Building a Restaurant Voice Ordering System with Amazon Connect

Many restaurants still receive a significant portion of orders via phone calls, often handled by staff who are busy with in-person customers. To streamline this process, AWS has demonstrated a solution that enables customers to place orders through a voice system that answers calls, interacts naturally, and confirms orders without human intervention. The system uses Amazon Connect for telephony, Amazon Lex V2 with Amazon Connect Agentic Voice for real-time speech recognition and synthesis, and an Amazon Connect AI agent for reasoning and decision-making. Backend services are connected via the Model Context Protocol (MCP) and Amazon Bedrock AgentCore Gateway, allowing flexible integration with restaurant systems.

The architecture separates call handling, conversation orchestration, and backend data management. When a call arrives, Amazon Connect manages the telephony channel, routing the call through a contact flow that initiates an AI-powered conversation. The contact flow captures caller information, plays greetings, and connects the caller to a speech-enabled bot that handles the interaction. The AI agent, powered by Anthropic Claude Haiku 4.5 in Amazon Bedrock, manages the dialogue, fetches menu data, manages carts, and confirms orders, all while calling backend services through MCP tools exposed by AgentCore Gateway.

The solution emphasizes real-time speech recognition and expressive text-to-speech, enabling natural and efficient interactions. It also incorporates AI Guardrails to ensure conversations remain safe and on-topic, filtering out harmful or off-limits content. The backend infrastructure uses Amazon DynamoDB for data storage, Amazon Location Service for geocoding and routing, and AWS Lambda for business logic, all deployed via AWS Cloud Development Kit (CDK).

Deployment involves setting up the telephony channel, configuring the AI agent, and connecting to backend services. The entire system can be tested by dialing the provided phone number, where users can speak naturally to place orders, inquire about menus, or specify pickup locations. The system handles multiple concurrent calls, maintains individual carts, and logs all interactions for monitoring and debugging.

Cost considerations are primarily based on usage of Amazon Connect call minutes, Amazon Lex speech requests, and Anthropic Claude tokens. The solution is designed to be scalable and adaptable, allowing restaurants to modify menus, locations, and backend integrations without altering the core AI logic. Cleanup procedures are available to remove resources and avoid ongoing charges.

This approach demonstrates how AI and telephony services can be integrated to automate restaurant order-taking, reducing staff workload and improving customer experience. The sample code and deployment instructions are available on GitHub for adaptation to specific restaurant needs.