How to Build a WhatsApp AI Agent With n8n

Updated 2026-10-07 · 1356 words

This page may contain affiliate links. If you buy through them we may earn a commission at no extra cost to you.

Direct answer: building a WhatsApp AI agent in n8n means connecting a WhatsApp Trigger node (on the official Cloud API) to an AI Agent node with a connected LLM, one or two tools it can call to actually do something, and a memory node keyed to the customer's phone number. That combination is what separates an agent from a scripted reply bot: it decides for itself which tool to use and holds context across a conversation. Plan on roughly an hour for a working version, most of it spent on the system prompt and testing the agent against messages it shouldn't be able to answer.

Step 1: Get access to the official WhatsApp Business Platform

Before any of the agent logic, you need the same prerequisite as any WhatsApp automation: a Meta Business Account with a verified phone number, registered through the WhatsApp Business Platform (Cloud API). This isn't optional scaffolding; unofficial libraries that puppet a personal WhatsApp number violate Meta's terms and risk a ban with no appeal. We cover this registration step, including the Business Solution Provider alternative, in full in our guide to automating WhatsApp with AI. If you already have that access, skip straight to Step 2.

Step 2: Add the WhatsApp Trigger and credential in n8n

In n8n, create a WhatsApp credential using your Cloud API access token and phone number ID from Meta's developer console, then add a WhatsApp Trigger node to a new workflow. Select your credential and subscribe to the messages field. Save and activate the workflow; like every n8n trigger, it does nothing while left inactive, a detail that trips up more builds than any logic error does.

Send a test message from your own phone to confirm the trigger fires and you can see the payload, including the sender's number at $json.messages[0].from and the message text, before adding any agent logic on top.

Step 3: Add the AI Agent node and connect an LLM

Drop in an AI Agent node downstream of the trigger and connect it to a Chat Model node, OpenAI, Anthropic, or another supported provider, using a credential you've set up. Write the system prompt before anything else: define the agent's role narrowly ("you are a support agent for [business], answering questions about order status, shipping, and store hours only") rather than a generic "you are a helpful assistant." In our testing, prompt specificity was the single biggest factor separating an agent that behaved predictably from one that didn't; this is the same lesson from our broader guide to building an AI agent in n8n, and it applies at least as much on a customer-facing channel like WhatsApp.

Map the agent's text input to the incoming message ({{ $json.messages[0].text.body }}) and its output to a WhatsApp send node that replies to the same sender.

Step 4: Give the agent tools so it can act, not just talk

A WhatsApp agent that only chats is a worse version of a canned-reply bot. Connect tool sub-nodes to the AI Agent's tool input: an HTTP Request Tool to check order status against your store's API, a Vector Store Tool over your FAQ or policy documents, or a Workflow Tool calling another n8n workflow entirely. Keep the list to two or three tools with clearly distinct purposes. We tested an agent with five overlapping tools (order lookup, shipping lookup, two different FAQ sources, and a calendar check) and it misrouted noticeably more questions than a version trimmed back to three; a narrower, well-labeled toolset beat a broader one every time.

Step 5: Add memory, scoped to the phone number

Attach a memory node, Simple Memory for a fast build, or a database-backed option like Postgres Chat Memory if you want conversations to survive a server restart, and set its session key to the sender's number: {{ $json.messages[0].from }}. This is the step people skip because the agent "seems to work" in a single-message test. It doesn't actually hold a conversation until memory is both attached and correctly isolated per customer; without the per-number key, two customers messaging at once can end up sharing context, which reads as the agent confusing one person's order with another's.

Step 6: Respect the 24-hour window for anything the agent initiates

Everything from the general WhatsApp automation rules still applies here. Inside Meta's 24-hour customer service window, the agent can send a free-form reply. Outside it, for example if the agent needs to follow up on an order update two days later, the API rejects a free-form send and the message has to go out as a pre-approved template instead. Build a branch that checks the timestamp of the customer's last inbound message and routes agent-initiated follow-ups to a template-send path rather than the agent's direct reply.

Step 7: Test adversarially before it touches real customers

Run the obvious happy-path question first, then deliberately try to break it: ask something outside its scope, ask it to look up an order that doesn't exist, send a message in a different language than you tested with. This is where a too-permissive system prompt shows up, an agent that confidently answers a billing question it was never meant to handle is a worse outcome than one that says "let me get a person for that." Add a human-handoff path (a Slack or email alert) for anything the agent flags as low-confidence or clearly outside its defined scope, and keep that handoff in place for refunds, complaints, or account changes regardless of how well testing goes.

Agent vs. simpler alternatives

When NOT to build a WhatsApp agent in n8n

Skip this if your WhatsApp volume doesn't justify a conversational agent at all, a handful of daily messages a person already answers in minutes don't need tool-calling AI behind them. It's also the wrong starting point if you haven't yet validated a simpler triage-and-draft pattern on this channel; the plain automation in our WhatsApp guide is faster to ship and easier to debug, and it's worth running that first to confirm the underlying Business Platform setup works before adding agent logic on top. If you'd rather not build and maintain this in n8n at all, a managed option like Lindy or our broader no-code AI agent builder roundup gets to a working agent faster at the cost of some control over tools and logic.

Common pitfalls

The most common failure is an inactive workflow, n8n doesn't auto-activate a saved workflow, so a WhatsApp trigger that looks correctly wired will catch nothing until you flip it on. The second is memory that isn't scoped per number, which surfaces as the agent mixing up customers rather than an obvious error. The third is forgetting the 24-hour window applies to the agent too; an agent that tries to follow up outside the window will have its message silently rejected unless you've branched to a template send.

For the official platform reference, Meta's WhatsApp Business Platform documentation covers Cloud API setup and template approval, and n8n's own docs cover the AI Agent node's tool and memory options in more depth than fits here. If customer service is the agent's main job, our roundup of the best AI agents for customer service is worth a look for comparison against a purpose-built platform rather than a custom n8n build.

Get the Business Platform access and a plain triage workflow working first, then layer in the AI Agent node, tools, and memory once you know the underlying channel is reliable. Building the agent logic on top of an unverified trigger just means debugging two problems at once.

Save this to Pinterest