How to Build a WhatsApp AI Customer Service Bot for Your Bangladesh Business

WhatsApp is not optional for Bangladesh businesses anymore — it is the primary customer communication channel. If someone has a question about your product, a complaint about an order, or needs your business hours at 11 PM, they are going to WhatsApp you. The question is whether a human has to be there to answer, or whether an AI can handle it.

This guide walks through exactly how to build a WhatsApp AI customer service bot — from the API layer down to the AI backend. It assumes you are technical enough to follow a setup guide but are not necessarily a developer.

Understanding the Stack

Before touching any tools, it helps to understand the layers involved:

1. WhatsApp Business API — the official way to send and receive WhatsApp messages programmatically. This is the channel layer. Without this, you cannot receive messages in code.
2. A workflow or agent platform — something that receives the incoming message, figures out what to do, and sends a reply. This is where the logic lives. Options include OpenClaw, n8n, or custom code.
3. An AI model — the language model that actually understands the customer's message and generates a response. This could be GPT-4 via API, or a self-hosted model like Hermes.
4. Your business data — product catalog, order database, FAQ document, pricing list. The AI needs access to this to give accurate answers.

These four layers need to be connected. There is no single product that handles everything out of the box — though platforms like OpenClaw come close for the middle two layers.

Step 1: Get WhatsApp Business API Access

The official route is through Meta's WhatsApp Business Platform. You register at developers.facebook.com, create a Meta app, and apply for WhatsApp Business API access.

For most Bangladesh small businesses, the simplest path is through a Business Solution Provider (BSP) — companies like 360dialog, Twilio, or Wati. They handle the Meta approval process and give you a simpler API to work with. Costs typically run $30–$80/month for the BSP plus WhatsApp's per-conversation fees (around $0.004–$0.008 per conversation for utility messages in Bangladesh).

What you will need for registration:

• A Meta Business account (business.facebook.com)
• A verified business phone number not already linked to a personal WhatsApp account
• A registered business name
• A brief description of your use case

Approval typically takes 3–7 business days. During this time, you can build and test everything else.

Step 2: Choose Your Agent Platform

Once you have WhatsApp API access, you need something to receive the messages and respond intelligently.

Option A: OpenClaw (Recommended for AI Agents)

OpenClaw is an autonomous AI agent platform that connects natively to WhatsApp, Telegram, and Discord. It is self-hosted — you run it on a VPS — which means your customer conversations stay on your server, not on a third-party platform.

OpenClaw uses Qdrant (a vector database) to give your agent long-term memory. This means it can remember that a specific customer complained about a delayed order two weeks ago, and factor that context into today's reply.

Setup overview for OpenClaw + WhatsApp:

1. Deploy OpenClaw on a VPS (minimum 2GB RAM, 1 vCPU recommended)
2. Configure your WhatsApp Business API webhook URL to point to your OpenClaw instance
3. Define your agent's persona, knowledge base, and permitted actions in OpenClaw's config
4. Connect Qdrant for memory (OpenClaw has built-in Qdrant integration)
5. Test with a staging phone number before going live

Option B: n8n (Best for Multi-Step Workflows)

If your bot needs to do more than just answer questions — for example, check order status in WooCommerce, update a Google Sheet, and then reply — n8n is worth considering. It is a visual workflow builder with 400+ integrations including WhatsApp, AI models, databases, and common business tools.

An n8n WhatsApp bot workflow looks like this: WhatsApp webhook triggers → extract message text → send to AI (GPT-4 or self-hosted model) → optionally query your order database → format reply → send back via WhatsApp API.

Option C: Custom Code

If you have a developer, you can build directly on the WhatsApp API with Node.js or Python. More flexible, but more maintenance burden. Only choose this route if the off-the-shelf platforms genuinely cannot do what you need.

Step 3: Set Up Your AI Backend

The agent platform handles the flow — the AI model handles the actual understanding and generation. You have two main options:

OpenAI API (GPT-4 or GPT-4o)

Easiest to start with. Sign up at platform.openai.com, get an API key, and point your agent at it. GPT-4o is fast and accurate for customer service tasks. Cost for a small Bangladesh business handling 500 conversations/month: roughly $15–$40 depending on message length.

