Project Case
Building a WhatsApp AI Booking & Customer Support System
How I designed a WhatsApp-based customer interaction platform that combines conversational AI with deterministic backend workflows for bookings, confirmations, support, and CRM escalation.
Context
The goal was to allow customers to interact with businesses through WhatsApp instead of requiring them to use a traditional application for every task. A simple message such as a request for a ride still needed to become structured information that the backend could validate and process. The platform also needed to support customer service conversations and escalate cases to human agents when automation could not resolve them.
Constraint
The main challenge was defining the boundary between conversational AI and deterministic business logic. Customer messages could be incomplete or ambiguous, while important operations such as booking creation, pricing, availability, and customer updates could not safely depend on unconstrained model output. The system also needed to recover from failed APIs, unsupported requests, incomplete information, and model errors without leaving the customer stuck.
Approach
- Used the LLM primarily as an interface for understanding customer intent and collecting information while keeping business decisions and persistence inside validated backend services.
- Converted conversational requests into explicit workflows with required information, validation rules, backend actions, confirmation states, and fallback paths.
- Used structured WhatsApp interactions such as buttons, lists, and Flows where predictable input was more appropriate than open-ended conversation.
- Connected the conversational layer to backend capabilities such as booking creation, booking lookup, customer information, service availability, pricing-related operations, support workflows, and CRM escalation.
- Validated important model-generated tool inputs before allowing them to reach downstream business services instead of treating model output as trusted data.
- Designed explicit recovery paths for missing information, unavailable services, backend failures, unsupported requests, and conversations requiring human assistance.
- Reduced unnecessary AI processing by using deterministic WhatsApp interactions for steps that did not require natural-language reasoning, lowering AI-assisted WhatsApp booking costs by approximately 75–80% in the relevant workflow.
- Added workflow-level logging and diagnostics so the engineering team could trace failures across customer input, model behavior, tool execution, backend validation, and external service communication.
Result
The platform enabled customers to initiate bookings, receive confirmations, interact with business services, and escalate support issues directly through WhatsApp. The same architecture also provided a foundation for extending WhatsApp automation into other workflows, including delivery and customer-service operations.
Lesson
Reliable AI products require a clear boundary between probabilistic and deterministic systems. The model can help understand what the customer wants, but the backend should remain responsible for validating the request, deciding what actions are allowed, and determining what gets persisted.