Building AI WhatsApp Bots

WhatsApp Business API Access

Unlike Discord or Telegram where anyone can create a bot, WhatsApp requires a verified business account and access to the Business API. There are two paths to API access.

Meta Cloud API is Meta's directly hosted solution. You set up through the Meta Developer Portal, verify your business, and register a phone number for your bot. The Cloud API is free to use (you only pay per-conversation fees), handles hosting and scaling, and provides the latest features first. The trade-off is that your messages flow through Meta's infrastructure, which may not meet all data residency requirements.

Business Solution Providers (BSPs) like Twilio, MessageBird, Vonage, and Infobip offer managed WhatsApp API access with additional features like analytics dashboards, multi-channel integration, and dedicated support. BSPs charge their own fees on top of Meta's per-conversation pricing. The advantage is a more complete platform with better tooling, the disadvantage is added cost and an additional dependency.

Choosing between Cloud API and a BSP depends on your technical capacity and scale. Cloud API is the right choice if you have developers who can build the integration directly and you want to minimize per-message costs. A BSP is better if you want turnkey setup, need compliance certifications that the BSP provides, or plan to use multiple messaging channels through the same provider. Many BSPs offer free tiers or trials that let you test the integration before committing to a paid plan.

Business verification through Meta requires submitting business documentation and completing a review process. This typically involves providing a business name, address, website, and category, then verifying through a phone call or document upload. The verification process can take from a few days to a few weeks depending on the business type and documentation quality.

Conversation Pricing Model

WhatsApp uses a conversation-based pricing model that directly affects how you design your bot. Each conversation is a 24-hour messaging window between your business and a user. The price per conversation varies by country and conversation category.

User-initiated conversations start when a user messages your business first. These are the cheapest category because the user has shown intent. Once a user sends a message, you have a 24-hour window to send unlimited messages in the conversation. If 24 hours pass without a new user message, the window closes and you can only re-engage through a template message (which starts a new, business-initiated conversation).

Business-initiated conversations start when you send a template message to a user outside of an active conversation window. These are more expensive and require pre-approved message templates. Templates must be submitted to Meta for review and approved before use. They can include variables for personalization but the overall structure must be approved.

For AI chatbots, the user-initiated model aligns well: users message your bot with questions, the bot responds within the session window, and the per-conversation cost is relatively low (typically $0.005-$0.08 depending on region). The key design implication is that you want to handle as much of the conversation as possible within a single 24-hour window rather than needing to re-engage later with template messages.

Message Types and Formatting

WhatsApp supports several message types that affect how your AI bot can present information.

Text messages are the most basic format and support simple formatting: bold (asterisks), italic (underscores), strikethrough (tildes), and monospace (backticks). Messages have a 4,096 character limit. For AI-generated responses, configure your LLM to use WhatsApp-compatible formatting and keep responses concise since users expect brief, mobile-friendly messages.

Interactive messages include list messages (presenting up to 10 options in a structured list) and reply buttons (up to 3 quick-reply buttons). These are valuable for AI chatbots because they guide user input into structured choices, which reduces ambiguity and improves the quality of subsequent AI responses. For example, after identifying a user's general topic, the bot can present a list of specific options to narrow down their need.

Media messages support images, videos, audio, documents, and location sharing. AI bots can send images (product photos, charts, diagrams), documents (PDFs, invoices), and location pins. Receiving media from users opens up use cases like image recognition, document processing, and voice transcription.

Designing for WhatsApp Constraints

WhatsApp's mobile-first design and Meta's strict policies create specific design constraints that affect AI chatbot behavior.

Response length should be shorter than on other platforms. WhatsApp users are typically on mobile devices and expect quick, scannable messages. Configure your LLM to produce concise responses (2-4 short paragraphs maximum) and break longer content across multiple messages with clear structure.

Template message restrictions mean you cannot send arbitrary messages to users who have not messaged you first or whose 24-hour window has closed. This affects features like proactive notifications, follow-ups, and re-engagement. Every template must be approved by Meta, which takes 24-48 hours and must follow strict formatting guidelines that prohibit promotional language in certain categories.

Quality ratings affect your ability to send messages. Meta monitors user feedback (blocks, reports) and adjusts your messaging tier accordingly. Bots that generate low-quality or unwanted responses may have their messaging limits reduced. This creates a strong incentive to ensure your AI chatbot produces high-quality, relevant responses and provides easy opt-out mechanisms.

Commerce policy compliance is required for any bot that facilitates sales or payments. WhatsApp has specific rules about product catalogs, payment processing, and promotional messaging that vary by region. Understanding these policies before building is essential to avoid compliance issues.

Integration Architecture

The typical WhatsApp AI bot architecture includes a webhook endpoint that receives messages from the WhatsApp API, a message handler that processes incoming messages and manages conversation state, the LLM integration for generating responses, and a message sender that formats and delivers responses through the API.

Webhook verification is required when setting up your endpoint. Meta sends a verification request with a challenge token, and your endpoint must respond correctly to confirm ownership. After verification, all incoming messages are delivered to your webhook as JSON payloads.

Message delivery is asynchronous. When you send a message through the API, you receive a message ID but not delivery confirmation. Delivery and read receipts come as separate webhook events. Your system should track message status to handle retries for failed deliveries and to understand conversation flow.

Handling Media and Voice Messages

WhatsApp users frequently send voice messages instead of typing, especially in markets like Latin America, the Middle East, and South Asia where voice messaging is the dominant communication mode. Your AI bot should handle voice messages by downloading the audio file from the WhatsApp API, transcribing it using a speech-to-text service like OpenAI Whisper or Deepgram, and processing the transcribed text through your normal conversation pipeline. Without voice message support, your bot effectively ignores a large percentage of user messages in these markets.

Image processing opens up powerful use cases. Users can send photos of products they want to find, screenshots of error messages they need help with, or documents they want summarized. Vision-capable LLMs like GPT-4o and Claude can analyze these images directly. The workflow is straightforward: download the image from the WhatsApp media endpoint, send it to the vision model with the user conversation context, and return the analysis as a text response. This capability is particularly valuable for e-commerce bots where users can snap a photo and ask "Do you sell something like this?"

Document handling lets users share PDFs, spreadsheets, and other files with your bot. A customer support bot can accept warranty documents, receipt photos, or order confirmations and extract relevant information automatically. Process documents by downloading from the WhatsApp media endpoint, extracting text using a document processing service, and feeding the extracted content into your LLM as context for the conversation.