The shops that move from "AI nothing" to "AI answers every call" in 90 days do not do it by signing five contracts in week one. They do it by deploying one workflow per month, building the verify discipline before scaling, and protecting the dispatch board through the rough first 30 days when every new tool produces visible mistakes. This is the Crawl, Walk, Run pattern documented across multiple 2026 shop deployments — the ACHR News reporting, the Avoca HL Bowman case, the Rilla Climate Experts case, the Housecall Pro and Sera 2026 case-study sets. Month 1: missed-call answering live. Month 2: CSR call summaries live. Month 3: first ride-along scorecards in production. By day 91, the shop has crossed the 12%-embedded threshold ServiceTitan's 2026 State of AI report defines, has documented metric movement on at least three numbers (booking %, ACW time, dispatch yield or close rate), and has the operational rhythm to keep going. This lesson is the actual cadence — what gets done each week, what each role learns, what the verify pass catches, and where shops have broken the dispatch board by going too fast.

Why Crawl, Walk, Run Exists — The Two Ways AI Rollouts Fail

The two failure modes that destroy AI rollouts in trades shops are well-documented by 2026. Failure mode one: sign everything at once. The owner reads three case studies in a weekend, schedules a Monday call with Avoca, Rilla, Hatch, and ResponsiBid all in the same week, signs four contracts by Friday, and tries to deploy four tools simultaneously across a CSR floor, a dispatch board, a sales-advisor team, and a marketing function. By day 30, two of the four tools have broken integrations the team can't troubleshoot, three of the four have CSRs or advisors complaining the AI is "making mistakes," and the owner cancels two contracts in panic by day 45. Failure mode two: pilot forever. The owner signs Avoca, runs a "pilot" for six months without ever defining what success looks like or building the daily verify discipline, declares the pilot "inconclusive," and concludes AI is not ready for the trades. Both failure modes are operator errors, not tool errors. Crawl, Walk, Run is the documented antidote.

The cadence is built around three principles. First, one workflow per month — one tool deployed, one CSR or dispatcher or advisor trained, one metric tracked. Second, verify pass before scale — the team learns to catch the tool's mistakes in week one before the volume scales in week three. Third, dispatch board sacrosanct — nothing in the AI rollout breaks the team's ability to ship calls and run the day. The dispatch board running unbroken on day 30 is more important than any AI metric on day 30. Shops that violate these three principles fail the rollout; shops that hold them ship the 90-day deployment that crosses the embedded threshold.

The Crawl, Walk, Run nomenclature comes from the ACHR News 2026 reporting on how HVAC contractors are successfully adopting AI. It maps to the three months and three difficulty tiers: Crawl (month 1, missed-call answering — the lowest-risk, highest-ROI workflow), Walk (month 2, CSR call summaries — moderate risk, broad team adoption), Run (month 3, ride-along scorecards or marketing recap — higher risk, deeper workflow change). The cadence is replicable. The cadence is documented. The cadence works.

Month 1 — Crawl: Missed-Call Answering

The Crawl month is missed-call answering because it is the highest-ROI workflow with the lowest team-change footprint. Avoca, Jobber AI Receptionist, Housecall Pro AI Agents, or ServiceTitan Voice handles the inbound call when the CSR is on another line, after hours, or during overflow. The CSR floor's workflow barely changes — they handle the calls they handle today, the AI handles the ones that would have been missed. The dispatch board is unaffected; the AI books into the same ServiceTitan / Sera / HCP slots the CSR books into. The verify discipline is light — the CSR reviews AI-booked calls at 4 p.m. daily for slot accuracy, address spelling, and any commitment the AI made that the shop cannot keep.

Week 1: vendor selection and contract. The owner runs a 5-day vendor evaluation against Avoca, Jobber AI Receptionist, Housecall Pro AI Agents, and ServiceTitan Voice. Decision criteria: integration depth with current FSM, booking-rate test on a 50-call sample, after-hours capture demonstration, transfer-to-CSR accuracy, cost per booked call. Most $5M shops on ServiceTitan land on Avoca because of the bolt-on lift advantage and ServiceTitan integration depth; HCP shops often pick HCP AI Agents for the in-platform integration; Jobber shops land on Jobber Copilot AI Receptionist. Contract signed by end of week 1.

Week 2: integration and pilot configuration. The vendor implementation team plumbs the AI into the FSM, configures the AI prompts to match the shop's voice (job types, service area boundaries, dispatch-fee handling, after-hours emergency protocol), and runs internal-test calls with the owner and one CSR. The CSR row gets a 30-minute training on what the AI is doing, what to verify on AI-booked calls at the 4 p.m. review, and what the escalation path is when the AI gets something wrong. The dispatch board is untouched.

Week 3: soft launch. The AI handles after-hours calls only — the lowest-risk window. The CSR floor reviews the AI-booked calls each morning. Most calls book cleanly. Some have minor errors (a misspelled customer name, an ambiguous job type, a slot that conflicts with a tech's existing route). The CSRs fix the errors and feed examples back to the vendor for prompt tuning. By end of week 3, after-hours booking rate has moved from 0-15% baseline to 70-85% with cleaner workflow than week 1.

Week 4: full deployment. The AI handles overflow during business hours — calls that come in while CSRs are on other lines or during peak windows. The CSR floor's role evolves to handle the calls AI can't close (price-shoppers requiring rebuttal, complaint calls requiring empathy, complex re-bookings requiring judgment). By end of week 4, missed-call rate has moved from 22% baseline toward the under-5% target. The shop has crossed the threshold from "AI nothing" to "AI handles one workflow daily." That is the Crawl outcome.

The metrics that move in month 1: missed-call % (22% → 8-12% by day 30, trajectory toward under 5% by day 60). After-hours capture rate (0-15% → 70-85%). Booking % (65% → 72-78%). Cost per booked call (depends on prior baseline, typically improves 15-25%). The dispatch board: unaffected. The CSR floor: trained, adjusted, on the new daily verify habit. The owner: has documented metric movement and the basis for month 2.

Month 2 — Walk: CSR Call Summaries

Month 2 is CSR call summaries because it is the natural next workflow — same team (CSR floor), same tool family (call-handling AI), expanded use case (not just missed-call but every call). CallRail Conversation Intelligence, ServiceTitan's in-platform call summarization, Avoca's post-call notes, or Housecall Pro's AI Team module all produce a structured summary of every call: customer name, issue, equipment mentioned, slot booked, follow-up action, sentiment. The CSR's after-call work drops from 4 minutes per call to 15 seconds per call. ACW reduction is 95%+. The CSR floor reclaims 5+ hours per day of phone-row time.

Week 5: tool selection. Most shops already have access via existing CallRail or FSM subscription; the decision is whether to activate the in-platform summary feature or add CallRail Conversation Intelligence as the dedicated layer. Decision criteria: existing FSM AI summary quality (Avoca is excellent here for shops already using Avoca; ServiceTitan's in-platform is competent; HCP AI Team is solid), need for cross-channel sentiment monitoring (CallRail wins), and existing CallRail relationship. Activation typically happens in 1-3 days; no new contract needed for shops on existing CallRail.

Week 6: CSR training on the verify discipline. The CSR's new workflow: after each call, review the AI-generated summary in 5 seconds (customer name correct, issue captured, slot/follow-up accurate), paste into ServiceTitan / Sera / HCP customer record, move to next call. The verify pass is 5 seconds. Pre-AI, ACW was 4 minutes; post-AI with verify, it's 15-20 seconds. The CSR floor learns the new rhythm in 3-5 days. The trickiest part: trusting the AI's summary enough to paste without rewriting. Veteran CSRs sometimes rewrite the summary out of habit; the manager coaches them off this behavior because the rewriting eats the ACW savings.

Week 7: full deployment + measurement. AI summaries running on every inbound call. CSR floor on the 5-second verify discipline. ACW data tracked: average ACW per call dropped from 4 min to 15-20 sec; CSR floor capacity reclaimed roughly 18-22 hours per week (5+ hours per day across 4 CSRs). The capacity gets reinvested: 70-100 additional calls handled per week (lifting booking %), or a CSR seat consolidated (reducing headcount cost), or higher-quality handling of complex calls (lifting close rate on booked calls).

