A ten-second question took three minutes.
Now an AI agent answers it during the call.

With a week to go, the tempting move is to patch whatever is failing loudest. We rebuilt the Amazon Connect contact flows from the ground up instead, including the routes and handoffs that decide where each request goes, because routing was broken across the whole deployment and patches would have left the next failure hidden.

Four problems threatened the launch. We fixed each one at its cause.

• The routing and queue logic was broken, so we rebuilt the contact flows. From the first hour, customers reached the right place and people got the right calls.
• Escalated calls were arriving with no context, so we repaired the Amazon Connect–ServiceNow integration all the way through, because the receiving team needed the conversation history before taking over.
• Integration failures were happening silently. We rewrote the AWS Lambda functions behind live lookups, routing decisions and CRM updates, then tested them under production conditions.
• There was no view of system health. We set up Amazon CloudWatch monitoring, structured logs and alerts, which meant the operations team could see a failure as it happened.

The contact center launched on time. The work continued afterward as scoped pieces of work for voice AI agents, chat and observability, all on the customer's live systems. Every AI capability that came later stands on that week, because an AI agent running on a broken contact flow only reaches the wrong place faster.

A better menu, something like 'press 3 for showtimes', would have been the easy fix. We built three voice AI agents on Amazon Connect and Amazon Nova Sonic, because a customer should just be able to say what they need.

Nova Sonic handles speech-to-speech conversation as one continuous stream, with no text step in between. That keeps the tone of the caller's voice, lets it notice frustration and adapt, and lets it deal with interruptions and follow-up questions the way a person would.

We built the voice side on a custom container-based architecture that keeps call transport and business-system access apart from speech processing:

• Amazon Connect manages the contact flows, queues and routing, and passes the call into an Amazon Chime SDK session. The contact ID links that session to the contact-center interaction.
• A containerized SIP service connects the voice session to Nova Sonic, with an Application Load Balancer handling the service connections.
• An MCP server in a private subnet gives the AI agents their business capabilities. VPC and VPN connectivity link it to the customer's protected APIs.
• Lambda functions and session state record call information and outcomes, keeping each business action tied to its call.

Each AI agent has one job, and a defined set of requests:

• The refund AI agent checks eligibility against live systems and issues qualifying refunds automatically, during the call. It also gives refund status and gift card balances.
• The showtime AI agent gives current schedules, theater details and seat formats, and sends an SMS follow-up when that helps.
• The loyalty AI agent gives rewards, tier status and account information from live APIs, replacing the stale exports.

A refund is the one request where the AI agent moves money, so its design gets the most care. The workflow checks live eligibility before issuing anything, and the integration is split into separate functions for looking up, issuing and handing over:

• Refund Lookup gets what the request needs from the on-premises refund service.
• Refund Issue calls the service to carry out the refund.
• Both Lambda functions run in a private subnet and reach the on-premises network over VPN.
• Transfer Call is a separate path for handing the call to a person, and Amazon DynamoDB records call-action state.

Keeping the lookup separate from the issue means the function that moves money has its own permission, can be tested on its own, and never runs by accident while the AI agent is only answering a question. Anything the AI agent can't resolve follows the transfer path into Amazon Connect, with ServiceNow supplying the conversation so far.

The chat AI agent runs on Amazon Bedrock and covers refund eligibility, loyalty lookups, showtimes, theater information and account questions. A chat session starts like this: the website sends a request carrying the user and page context; Amazon API Gateway invokes Lambda, which starts the chat in Amazon Connect and returns a signed chat URL; and a containerized chat service runs the conversation, reaching business capabilities through an MCP server.

The decision that matters is where each kind of answer comes from:

• Policy rules come from a retrieval-augmented generation pipeline on Amazon OpenSearch, which draws on curated refund eligibility rules, loyalty terms and service content.
• The facts come straight from the systems themselves. Booking information, refund status and rewards balances come live from the CRM, ticketing and loyalty APIs through Lambda, replacing batch exports.
• Some questions need both. 'Can I get a refund?' needs the policy and the customer's booking together, and the chat AI agent combines them.
• Multi-turn conversation management keeps the earlier exchanges, so a follow-up question makes sense.

This matters because a wrong answer on refund terms or loyalty rules carries a real business and regulatory cost, so the model explains the rule while the systems supply the balance.

The handoff to a person is where most AI contact centers fall apart, because the customer has to explain everything again to someone who knows nothing. We carried the context through every change of channel:

• Amazon DynamoDB holds call state, routing data and conversation persistence.
• Amazon ElastiCache keeps the session context across channels, including a move from chat to voice.
• On escalation, ServiceNow receives the transcript, the detected intent and the sentiment, so the person taking over reads the conversation before saying hello.

We built the observability around the conversations as well as the systems:

• Contact Lens for Amazon Connect supplies transcription, sentiment analysis and intent classification on every call.
• Automated call evaluation covers every call, well beyond the small sample supervisors could listen to by hand.
• CloudWatch dashboards show queue depth, handling times, bottlenecks and escalation patterns as they happen.

Supervisors use the conversation record and the evaluations to find the calls that need attention and coach their people on this morning's calls rather than last week's.