Build a voice travel concierge with Amazon Bedrock AgentCore, Managed Knowledge Base and Nova Sonic
Amazon Web Services Ravi Kumar
AWS shows a voice concierge for airline apps using Bedrock AgentCore, Nova Sonic, and Knowledge Bases. It lets travelers talk to bookings, policy docs, and live agents without leaving the app.
Based on reporting by Amazon Web Services, Ravi Kumar — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Airline apps already handle the basics: flights, seats, bookings. AWS is now pushing them one step further with a voice layer that lets travelers ask for a seat change, check a delay, or get policy answers by speaking instead of tapping through screens.
The company’s demo builds that experience on three managed services. Amazon Bedrock AgentCore hosts the agent, Amazon Nova 2.5 Sonic handles real-time speech, and Amazon Bedrock Knowledge Bases answers policy questions from airline documents. The pitch is less about a flashy front end than about the plumbing: keep the conversation flowing, reach backend systems without tightly coupling them, and survive traffic spikes when everyone remembers they need to fly somewhere at the same time.
The architecture splits into clean layers. A React app sits on AWS Amplify. Amazon Cognito handles sign-in and temporary credentials. AgentCore runtime runs the voice agent with microVM isolation per session, while AgentCore Gateway exposes backend systems as MCP tools. Behind that, API Gateway, AWS Lambda, and DynamoDB handle the actual airline work: itineraries, seat maps, passenger updates, flight status, loyalty data, and escalation. Amazon SES sends email notifications. AWS CloudWatch watches the whole thing.
The knowledge base side is more interesting than it first sounds. Travelers can ask about baggage, change fees, pet travel, refunds, special assistance, loyalty terms, and upgrades. AWS says the documents are uploaded to Amazon S3, then Bedrock handles embedding, chunking, indexing, storage, and retrieval, with citations coming back to the agent. When those policies change, you sync the knowledge base and the updates are available right away. No redeploy dance.
The voice flow itself leans hard on Nova 2.5 Sonic. Audio streams over a WebSocket as 16 kHz PCM. The model transcribes speech, triggers tools in parallel when needed, and can keep talking while it waits for results. It also supports barge-in and natural turn-taking, which matters because nobody wants a robot to monologue while their flight boards without them. For write actions, the sample uses a confirm-before-write pattern, and when a traveler asks for a human, Lambda returns a reference number and the app passes them to a live agent with an estimated wait time.
My take — AI-written commentary, not fact-checked reporting
This is the right direction: voice only matters when it reaches real systems, not when it just chatters politely over a fake FAQ. The bigger signal here is how normal the stack is getting — auth, tools, retrieval, guardrails, handoff — which is exactly why the closed-model crowd keeps winning enterprise deals while everyone else debates poetry.
Read more about this at: Amazon Web Services