The AI Agent That Answers Your WhatsApp While You Sleep (No Code Required)

Customers message at 11pm. They message on Sunday. They message in the gap between "I'm interested" and "I bought from someone who answered." Here's how to put an AI agent on your WhatsApp Business number that handles the first reply around the clock—with no code, and with a human still in charge of anything that matters.

First, the Rules WhatsApp Sets

WhatsApp is not an open inbox, and the agent has to respect that. Meta's WhatsApp Business Platform works on a customer service window: per Meta's pricing documentation, a 24-hour window opens when a WhatsApp user messages a business, and within it "all non-template messages are free"; once it closes, a business "can no longer send non-template messages" and must use approved message templates, which Meta charges per message in categories such as marketing, utility and authentication. Meta's documentation dates the per-message pricing model to July 1, 2025.

The practical consequence: an agent that replies to inbound messages inside the window is on the cheapest, simplest path. An agent that initiates conversations needs templates and a budget. Build the first kind first.

What You're Building

A workflow with four nodes:

1. A Chat Message trigger listening on WhatsApp.
2. An AI Agent node with your support Context attached and a short allow-list of tools.
3. A Condition that decides whether the agent's answer can go out directly or needs a person.
4. A Chat Interface node that sends the reply back on WhatsApp, plus a Slack node that posts escalations to your team.

Runs execute on MultiverseAI's server, so the workflow is awake when you aren't.

Step 1: Connect WhatsApp

In the Connections tab, link your Meta business account; WhatsApp Business is one of the 50 integrations and uses your own account, not a shared platform number. Inbound messages to your number are what will trigger the workflow.

Step 2: Write the Context (This Is the Whole Job)

The agent is only as good as what you tell it. Create a Context with:

• What you sell and what you don't. Products, prices if they're public, service area, opening hours.
• The answers to the ten questions you get most. Delivery times, returns, payment methods, where to find the order number.
• What it must never do. Promise a refund, quote a custom price, confirm an appointment, give medical or legal advice. Tell it to hand those to a person.
• Tone. Short messages, one question at a time, no walls of text—WhatsApp is a phone screen.

Keep it under a page. A focused Context beats a sprawling one.

Step 3: Build the Graph

Open the Workflow Builder and either place the four nodes by hand or type the plan into the AI Workflow Builder: "When a WhatsApp message arrives, have the support agent draft a reply using the Support Context; if the reply mentions refunds, prices or appointments, post it to #support-queue in Slack for approval instead of sending; otherwise send it back on WhatsApp." It drafts the graph; you check the wiring and save.

On the AI Agent node, choose a fast model for the default path (replies should arrive in seconds) and give it only the tools it needs—usually a knowledge search over your Context and, if you track orders in a sheet, Google Sheets read access. The 88-tool toolkit is there, but a support agent that can post to Threads is a liability, not a feature.

Step 4: Put the Approval Gate in the Right Place

The Condition node is your safety line. Route to approval when the draft contains money, dates, or anything the Context told the agent to escalate. Approved replies are sent by the Chat Interface node; rejected ones are answered by the human directly in Slack, and the agent learns nothing it shouldn't. Start with the gate catching more than you think you need. Loosen it after a week of reading the logs.

Step 5: Test on Your Own Phone

Message the business number from your personal WhatsApp. Ask the easy questions, then the trick ones ("can I get 20% off if I order ten?"). Read the run log for each message—every tool call and every charge is listed—and adjust the Context, not the prompt, when it gets something wrong. Each run costs a few credits (1 credit = $0.01, billed at provider cost plus 50%), so a long test session is cheap.

Step 6: Switch It On, Then Watch the First Week

Enable the workflow and keep the Slack escalation channel open for a week. You'll see three kinds of messages: ones the agent handled well, ones it escalated correctly, and ones it should have escalated. The third kind writes your next Context edit. After the first week, most teams find the escalation channel quiet enough that the human on duty can be the person who's already awake.

What Not to Automate Yet

Don't let the agent send first messages—template pricing and consent rules make outbound a different project. Don't let it confirm anything that costs money without a person approving. And don't skip the "I'm an automated assistant" line in the opening reply; customers accept an honest bot far better than a confusing one.

Sleep

The agent doesn't replace your support team. It replaces the 11pm gap where the customer gave up. Start free, connect the number, write the Context tonight, and let the first reply go out while you're not looking.

Sources

1. Meta for Developers, "Pricing — WhatsApp Business Platform" — https://developers.facebook.com/docs/whatsapp/pricing