Build a WhatsApp Customer Service Bot That Handles 80% of Inquiries

The Problem

An e-commerce company with 5,000 monthly orders gets 200+ WhatsApp messages daily from customers. Three support agents spend their entire day answering the same questions: "Where is my order?" (45%), "How do I return this?" (20%), "Do you have this in stock?" (15%), and "What are your business hours?" (10%). Only 10% of inquiries actually require a human being.

The agents are burning out. Response times average 4 hours because each agent is juggling 70+ conversations. Customers who do not get a quick reply abandon their carts — the company estimates 15% customer churn is directly tied to slow support. And every "where's my order?" message that a human answers is four minutes wasted on a lookup that could be instant.

The Solution

Using the whatsapp-bot-builder, coding-agent, and template-engine skills, the agent builds a WhatsApp Cloud API bot with interactive menus for order tracking, a multi-step returns flow, keyword-based FAQ matching, intelligent human handoff with queue management, and proactive notification templates for order updates.

Step-by-Step Walkthrough

Step 1: Set Up the WhatsApp Cloud API and Welcome Flow

Set up a WhatsApp Business API bot using the Meta Cloud API. When a customer
sends any message, respond with a welcome message and an interactive button
menu:

"Hi! I'm the assistant for [store]. How can I help?"
Buttons: [Track Order] [Returns] [Talk to Agent]

If they haven't messaged in 24+ hours (outside the reply window), send
the approved "welcome_back" template message instead. Store conversation
state in Redis with a 24-hour TTL matching the WhatsApp conversation window.

The bot scaffolds a Node.js Express server with Meta's webhook verification, message routing based on conversation state stored in Redis, and interactive button responses. The 24-hour conversation window is the key constraint with WhatsApp Business API — outside that window, only pre-approved template messages can be sent. The bot detects expired windows and falls back to the welcome_back template automatically.

Conversation state in Redis uses a 24-hour TTL that matches the WhatsApp window exactly. When a customer taps "Track Order," the state machine knows they are mid-flow and routes follow-up messages to the order tracking handler instead of showing the welcome menu again.

Step 2: Build the Order Tracking Flow

When the customer taps "Track Order", start a conversation flow:
1. Ask for order number or email address
2. Look up the order in the database (query by order number or last order by email)
3. Return: order status, items, tracking number with carrier link, estimated delivery
4. If order is "shipped", show a button: [Track on Map]
5. If order is "processing", show: [Notify When Shipped]
6. If no order found, offer to connect with an agent

Handle typos in order numbers — suggest closest matches if the format looks
right but the exact number isn't found.

This is the flow that eliminates 45% of the support workload on day one. The bot asks for an order number or email, looks it up, and returns a formatted status message with emoji indicators for each stage:

• Processing: ordered, payment confirmed, preparing shipment
• Shipped: tracking number with a tappable carrier link
• Out for Delivery: estimated arrival time
• Delivered: confirmation with a "Rate us" button

Fuzzy matching handles the inevitable typos. If a customer types "ORD-1284" but the real order is "ORD-1248," the bot calculates Levenshtein distance and suggests the closest match: "Did you mean order ORD-1248?" This alone prevents dozens of "order not found" escalations to agents each week.

The "Notify When Shipped" button subscribes the customer to a push notification via template message when the order status changes. No more customers messaging back every few hours to check if anything has changed.

Step 3: Automate Returns and FAQ

Build the returns flow:
1. Ask for order number
2. Show order items as a numbered list — customer selects which item to return
3. Ask for return reason (interactive list: Wrong size, Defective, Changed mind, Other)
4. If "Defective" — ask for a photo (handle image message type)
5. Generate a return label (PDF) and send as a document
6. Confirm return with estimated refund timeline

Also build an FAQ handler: if the customer types a free-text question,
use keyword matching to detect common questions and respond with the right
answer. If no match, offer the agent handoff button.

The returns flow walks customers through the entire process without human intervention. Interactive lists let them pick which item to return and why. If they select "Defective," the bot asks for a photo — handling WhatsApp's image message type — and attaches it to the return request for the warehouse team.

The bot generates a return shipping label as a PDF and sends it as a document message. The customer gets the label, the estimated refund timeline, and a confirmation — all without waiting for an agent to process the request.

The FAQ engine matches free-text questions against 20 common topics using fuzzy string matching:

Unmatched questions get escalated with the original message preserved, so the agent sees exactly what the customer asked without having to ask again.

Step 4: Implement Human Agent Handoff

When a customer needs a human agent:
1. Add them to a queue with their conversation history
2. Notify the support team group: "New customer waiting — [name], topic: [returns],
   wait time: ~3 min"
3. An agent claims the conversation by reacting or clicking
4. Bridge messages: customer messages go to the agent's chat, agent replies go back
   to the customer's WhatsApp
5. When the agent sends /close, end the conversation, ask the customer to rate
   the experience (1-5 stars via buttons), and log the session

The 10% of inquiries that actually need a human get a smooth handoff. The customer sees their queue position and estimated wait time: "You're #2 in line, estimated wait: 5 minutes." The support team — on Slack or Telegram, configurable — gets a notification with the customer's name, topic, and conversation history so far.

An agent claims the conversation with a click. From that point, bidirectional message bridging routes customer WhatsApp messages to the agent's interface and agent replies back to the customer. The customer does not know they have been transferred — the conversation feels continuous.

When the agent sends /close, the bot asks the customer for a 1-5 star rating via interactive buttons and logs the entire session: conversation transcript, wait time, resolution time, and rating. This data feeds into weekly quality reports.

Step 5: Set Up Proactive Notifications with Templates

Create approved WhatsApp template messages for:
1. Order confirmation with details
2. Shipping notification with tracking link
3. Delivery confirmation with rating prompt
4. Abandoned cart reminder
5. Back in stock alert

Build a notification service that triggers these based on order status
changes and cart events. Respect opt-in/opt-out preferences. Include
rate limiting to stay within WhatsApp's messaging tier limits.

Template messages are pre-approved by Meta and can be sent outside the 24-hour conversation window. The notification dispatcher maps order events to templates automatically:

• Order confirmed triggers the confirmation template with order details
• Status changed to "shipped" sends tracking info to customers who opted in
• Delivered asks for a rating (and the positive ratings feed into social proof)
• Cart abandoned for 2+ hours sends a gentle reminder with a direct link back
• Wishlisted item back in stock sends an alert with a purchase link

Opt-in/opt-out management uses a database table — customers can text "STOP" at any time to unsubscribe. Rate limiting tracks the current WhatsApp messaging tier (1K, 10K, or 100K conversations per day) and queues excess messages to send within limits.

Real-World Example

The store owner deploys the bot on a Monday. Order tracking queries get instant responses starting day one, eliminating 45% of the agent workload immediately. Customers who used to wait 4 hours for a "your order shipped yesterday" message now get the answer in 3 seconds.

The automated returns flow handles 80% of return requests without human intervention. FAQ matching catches another 25% of common questions. The agents' daily queue drops from 200+ messages to about 30 — and those 30 are the genuinely complex issues that benefit from human attention.

After 2 months: average response time drops from 4 hours to 30 seconds for automated queries and 15 minutes for agent-handled ones. Customer satisfaction scores improve from 3.2 to 4.6 out of 5. The abandoned cart reminders recover an estimated 8% of abandoned carts. The team starts discussing whether they can operate with 2 agents instead of 3 — not to cut costs, but to redeploy the third person to a role where they are not answering "where's my order?" sixty times a day.