Week 8: integration with sentiment alerts. CallRail Conversation Intelligence's sentiment monitoring flags calls where the customer's sentiment crossed a threshold (frustrated, escalation language, complaint pattern). These flagged calls escalate to the owner's daily dashboard. The CSR row begins seeing patterns: which call types produce frustrated sentiment, which CSRs handle frustration well, which scripts need updating. The patterns feed into the L2 Chapter 2 CSR coaching workflow, but the month-2 outcome is the ACW reclamation plus the sentiment-alert feed live.

The metrics that move in month 2: ACW (4 min → 15-20 sec; 95% reduction). CSR floor capacity (+18-22 hours/week reclaimed). Booking % (continuing the climb from month 1, 72-78% → 78-82%). CSR show rate (84% baseline → 87-89% as cleaner handoffs reduce no-show miscommunication). Sentiment alerts: live, feeding owner dashboard. The dispatch board: still unaffected. The CSR floor: in the rhythm. The owner: two embedded workflows running with measured lift.

Month 3 — Run: Ride-Along Scorecards (or Marketing Friday Recap)

Month 3 is the higher-risk workflow because it touches a different role (sales advisors or marketing manager) and a deeper workflow change (kitchen-table sales coaching, or weekly marketing reporting). The decision tree: shops with active replacement business (HVAC, electrical, roofing, water-treatment, garage-door replacement, plumbing re-pipes) where Comfort Advisors run 4-8 in-home quotes per week deploy Rilla. Shops without significant replacement business deploy the AI-drafted marketing Friday recap as the month-3 workflow instead. Both are valid Run-month choices; both produce documented metric movement.

Run Path A — Rilla Ride-Alongs

Week 9: Rilla contract and setup. Rilla runs $200-$400 per advisor seat per month. The setup: each Comfort Advisor gets a lapel mic and a Rilla account; the service manager gets the scorecard dashboard; the rubric (intro, system-condition narrative, repair-vs-replace pivot, options presentation, financing pivot) is customized to the shop's sales process. Contract through deployment typically 5-10 business days. The advisors get a one-hour orientation; the service manager gets a two-hour rubric-tuning session.

Week 10: pilot with two advisors. The two early-adopter advisors wear the lapel mic on every in-home quote for the week. Rilla transcribes every kitchen-table conversation, scores against the rubric, produces a one-page scorecard per advisor per day. The service manager reviews each scorecard at the next morning's 15-minute advisor huddle. The early-adopter advisors see their patterns; the rubric tunes against their actual conversations; the comp-plan conversation begins (does Rilla's score factor into bonus? when?).

Week 11: expand to all advisors. The remaining advisors join the program. The full team is on Rilla. The service manager runs 30-40 virtual ride-alongs per day across the team — vastly more coverage than the 2-3 in-person ride-alongs that the role used to deliver. Pattern recognition emerges: which advisors are weak on the repair-vs-replace pivot, which are weak on financing introduction, which close consistently. Coaching cards get specific.

Week 12: measurement and comp-plan alignment. Close rate baseline (38-45% typical) vs. close rate after 4 weeks of Rilla coaching (typically +5-12 points by week 12, on track to documented 18% lift by month 6). Average ticket lift (good/better/best presentation discipline lifts mid-tier and high-tier selection). Financing close rate lift (14% baseline → 18-24% by week 12, on track to 28-40% target). The comp plan adjusts to reward RGA improvement alongside close rate. The advisor team transitions from "Rilla is watching me" to "Rilla is helping me close more deals." That transition is the critical leadership challenge of month 3.

Run Path B — AI-Drafted Marketing Friday Recap

Week 9: marketing manager builds the voice prompt. The voice prompt is the proprietary asset — built from the manager's prior memos, capturing tone and structure. Two days of work; informed by L3 Chapter 6 patterns. The data pipes (CallRail, GLSA spend report, ServiceTitan booking conversion, Hatch nurture stats, NiceJob review counts) get plumbed into a model the marketing manager can paste into weekly.

Week 10-12: weekly recap cycle. Each Friday afternoon: paste data, run model with voice prompt, get one-page recap with channel ROAS, booking-% drag, cost-per-booked-call by source, lead-source-to-revenue waterfall, next-week narrative. Marketing manager edits in 5 minutes. Ships to owner at 4:30 p.m. Friday. Owner reads at 5:00 p.m. Friday. Monday standup runs against fresh data. Spend reallocation happens Monday rather than the following Friday. Drafting time: hours pre-AI, 8 minutes post-AI. The first three weekly cycles surface voice-prompt tuning needs; by cycle 4-5, the recap reads as the manager wrote it and the owner can no longer tell which sentences were AI-drafted.

