This is the most important and most often underestimated constraint. WhatsApp imposes a 24-hour conversation window: once a user sends you a message, you can reply freely (free-form messages) for 24 hours. After that period, you can only send template messages — messages pre-approved by Meta in a fixed format.

In practice, this means your agent must capture intent and qualify the prospect within this 24-hour window. If the prospect sends a message at 10 PM, your agent doesn't respond (or responds poorly), and the prospect returns at 10 AM the next day... the window is closed. You can only send them a template.

Templates must be submitted to Meta for approval (24 to 48 hour turnaround). They cannot contain dynamic content beyond a few predefined variables (name, date, amount).

Per-conversation costs

The WhatsApp Business API charges per conversation (not per message). In 2026, Meta's rates for the Europe zone are approximately:

User-initiated conversation: 0.03 EUR per 24-hour conversation.
Business-initiated conversation (template): 0.06 EUR to 0.10 EUR depending on the template category.
The first 1,000 conversations per month are free.

On top of this come LLM costs (approximately 0.01 to 0.05 EUR per conversation depending on length), n8n hosting costs, and BSP fees. For an SMB handling 500 conversations per month, the total budget sits between 30 and 80 EUR/month.

What the agent cannot do

A WhatsApp AI agent, however capable, has intrinsic limitations:

It cannot process payments directly within WhatsApp (payment buttons are limited to certain countries and merchants).
It cannot send large files — the limit is 100 MB per media.
It doesn't replace an advisor for complex decisions (price negotiation, legal advice, technical diagnostics). Its role is to qualify, not to close.
It doesn't handle prolonged ambiguity well — if the prospect changes topics three times, the LLM may lose track.
It cannot initiate a conversation with a user who has never contacted you (except via an approved template, and only if the user has given consent).

FAQ

How much does a WhatsApp AI agent cost per month?

For an SMB handling between 200 and 1,000 monthly conversations, the total cost (WhatsApp API + LLM + n8n hosting + BSP) ranges from 30 to 150 EUR/month. This is significantly less than a part-time human agent, and the AI agent is available 24/7.

Can I use personal WhatsApp instead of the Business API?

No. The unofficial API (via undocumented libraries like whatsapp-web.js) violates WhatsApp's terms of service. Your number risks being permanently banned. The Business API is the only legal and sustainable path for automating WhatsApp.

Does a WhatsApp AI agent work in French?

Yes. Modern LLM models (GPT-4o, Claude, Mistral) master French with near-native quality. Simply write the system prompt in French and specify the response language. Mistral, a French model, actually offers an excellent quality-to-price ratio for this use case.

What happens after the 24-hour window?

After 24 hours without an exchange, you can only contact the user with a template message approved by Meta. These templates are limited in customization and must comply with WhatsApp policies (no aggressive marketing, no misleading content). This is why qualification must be fast and complete.

How do I transfer the conversation to a human?

Configure a transfer keyword in your n8n workflow (for example, if the LLM detects an escalation intent, or if the prospect explicitly asks to speak to someone). The workflow then sends a Slack/Teams notification to the team, tags the contact in the CRM, and disables automatic responses for that number for 24 hours.

Update — April 2026

Several recent developments impact the implementation of a WhatsApp AI agent:

Meta expanded template categories in February 2026, adding "service" and "order update" templates with reduced rates (0.03 EUR instead of 0.06 EUR).
WhatsApp Interactive Flows are now available in Europe. They let you display native forms within the app, combining the flexibility of an LLM with the structure of a form — ideal for lead qualification.
Mistral Large 3 offers a competitive European alternative to GPT-4o for French-language conversations, with a per-token cost approximately 60% lower.
n8n introduced native AI Agent Nodes in version 1.60+, significantly simplifying the configuration of conversational workflows without custom code.

Conclusion

A WhatsApp AI agent is no longer an experimental luxury — it's an operational asset accessible to any SMB in 2026. The combination of WhatsApp Business API + LLM + n8n + CRM forms a mature, proven, and affordable stack. The results are concrete: response rate multiplied by 3 to 4, automated appointment scheduling, customer support relieved of 70% of its workload.

But implementation requires rigor. The 24-hour window demands fast qualification. Meta templates require upfront planning. And an agent that's too fast, too mechanical, loses the interlocutor's trust.

One often-overlooked detail: a WhatsApp agent that responds too quickly loses credibility. A/B tests conducted across several deployments show that adding a 2 to 3 second delay with the "typing..." indicator increases perceived trust and conversion rates by 18%. The human user needs to feel they're conversing with an entity that "thinks" — even if the response is generated in 800 milliseconds.

The WhatsApp AI agent is a powerful tool. Like any tool, its effectiveness depends on the quality of its configuration, the clarity of its scenarios, and the relevance of the system prompt that guides it. Start with a simple scenario (welcome + qualification + appointment scheduling), measure the results, then iterate.