For the group conversation · Prepared by Will Robbins (GSM, Manor) · data as of 10/15/26
5/22 → 8/21 = 91 days, from the first Chloe message to Will's switch-off. No written trial, contract term or price is on file anywhere we can read. For 6 weeks (5/26–7/6) Chloe was the only AI on our leads, and nearly so through 7/27.
| Impel's own counts | to 6/23 (rows from 5/27) | July | 7/23–8/21 (Will's read) |
|---|---|---|---|
| Leads worked | 337 | 542 (400 new) | 562 |
| Messages | 2,060 | 4,134 (69 bounced) | 4,166 |
| New leads engaged | 101 (30%) | 113 (28%) | 86 |
| Handoffs to a person | 126 | 126 (41 call requests) | n/a |
| AI-set appointments | 4 | 3 | 5 |
| Sales Impel credits itself | 0 AI-assisted | tile blank | 0 |
"542 worked" is nearly every Internet lead in the CRM that month (real July Internet leads ~425 after removing a 2,081-record Ford mailing import). A 6/30–8/17 API pull agrees: 616 worked, 171 handoffs, 6 appointments. Will's 7/23–8/21 read was recorded in the 8/25 comparison, not re-pulled.
Sold = name match to VS / Axcessa; a handoff that sold may have been sold by a rep who never saw the handoff.
| Full census, VS comm history | Chloe | Our machine |
|---|---|---|
| Conversations we could verify in VS | 401 (of 600 Impel claims) | 556 |
| Customer replied (generous) | 7.0% | 10.4% |
| Replied to that AI (strict) | 4.2% | 10.3% |
| Real back-and-forth, 2+ reply days | 2 | 2 |
| Sold in the window | 0 | 2 |
Strict = the customer's last message before the reply was that AI's, and the other AI hadn't sent anything in the 15 minutes before. 183 of Impel's 600 claimed conversations matched no VS lead at all. An earlier 41-lead sample said Chloe got 71% replies; it was pre-filtered toward engaged leads and is withdrawn.
| Period (lead created) | Which AI | Leads | Reached ≤1h (VS clock) | After-hours first touch, median | Person ≤1h | Appt set | Shown | Sold |
|---|---|---|---|---|---|---|---|---|
| Jan–Mar 2026 | neither (people only) | 1,222 | n/a | n/a | thin notes | 12.0% | 7.0% | 11.3% |
| 4/7–5/25 | ours alone | 586 | n/a | n/a | 55.9% | 16.2% | 11.4% | 12.3% |
| 5/26–7/27 | Impel alone (ours paused) | 720 | 94%* | 4 min* | 43.6% | 17.8% | 11.7% | 12.4% |
| 7/28–8/21 | both | 289 | 95.5% | 4 min | 37.6% | 16.3% | 9.3% | 11.4% |
| 8/22–9/12 | ours alone (Chloe off) | 289 | 70.2% | 241 min | 42.8% | 18.7% | 11.8% | 12.8% |
| 9/13–10/13 too young | ours (+ Impel's 10/7–10/9 burst) | 305 | 83.3% | 12 min | 43.6% | 13.4% | 8.9% | 9.2% |
Same months last year (people only): Jan 7.3 · Feb 6.7 · Mar 9.5 · Apr 9.6 · May 9.8 · Jun 9.6 · Jul 10.2 · Aug 10.8. September 2026 leads are still closing.
Jan 11.2 · Feb 12.2 · Mar 12.8 · Apr 17.6 · May 15.8 · Jun 18.4 · Jul 16.7 · Aug 16.9 · Sep 17.6
Jan 6.4 · Feb 7.9 · Mar 6.9 · Apr 12.6 · May 10.6 · Jun 11.7 · Jul 10.4 · Aug 12.1 · Sep 9.5
| Internet leads Sep 1–Oct 12 (census) | Austin | Buda | Manor |
|---|---|---|---|
| Leads read | 743 | 738 | 395 |
| Impel first touch on the lead | 99.3% | 96.1% | 7.1% (off) |
| Impel median minutes, all hours | 3.7 | 3.8 | 4 |
| …after hours | 4.6 | 4.5 | 3.3 |
| Impel leads with 2+ messages logged | 46.5% | 45.8% | 75% |
| Impel lead never touched by a person | 65 | 3 | 0 |
Personas: Emma Reyes (Austin), Ella Moore (Buda), Chloe Davis (Manor). Customer replies to Impel are not logged in VS at either store (0 before any staff touch), so Impel's reply rate there is unavailable.
| Austin | Buda | Manor* | |
|---|---|---|---|
| Median minutes, all hours | 164.2 | 41.3 | 128 person-only (5.6 if our machine's texts count) |
| …leads created 8a–8p | 61.6 | 15.1 | 3 |
| …leads created after hours | 672.6 | 617.8 | 16 h for 5–9pm arrivals, person-only |
| Staff touch within 1h | 34% | 53% | 42% person-only (77% blended) |
| No staff touch in 24h (or ever) | 24.4% | 5.8% | 1% |
| Customer replied after first staff outbound | 29% | 37% | 55% |
*Manor blends our machine with staff (machine texts post under user names); Austin and Buda exclude Impel. Manor person-only: 14.5 min median for leads landing 8:30–5, 104 overall (nights and evenings). Not like-for-like on speed; the 24h and reply rows are the fair read.
With a bot in front, the VS response tile stops measuring people. Judge staff speed on comm history, store by store, whatever we decide about Impel.
| Per 100 leads, last 30 days (all types) | Manor | Austin | Buda |
|---|---|---|---|
| Appointments set | 28.3 | 19.7 | 14.9 |
| Shown | 17.1 | 9.1 | 9.0 |
| VS sold | 17.8 | 19.7 | 13.4 |
Counts: 62/498 · 110/888 · 59/852. All non-bulk leads: 16.1% · 20.5% · 13.5% (Manor carries repeat-customer leads, Buda service referrals).
The data can't say Impel helps or hurts closing, at any store. What it does show is speed and coverage. What decides the sale is what people do after the first touch.
| Per store, per month | Austin | Buda | Manor (July, live) |
|---|---|---|---|
| Internet leads Impel first-touched | ~539 | ~511 | 542 claimed · 263 verified |
| Cost per lead touched | ~$5.60 | ~$5.90 | $5.50 claimed · $11.40 verified |
| AI-set appointments | dashboard not read | dashboard not read | 3 (≈ $1,000 each) |
| Sales it can show it caused | 0 shown | 0 shown | 0 |
Austin / Buda leads = last 30 days of Internet leads × Impel coverage (X §1, §4). No Impel invoice, contract amount, term or renewal date is on file anywhere we can read; VS lead-source ROI has Impel's cost set to $0. The "$3,000" on Impel's own report page is estimated gross on its one sale, not a fee.
Neither AI has a measured cost per sale. Formula once both bills are on file: monthly cost ÷ shown appointments it touched.
| Impel (Austin, Buda; Manor until 8/21) | Manor's machine | |
|---|---|---|
| Speed, store hours | 3.4–3.6 min median | 3 min first outbound (blended); first-touch 4.6 min |
| Speed, after hours | 4.5–4.6 min, texts any hour (one at 11:30pm) | Email in ~12 min overnight (median, 9/13–10/13); texts wait for the 8am window; staff first outbound 94 min. Since 10/13, customers who write back 9pm–9am get an answer in ~5 min |
| Coverage | 96–99% of Internet leads | 81% (by design skips leads a rep is already working, recent buyers, credit apps) |
| Replies | 4.2% strict at Manor (Jul–Aug census); not logged in VS at Austin/Buda | 10.3% strict, same census; first-touch 23.6% attributable |
| Appointments | 3–6 AI-set a month at Manor; sister stores unread | 7 strict "assisted" in 90 days; setter not recorded in VS |
| VS clock | Stops it at Austin (tile looks great); not at Buda | Doesn't stop it |
| Errors seen | STOP not honored for 5 days · quoted $1,232 over the feed · deposit-hold answer · same copy by text and email at once · 23 messages to one silent customer · notes cut off at ~450 characters | Told 8 customers a sold unit was "still here" (14 sends; sold-unit gate added 10/15, earlier sends being corrected) · cadence inside live rep threads 26.6% → 0 after the 10/14 gate · trade pitch after "no trade" 2 of 1,331 |
| Guards | No master off, no quiet hours, no daily cap. Per-source switches (it texted anyway). 13-step cadence to day 51 | Opt-out and STOP block · 120-day recent-buyer gate · no holds or deposits · no rates or negotiation (0 breaches in 391 sends since 10/10) · never names a sister store · steps back when a person is talking · quiet hours |
Both AIs on the same lead is the worst case: 78 near-simultaneous double texts at Manor in Jul–Aug and 4 duplicate customers in Sept–Oct. Whatever the group decides: one AI per lead.
Will's call stands: Manor stays off Impel (8/21, 10/13). Still owed by Impel, in writing: lead nurture off account-wide · billing end date and last invoice · final data export · Impel media off our vehicle pages. Website chat ("Chat-Impel") still sends leads (latest 10/14): keep it or cancel it?
The sister-store call is the group's.
| Setup, 90 days | Leads per arm | Smallest difference it can see |
|---|---|---|
| Austin only | ~815 | Shown 9 → 13 per 100 · close 12.4% → 17.4% |
| Austin + Buda | ~1,610 | Shown 9 → 12 per 100 · close 12.4% → ~16% |
80% power, 5% false-alarm rate. Internet leads, last 30 days: Austin 543, Buda 532. A 1–2 point sales difference needs 6+ months.