AWS's restaurant AI host using Bedrock AgentCore and Nova 2 Sonic cuts missed calls
Restaurants miss 150 calls a month – AI answers instantly, taking orders by voice.
AWS demonstrates a solution for restaurants using Amazon Bedrock AgentCore and Amazon Nova 2 Sonic to handle the nearly 150 missed phone calls per location each month—60% of which are customers trying to place orders or book tables. The system is built in three decoupled layers: telephony (Amazon Chime SDK, ECS/Fargate SIP gateway), agent (AgentCore Runtime with Nova 2 Sonic for real-time speech), and backend (menu, carts, orders via API Gateway, Lambda, DynamoDB). MCP (Model Context Protocol) connects the agent to backend tools, enabling channel independence. A key innovation: the agent session is warmed while the phone rings, ensuring zero latency when the caller speaks. The full stack deploys via AWS CDK, with container images built by CodeBuild and stored in ECR.
The telephony layer handles inbound SIP calls through a Network Load Balancer and Amazon ECS on Fargate. The agent layer runs each call in an isolated microVM, with Nova 2 Sonic handling bidirectional speech-to-speech (including interruptions and turn-taking). The backend layer uses Amazon Location Service for geocoding and DynamoDB for customer/order data. IAM secures all API endpoints. This architecture means the same agent can later be reused for mobile apps or kiosks without rewriting the backend. By automating these calls, restaurants can capture lost revenue without pulling staff off the floor.
- Restaurants average 150 missed calls per month per location, with 60% being orders or reservations.
- Uses Amazon Nova 2 Sonic for real-time speech-to-speech and Amazon Bedrock AgentCore for conversational logic.
- Architecture decouples telephony, agent, and backend via SIP gateway and MCP, enabling reuse across channels.
Why It Matters
Restaurants can automate 60% of missed calls, turning lost orders into revenue without adding staff.