The shop with the architected AI receptionist stack from Lesson 1 has its inbound topology working: Tier One AI-handles 75% of calls, Tier Two warm-transfers 12%, Tier Three pages on emergencies, the 11-field handoff lands at 96% completion, and after-hours capture climbs to 70% by month two. The shop still loses calls. After-hours overflow at 9:14 p.m. that the AI's confidence threshold rejected to the recovery queue. In-hours abandoned calls — the homeowner who waited 47 seconds on hold and hung up at 10:32 a.m. Voicemail-tag calls where the homeowner left a message and the morning CSR did not call back within 4 hours. Hold-abandonment spikes during a 1:00 p.m. CSR break gap. These are the calls Lesson 1's architecture surfaces; this lesson is the named workflow that recovers them. The lost-call recovery loop is the daily AI report on missed, abandoned, and lost calls, the CSR re-outreach playbook against each kill reason, the AI-drafted follow-up text library, and the cadence that converts the residual 20-25% leak into a 60-80% recovered-call workflow. Target: recover 60-80% of after-hours leakage past the AI receptionist's primary capture, recover 25-35% of in-hours abandoned calls, and lift the shop's effective booking floor from the architecture's 75% to a recovery-augmented 85-88% inside 90 days. The recovery loop is not optional — it is the second half of the L3 Ch2 workflow that turns architecture into compounded revenue.

Why the Recovery Loop Is the Second Half of the Receptionist Architecture

The L3 service manager who completes Lesson 1's architecture without Lesson 2's recovery loop captures 70-75% of inbound and stops there. The remaining 25-30% is structurally invisible because the AI receptionist captured the primary attempt, the metrics dashboard shows "booking %" rising, and the shop's owner concludes the AI deployment is complete. Inside that residual 25-30% sits the $200K-$450K of annualized leak per 7-truck residential shop that the recovery loop is built to capture. The leak surface has four sources, each with its own kill-reason signature.

First, the AI's confidence-threshold rejections — calls where the AI's handling confidence dropped below the architect's 78-82% threshold and the call routed to warm transfer, but Tier Two pickup failed or the receiving CSR booking-% on that variant landed at 50-55%. Second, in-hours abandoned calls — the homeowner who waited 30-47 seconds on hold and hung up because the CSR row was on other calls. Third, voicemail-tag — the homeowner who left a message on the rare voicemail backstop, did not get a callback within 4 hours, and called the next shop. Fourth, hold-abandonment spikes — recurring volume spikes during shift gaps (the 1:00 p.m. CSR break gap, the 7:30 a.m. dispatch-board reset, the 5:00 p.m. shift-change) where the AI is live but the warm-transfer destination is structurally understaffed. The recovery loop instruments all four sources, surfaces them on a daily AI report, and runs the re-outreach against each one with kill-reason-specific text and callback discipline.

The math defends the build. At a 7-truck residential shop with 240-310 weekly inbound, the architecture's 75% Tier-One capture lands ~180-230 booked calls. The remaining 60-80 calls split: 20-28 after-hours overflow past primary capture, 18-25 in-hours abandoned, 10-15 voicemail-tag, 12-18 hold-abandonment-spike. The recovery loop captures 60-80% of after-hours overflow, 25-35% of in-hours abandoned, 55-70% of voicemail-tag, 35-50% of hold-abandonment-spike — net 28-42 recovered calls/week. At $387 average ticket: $11K-$16K/week, $565K-$830K annualized. Structurally additive to the architecture's primary capture and earns its own ROI line on the dashboard.

The Daily AI Report on Missed, Abandoned, and Lost Calls

The recovery loop opens with the daily AI report. At 7:00 a.m. each morning, the AI compiles the prior 24-hour log of lost calls, segments by kill-reason, surfaces the recovery priority list, and pushes the report to the CSR-floor lead's screen before the shift starts. The report is built from CallRail Conversation Intelligence, the AI receptionist's post-call audit log (Avoca, Jobber AI Receptionist, HCP AI Agents, or ServiceTitan Voice depending on stack), and the FSM platform's appointment-template gap log. Three named report sections drive the workflow.

