Change Airline Seats by Voice: AWS Nova 2.5 Sonic Travel Concierge Explained

On October 6, 2026, AWS published a reference architecture for a real-time voice concierge in airline applications. Amazon Nova 2.5 Sonic, Amazon Bedrock AgentCore, and Knowledge Bases enable spoken itinerary checks, seat changes, meal preferences, policy questions, and escalation to a human. The sample uses synthetic airline data rather than reporting a commercial deployment.

VOICE AS AN ADDITIONAL INTERFACE: Travelers traditionally navigate several screens to inspect reservations or select seats. A voice layer lets them ask whether a window seat is available or whether a flight is delayed. AWS deliberately keeps the existing screens, allowing people to switch between tapping and talking in one session.

THREE CORE COMPONENTS: Nova 2.5 Sonic processes incoming and outgoing speech. AgentCore Runtime hosts the agent, while AgentCore Gateway exposes airline operations as MCP tools. Amazon Bedrock Knowledge Bases retrieves relevant policy documents so the agent can answer baggage, cancellation, pet-travel, and loyalty questions using source material.

FROM THE APP TO THE BACKEND: A React front end hosted on AWS Amplify authenticates users with Amazon Cognito. It opens a signed WebSocket connection to AgentCore Runtime. When the agent needs business data, Gateway translates MCP tool calls into API Gateway requests, Lambda executes business logic, and DynamoDB supplies booking and flight records.

WHY MCP MATTERS: Model Context Protocol standardizes how an AI application discovers and invokes external tools. Airlines can expose existing booking and customer-service endpoints without rebuilding them around the voice model. Separating tool definitions from conversational logic improves maintainability.

BIDIRECTIONAL AUDIO STREAMING: The browser sends 16 kHz PCM audio over WebSocket and receives spoken responses through the same connection. Nova 2.5 Sonic can interpret speech, request data, and respond while the session remains active. This supports a conversational experience rather than a sequence of disconnected voice recordings.

HANDLING LATENCY AND INTERRUPTIONS: Asynchronous tool calls let the model keep a conversation moving while backend requests execute. Interim speech can mask some waiting time, and barge-in handling allows travelers to interrupt. Actual responsiveness still depends on network conditions, backend APIs, and load.

CONFIRM BEFORE WRITING: The sample follows a confirm-before-write pattern. Before changing a seat or meal preference, the concierge asks the traveler to approve the intended change. Reading a reservation and modifying it carry different risks, so speech-recognition mistakes must not silently become transactions.

AUTHENTICATION IS NOT VOICE RECOGNITION: The traveler signs in through Cognito, obtains tokens and temporary AWS credentials, and uses SigV4-signed connections. Backend APIs enforce IAM authorization. The design does not treat a matching voice or knowledge of a booking number as sufficient identity proof.

GROUNDING POLICY ANSWERS: Baggage and refund policies vary by fare and can change over time. Knowledge Bases retrieves relevant passages from airline documents and provides citations to the agent. Updated source documents can be synchronized without redeploying the entire conversational application.

ESCALATION TO A PERSON: When the concierge cannot resolve a request or the traveler asks for a human, it records an escalation and returns a reference number and estimated wait time. The front end can place a call to a configured support number. The sample does not include a complete contact-center platform.

USER-EXPERIENCE CONSIDERATIONS: The demonstration reads flight numbers and confirmation codes character by character to improve clarity. Real deployments must test airport noise, connectivity, language support, accessibility, and user preferences. Keeping screen controls available is an important part of inclusive design.

SECURITY AND OPERATIONS: AgentCore Runtime provides per-session microVM isolation. CloudWatch supports monitoring, while KMS encrypts data at rest. AWS recommends adding Bedrock Guardrails for production use. Organizations must also decide how long voice transcripts and personal information are retained and who can access them.

METRICS FOR REAL DEPLOYMENTS: A working demo does not prove customer-service improvement. Airlines should measure task completion, recognition errors, repeated confirmations, human handoff rates, handling time, and satisfaction. For consequential booking changes, correctness and recoverability may matter more than raw speed.

IMPLEMENTATION AND COST: AWS provides CDK-based deployment for the front end, Gateway, Runtime, and backend. Developers need a supported Nova 2.5 Sonic Region, AWS permissions, and the required development tools. Cloud resources may continue generating charges after testing unless they are cleaned up.

THE BIGGER PICTURE: The architecture treats voice as a layer over existing APIs, identity systems, policy documents, and human support rather than a disconnected chatbot. The same principles—multimodal choice, verified identity, grounded answers, and explicit confirmation—could inform voice experiences in other service industries.