The metrics that move in month 3 (Run path A — Rilla): close rate (38-45% → 43-50% by day 90, on track to 56% by day 180). Financing close rate (14% → 18-24%). Average ticket on closed proposals (+$200-$600 from good/better/best discipline). RGA (ride-along grade): live, weekly trend. The metrics that move in month 3 (Run path B — marketing recap): GLSA ROAS (3.2x → 3.8-4.5x as faster reallocation cycle compounds). Marketing manager time reclaimed (8-15 hours/week from recap drafting). Owner decision velocity (Monday reallocation vs. following Friday).

Day 91 — The Embedded Threshold Crossed

By day 91, the shop has three workflows live with measured metric movement. Missed-call answering (Avoca or equivalent) running at 90%+ booking rate on AI-handled calls, missed-call % under 8%. CSR call summaries (CallRail or in-platform) running on every call, ACW under 30 seconds, sentiment alerts feeding owner dashboard. Either ride-along scoring (Rilla) with first close-rate lifts visible or AI-drafted marketing recap shipping Friday at 4:30 p.m. The shop has crossed from 0% embedded to firmly above the 12% threshold ServiceTitan's 2026 State of AI report defines.

What the owner can defend at day 91: documented missed-call recovery in dollar terms (50+ additional bookings per month × average ticket × 250 days extrapolated = $80-$150K annual margin contribution from Avoca alone). Documented ACW reclamation (18-22 hours/week of CSR floor capacity reinvested). Documented close-rate lift or marketing decision-velocity lift. The owner has a quarterly tool-ROI roll-up showing AI stack cost of $4-$8K/month vs. margin contribution of $15-$50K/month. The math defends the next 90 days of deployment expansion.

What the team has learned by day 91: how to verify AI output (the 5-second summary check, the 4 p.m. AI-booked-call review, the daily advisor scorecard huddle or the Friday recap voice-prompt tuning). How to escalate when the AI gets something wrong. How to coach off AI-generated artifacts (scorecards, summaries, recaps). The daily and weekly rhythms that compound into the 12-month operating cadence the L4 program will build out.

What the dispatch board looks like at day 91: unchanged operationally but informed by better data. AI summaries feed the dispatcher cleaner customer-history context. Sentiment alerts flag re-dispatch candidates (the customer who escalated yesterday gets the senior tech today). The dispatch board itself hasn't been "AI-ized" yet — that is the L2-L3 work in the next quarter. But the inputs the dispatcher is working with are dramatically richer.

The Three Failure Modes That Derail the 90-Day Cadence

Failure mode one: deploying month-2 (CSR summaries) before month-1 (missed-call answering) stabilizes. Tempting because activation is easier; destructive because the CSR floor is learning two new disciplines simultaneously (verify AI-booked calls AND verify AI-generated summaries). Cognitive overload. Errors get missed. Trust erodes. Avoid by completing the missed-call workflow's 4-week deployment before activating the summaries workflow.

Failure mode two: deploying Rilla without the comp-plan conversation. Advisors hear "Rilla scores every conversation" and assume the score gates their bonus. Trust collapses on day 3. The fix is the comp-plan conversation in week 9 — the service manager explains explicitly that Rilla scores are coaching inputs, comp impact is gradual and gated by manager judgment (not raw score), and the rubric is open and amendable. Advisors who understand the comp framework adopt; advisors who are unclear about it resist. Owners who skip this conversation lose advisors in month 3.

Failure mode three: dispatch-board pressure breaking the rollout cadence. A busy week 3 of month 1 lands during a heat-wave or a cold snap; the dispatch board is on fire; the CSR floor is on every call; the AI verify discipline gets skipped because nobody has time. By week 4, the AI is producing errors nobody is catching. The owner blames the tool. The fix is the dispatch-board sacrosanct principle — if the board is on fire, the AI verify discipline doesn't get skipped, it gets sustained by reducing other commitments (push the new advisor onboarding back a week, defer the marketing campaign launch, accept the temporary friction). The 90-day cadence has slack built in; protect it.

