A large beverage CPG company is running an AI voice coach for its field sales representatives on serverless AWS infrastructure, at roughly $0.007 per five-minute session with zero idle costs. The system lets reps hold natural voice conversations with an AI that knows their accounts, products, and selling scenarios - an approach aimed at a persistent problem: coaching quality varies by geography and manager availability.
The architecture uses AWS AppSync Events as a real-time audio transport layer between a rep's browser and a voice agent running on Amazon Bedrock AgentCore with Amazon Nova Sonic. It's in production at the beverage company, where field reps use voice conversations to prepare for account visits, get personalized product recommendations, practice cross-sell strategies, and work through objection handling.
What the system does
Reps talk to the AI coach the way they'd talk to a manager. Two dedicated channels carry audio in each direction: one for the rep's microphone audio going to the AI, one for the AI's spoken responses coming back. AWS AppSync Events handles connection management, message routing, fan-out, and auto scaling - no WebSocket servers to provision or manage.
That decoupling gives the architecture flexibility. Clients can publish audio over HTTP and subscribe over WebSocket. Authentication is built in through IAM, Amazon Cognito, API keys, or Lambda authorizers. For a sales organization, the practical result is a coaching tool that's available on demand, from any browser, without IT standing up dedicated infrastructure.
What it costs
The pricing model is what makes this viable across an entire sales force. AppSync Events charges $1.00 per million event operations. A five-minute session produces roughly 7,000 event operations, putting the cost at about $0.007 per session.
The alternatives cost more before any sessions happen. An always-on Fargate deployment with an Application Load Balancer runs $50 or more per month in base infrastructure. API Gateway WebSocket requires Lambda and a DynamoDB connection table. LiveKit and Pipecat need a framework server plus TURN/STUN infrastructure.
For a team of 50 reps averaging three sessions per day, AppSync Events costs about $31.50 per month - and nothing when the system isn't in use. The Fargate approach costs $50 per month before a single coaching session occurs.
Production lessons
During deployment, every voice session terminated at roughly 112 seconds. The team eventually traced it to two independent bugs: a missing health check endpoint and a wrong AppSync endpoint for WebSocket subscriptions. The container was being killed by AgentCore with SIGKILL, which bypasses all signal handlers and produces no error logs - so neither fix alone resolved the problem.
The team published a checklist from the experience: implement the /ping health endpoint, use the API-specific AppSync DNS for WebSocket subscriptions, pass model IDs through environment variables, and disable HTTP/2 timeouts for long streams. They also recommend bumping a build version variable to force container updates and logging audio chunk counts per session for diagnostics.
Why this matters for sales teams
The takeaway for sales leaders is that AI coaching no longer has to be rationed by infrastructure costs or manager availability. At $0.007 per session, an organization can offer unlimited coaching to every rep - not just top performers or those with a manager nearby. The same serverless audio transport pattern applies to customer service, clinical documentation, real-time transcription, field service assistance, and training simulations.
For reps and managers evaluating AI coaching tools, the AI Learning Path for Sales Representatives covers the practical skills for working with AI tools. The AI for Sales tag tracks related developments in this space.
Your membership also unlocks:
