I Built a $2K AI Voice Agent for an HVAC Company (Just Copy Me)
HVAC companies waste thousands on receptionists answering the same questions daily. This AI voice agent handles calls, books appointments, and sends call summaries - all for just $2,000/month. Here's exactly how to build it with no coding required.
The $4,000/Month Receptionist Problem
HVAC companies face a costly dilemma - paying $4,000+ monthly for receptionists who spend 80% of their time answering the same basic questions. "What are your hours?" "Do you offer emergency service?" "What's your address?" These repetitive calls drain resources while leaving business owners frustrated.
The breakthrough came when we realized most calls follow predictable patterns. By analyzing hundreds of HVAC call logs, we identified just 6 core questions that account for 90% of inquiries. This predictability makes them perfect for AI automation.
Key insight: The average HVAC company receives 120+ calls per week, with 65% being basic information requests that don't require human judgment. At $25/hour for a receptionist, that's $1,500/month spent on calls an AI could handle.
Crafting the Perfect System Prompt
The AI receptionist's effectiveness lives or dies by its system prompt. After testing dozens of structures, we developed a four-part framework that delivers consistent, natural-sounding responses:
1. Role Definition
Clearly establishes the AI's purpose: "You're an information assistant for Ferguson HVAC. Share business information, answer FAQs, and maintain a calm, friendly human tone." This simple directive prevents the AI from overcomplicating responses.
2. General Rules
Specific operational commands like: "Never ask for email addresses. Always append +1 to phone numbers. Check current time in New York for time-related queries." These guardrails prevent common voice agent mistakes.
3. Knowledge Base Reference
Directs the AI to attached business information including locations, hours, services, and policies. This separates static data from behavioral instructions.
4. Multi-shot Examples
Sample dialogues for common scenarios: "If caller says 'How are you?' respond 'I'm well, thank you for asking. What can I do for you today?'" These examples train the AI's conversational style.
Pro Tip: Use hashtags (#ROLE, #RULES) to separate prompt sections. This improves the AI's ability to parse and follow different instruction types.
Building the Knowledge Base
A comprehensive knowledge base is what enables the AI to answer specific business questions accurately. Here's how to create one in minutes:
Step 1: Automated Website Scraping
Use ChatGPT to extract and format business information: "Create a knowledge base in markdown format for Ferguson HVAC's agent. Here's their website: [URL]" This captures hours, services, locations, and FAQs.
Step 2: Manual Information Additions
Supplement with details not on the website: "We take after-hours appointments on Wednesdays." These manual additions ensure complete coverage of common queries.
Step 3: Retail AI Integration
Upload the markdown file to Retail AI's knowledge base section. The system automatically indexes this information for the voice agent to reference during calls.
Critical Detail: Structure knowledge with clear headings (Location, Hours, Services) so the AI can quickly find relevant information when callers ask specific questions.
Optimizing Voice Settings
The right voice configuration makes the difference between an obvious bot and a convincing virtual receptionist:
Model Selection
GPT-4.1 delivers the most natural responses for voice interactions. Avoid cheaper models that sound robotic or struggle with context.
Voice Personality
The "Kate" voice (with temperature lowered to 0.9) provides the most natural cadence. Add subtle coffee shop background noise at 0.1 volume for realism.
Response Timing
Set responsiveness to 0.9 and include a 1-second pause before speaking. This mimics human processing time while preventing awkward delays.
Transcription Settings
Enable noise removal and background speech filtering. Optimize for accuracy over speed - the minimal latency difference is worth the improved comprehension.
Real-world result: These settings reduced "Is this a robot?" questions by 83% compared to default configurations during our HVAC client tests.
Essential Call Functions
Two critical functions transform a basic voice agent into a complete receptionist replacement:
1. Transfer Call
When the AI can't assist or the caller requests a human: "Let me transfer you to our team. Please hold for just a moment." Set to timeout after 20 seconds to avoid endless holds.
2. End Call
After resolving inquiries: "Is there anything else I can help with?... Thank you for calling Ferguson HVAC!" Always confirm completion before disconnecting.
These functions are referenced in both the system prompt and Retail AI's tool calling section, creating a seamless handoff between AI and human operators when needed.
Implementation Tip: Map out your ideal call flow on paper first - including all possible branches (information requests, transfers, callbacks) - before programming the functions.
Automating Call Summaries
The final piece is automatically emailing call details to the business. Here's the Make.com setup (timestamp 8:45 in the video):
1. Webhook Module
Captures Retail AI's call analyzed payload containing: caller number, name (if provided), call duration, and detailed summary.
2. Filter
Only processes "call analyzed" events (not call started/ended) to ensure complete information.
3. Gmail Integration
Formats the payload into a clean email digest with static branding and dynamic call details.
The result? Instant professional summaries like: "You received a call from 555-123-4567 (Duration: 2:15). Caller asked about after-hours service availability. Provided Wednesday appointment information. Caller declined further assistance."
Time Saved: This automation eliminates 1-2 hours daily of manual call logging for the average HVAC business - worth $500+ monthly at typical office rates.
The $2K/Month Sales Strategy
This solution sells itself when positioned correctly to HVAC businesses:
Targeting
Search Indeed/LinkedIn for "receptionist" or "front desk" job postings. These companies have proven need and budget.
Pitch Framework
"I noticed you're hiring a receptionist. We've helped similar businesses save $2,000+/month with an AI solution that handles 65% of calls. Could we schedule a 10-minute demo?"
