The L1 Cardinal Rule says verify everything AI generates that touches a customer. Lesson 1 of this chapter installs the 30-second habit. Lesson 2 makes the discipline visible on the bulletin board. This lesson draws the third operational boundary — when AI is sufficient and when the human picks up the phone. The boundary matters because universal verification with no calibration becomes universal friction; universal trust with no calibration becomes universal exposure. The right answer is role-specific decision rules with named thresholds. The owner has the final word on financing language, regulatory filings, dispute language, and any customer email above $5K in dispute. AI handles routine acknowledgment and routing; humans handle judgment, regulated language, escalation, and any communication where the cost of being wrong exceeds the cost of being slow. This lesson is the decision rules by role — CSR, dispatcher, tech, Comfort Advisor, service manager, marketing manager, owner — and the named phone calls that escalate uncertainty to source-of-truth in five minutes instead of five days of recovery.

Why Decision Rules and Not Discipline Alone

Lesson 1 installs the verify habit. Lesson 2 makes it visible on the bulletin board. Neither answers the question every role asks 50 times a day: "Is this artifact verify-and-ship, or is this human-decides?" Not "always verify" (universal friction collapses throughput); not "AI handles it" (universal trust collapses defensibility). The answer is decision rules — by role, artifact type, dollar threshold, regulatory exposure — that tell the team when to ship AI and when to pick up the phone.

Rules convert verify discipline from cognitive load to automatic routing. Without rules, the CSR decides per booking whether the Avoca confirmation is sufficient. With rules, routing is automatic — under $250 in-ZIP residential, AI ships; any commercial booking, human reviews; any complaint flag, human callback in 30 minutes. Rules eliminate per-artifact decision overhead; team operates at speed without sacrificing discipline. Rules also build the regulatory defense — a documented decision-rule table by role, threshold, and escalation trigger demonstrates intentional governance design, not ad-hoc improvisation. Investigators credit intentional design; absence of design produces the "shop relied on AI without bounded authority" finding that drives sanctions.

The $5,000 Line — The Owner's Final Authority

The most-cited rule in 2026 trades AI is the $5,000 threshold. Above $5K in customer commitment, owner or owner-designate signs off before ship; below $5K, role-distributed verify discipline applies. The threshold matches three realities: (a) bottom-end of material commercial commitment in residential trades — large enough to attract regulator attention if mishandled, small enough not to capture routine service; (b) captures every typical replacement proposal, financing arrangement, recall escalation, customer-recovery commitment; (c) correlates with the dollar threshold at which state contractor boards take a formal posture on misrepresentation complaints.

The owner has final authority on four categories above the $5K line. Financing language — every customer-facing artifact quoting APR, term, fee, payment, or adverse-action language gets owner or owner-designate (sales manager / service manager) signoff. Signoff is a timestamp, not a re-write — the advisor pulled portal output verbatim per Lesson 1; signoff confirms the paste happened and surrounding context is shop-voice.

Regulatory filings — EPA 608 documentation, state contractor board responses, permit applications, OSHA filings, employee classifications, IRS Form 4136, Section 179 memos. Owner reviews and signs every regulatory filing. AI may draft; the owner verifies and signs. The owner's signature is the licensed-individual attribution the regulator credits.

Dispute language — every customer-recovery email above $5K gets owner signoff before send. Disputes are the highest-stakes customer-facing artifacts; AI-drafted dispute language is the highest-frequency category in shops that have rolled back AI at 90 days. Signoff catches commitment-language slips (we'll refund, we guarantee, this won't happen again) that produce FTC and state-AG exposure.

Customer email above $5K in dispute — any email where the disputed amount exceeds $5K routes to owner before send. The threshold is asymmetric because email is a binding written record the customer can produce in litigation; the dollar-threshold signoff matches the litigation-discovery threshold plaintiff firms target.

The CSR Decision Rules

The CSR runs the highest-volume verify surface — 60-120 inbound per day plus AI-handled bookings via Avoca / Jobber AI Receptionist / Housecall Pro AI Agents / ServiceTitan Voice. Universal verification is impossible at that volume; rules are dollar and complexity threshold.

AI ships without CSR review when: routine residential under $250 estimated; in-ZIP territory; no complaint flag; standard window/equipment/dispatch fee; existing customer matches FSM record. These are 60-75% of bookings; the dispatcher picks them up with the 5-second skim during acceptance.