Section One: The Lost-Call List with Kill Reasons

Every call the AI did not book lands tagged by AI-inferred kill reason. Twelve standardized categories: AI-confidence-threshold rejection, abandoned-in-hold, voicemail-tag, service-area-out, price-shock, slot-mismatch, language-barrier, multi-property-complexity, high-emotion-deflection, warm-transfer-bounce, already-have-a-tech-firm, call-back-tomorrow-no-confirm. Each call shows: timestamp, kill reason, AI transcript link, FSM customer record link, and recovery priority score (1-10, calculated from estimated ticket value × recovery probability × time-sensitivity).

Section Two: The Recovery Priority List

The AI re-ranks the lost-call list by recovery priority and surfaces the top 15-25 for the morning CSR shift. Priority is calculated from: estimated ticket value (HVAC no-cool in July scores higher than a tune-up inquiry), recovery probability per kill reason (after-hours overflow and voicemail-tag score higher than already-have-a-tech-firm), time-sensitivity (calls from the prior evening score higher than calls from 36 hours back; the recovery curve decays sharply after 24 hours), and homeowner availability signal (the AI's classification of whether the homeowner indicated immediate need vs. researching). The CSR-floor lead reviews the top 15-25 at 7:30 a.m., assigns them to CSRs by call type and CSR strength on each kill reason (rookie CSRs do not handle high-emotion-deflection re-outreach; senior CSRs prioritize price-shock recovery), and the floor starts re-outreach at 8:00 a.m.

Section Three: The Trend Analysis

The trend section compares the prior 24 hours to the trailing 7- and 28-day baselines. Kill-reason distribution shifts surface architecture-tuning opportunities: a spike in abandoned-in-hold means the in-hours staffing gap needs covering; a creeping rise in AI-confidence-threshold rejections means the system prompt's rebuttal library has drifted and needs the L2 Ch2 calibration refresh; a recurring 1:00 p.m. hold-abandonment-spike means the CSR break schedule has not been adjusted for the architecture. The trend section is the input to the Friday tune-and-update cadence.

The CSR Re-Outreach Playbook Against Each Kill Reason

The recovery loop's central artifact is the kill-reason-specific re-outreach playbook. Generic "call them back" produces 25-30% recovery; kill-reason-calibrated re-outreach produces 55-75% recovery. Each kill reason has a documented re-outreach approach with timing, channel (call vs. text first), opener language, and the AI-drafted follow-up text that goes out when the live call does not connect.

After-Hours Overflow and Voicemail-Tag Recovery

The highest-recovery kill reasons. After-hours overflow recovery: AI-initiated callback within the first 30 minutes of the morning shift (8:00 a.m. for an 8 a.m. shift opener), AI dials out, identifies the call ('I see you reached out last night about your no-heat — I have an opening this morning between 10 and 12 or 1 and 3'). Recovery rate at 8:00 a.m. callback: 65-78%. Voicemail-tag recovery: CSR re-outreach within 4 hours of the original voicemail, AI-drafted text first followed by a live call attempt if the text does not produce a reply within 20 minutes. Text-first lifts recovery 12-18 points over call-first because homeowners screen unknown numbers in the morning. Combined recovery rate for after-hours overflow and voicemail-tag: 60-80% of the leak this segment otherwise would lose entirely.

In-Hours Abandoned-Call Recovery

The homeowner who waited and hung up. The AI flags the abandoned call within 90 seconds of the hang-up, the CSR row gets the priority alert, and re-outreach goes out within 5-12 minutes via AI-initiated callback with shop voice ('I see we missed your call a few minutes ago — I'm calling back right now. Were you reaching out about a no-cool or something else?'). Recovery rate at 5-12 minute callback: 38-50%. Recovery rate drops to 18-25% if the callback waits more than 30 minutes. The recovery loop's discipline is the sub-15-minute callback window for in-hours abandonment, mirroring Lesson 1's 15-minute architectural callback for the recovery queue. Without the discipline, in-hours abandoned-call recovery collapses to 8-12% as homeowners contact the next shop.

