AWS builds AI phone host for restaurants using Bedrock and Nova 2 Sonic

AWS published a technical guide and working code sample for deploying a restaurant voice ordering AI that answers phone calls, takes orders, and books tables. The system uses three core AWS components: Amazon Bedrock AgentCore to host and run the conversation logic, Amazon Nova 2 Sonic to handle real-time speech processing, and the Model Context Protocol (MCP) to connect the agent to a restaurant backend that holds menus, carts, orders, and location data.

The architecture is organized into four layers. The restaurant backend (Section A) stores customer, menu, order, and location data in Amazon DynamoDB and Amazon Location Service, with AWS Lambda handling business logic exposed through Amazon API Gateway. The AgentCore Gateway (Section B) exposes those backend endpoints as tools the agent can call by name, decoupling the agent from backend changes. The agent itself (Section C) runs in Amazon Bedrock AgentCore Runtime, with each call isolated in its own microVM. The phone layer (Section D) uses Amazon Chime SDK Voice Connector to accept inbound calls and a SIP gateway running on Amazon ECS Fargate to translate between the phone network's RTP media format and the agent's WebSocket protocol.

When a customer calls the restaurant's toll-free number, Amazon Chime SDK answers and invokes a Lambda function that creates a session identifier and warms up the agent's microVM while the phone is still ringing. This avoids a cold start when the actual media stream arrives. The SIP gateway then bridges the call by opening a signed WebSocket connection to AgentCore Runtime, attaching the caller's audio to the warmed session. The agent runs the conversation using Amazon Nova 2 Sonic for speech-to-speech interaction, calling backend tools through AgentCore Gateway when it needs menu, cart, order, or location information. The agent supports both text and audio input and output, handling transcription, turn-taking, and interruptions in a single bidirectional stream.

The full solution deploys via AWS CDK using a provided GitHub repository. Deployment requires an AWS account with access to Amazon Bedrock (specifically Amazon Nova 2 Sonic), Amazon Chime SDK PSTN Audio, AgentCore Runtime, and optionally AgentCore Gateway in the target region. The deployment script builds the backend first, then the AgentCore Gateway, then the agent container image (built in AWS CodeBuild and stored in Amazon ECR), and finally the SIP gateway and phone provisioning. Once deployment completes, the script prints the toll-free number and seeds sample menu and location data so a real order can be placed immediately. All data at rest is encrypted using AWS Key Management Service, and monitoring and logging are centralized through Amazon CloudWatch.

The problem this solution addresses is concrete and widespread: restaurants lose business because they miss inbound calls during peak service hours, when answering the phone becomes logistically impossible. The article notes that approximately 60 percent of the 150 calls per location monthly are order or reservation requests—precisely the calls that represent lost revenue. Adding a website or mobile app helps customers who prefer to order online but does nothing for callers who prefer the phone, creating a gap that no single digital channel has closed.

The technical architecture reflects a deliberate design principle: separating the phone layer (SIP gateway, media handling), the agent layer (conversation logic and speech processing), and the restaurant backend (menus, orders, locations). This separation means the ordering logic is decoupled from the telephony channel, allowing the same agent and backend to serve multiple interfaces (phone, app, kiosk) without duplication. The use of the Model Context Protocol as an open standard between the agent and backend further enables the restaurant to change backend systems without redeploying the agent itself.

The system also solves a user-experience problem that is often overlooked: cold start latency. By warming the agent's microVM session while the phone is still ringing, the solution ensures the caller hears a greeting immediately, avoiding the awkward silence that would damage trust in an automated system. This detail reveals an understanding that voice ordering systems are sensitive to timing in ways web-based systems are not.