Picked from the software, topic, and level of “Agentic AI Without Coding Class #40: Practice: Connecting WhatsApp and Running 24-Hour Customer Service”.
At most one email a week. Unfollow anytime with one click.
Transcript
The full narration of this video. The current sentence is highlighted while the video plays; click any sentence to jump there.
Show transcript
Introduction
Hi, and welcome to the closing episode of Session 8 of Agentic AI Without Coding. The Sunset Coffee agent is already serving smoothly on Telegram. Now we move it to where most customers are: WhatsApp, through the official route. Then we run it as 24-hour customer service, complete with escalation and monitoring.
Step 1: Setting up the official route: WhatsApp Cloud API
We start with the official route we chose in Episode 36: the WhatsApp Business Platform, through the Cloud API directly from Meta. The first step is a Meta Business account for Sunset Coffee, with the business details that will be verified later. Second, open Meta for Developers and create a new business-type app. This app is the container that connects WhatsApp to other systems. Third, add the WhatsApp product to that app. Meta immediately gives you a free test number. This test number can only send messages to a few recipient numbers you register, which is enough for trying it out. Fourth, to serve real customers, add your business number and complete its verification. Remember, a number used with the API generally can't also be used in the regular WhatsApp app at the same time. Fifth, create a permanent access token through the system user feature in the business settings. The temporary token on the starting page only lasts about a day, so don't use it for a twenty-four-hour service. Meta's menu names change often, so follow the latest official guide if the screens look different. From the WhatsApp settings page in the Meta app, note five items. First, the Phone Number ID, the identity of the sending number. It's not the phone number itself, but a long numeric code. Second, the WhatsApp Business Account ID, the identity of your WhatsApp business account. Third, the App ID and App Secret from the app's basic settings. n8n uses both to register the receiver for incoming messages. Fourth, the permanent access token. Like the Telegram token, it's secret. Never paste it in a group or a shared document. Fifth, while you're still using the test number, register your own phone number as a test recipient. All of this data is stored only in the n8n Credentials menu. Before moving on, let's be honest: the Cloud API setup is longer than Telegram's. If it feels like too much, there are other routes. First, through an official provider or BSP, which helps with sign-up and provides a team dashboard. There's a subscription fee, and make sure the provider is truly an official Meta partner. Second, unofficial services based on WhatsApp Web. As we covered in Episode 36, these are only appropriate for learning on a separate number. The number can be blocked, and customer data passes through someone else. Never use them for your main business number. Third, stay on Telegram for now while you prepare the Meta verification. That makes sense if your customers are willing to move there, even though the reach is smaller. Fourth, the WhatsApp Business App stays useful as a manual backup on another number, for times when the automated system has problems. In this episode we continue with the recommended route: the official Cloud API.
Step 2: Connecting WhatsApp to the agent in n8n
Switch to n8n. Duplicate the Telegram workflow from Episode 39 and name it Sunset Coffee WhatsApp CS Agent. Delete the Telegram Trigger, then add a WhatsApp Trigger node. Create a new credential for this trigger, with the App ID and App Secret you noted earlier. For Trigger On, choose Messages, meaning this node runs every time there's a message event on Sunset Coffee's WhatsApp number. Just like with Telegram, the webhook address is registered automatically to the Meta app. Click the test button, then send a message from your phone to the test number. The test message was caught. The data is similar to Telegram, just with different names. The sender's profile name is in contacts profile name, here it's Joko. The sender's WhatsApp number is in messages from. On WhatsApp, this number replaces the chat id: for replying, and as the memory key. The message itself is in messages text body. And notice the type field, which says text. This field matters for the next step. Now let's build the flow. After the WhatsApp Trigger, add an If node. Why? Because WhatsApp also sends other events to the webhook, such as sent and read message statuses. These events don't contain a customer message. The If condition is simple: messages type equals text. Only text messages are passed to the agent through the yes branch. Without this node, the agent might try to answer a blue check mark. The yes branch connects to the same AI Agent. In the prompt settings, change the expression to messages text body. Then click the memory and change its key to messages from, the sender's number. Every customer still has their own memory drawer. Finally, replace the Telegram sending node with a WhatsApp Business Cloud node, Send action. Look at the right panel. Sender Phone Number ID is set to Sunset Coffee's number. Recipient uses the messages from expression from the WhatsApp Trigger, meaning reply to the number that sent the message. Message type is Text, and Text Body uses the agent's output. The pattern is exactly the same as Telegram.