Warm-Transfer-Bounce and Slot-Mismatch Recovery

Warm-transfer-bounce: the homeowner cleared the AI to Tier Two but Tier Two did not pick up in time. Re-outreach is AI-initiated callback with apology language ('I see your call came through but we couldn't connect you live — happens when we get a rush. I'd like to make it right with a priority slot.'). Recovery rate at 30-min callback: 55-68%. Slot-mismatch: the homeowner needed a slot the board could not offer; recovery requires the dispatcher to flex capacity (split a stacked install, add a half-day overflow tech, lean on a partner shop) and the CSR to re-engage with a window that fits. Recovery rate with dispatcher flex: 60-72%; with rigid capacity, 20-28%. Slot-mismatch is where dispatch and CSR collaboration produces the largest joint metric movement.

Price-Shock and Call-Back-Tomorrow Recovery

Price-shock recovery: the homeowner declined the dispatch fee on the AI's read; re-outreach reframes the dispatch fee as credited-against-repair (the L2 Ch2 price-shopper pivot adapted to a recovery context). AI-drafted text first: 'Just wanted to follow up on your call earlier — I should have made clearer that the $79 dispatch is fully credited against any repair you approve, so it's essentially a free diagnostic when you go with us. Open to me holding a slot this afternoon?' Recovery rate via text-first reframe: 28-42%. Call-back-tomorrow recovery (homeowner refused held slot with text confirmation in the original call) requires AI to wait 24-48 hours, then re-outreach with an open-ended check-in rather than a slot push. Recovery rate at 24-48 hour respectful re-outreach: 22-35%; pushing harder than that converts the homeowner into a permanent no.

The AI-Drafted Follow-Up Text Library

Each kill reason has a text-library entry that the AI drafts on demand based on the original call's transcript, the homeowner's stated concern, and the recovery playbook's shape for that kill reason. The library is the second-highest-value artifact in the recovery loop (after the daily AI report itself). The text is what the homeowner reads at 11:00 a.m. while they're at work, on their phone, deciding whether to re-engage with the shop or stay with the next-shop they called instead.

The Text Library Rules

Six rules govern every recovery text the AI drafts. Rule 1: under 220 characters so the text fits one SMS segment and reads in 8 seconds. Rule 2: opens with a specific acknowledgement of the original call ('I see you reached out last night about the water heater'), not generic ('Following up on your call'). Rule 3: offers concrete action, not a question ('I have openings between 10-12 or 1-3 today' beats 'When are you available?'). Rule 4: includes one trust signal — a tech's name the homeowner referenced, the shop's licensed/insured status, the on-time guarantee, the credited-fee policy. Rule 5: closes with a low-friction reply path ('Reply YES to lock it' or 'Reply with a better time'). Rule 6: signed with a CSR first name and a callback number, never 'Shop Team' or a bot signature. Texts following all six rules produce 45-65% reply rates; texts violating any rule fall to 18-25%.

Kill-Reason-Specific Text Templates

The library carries one calibrated text per kill reason. After-hours overflow: 'Hi {name}, this is Sarah from {shop}. I saw your call last night about the no-heat — sorry we missed you. I have an opening this morning between 10-12 or 1-3. Tech is Marco, same one who did your install in 2024. Reply YES for either. — Sarah, 555-1234.' In-hours abandoned: 'Hi {name}, this is Jen from {shop}. I see we missed your call about 10 minutes ago — calling and texting because I want to make sure we connect. AC issue? Openings at 2 and 4 today. — Jen.' Slot-mismatch: 'Hi {name}, I dug into the schedule — I can get a tech out Wednesday 8-10 or Thursday 2-4. Either work better? — Mark.' Each template loads with the CSR's actual name, the shop's phone number, and the homeowner's actual concern from the original transcript.

The Text Tuning Cadence

The text library tunes weekly from the prior week's reply-rate data. The CSR-floor lead pulls the reply-rate distribution by kill reason and by CSR Friday at 4 p.m., flags any text-template variant where reply rate dropped below 35%, drafts the v(n+1) for the lower-performing template, and rolls it out at Monday's 8:00 a.m. shift. Texts compound: a library that started at 40% average reply rate climbs to 55-65% within 8 weeks of disciplined tuning. The 4 p.m. Friday cadence mirrors L2 Ch2's daily-review cadence, scaled down to weekly because the text library is a smaller surface than the in-call rebuttal library.