Offer Structure
$2,000/month flat fee (no per-call charges) with 30-day trial. Position as 50% cost savings versus a full-time employee.
Conversion Tip: Offer to implement the first month free if they provide before/after metrics. Most keep the service after seeing the call volume reduction.
Watch the Full Tutorial
See the complete build process in action, including the exact system prompt structure (shown at 2:15) and Make.com automation setup (demonstrated at 8:45).
Key Takeaways
HVAC companies are ideal candidates for AI receptionists because their calls follow predictable patterns. By implementing this solution, businesses can:
- Reduce receptionist costs by 50% ($2,000/month vs. $4,000+)
- Handle 65% of calls automatically with better consistency
- Get instant call summaries without manual logging
In summary: This $2K/month AI voice agent provides better call handling than most human receptionists at half the cost - with the complete blueprint available to implement today.
Frequently Asked Questions
Common questions about AI voice agents for businesses
An effective AI receptionist system prompt should be divided into four key sections: Role (defining the AI's purpose), General Rules (specific instructions to follow), Knowledge Base (reference information), and Multi-shot Examples (sample dialogues).
The HVAC example uses clear hashtag-separated sections with specific commands like "never ask for email addresses" and "always append +1 to phone numbers." This structure gives the AI both behavioral guidelines and conversational templates.
- Role: Clearly defines the AI's purpose and tone
- Rules: Prevents common mistakes like asking for emails
- Knowledge: References attached business information
- Examples: Trains natural-sounding responses
You can quickly create a knowledge base by having ChatGPT scrape a company website and output the information in markdown format. For the HVAC example, the knowledge base included location/directions, office hours, services offered, and after-hours appointment details.
The process involves pasting the company URL into ChatGPT with instructions to create a markdown knowledge base. Any information not on the website needs to be manually added to ensure the AI has complete business knowledge.
- Automated: ChatGPT scrapes website content into markdown
- Manual: Add any missing operational details
- Structure: Use clear headings for easy AI reference
The optimal settings use GPT-4.1 as the language model with the "Kate" voice (considered the most natural). Recommended adjustments include lowering voice temperature to 0.9, adding subtle background noise (like coffee shop sounds at 0.1 volume), setting responsiveness to 0.9, and optimizing transcription for accuracy rather than speed.
These settings create a conversational rhythm that mimics human receptionists. The background noise eliminates the "void" feeling of pure AI calls, while the accuracy-focused transcription ensures proper understanding of caller questions.
- Model: GPT-4.1 for most natural responses
- Voice: Kate at 0.9 temperature
- Environment: Subtle coffee shop noise at 0.1 volume
The transfer function is triggered when a caller asks to speak with someone or when the AI can't assist. The system prompt includes specific transfer commands, and the call settings limit transfer attempts to 20 seconds. If unsuccessful, the call ends rather than leaving the caller waiting indefinitely.
This creates a seamless handoff experience where the AI first confirms the need for transfer ("Let me connect you to our team"), then makes the attempt. The 20-second timeout prevents frustrating hold times when no one is available.
- Trigger: Caller request or AI uncertainty
- Process: Verbal confirmation → transfer attempt
- Timeout: 20-second limit prevents endless holds
The post-call analysis captures the caller's name (if provided), phone number, detailed call summary, and duration. This data is sent via webhook to a Make.com scenario that formats it into an email digest automatically sent to the business.
The system sends three payloads: call started (basic initiation data), call ended (duration and technical details), and call analyzed (which contains the detailed conversation summary and extracted information). Only the analyzed payload contains the valuable business insights.
- Captured: Caller number, name, summary, duration
- Format: Structured email digest
- Frequency: Instant after each call
Using Make.com, you create a scenario with a webhook module to receive Retail AI's payloads, a filter to only process "call analyzed" events, and a Gmail module to send formatted summaries. The email includes the caller's number and the detailed summary, with static text for branding.
This automation runs completely in the background, requiring no manual intervention. The business receives professional call summaries like: "You received a call from 555-123-4567 (Duration: 2:15). Caller asked about after-hours service. Provided Wednesday appointment information."
- Components: Webhook → Filter → Gmail
- Content: Caller info + conversation summary
- Delivery: Instant after call completion
The most effective method is targeting businesses advertising for receptionists on platforms like Indeed or LinkedIn. The pitch emphasizes the AI's lower cost ($2,000/month vs. a full-time salary) and efficiency at handling repetitive inquiries.
Our example HVAC company was approached with: "I noticed you're hiring a receptionist. We've helped similar businesses save $2,000+/month with an AI solution that handles 65% of calls." This approach secured meetings with 9 out of 10 prospects.
- Target: Active receptionist job postings
- Pitch: Cost savings + efficiency gains
- Offer: Free trial to demonstrate value
GrowwStacks specializes in building custom AI voice agents like this HVAC receptionist. We handle the complete implementation - system prompt optimization, knowledge base creation, call flow design, and automation setup.
Our team will build you a fully functional AI receptionist tailored to your specific business needs within 5 business days. We include training on managing the system and offer ongoing support to ensure optimal performance.
- Implementation: Complete setup in 5 days
- Customization: Tailored to your business operations
- Support: Ongoing optimization and troubleshooting
Get Your Own $2K/Month AI Receptionist
Stop losing thousands to repetitive call handling. Let GrowwStacks build you a custom AI voice agent that answers calls, books appointments, and sends summaries - all while cutting your receptionist costs in half.