WhatsApp is the dominant communication channel in the Netherlands. 13 million users, active daily. Yet most businesses still communicate via email or phone. The WhatsApp Business API changes that, but the implementation is less straightforward than you might expect. We've built six WhatsApp integrations for clients in healthcare, retail, and services. Here's what we learned.
WhatsApp Business App vs. Business API
The free WhatsApp Business App is fine for a sole proprietor. You get a profile, automated greetings, and labels for conversations. But as soon as you want multiple employees responding, want to connect conversations to a CRM, or want to send messages based on a trigger in your system, you need the API.
The API isn't an app. It's an interface that lets you send and receive WhatsApp messages programmatically. You need a Business Solution Provider (BSP) as intermediary: 360dialog, Twilio, or MessageBird are the most commonly used. We work with 360dialog by default due to lower costs and direct Meta integration without a middleman.
What it costs
Meta charges per conversation, not per message. A 24-hour conversation with a customer in the Netherlands currently costs around 0.08 euros for service conversations (customer initiates) and 0.12 euros for marketing conversations (business initiates). On top of that, you pay the BSP: between 50 and 200 euros per month, depending on volume and provider.
For an average SME with 500 conversations per month: budget 150 to 250 euros per month all-in. Compare that to the cost of phone availability: a receptionist easily costs 2,500 to 3,000 euros per month. Even if WhatsApp takes over just 30% of phone traffic, the business case is clear.
The architecture we use
Our standard setup: WhatsApp Business API via 360dialog, connected to n8n as middleware. n8n receives incoming messages via webhook, classifies them (appointment, question, complaint), and routes them to the right system. For a dental practice, an appointment request goes to Calendly. For a webshop, an order question goes to the WMS. For a real estate agent, a viewing request goes to HubSpot.
The advantage of this setup: the logic lives in n8n, not in WhatsApp. You can adjust the flow, add new routes, or refine the classification without touching the API connection. The WhatsApp side stays stable while the business logic evolves.
For message classification, we use a combination of keyword matching and GPT-4. Simple patterns (the word "appointment", "opening hours", "price") are handled directly without AI. More complex messages go through GPT-4 with a compact system prompt that categorizes the message. The AI cost per message is negligible: less than one cent.
The 24-hour window: the biggest pitfall
WhatsApp enforces a strict 24-hour window. Once a customer sends you a message, you have 24 hours to reply freely. After those 24 hours, you can only send template messages, and those need pre-approval from Meta. This sounds like a detail, but it has major implications for your architecture.
Example: a customer sends a question on Monday afternoon. Your bot responds immediately. Tuesday you want to send a follow-up with additional information. That's no longer possible with a free-form message. You have to use an approved template. And those templates can't contain promotional content in the service window.
Our approach: we process all incoming messages immediately. No queue, no "we'll get back to you." The bot responds within seconds, and if it's a complex request, it escalates directly to a human with all context attached. This way you stay within the 24-hour window.
Common mistakes
The three mistakes we encounter most often. One: sending too many automated messages. Meta is strict about spam and blocks numbers that send messages too aggressively. We had a client whose number was blocked for three days because they sent a bulk reminder to 2,000 customers at once. Spread messages out over a few hours.
Two: not building a fallback to a human. Customers stuck in a bot loop become furious. Every flow we build has an escape: if the bot doesn't understand what the customer wants twice, the conversation goes to a human.
Three: forgetting that template messages need approval. This can take 24 to 72 hours. Plan your templates well in advance. And keep them short and business-like, because Meta rejects anything that sounds even slightly promotional in a service template.