CSR reviews before dispatch when: any new customer; any commercial-coded call; any after-hours booking; estimated service over $250; equipment flagged for skilled-tech-only; customer history flag (prior complaint/payment issue/cancellation); out-of-ZIP request; transcription confidence below threshold. These are 20-35% getting a 30-second CSR pass.

CSR escalates and picks up the phone when: customer expresses urgency the booking doesn't reflect ("basement flooding" booked as 48-hour drain inspection); medical emergency, fire risk, gas smell, water active leak, electrical arcing; quote dispute or pricing pushback in call audio; competitor or "calling around" mention — CSR calls back in 10 min for save-the-call; commercial commitment ("9 a.m. guaranteed") in AI confirmation that doesn't match the dispatch board.

Phone source-of-truth surfaces: when verify is uncertain, CSR calls customer (re-confirm address/equipment), dispatcher (capacity reality), on-call tech (skill match), supplier (after-hours parts), manufacturer (warranty status on recall claim). Phone produces deterministic answers in 5 minutes; re-prompting AI produces different fluent fabrications with the same source-of-truth gap.

The Dispatcher Decision Rules

The dispatcher's primary AI surface is ServiceTitan Dispatch Pro, Sera Systems, or FieldEdge AI routing plus the Lesson 1 override-log (5-15% target, documented reasons).

Dispatch Pro ships when: suggested tech matches skill tag; within travel-time band; tech's day capacity supports the slot; no callback-originator conflict; no install-crew preservation conflict. Dispatcher accepts; on-my-way text fires.

Dispatcher overrides when: comp-plan considerations (residual commission ceiling, spiff status); install crew preservation; recall originator (recall must go to install tech); customer-requested tech; high-revenue call to high-MPR tech for membership pitch; rookie protection on complex commercial. Each override logged with reason.

Dispatcher escalates and picks up the phone when: re-route breaks install crew or splits a high-revenue commercial job; suggested tech has multiple recent recalls on similar work; call requires permit or licensed-electrician signoff and Dispatch Pro routed to unlicensed tech; emergency (active leak, gas smell, no-heat subzero, no-cool 100F+) where slot doesn't match urgency; multi-day install scheduling conflict. Escalation to ops manager or owner; phone call resolves in 5-10 min.

The Tech Decision Rules

The tech's AI surfaces are voice notes, repair-vs-replace prompts, and AI-tagged photo annotations.

Tech reads AI directly when: voice-note structured output matches observation (5-sec confirm against data plate + diagnosis); repair-vs-replace numbers match input; photo annotations match actual conditions. Customer conversation proceeds with AI-supported talk-track.

Tech verifies before reading when: AI-generated refrigerant charge weight (always against EPA 608 binder + manufacturer table — never read to customer or use on gauges without verify); part number (against supplier catalog); warranty year (against matrix); rebate (against current utility/state page); NEC/IPC/IMC/IRC/IFGC code section (against current adopted code — better, omit citation and use "meets current code" general language).

Tech escalates and picks up the phone when: repair-vs-replace economics above $5K and customer uncertain — call Comfort Advisor or service manager for owner-level engagement; warranty status uncertain on coil/compressor failure within manufacturer term — call warranty desk (Goodman, Trane, Carrier, Lennox, Bryant, Rheem, York) for current term and registration; EPA 608 violation suspected on prior service — call service manager and certified-tech-of-record; permit required and not pulled — stop work, call licensed qualifying individual; fire risk, active gas leak, structural electrical hazard — stop work, call shop, call utility shut-off as appropriate.

The Comfort Advisor Decision Rules

The advisor faces the highest-stakes surface — $8K-$45K residential and $25K-$250K commercial proposals. The $5K threshold lands directly in daily work; almost every advisor proposal sits above it.

Advisor ships AI-drafted proposal when: total under $5K (rare on replacement); all five Lesson 1 checkpoints clean; no code citations in customer-facing language; no commitment language outside shop policy; financing is portal-verbatim or absent; no customer signal of credit/payment/dispute concerns.

Advisor gets service-manager or owner signoff before ship when: $5K-$14K (sales-manager); $14K+ (owner); any commercial; financing above $10K; rebate stack the advisor hasn't previously assembled at that combination; customer surfaced credit-pull concerns in call audio; heat-pump-plus-panel-upgrade (cross-trade integration risk); commercial replacement with phased delivery.

