Loading page…
The sample shows how a voice agent can take an order from a phone call through backend tools, but restaurants must supply and operate the systems holding their own menus, carts, orders and locations.
Listen to this story
AWS has released a deployable CDK reference pattern for voice-based restaurant ordering, combining Amazon Connect, Lex V2, Claude Haiku 4.5 on Bedrock, and MCP-connected backend tools. The sample demonstrates calls, menus, carts, locations, profiles, and order placement, but it is not a managed restaurant system: operators must supply and govern the live data and APIs.
The sample’s restaurant backend uses API Gateway, Lambda, DynamoDB, and Amazon Location Service for menus, carts, profiles, orders, and locations.
AgentCore Gateway exposes backend functions as discoverable MCP tools, keeping ordering logic separate from the Amazon Connect phone channel.
AWS provides sample code and CDK deployment—not a managed data layer—so operators must connect their own menus, orders, and business rules.
AWS has published a deployable reference architecture for a restaurant AI host that answers phone calls, handles menu questions, recommends pickup locations, manages a cart and confirms an order aloud. The useful distinction is not a finished restaurant product: it is an integration pattern that joins Amazon Connect telephony, an AI agent and an operator-managed ordering backend.
The design divides the work. Amazon Connect receives the inbound call and runs its contact flow. Amazon Lex V2, using Amazon Connect Agentic Voice, converts speech to text and generates spoken replies. An Amazon Connect AI agent drives the conversation with Anthropic Claude Haiku 4.5 in Amazon Bedrock.
When the agent needs restaurant information or must change an order, it calls named tools exposed by Amazon Bedrock AgentCore Gateway through the Model Context Protocol, or MCP. The gateway passes those requests to backend APIs, which the sample implements with API Gateway and Lambda functions backed by DynamoDB and Amazon Location Service.
AWS’s sample backend contains the restaurant-facing functions: menu retrieval, cart management, order placement, customer profiles and location lookup. But AWS supplies the CDK deployment and sample code, not a managed restaurant data system. An adopter must connect and operate the backend that reflects its own menu, orders, locations and customer records.
That separation is the architecture’s central design choice. AWS says MCP is an open standard for connecting agents to external tools, allowing the backend to change without altering the agent. The pattern therefore aims to keep ordering logic independent of the phone channel rather than embed it inside the telephony workflow.
The reference design attaches an Amazon Connect AI Guardrail to the agent, with content filters, denied topics and profanity filtering. Its backend gateway validates custom JWT authorization against the Amazon Connect instance, while the sample places its REST endpoints behind IAM authorization.
The full implementation is available as a sample repository and deploys its stacks through AWS CDK. For restaurant operators, the practical test is less whether the call can reach an agent than whether their live order data and business rules can safely become the tools that agent is allowed to use.
Loading discussion...
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.