The Second 90 Days — What Comes After Embedded

Day 92-180 is the consolidation phase. The three workflows from month 1-3 deepen. Missed-call recovery moves from 70-85% AI booking rate to 90-95%. ACW workflow becomes invisible — nobody thinks about it, the rhythm is automatic. Rilla coaching produces close-rate lift trending to the documented 18% by month 6. The marketing Friday recap voice prompt is well-tuned; the owner trusts the recap completely; spend reallocation cycle is rapid.

Day 92-180 is also when the next workflows layer in. Hatch for stale-lead nurture is a natural Q2 addition — the dormant-lead pile gets reactivated, the CSR row handles the re-engagement follow-up calls, the marketing manager owns the sequence design. ResponsiBid for proposal drafting is another Q2 candidate where shops with significant replacement business cut proposal time from 25-45 minutes to 8-12 minutes and lift average ticket $200-$600 per closed deal. Dispatch Pro tuning intensifies as month-2 baseline data accumulates and the dispatcher learns the override discipline. The L2-L3 workflow-design lessons walk these in detail.

Day 181-365 is the optimization phase. The AI stack is in place. The team has the rhythm. The metrics are moving. The work shifts to deepening per-workflow lift: tuning Avoca's prompts against shop-specific call patterns, building per-tech and per-advisor Rilla scorecards, refining the GLSA AI bidding feedback loop, expanding AEO publishing cadence. By end of year one, the shop has typically captured 60-80% of the projected $700K-$2.7M margin contribution. Year two captures the remainder plus the compounding effects of operating-cadence discipline.

Key Takeaways

• Crawl, Walk, Run is the documented 90-day cadence (ACHR News 2026, multiple shop case studies) that moves a shop from 0% embedded to above the 12% threshold. Three principles: one workflow per month, verify pass before scale, dispatch board sacrosanct.
• Month 1 (Crawl): Missed-call answering. Avoca, Jobber AI Receptionist, Housecall Pro AI Agents, or ServiceTitan Voice. Missed-call % from 22% baseline to 8-12% by day 30, after-hours capture from 0-15% to 70-85%, booking % from 65% to 72-78%. CSR daily 4 p.m. verify pass on AI-booked calls.
• Month 2 (Walk): CSR call summaries. CallRail Conversation Intelligence, ServiceTitan in-platform summaries, Avoca post-call notes, or HCP AI Team. ACW from 4 minutes to 15-20 seconds (95% reduction). CSR floor capacity reclaim 18-22 hours/week. Sentiment alerts feed owner dashboard. Critical discipline: 5-second verify, no rewriting.
• Month 3 (Run): Ride-along scorecards OR marketing Friday recap. Rilla for replacement-heavy shops ($200-$400/seat/mo, close-rate lift trending to documented 18% by month 6). AI-drafted marketing recap for service-heavy shops (8 minutes drafting + 5 minutes editing, ships Friday 4:30 p.m.). Comp-plan conversation in week 9 is non-negotiable for Rilla.
• Day 91 outcome: three workflows live, documented metric movement, AI stack ROI 3-12x already evident, shop firmly above 12% embedded threshold. Owner can defend the next 90 days of expansion.
• Three failure modes: (1) deploying month 2 before month 1 stabilizes — cognitive overload, trust erosion; (2) deploying Rilla without comp-plan conversation — advisor trust collapse; (3) dispatch-board pressure breaking the rollout — verify discipline gets skipped, AI errors compound. The cadence has slack built in; protect it.
• Day 92-180 consolidation: depth on the three deployed workflows + layer in Hatch (stale-lead nurture), ResponsiBid (proposal drafting), Dispatch Pro tuning. Day 181-365 optimization: per-workflow tuning, GLSA AI bidding refinement, AEO publishing cadence. End-of-year-one: 60-80% of projected $700K-$2.7M margin contribution captured.
• The dispatch board stays sacrosanct. The 90-day cadence moves around the board, not through it. Day 91 board running unbroken is more important than any AI metric on day 91. Owners who violate this lose the team's trust in the rollout.
• What the team learns by day 91 is more valuable than what the metrics move: verify discipline, escalation rhythm, coaching-off-AI-artifacts habit, daily/weekly cadence. The metrics compound; the operating discipline compounds harder.