Advisor escalates and picks up the phone when: Wisetack/GreenSky/Synchrony portal output unusual (unexpectedly low APR, long term, unexpected required disclosure) — call lender's contractor support before showing customer; manufacturer warranty term doesn't match matrix — call warranty desk; jurisdictional code citation looks wrong or amendment may apply — call licensed electrician or building inspector; competing bid includes language outside industry norm — call service manager for sanity check; spouse not present and proposal crosses joint-decision threshold — pause, schedule return, do not close half a marriage. Verify at truck; signoff before kitchen table; phone call when verify is uncertain. None of this slows the close — it ensures the close lands on solid ground rather than producing 90-day rollback.

The Service Manager Decision Rules

The service manager's AI surfaces are Rilla scorecards and AI-drafted ride-along commentary, plus AI-tagged RC&D triage.

Coach from AI commentary directly when: flagged moment matches the ride; talk-track reads shop-voice not generic; coaching point consistent with advisor's scorecard pattern; advisor's psychological readiness assessed (no fresh personal crisis, no pending disciplinary conversation).

Re-write AI commentary before coaching when: tone is generic or scripted; flagged moment is out of context from the ride arc; talk-track wouldn't land with this customer demographic or advisor selling style; coaching requires nuanced context AI cannot capture (mid-divorce homeowner, recent bereavement, prior negative shop experience).

Escalate and pick up the phone when: true recall (24h SLA) on tech with three prior recalls in 90 days — call ops manager/owner for performance escalation; warranty dispute with manufacturer refusing coverage — call manufacturer warranty desk then customer for transparency; AI commentary flags coaching that touches employee-performance review territory (PIP, termination consideration) — pause coaching, call HR or owner; multiple ride-alongs flag same advisor same point across 30 days — escalate to comp-plan or development-plan conversation.

The Marketing Manager Decision Rules

The marketing manager operates AI review responses (NiceJob, Podium AI Employee, Birdeye AI Employee, Yelp AI), Hatch outbound, and the AI-built Friday marketing recap.

AI ships review responses when: response acknowledges specific complaint; no commitment language ("we will refund," "we guarantee," "this won't happen again"); no fabricated remedies; tone matches shop voice; star rating 4-5; no unresolved customer-recovery case mentioned.

Reviews before publication or 24-hour hold when: 1-3 star; references specific dollar amount, tech name, equipment failure, or timeline; mentions safety, fire risk, water damage, gas leak; from customer with active recovery ticket; from high-membership-tier customer; commitment language slip the system prompt should have blocked.

Escalate and pick up the phone when: 1-star with multiple specific allegations (call owner before any public response); regulatory complaint mentioned in review (state contractor board, BBB, CFPB, FTC) — call owner, do not publish AI response until counsel-reviewed; active media coverage referenced — call owner and external comms; Hatch flags repeat opt-out violations (cache-drift) — call vendor support and freeze sequence pending audit; AI marketing recap surfaces channel anomaly dashboard can't explain (40% spend drop, 60% lead-source shift) — call agency or analyst before relying on recap.

The Owner's Decision Rules

The owner is final authority on four categories — financing language, regulatory filings, dispute language, and customer email above $5K in dispute. The owner does not verify every artifact; the owner verifies that the verify discipline runs and signs where signature is the operative legal act.

Owner signs directly when: financing arrangements above $10K (below $10K delegated to sales manager); customer-recovery commitments above $5K; regulatory filings (every EPA 608, contractor board, permit, OSHA, employee classification); license-renewal materials; PE quarterly submissions; franchisor compliance memos; insurance underwriter responses; CPA tax approvals.

Owner escalates and picks up the phone when: any state contractor board complaint — call counsel and franchisor before response; CFPB or state-AG inquiry — call counsel; EPA inspector visit or notice — call counsel and certified-tech-of-record; TCPA class-action notice — call counsel and freeze all AI outbound pending audit; FTC inquiry — call counsel; major-account customer dispute above $25K — call counsel and customer's executive contact; catastrophic field event (injury, fatality, structural damage) — call counsel, insurance, OSHA reporting line as required, and the family.