The Recovery Loop Metrics and the 90-Day Rollout

The recovery loop moves five headline numbers and three diagnostic numbers. The owner reads them on the dashboard daily.

Recovery rate by kill reason. The single most important metric. After-hours overflow recovery: 65-78%. Voicemail-tag: 55-70%. In-hours abandoned: 38-50%. Warm-transfer-bounce: 55-68%. Slot-mismatch: 60-72% with dispatcher flex, 20-28% without. Price-shock: 28-42%. Call-back-tomorrow: 22-35%. The dashboard segments by kill reason because recovery cadence differs per segment.

Aggregate recovered calls per week. Target: 28-42 recovered calls/week at a 7-truck residential shop. At $387 average ticket, $11K-$16K weekly recovery, $565K-$830K annualized. The number is purely additive to the architecture's primary capture.

AI-recovered revenue per week. The dollar value above translated into the FSM-tracked completed-job revenue from recovered bookings. Owner's preferred metric because it lands in the P&L directly.

Effective booking floor. Architecture booking-% (75% Tier-One mature) plus recovery-augmented additional bookings (recovery rate × leak share). Target: 85-88% effective booking floor inside 90 days. The metric the owner reports to the coach, the PE partner, or the franchisor.

Text reply rate by kill reason and by CSR. Library performance signal. Target above 45%; below 35% triggers Friday tuning. Per-CSR distribution surfaces who needs coaching on text-following voice continuity.

Three diagnostic numbers: callback timing distribution (95% of in-hours abandoned recoveries should fire within 15 minutes — the recovery-loop equivalent of Lesson 1's architectural callback window); kill-reason-mix shift over 28 days (any segment moving more than 25% week-over-week needs architecture-level tuning, not just recovery cadence tuning); recovery-call show rate (target 90%+ — recovered calls should show within 4 points of standard bookings; lower means the recovery is over-promising or the homeowner re-engaged the next shop in parallel).

The 90-day rollout. Week 0: configure the daily AI report against CallRail and the AI receptionist's post-call log, build the kill-reason taxonomy in the FSM, draft the v1 text library across all 12 kill reasons, train the CSR row on the re-outreach playbook. Week 1: daily report at 7:00 a.m., re-outreach starts at 8:00 a.m. against the top-15 priority list, Friday tuning cadence opens. Weeks 2-4: per-kill-reason recovery rates surface, text library tunes weekly, dispatcher flex policy formalizes. Aggregate recovery climbs from 28-35% at week 1 to 48-58% at week 4. Weeks 5-8: depth-phase tuning, edge cases land, text reply rate climbs to 50-58%, aggregate recovery 58-68%. Weeks 9-12: polish phase, dispatcher-CSR collaboration on slot-mismatch matures, call-back-tomorrow re-outreach stabilizes, recovery compounds to target 60-80%. Day 90: effective booking floor 85-88%, recovery-loop revenue lands as a distinct line on the owner's dashboard.

The Failure Modes the Recovery Loop Defends Against

Three failure modes kill the recovery loop's compounding. First, the "generic callback" failure: the CSR row treats every lost call the same — call them back, leave a voicemail, move on. Recovery rate stalls at 25-30% across kill reasons because the re-outreach is not calibrated. Fix: the kill-reason playbook deployed and the text library loaded by Monday of week 1. Second, the "no daily report" failure: the floor lead skims their own intuition for which calls to recover and the priority list is unsorted. Recovery rate drops 18-25 points because the high-priority calls get worked in random order and the low-priority calls absorb time that should have gone to the high-value recovery. Fix: the 7:00 a.m. AI report on the floor lead's screen before shift start, the priority score visible per call, the top-15-25 list scoped for the morning. Third, the "no Friday tuning" failure: the text library v1 ships, reply rates stagnate at 38-42%, and the floor lead does not pull the weekly distribution. Recovery rate ceiling locks at 45-55% instead of compounding to 60-80%. Fix: the Friday 4 p.m. cadence held weekly, the lower-performing templates flagged, the v(n+1) draft rolled out at Monday 8:00 a.m. shift.

