Voice AI for Roofing Contractors: Full Retell AI Setup Guide

A roofing contractor in the DFW area came to me with a straightforward problem: after every hailstorm, his phone rings non-stop for 72 hours. He misses about 60% of those calls. Each missed call is a potential $8,000–$25,000 job. His answering service costs $380/month and takes 4–8 hours to follow up. By then, the homeowner has already booked someone else.

This isn't unique to him. Roofing is one of the highest-CAC industries in home services because lead response time determines who wins the job. The contractor who calls back first closes 40–60% more deals. That's the entire game.

The solution wasn't more staff. It was a voice AI agent that answers every call instantly, asks the right qualification questions, captures lead data to the CRM, and books an inspection slot — all before the homeowner can dial the next contractor on Google.

I've tested all four platforms in production. Here's the honest breakdown for a home services contractor use case:

Retell wins for three reasons: the LLM websocket architecture gives direct control over the Claude system prompt without going through Retell's own prompt layer, the built-in Twilio integration eliminates separate phone infrastructure management, and the post-call webhook fires reliably on every call. With Bland I had webhook delivery failures on 2–3% of calls. For a roofing contractor at 200 calls/month during storm season, that's 4–6 lost leads per month — unacceptable.

Create an LLM Agent in Retell — not a Simple Agent. You need full conversation logic control. Key settings for roofing:

• Voice: ElevenLabs "Josh" or "Adam" — male voices test better for contractor brands in the South. "Rachel" works well for female-presenting brands.
• Interruption sensitivity: Set to 0.3 (low). Homeowners ramble when anxious about roof damage — don't cut them off.
• Turn detection threshold: 1000ms. Slightly longer pause detection prevents jumping in while the caller is thinking.
• Max call duration: 10 minutes. Anything longer is a transfer scenario.

Agent creation via the Retell API:

curl -X POST https://api.retellai.com/v2/create-agent \
  -H "Authorization: Bearer YOUR_RETELL_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "agent_name": "Alex — DFW Roofing",
    "voice_id": "11labs-Adam",
    "response_engine": {
      "type": "retell-llm",
      "llm_id": "YOUR_RETELL_LLM_ID"
    },
    "language": "en-US",
    "interruption_sensitivity": 0.3,
    "enable_backchannel": true,
    "max_call_duration_ms": 600000,
    "end_call_after_silence_ms": 8000,
    "reminder_trigger_ms": 15000,
    "reminder_max_count": 2
  }'

The enable_backchannel flag adds natural "mm-hmm" and "got it" responses while the caller speaks. This dramatically improves how human the agent sounds during longer explanations of damage — homeowners don't feel like they're talking at a void.

The system prompt is where most voice AI deployments fail or succeed. For a roofing contractor, the agent needs to handle four call types: storm damage, active leak, new roof quote request, and wrong number. It also needs hard guardrails — never quote prices, never promise timelines, always redirect to the inspection offer.

Here's the production system prompt I used:

You are Alex, a friendly scheduling assistant for [Contractor Name]
Roofing, a licensed contractor serving the DFW metroplex.

Your ONLY job: qualify the caller and schedule a free roof inspection.

## TONE
- Warm, confident, professional. Never robotic.
- Natural speech: "absolutely", "of course", "let me grab that for you"
- Mirror the caller's energy — calm if anxious, quick if rushed

## QUALIFICATION FLOW (one question at a time)

1. "Are you the homeowner at the property?"
   → If no: "No problem — can I get the best number for the homeowner?"

2. "What kind of roofing concern brought you to us today?"
   → Storm/hail → ask: "Did this happen in the last 12 months?"
   → Active leak → treat as urgent, offer next-day inspection
   → New roof → standard inspection flow

3. "What's the full address of the property?"

4. "Have you already filed an insurance claim, or are you still deciding?"
   → Filing: "Our inspectors document everything for the adjuster — no cost to you"
   → Undecided: "We can do a free inspection first and let you know if it's claim-worthy"

5. "What days and times work best for you this week?"
   → Offer: morning (8–12), afternoon (12–5), or specific day

6. "And the best number to confirm the appointment?"

## GUARDRAILS
- NEVER quote a price or cost estimate
- NEVER promise same-day service unless it's an active leak emergency
- NEVER discuss competitors
- Insurance questions: explain the free inspection, don't give legal/claim advice

## END OF CALL
After your closing statement ("Perfect — I've got you scheduled..."),
output this JSON block silently (do NOT read it aloud):
{
  "lead": {
    "is_homeowner": true,
    "damage_type": "storm|leak|new_roof|other",
    "property_address": "123 Main St, Plano TX 75023",
    "insurance_status": "filed|considering|none",
    "preferred_time": "Saturday morning",
    "callback_number": "+19725551234",
    "urgency": "high|medium|low"
  }
}

The JSON block is how I extract structured data from the conversation. Claude outputs it after the closing statement — Retell captures it in the transcript and my webhook handler parses it to create the CRM record.

