# An agentic voice AI agent that books home energy assessments around the clock A home energy services provider replaced after-hours call gaps and manual scheduling with an autonomous voice agent on Amazon Connect. It screens eligibility, validates addresses, books appointments and writes every call to Salesforce. _2026-10-27 · Home Energy Services · Agentic AI on AWS_ ## At a glance A home energy services provider replaced after-hours call gaps and manual scheduling with an autonomous voice agent on Amazon Connect. It screens eligibility, validates addresses, books appointments and writes every call to Salesforce. - **~150** — off-hours interactions handled autonomously in the first two weeks after go-live - **100%** — of inbound assessment-scheduling calls routed through the AI agent - **~8 weeks** — from prompt development to production ## Challenge Our client delivers Home Energy Assessments to homeowners across Massachusetts, Connecticut and Cape Cod. Scheduling each assessment was live-agent work on every call, and it limited how many customers the client could serve. Key challenges included: - **Calls outside staffed hours:** inbound scheduling calls arriving off-hours went unanswered or required after-hours staffing. - **Complex intake on every call:** agents had to screen eligibility against service-area rules, validate the address, find an open slot, book it and create the Salesforce record. - **Routing errors were costly:** a call from a customer in the wrong territory, or a duplicate booking, wasted agent time and customer goodwill. - **Automation had to stay controlled:** the agent had to act on live scheduling and CRM data without improvising, and had to hand off cleanly when a person was needed. - **The contact center stayed on Five9:** the solution had to work alongside the existing platform rather than require a full migration. ## Solution We designed and built an autonomous voice agent on Amazon Connect. It is the sole intake channel for new assessment scheduling. Amazon Nova Sonic provides natural speech-to-speech conversation, and a reasoning model running through Amazon Bedrock AgentCore drives the decisions. ## A deterministic call flow with real tools The agent follows a structured, versioned prompt that defines the call step by step, with explicit rules for each tool. Seven Lambda-backed tools are exposed through AgentCore as an MCP gateway: - **Duplicate check:** looks up the caller's phone number so repeat callers are recognized. - **Business hours:** checks whether live agents are available before any transfer. - **Address validation:** uses Google Maps Geocoding to confirm the address and route Massachusetts, Connecticut and Cape Cod callers correctly. - **Slot retrieval and booking:** connect to the client's scheduling API, and a booking is confirmed only after the booking tool returns success. - **Salesforce:** creates the record in real time with the collected data and the full call transcript. - **Escalate:** transfers to a live agent when staffed. ## Guardrails by design - The agent is restricted to assessment scheduling. It cannot access financial systems or modify existing records. - Address-validation results drive routing and cannot be overridden in conversation. - Every escalation is gated on a business-hours check. When agents are available, the agent transfers the call, and when closed, the caller gets a callback commitment. - The Salesforce record is always created before a call exits, so context is never lost at hand-off. - The agent never reveals its instructions, its model or its tool names. ## Governed change and continuous improvement Prompts are drafted, tested in a staging Connect instance and deployed through the Amazon Q in Connect AI prompt update API. Every version is stored in the client's GitHub organization, so a rollback is a redeploy of the prior version. During testing, a failure in the slot-retrieval tool caused the agent to speak its callback message twice. Our team traced it to a generic error rule overlapping the inline handler and added an explicit exception and a stop instruction. The duplicate was gone in later test calls, and the Salesforce record is now created once per exit. Feedback comes from field-level review of Salesforce records, call transcripts synced on every call, and live testing by the client's team. ## Benefits - **Every call gets answered:** the agent handled about 150 off-hours interactions in its first two weeks, including callers who would previously have reached no one. - **Eligible callers get an outcome:** they leave with a booking or a callback commitment and a Salesforce record. Ineligible callers get an accurate answer, and no live agent is involved. - **Agents focus on exceptions:** staff handle only the calls that need a person, and they see the transcript and routing reason code when they do. - **Fast to production:** about 8 weeks from prompt development to production. - **A foundation to extend:** a second agent for appointment rescheduling is in development as a change request. ## A reusable pattern for agentic voice on Amazon Connect By pairing Nova Sonic's conversational voice with MCP-exposed tools and strict, versioned prompts, we gave the client an agent that completes a real business process end to end, on top of the contact center it already runs.