One architectural failure mode the service manager has to guard against: confusing recovery-rate movement with architecture-level capture-rate movement. A pilot where the daily AI report shows a creeping rise in AI-confidence-threshold rejections but the recovery rate stays at 55-65% looks like the recovery loop is working — and it is, but the architecture is decaying upstream. Fix: the monthly architecture review from Lesson 1. If kill-reason mix shifts more than 25% in 28 days toward upstream-driven categories (confidence-threshold rejections, warm-transfer-bounces), the architect re-tunes the system prompt and confidence threshold rather than leaning harder on the recovery loop.

Weekly Reporting to the Owner and the Quarterly Recovery Review

The recovery loop reports weekly to the owner with three numbers and one chart. Friday 5 p.m. email: (1) recovered calls this week vs. trailing 8-week average and dollar value at $387 average ticket; (2) effective booking floor versus 85-88% target; (3) text reply rate distribution and the top-three kill reasons driving the week's wins. The chart is kill-reason mix over the trailing 4 weeks, color-coded by recovery rate — green at target, yellow 5-10% under, red more than 10% under. Owner reads in 4 minutes and decides whether to escalate any kill reason to the quarterly architecture review or trust the Friday tuning cadence.

The quarterly recovery review (90 min, service manager plus ops manager) re-evaluates the kill-reason taxonomy, the text library against the prior quarter's reply-rate distribution, the dispatcher-flex policy, and the integration between the recovery loop and the receptionist architecture. The review catches slow drift the weekly cadence misses — a kill-reason becoming obsolete (after-hours overflow shrinking as architecture matures), a new kill-reason emerging (multi-property landlord rising to 8-12% as commercial mix grows), or a structural homeowner behavior shift (rising text-screening defeating call-first re-outreach in certain ZIPs). Output: refreshed playbook for the next quarter, plus a memo paragraph the service manager appends to the Lesson 1 architecture memo at quarterly review.

Key Takeaways

• The recovery loop is the second half of the L3 Ch2 workflow. Lesson 1 builds the architecture that captures 70-75% of inbound. This lesson recovers 60-80% of the remaining 25-30% leak — $565K-$830K annualized at a 7-truck residential shop.
• Four leak sources, twelve kill reasons. AI-confidence-threshold rejection, abandoned-in-hold, voicemail-tag, service-area-out, price-shock, slot-mismatch, language-barrier, multi-property-complexity, high-emotion-deflection, warm-transfer-bounce, already-have-a-tech-firm, call-back-tomorrow-no-confirm. Each has a calibrated re-outreach approach with different recovery rates.
• The 7:00 a.m. daily AI report is the workflow's anchor. Three sections: lost-call list with kill reasons, recovery priority list (top 15-25 ranked by ticket × probability × time-sensitivity), trend analysis vs. trailing 7- and 28-day baselines. CSR-floor lead reviews at 7:30 a.m., re-outreach starts at 8:00 a.m.
• Recovery rates by kill reason hit specific targets. After-hours overflow 65-78%, voicemail-tag 55-70%, in-hours abandoned 38-50% (only if callback within 15 minutes), warm-transfer-bounce 55-68%, slot-mismatch 60-72% with dispatcher flex, price-shock 28-42%, call-back-tomorrow 22-35% with respectful 24-48 hour re-outreach.
• The AI-drafted text library follows six rules. Under 220 chars, specific opener, concrete action not question, one trust signal, low-friction reply path, CSR first-name signature. Texts following all six produce 45-65% reply rates; violating any rule drops to 18-25%.
• Friday 4 p.m. tuning cadence compounds the library. Pull reply-rate distribution by kill reason and CSR, flag any template under 35%, draft v(n+1), roll out at Monday 8 a.m. Library climbs from 40% average reply rate at v1 to 55-65% inside 8 weeks.
• Effective booking floor is the owner-facing target. Architecture primary capture (75% Tier One) plus recovery-augmented additional bookings = 85-88% effective booking floor by day 90. The number the owner reports to the coach, PE partner, or franchisor.
• Three failure modes to defend against: generic callback (recovery stalls at 25-30%), no daily AI report (random priority order, 18-25 point drop), no Friday tuning (text library ceiling locks at 45-55%). One architectural failure mode: confusing recovery-rate movement with architecture-level capture; a 25%+ kill-reason-mix shift triggers monthly architecture review from Lesson 1, not just recovery-loop tuning.