Step 3: Escalation, operating hours, and the log
This is the complete workflow. The top path: the WhatsApp Trigger, If, AI Agent, and the WhatsApp sender. Below the agent, the AI model and the Sheets tools from Session 7 stay the same. The Escalation tool from Episode 39 is reused too. Notifications still go to Dimas through Dimas's personal Telegram bot. One notification door for every channel, so Dimas only has to watch one place. A new tool: Write log. This is a Google Sheets node that adds a row to the Conversation Log sheet, with columns the agent fills in: name, channel, category, summary, and status. And memory per sender number. Nine nodes, without a single line of code. Let's compare with Telegram, so it's clear what actually changes. The trigger becomes the WhatsApp Trigger, plus an If node to filter text messages. The message text comes from messages text body, not message text. The customer identity uses the sender's WhatsApp number, not the chat id. The reply sender is the WhatsApp Business Cloud node. And there's an extra rule: the twenty-four-hour service window. Because in customer service the customer always starts the chat, the agent's replies always fall inside the window. Only messages you start yourself need a template, for example an order update the next day. Everything else, the agent's brain, tools, instructions, escalation, and ethics, stays the same. The agent's instructions are also updated a little. The context block names the WhatsApp channel, the customer's profile name, and the current time, filled in automatically by n8n. Without the current time, the agent doesn't know whether the shop is open. The task block adds a new duty: after a conversation is finished or escalated, the agent uses the Write log tool with those five columns. The first limits block covers the night: orders are still logged and the agent says they'll be processed tomorrow at eight, and escalations get the status waiting for tomorrow. The second limits block keeps the ethics from Episode 38: the agent never sends promotions or starts a chat, it only replies. And the format block stays the same: short, polite, no tables, and the first message always introduces the bot as an automated assistant. One more thing. Our agent only replies. But sometimes Sunset Coffee needs to start a chat, for example to say an order is ready for delivery the next day. Outside the twenty-four-hour window, a message like this must use a template approved by Meta. It carries information the customer is actually waiting for: the order number, the delivery time, and the admin's contact. Not a promotional pitch. Remember, templates can be paid, with different rates for each type. Always check the latest official pricing page before turning it on. And it's best built as a separate workflow triggered by a status change in the Orders sheet, rather than left to the agent's judgment. For things with fixed rules, a regular automation is safer, exactly the principle from Episode 1.
Step 4: Testing day and night service, then activating
The first test, during the day. Joko asks on WhatsApp: can I order cold brew? The agent introduces itself as an automated assistant, gives the price and stock from Sheets, then shows how to ask for the admin. Joko orders four bottles, delivered to the office at three in the afternoon. The agent calculates two hundred twenty thousand rupiah, and because it's above two hundred thousand rupiah, delivery is free under the policy. After Joko confirms, the order is logged in the sheet. Same agent, new door. The second test, at ten past eleven at night, when the shop is closed and Dimas is already resting. Putri asks whether one hamper can be delivered tomorrow morning. The agent still answers with the price and delivery fee, and is honest that the order will be processed tomorrow from eight. The customer doesn't have to wait until morning just to know the price. Then Putri raises a complaint: yesterday's coffee tasted sour. This is a complaint, so escalation is required. The agent apologizes and makes an honest promise: Dimas will reply tomorrow from eight. Not right away, because nobody is on duty tonight. The notification reaches Dimas's personal Telegram bot that same night, complete with the time promise already given to Putri and its status: waiting for tomorrow. The next morning, at three minutes past eight, Dimas starts the day with this case. The agent's promise is kept, and Putri is contacted on time. Before opening it to customers, check five things. First, every test, by day and by night, has passed using Meta's test number. Also register Laras's and Dimas's phone numbers as test recipients, so they can try it as customers too. Second, the credential uses a permanent token, not the temporary one that expires within a day. This is the most common reason a WhatsApp bot suddenly goes quiet the next day. Third, the business number is added and verified, and the Phone Number ID in the sending node is switched to that number. Fourth, the workflow is activated, and the notification to Dimas has been tested and arrives. Fifth, Laras and Dimas know how to turn off the bot, simply by deactivating the workflow, if it ever starts giving strange answers. This emergency handle puts everyone more at ease.
