n8n and WhatsApp: how to automate it, step by step.

A booking workflow, node by node

A restaurant wants WhatsApp bookings without staff typing every reply. The workflow below is the shape of it; the booking check runs against whatever system the hosts already use.

1. WhatsApp Trigger (Messages). A guest writes “Table for four on Friday at 8?”.
2. Switch. Text goes to the agent; a button reply goes straight to the matching branch; a voice note goes through a media download first.
3. AI Agent with tools. It reads the request, checks availability through an HTTP Request to the booking system and drafts the answer. It may only act through those tools.
4. WhatsApp Business Cloud — Send. The reply goes back inside the customer service window: “Friday 8 pm is free for four. Shall I book it?”.
5. Send and Wait for Response. For a large group, the workflow asks the manager to approve before confirming, and resumes on their answer or when the time limit runs out.
6. WhatsApp Business Cloud — Send Template. The reminder the day before is a utility template, since the window may have closed by then (how templates work).

The 24-hour window

When a customer messages or calls the business, Meta opens a 24-hour customer service window, reset by each new message or call. While it is open, the workflow can send any free-form message: text, media, buttons. Once it closes, only a pre-approved template can be sent. Meta also notes a known issue: in rare cases a message arrives but the business cannot answer it within the window.

In n8n, this means a branch that sends plain text after a delay — a follow-up, a reminder, a staff approval request to someone who has not written recently — will fail unless it uses Send Template. Bots also have to follow Meta's terms and offer a way to reach a person; our article on WhatsApp bots covers both.

Pitfalls that break n8n WhatsApp flows

• “Bad request — please check your parameters” on Send Template. n8n's docs point to the template parameters: their type and their order. Map each variable explicitly with an Edit Fields (Set) node and give it a fallback value.
• Media operations failing. In “Input Data Field Name” or “Property Name”, write the field name itself, such as data, not an expression.
• A workflow that stops after a few hours. Check the token: Meta says user tokens from API Setup expire quickly and must be regenerated every few hours. Use the system user token from step 2.
• Silence after a webhook fix. Do not send a DELETE to subscribed_apps to reset the webhook: n8n warns it unsubscribes your app from the account and stops all delivery.
• Test and production fighting. One webhook per app: only one of the two URLs receives events at a time.

When a workflow tool is enough, and when it is not

n8n is a good fit for linear jobs: route a message, look something up, send a template, notify a team. It gets harder when the conversation itself is the product — memory across days, many tools, testing every change before customers see it, an audit of what the agent did and why. At that point, the agent can move into code while n8n keeps the jobs around it.

The agents we build for clients — bookings for a restaurant group in South Africa, voice-note bookings for gyms in East Africa — run on the business's own number on the official platform, act only through tools we write and hand over to staff. See WhatsApp AI agents, or the AI agents and automation we build.