I built an AI calling agent for a US-based HVAC service using GPT for conversation logic, ElevenLabs for text-to-speech, and Retell AI for telephony - achieving a 20% improvement in user satisfaction and significantly shorter support call durations.
HVAC companies field a high volume of phone calls - appointment scheduling, service status checks, and emergency triage. Many of these calls follow predictable patterns, making them ideal candidates for voice AI. The challenge is not automation alone; it is automation that customers actually prefer over waiting on hold.
The goal: sound human, resolve faster
The client did not want a robotic IVR menu. They wanted a conversational agent that could handle natural speech, remember context within a call, and resolve common requests without frustrating customers.
Architecture and integrations
GPT for conversation intelligence
GPT handles intent recognition, multi-turn dialogue, and response generation. Prompt engineering accounts for HVAC-specific terminology, common customer scenarios, and graceful escalation when the agent reaches its capability boundary.
ElevenLabs for voice quality
Text-to-speech quality makes or breaks phone agents. ElevenLabs provides natural prosody and pacing. I tuned voice parameters to match the brand tone - professional, warm, and efficient.
Retell AI for telephony
Retell AI manages the phone infrastructure - call routing, audio streaming, and latency optimization. Low latency is critical; delays over 500ms make conversations feel unnatural.
Make.com for orchestration
Make.com connects the pipeline: call events trigger GPT processing, responses flow to ElevenLabs for synthesis, and outcomes sync to scheduling systems. This no-code orchestration layer accelerated iteration during tuning.
Prompt engineering for phone conversations
Phone conversations differ from chat. Callers interrupt, change topics mid-sentence, and use informal language. The prompt system includes repair strategies ("Sorry, I did not catch that - could you repeat the address?"), confirmation loops for scheduling, and explicit escalation triggers for emergency HVAC situations.
Results
- 20% improvement in user satisfaction scores.
- Significantly shorter average call durations.
- Natural, engaging conversational style praised by callers.
- Reduced load on human agents for routine inquiries.
- Faster appointment scheduling during peak season.
Lessons for industry-specific voice AI
- Domain-specific prompts outperform generic assistants - HVAC has its own vocabulary.
- Always define escalation paths for emergencies and edge cases.
- Measure satisfaction, not just call deflection rate.
- TTS voice selection is a branding decision, not a technical default.
- Start with the top three call types and expand coverage iteratively.
Common questions
What does the HVAC AI calling agent do?
The AI calling agent handles inbound and outbound phone conversations for a US HVAC service - answering customer questions, scheduling service calls, and resolving common inquiries without requiring a human agent for every call.
What AI technologies power the HVAC voice agent?
The agent uses GPT for natural language understanding and response generation, ElevenLabs for realistic text-to-speech voice output, Retell AI for telephony infrastructure, and Make.com for workflow orchestration.
What results did the HVAC AI agent achieve?
The solution delivered a 20% improvement in user satisfaction and significantly reduced support call durations compared to the previous manual phone handling process.
How do you make an AI phone agent sound human?
Human-like phone agents require three things: sophisticated prompt engineering that accounts for conversational fillers and repair strategies, TTS voice selection and tuning matched to the brand, and context retention so the agent remembers what was said earlier in the call.
Is voice AI suitable for HVAC customer service?
Yes. HVAC services receive high volumes of repetitive calls - scheduling, status checks, and basic troubleshooting. AI agents handle these efficiently while escalating complex issues to human technicians.