The downside: your customer conversations are processed on OpenAI's US servers. For businesses handling sensitive customer data, this may be a concern.

Self-Hosted Model (Hermes, Llama 3, etc.)

Run the AI model on your own VPS. SNBD HOST's Hermes hosting gives you a managed self-hosted AI model with 1 vCPU and 2GB RAM — free for the first year during our beta. This keeps all data on servers you control.

The tradeoff: self-hosted models are slightly less capable than GPT-4 for complex tasks, but for customer service FAQs, order lookups, and standard queries, they perform well.

Step 4: Connect Your Business Data

An AI that only knows what is in its training data is not very useful for customer service. You need to give it access to your specific business information:

• FAQ document: Return policy, shipping times, business hours, product specs. Feed this into the AI's context or a vector database for retrieval.
• Order database: If customers will ask "where is my order?", the AI needs API access to your order management system. This typically means building a simple API endpoint your agent can call.
• Product catalog: Product names, prices, availability. Can be a simple CSV file loaded into a vector database.

The vector database layer (Qdrant is the most common choice) stores this information as embeddings, allowing the AI to retrieve the relevant chunks when answering a question without needing to read every document every time.

Step 5: Test Before Going Live

Test your bot thoroughly on a staging phone number before connecting it to your main business number. Specifically test:

• Questions your FAQ document covers — does it give accurate answers?
• Questions it should escalate to a human — does it do so gracefully?
• Rude or aggressive messages — does it stay professional?
• Messages in Bangla — does your model handle Bengali script accurately?
• Edge cases: what happens if someone sends a voice note, a photo, or just a sticker?

Realistic Expectations

A well-built WhatsApp AI bot should handle 60–80% of customer queries without human intervention within the first month of tuning. The remaining 20–40% will need human escalation — either because the query is complex, requires judgment, or involves something not in your data.

That is still a significant reduction in support workload. If your team currently handles 200 WhatsApp messages per day, an AI handling 70% means your team focuses on 60 genuinely complex cases instead of 200 mixed ones.

Handling Bangla Language

A practical concern for Bangladesh businesses: will the bot understand and reply in Bangla?

GPT-4o and GPT-4o-mini both handle Bangla well — including informal Banglish (mixed Bangla-English). Self-hosted models vary: Llama 3 and Hermes handle Bangla adequately for most customer service tasks, though complex idiomatic Bangla can trip them up.

You have two options for language handling:

• Auto-detect and reply in kind: Instruct the bot to "always reply in the same language the customer uses." This works for clear Bangla or English messages. Banglish (mixed language) gets a best-effort response, usually in Bangla.
• Fixed language: Define the bot language in the system prompt. "Always reply in Bangla" or "always reply in English" gives consistent output but may frustrate customers who write in a different language.

For most Bangladesh businesses, auto-detect is the right default. Test your bot with 20–30 sample messages in the languages your customers actually use before launching.

Ongoing Maintenance

An AI bot is not a set-and-forget deployment. Plan for ongoing maintenance:

• Knowledge base updates: When your prices, policies, or products change, update the bot's knowledge base. Outdated information is worse than no information — the bot will confidently give customers wrong answers.
• Reviewing escalated conversations: Read the conversations your human team handles after bot escalation. These reveal patterns — common questions the bot mishandles, gaps in your FAQ, types of queries you had not anticipated.
• Model updates: AI models improve. OpenAI releases new versions periodically. Self-hosted models get new releases from their developers. Periodic updates improve quality without major re-configuration.

Budget roughly 2–4 hours per month for maintenance once the bot is launched and stable. More in the first two months while you tune it based on real conversations.

Summary

Building a WhatsApp AI bot requires four things: WhatsApp Business API access, an agent platform (OpenClaw or n8n), an AI model (OpenAI API or self-hosted), and your business data in a retrievable format. Start simple, test rigorously, and expand scope only once the core FAQ layer is reliable.