Beyond the linear qualification flow, I designed four branch states the agent handles gracefully:

Transfer config for the escalation branch:

{
  "transfer_call": {
    "type": "warm_transfer",
    "phone_number": "+19725550001",
    "caller_id": "+19725550100",
    "handoff_message": "Hi, transferring a homeowner with
      an urgent roofing situation. Address: {{property_address}}.",
    "transfer_conditions": [
      "caller_requests_human",
      "urgency == high",
      "negative_sentiment_score > 0.8"
    ]
  }
}

Retell fires a call_ended webhook after every call. The payload includes the full transcript, call metadata, and dynamic variables. Here's the n8n Code node handler:

// n8n Code node — parses Retell post-call webhook
const body = $input.first().json;

// Extract JSON block from transcript
const transcript = body.transcript || '';
const jsonMatch = transcript.match(/\{[\s\S]*?"lead"[\s\S]*?\}/);

let leadData = {};
if (jsonMatch) {
  try {
    leadData = JSON.parse(jsonMatch[0]).lead;
  } catch (e) {
    // fallback to Retell dynamic variables
    leadData = {
      callback_number: body.retell_llm_dynamic_variables?.callback_number,
      property_address: body.retell_llm_dynamic_variables?.property_address,
      urgency: 'medium'
    };
  }
}

// Build GHL contact payload
const ghlContact = {
  phone: leadData.callback_number,
  address1: leadData.property_address,
  customFields: [
    { id: 'damage_type', value: leadData.damage_type || 'unknown' },
    { id: 'insurance_status', value: leadData.insurance_status || 'unknown' },
    { id: 'preferred_time', value: leadData.preferred_time || '' },
    { id: 'urgency', value: leadData.urgency || 'medium' },
    { id: 'call_id', value: body.call_id }
  ],
  tags: [
    'voice-ai-lead',
    `damage-${leadData.damage_type || 'unknown'}`,
    leadData.urgency === 'high' ? 'urgent' : 'standard'
  ],
  source: 'Retell AI Voice Agent'
};

return [{ json: ghlContact }];

From this n8n node, the workflow splits into two branches: one POSTs to the GHL Contacts API to create the lead record, the other sends an SMS confirmation to the homeowner via Twilio. The GHL contact creation:

curl -X POST https://services.leadconnectorhq.com/contacts/ \
  -H "Authorization: Bearer GHL_API_KEY" \
  -H "Version: 2021-07-28" \
  -H "Content-Type: application/json" \
  -d '{
    "locationId": "YOUR_GHL_LOCATION_ID",
    "phone": "+19725551234",
    "tags": ["voice-ai-lead", "damage-storm", "standard"],
    "customFields": [
      { "id": "damage_type", "value": "storm" },
      { "id": "insurance_status", "value": "considering" },
      { "id": "urgency", "value": "medium" },
      { "id": "call_id", "value": "retell_call_abc123" }
    ],
    "source": "Retell AI Voice Agent"
  }'

This contractor handles roughly 200 inbound calls/month during active season (March–June, October–November in Texas). Average call duration: 3.5 minutes.

The answering service also had a 4–8 hour follow-up lag. The AI agent creates the CRM record within 30 seconds of the call ending and fires the SMS confirmation while the homeowner is still in their driveway looking at their roof. Speed wins deals in this market.

Retell's call simulator lets you test the full conversation flow before forwarding the real number. I ran 40 test calls covering every branch state before launch. Issues caught in testing:

• Agent was cutting off callers who said "ummm" frequently — fixed by dropping interruption sensitivity from 0.5 to 0.3
• Claude was outputting the JSON block mid-conversation when it anticipated the call ending — fixed with explicit instruction: "output the JSON block ONLY after your closing statement"
• Transfer logic triggering on casual mentions of "the owner" unrelated to escalation — refined condition to require explicit request phrasing

For ongoing monitoring I built an n8n error workflow: if a Retell webhook arrives and the GHL contact creation fails, a Slack alert fires with the call transcript for manual recovery. In 30 days of production this triggered twice — both GHL API timeouts, not Retell failures.

The agent went live during a hailstorm week in DFW. Here's what happened:

The 11pm call story is worth telling. A homeowner in Plano called during a Thursday night storm. The AI agent qualified her, captured the address, and booked a Saturday morning inspection — at 11:23pm. The contractor closed a $4,200 repair job from that call. That's the ROI case in one data point: $73/month in AI costs captured a job the contractor would have missed entirely.