Step 5: Monitoring: conversation logs and routines
A twenty-four-hour service doesn't mean leaving it unsupervised. This is the Conversation Log sheet after one day of running. Joko's order and Rina's question from Telegram were handled fully by the bot. Budi's transfer issue was resolved by Dimas. Its status is done by Dimas. At night, Putri's order was logged by the bot, and the complaint has the status waiting for tomorrow. Early in the morning, Sari asked about delivery to Jakarta, which isn't in the data. The agent didn't guess; it escalated. Good. And at three minutes past eight, Dimas handled Putri's complaint and added the closing row. Look at the formula bar: Dimas's morning routine is just checking the rows with the status escalated and waiting for tomorrow. From the log, Laras can calculate the impact. Here's an illustration of the first week: about two hundred ten chats came in from WhatsApp and Telegram. About eighty percent were answered fully by the bot, without Dimas touching the phone. About thirty were escalated to Dimas. That number is reasonable, and it actually shows the agent knows its limits. And the bot replies within seconds, day or night. These numbers are an illustration. Count from your own log sheet, because in Session 10 we use them to measure the impact of your project. To stay healthy, Sunset Coffee has a monitoring routine. Daily, Dimas spends about ten minutes clearing every row with the status escalated or waiting for tomorrow. Also daily, open the execution history in n8n. A failed execution means a customer didn't get a reply, for example because the token expired. Weekly, Laras reads twenty random conversations from the log. Not to find Dimas's mistakes, but to judge the agent's tone and accuracy. Also weekly, update the stock and prices in the Products sheet. The agent is only as accurate as the data you give it. Then fix the instructions based on odd answers, exactly the ethics cycle from Episode 38. And monthly, check the Meta and AI model bills on their official pages, so costs stay under control. We go deeper into calculating subscription costs in Session 9.
Step 6: Session 8 output: automatic 24-hour customer service
Let's look back at the Session 8 journey. In Episode 36 we chose our doors: official WhatsApp, Instagram, and Telegram, and understood the risks of unofficial routes. In Episode 37 we designed when the bot answers and when Dimas steps in, including operating hours. In Episode 38 we set the ethics: be honest about the bot, no spam, protect privacy, and no promises outside policy. In Episode 39 we learned without risk on Telegram. And today, the same agent serves on WhatsApp through the official route, twenty-four hours a day. This is the Session 8 output you should have. First, an agent that answers customers on official WhatsApp, or at least on Telegram, twenty-four hours a day. Second, stock and price answers come from the data, and orders are logged automatically. Third, escalation to a human, complete with notifications and honest time promises. Fourth, the agent's instructions include ethics rules: honesty, no spam, and protecting privacy. Fifth, a conversation log and a monitoring routine. If you have all five of these for your own business, congratulations. You're now running automatic twenty-four-hour customer service. Imagine the difference for Laras. Before, night chats piled up until morning. Now customers get answers within seconds, orders are logged, and Dimas is only called for things that truly need a human. Keep your workflow, instructions, and log safe. We use them all again in Session 9, when we cover quality and security, and in the Session 10 final project.
Done!
Let's recap. We set up the official WhatsApp Cloud API, connected it to the same agent in n8n, added escalation, operating hours, and logging, tested it by day and by night, and then put together a monitoring routine. That's the Session 8 output: automatic twenty-four-hour customer service. In the next episode we start Session 9 and begin building a team of agents: a simple multi-agent setup with a researcher, a writer, and a reviewer working together. See you in Episode 41.
Practice exercises
Try these on your own after watching to lock in what you learned.
Connect your business's agent to the WhatsApp Cloud API using a test number, then test it with 10 questions from your list.
Create a Conversation Log sheet and follow the daily monitoring routine for one week. Note what percentage of chats the bot answered and how many were escalated.
Write one instruction improvement based on the first week's log, then retest with the questions that weren't answered well before.
Ask the class material
Type a question or keyword; we find the part of this episode that explains it and jump the video there.
My notes
Notes are saved with the current video time and are only visible to you.
Questions & answers
Stuck on a step? Ask here; the instructor and other learners can answer.
No questions yet.
Your instructor
Cerdas Inovasi Digital Team
Software developers, BIM & 3D engineers, and AI practitioners
53 video classes1,284 episodes
We build software, BIM models, games and AI automations for clients every day, and turn that hands-on work into step-by-step classes: every step is shown on screen and comes with a PDF guide you can keep.
