We analyzed every winning and losing roofing sales call across 13 campaigns. Then we built an AI voice agent that books appointments like the best human closer on the team.
Most AI voice agents are built from scripts someone wrote in a conference room. Ours was built from what actually works on the phone -- extracted from thousands of real conversations with real homeowners.
Every outbound call from 13 roofing campaigns across Alabama, Georgia, Nevada, and Texas. Wins, losses, hang-ups, do-not-calls, and wrong numbers -- all of it.
212 booked appointments, 424 losses over 40 seconds, and 76 edge cases. Each transcript enriched with objection tags, persona labels, timing markers, and outcome data.
The #1 closer -- 5.3% contact booking rate, 100% close rate in our 9-call deep sample -- was analyzed call by call. Every technique, every deviation from script, every recovery catalogued.
Full scorecards comparing opening style, close technique, objection handling, call duration, and booking efficiency across all agents and all four campaigns.
Every objection found in 10,794 calls, ranked by frequency and conversion rate. Winning and losing responses documented with exact quotes from real calls.
From "Easy Yes" to "Skeptic" to "Not-Interested Flipper" -- each persona identified with detection signals, recommended responses, and expected call duration benchmarks.
We read every booked appointment and hundreds of failures. The patterns are unmistakable -- and they have nothing to do with what's written in the script.
66.5% of all booked appointments fall in the 2-5 minute window. If you keep a homeowner on the phone past 60 seconds, booking probability jumps dramatically -- 98.6% of wins last past that threshold.
The top agent booked at 5.3% -- 2.65x the rate of the worst performer on the same lists and campaigns. We broke down every technique that made the difference, then taught the AI to replicate all of them.
If every agent performed at the top closer's rate, the team would generate 50% more appointments from the same call volume. The AI runs at the top closer's level, every single call.
82% of booked calls contained at least one objection. The difference between winning and losing is not whether objections appear -- it's how the agent responds in the 5 seconds after.
| Objection | Frequency | In Wins | In Losses | Book Rate | Verdict |
|---|---|---|---|---|---|
| Cost/Price Concern | 234 | 101 | 133 | 43.2% | Winnable with insurance pivot |
| "Not Interested" | 200 | 21 | 179 | 10.5% | Hardest -- 9 in 10 are lost |
| "No Damage" | 110 | 52 | 58 | 47.3% | Coin flip -- response decides it |
| "Roof Is Fine" | 105 | 37 | 68 | 35.2% | Winnable with age-based reframe |
| "Not Right Now" | 98 | 58 | 40 | 59.2% | Best conversion objection |
| "Who Is This?" | 57 | 20 | 37 | 35.1% | Trust gate -- fumbled intros kill |
| "Too Busy" | 29 | 13 | 16 | 44.8% | Winnable with reschedule |
| Warranty Questions | 21 | 10 | 11 | 47.6% | Buying signal, not objection |
| Spouse/Partner | 15 | 12 | 3 | 80.0% | Highest-converting objection |
| "Do Not Call" | 13 | 2 | 11 | 15.4% | Terminal -- comply immediately |
Top closers don't just book the appointment -- they qualify the lead with a systematic confirmation that creates psychological commitment and reduces no-shows.
The AI detects prospect personality within the first 15 seconds and adapts its pacing, tone, and close strategy accordingly. Two-thirds of bookings come from just two persona types.
These aren't projections or simulations. These are actual performance metrics extracted from 13 campaigns across 4 states.
| Agent | Total Calls | Contacts | Booked | Booking Rate | Avg Win Duration |
|---|---|---|---|---|---|
| Top Agent (637) | 1,544 | 806 | 43 | 5.3% | 284s (4m 44s) |
| Agent 240 | 2,268 | 864 | 43 | 5.0% | 204s (3m 24s) |
| Agent 636 | 1,464 | 767 | 36 | 4.7% | 174s (2m 54s) |
| Agent 690 | 1,403 | 812 | 34 | 4.2% | 189s (3m 9s) |
| Worst Agent (158) | 1,640 | 982 | 20 | 2.0% | 226s (3m 46s) |
| Metric | Booked (APPTBK) | Not Interested (NI) | Hang Up (HU) |
|---|---|---|---|
| Mean duration | 215 seconds | 34 seconds | 27 seconds |
| Median duration | 192 seconds | 30 seconds | 20 seconds |
| Calls past 60 seconds | 98.6% | 6.7% | 2.5% |
| Calls in 2-5 min sweet spot | 66.5% | 0% | 0% |
No text-to-speech chains or middleware layers. Audio is processed natively at the model level, delivering natural conversation with sub-200ms response times.
Audio goes directly into the model -- no speech-to-text, no text-to-speech pipeline. Eliminates the 300-500ms latency penalty of chained architectures. The voice sounds human because it is processed as sound, not text.
Not a decision tree. The AI understands conversation context, detects objection types in real-time, and selects the highest-converting rebuttal from 20 mapped objection categories with proven win/loss responses.
DNC detection within 2 seconds. Automatic removal. AI disclosure when asked. Deceased-person handling. Vulnerable population protection. Voicemail detection. TCPA-safe by design.
Detects Spanish within 2 seconds and delivers the full pitch in Spanish without transferring. Same structure, same objection handling, same close technique -- just in the prospect's language.
Traditional call centers charge $15-25/hour per agent and produce inconsistent results. The AI runs at top-closer level 24/7 at a fraction of the cost.
| Traditional Call Center | Offshore Call Center | AI Voice Agent | |
|---|---|---|---|
| Cost per hour | $18-25/hr | $8-12/hr | ~$0.10/min |
| Consistency | Varies by mood, fatigue, day | High turnover, accent issues | Same top-closer performance every call |
| Training time | 2-4 weeks per agent | 2-6 weeks per agent | Pre-trained on 10,848 real calls |
| Objection handling | Depends on agent skill | Often script-dependent | 20 types mapped with proven rebuttals |
| Compliance | Human error risk | Higher risk, less oversight | Built-in DNC, disclosure, edge-case handling |
| Scalability | Linear -- hire more agents | Linear -- hire more agents | Instant -- spin up 100 concurrent calls |
| Response time | ~500-800ms natural | ~500-800ms + latency | <200ms turn-by-turn |
| Bilingual | Requires separate team | Usually available | Automatic detection + seamless switch |
| Data capture | Manual CRM entry, often incomplete | Manual, error-prone | 100% automated -- every field, every call |
| Works weekends/evenings | Overtime rates | Overtime rates | 24/7, same rate |
Extracted from comparing every booked appointment against every failed call. These rules are baked into every AI call.
See the AI voice agent in action. We'll run a live demo call, walk through the data behind every decision it makes, and build a custom deployment plan for your campaigns.