Owner rules preserve optionality. Every escalation buys time, source-of-truth, and counsel before commitment. AI-drafted responses to regulatory or legal inquiry are categorically prohibited; owner signature with counsel review is the only defensible artifact.

The Phone as Source-of-Truth Port

The phone recurs across every role's rules because it produces deterministic source-of-truth answers in 5 minutes that re-prompting AI cannot. Calling Wisetack's contractor support gets the actual APR on this approval. Calling Goodman's warranty desk gets the current term on this serial. Calling Carrier Enterprise gets the current part number and availability. Calling the building inspector gets the jurisdiction-current code reading. Calling the licensed electrician of record gets the code-citation authority. Calling the manufacturer's regional rep gets the recall status on this equipment.

Each call is 5-10 minutes; each saves 5 hours to 5 days of recovery on an unverified artifact. The phone is the source-of-truth port; re-prompting AI is a different fluent fabrication with the same source-of-truth gap. When verify is uncertain, pick up the phone.

The Rules and the Shops That Survive AI

The 60-point gap in ServiceTitan's 2026 report (72% relevance vs. 12% embedded) is primarily a verify-discipline gap and secondarily a decision-rule gap. Tools exist; verify installs in 30 days; bulletin boards cost $40; decision rules are where shops fail because rules require operational design, not vendor capability. Shops designing rules in week three — after Lessons 1 and 2 have installed habit and board — capture published lifts and compound metric movement for twelve months. Shops operating AI without decision rules produce universal friction or universal exposure, both of which roll back at 90 days.

Three operating principles: AI handles routine acknowledgment and routing; humans handle judgment, regulated language, escalation, and any communication where being wrong costs more than being slow. The $5K threshold is the universal signoff line; above it, owner or designate signs every customer commitment, financing arrangement, regulatory filing, and dispute. The phone is the source-of-truth port; when verify is uncertain, the call to Wisetack/warranty desk/supplier/inspector/licensed individual/regulator contact resolves in 5 minutes what AI re-prompting cannot resolve at all. Build the rules in week three; compound for twelve months.

Key Takeaways

• Decision rules convert verify discipline from cognitive load to automatic routing — by role, by artifact type, by dollar threshold, by regulatory exposure. Eliminate per-artifact decision overhead; preserve discipline at throughput.
• The $5,000 line — Above $5K customer commitment, owner or owner-designate signs off. The threshold matches material commercial commitment, captures replacement-class proposals/financing/recall escalations/customer-recovery, and correlates with state contractor board's formal-investigation posture.
• Owner's four named categories — Financing language, regulatory filings, dispute language, customer email above $5K in dispute. AI may draft; owner signs.
• CSR rules — AI ships routine sub-$250 in-territory bookings (60-75%); CSR reviews commercial/after-hours/new-customer/over-$250 (20-35%); CSR escalates and calls back on urgency mismatches, save-the-call attempts, commitment-language conflicts.
• Dispatcher rules — Dispatch Pro / Sera / FieldEdge ships on skill+travel+capacity match; override 5-15% with documented reasons (comp plan, install crew, recall, customer request); escalate to ops manager or owner on emergency calls and crew/permit conflicts.
• Tech rules — AI direct read on voice-note + repair-vs-replace + photo annotations; verify before read on charge weights, parts, warranty, rebates, code; escalate and call on warranty disputes, EPA 608 violations, permit gaps, fire/gas/electrical safety.
• Comfort Advisor rules — Ship sub-$5K with all checkpoints clean; service-manager/owner signoff on $5K-$14K / $14K+ / commercial / unusual financing / heat-pump+panel combos; call Wisetack/GreenSky/Synchrony/manufacturer/inspector on uncertainty; pause and reschedule when spouse not present.
• Service Manager rules — Coach from AI commentary directly when it matches the ride; re-write when generic/out-of-context/wrong-tone/missing-context; escalate on recall-tech-pattern, warranty escalation, employee-performance territory, multi-30-day same-flag patterns.
• Marketing Manager rules — AI ships 4-5 star generic responses without commitment language; review or 24-hour hold on 1-3 star, specific allegations, active-recovery customers; escalate to owner on 1-star with regulatory complaint, media coverage, repeat opt-out violations.
• Owner rules — Direct sign on $10K+ financing, $5K+ recovery, every regulatory filing, license renewal, PE/franchisor/insurance submissions; escalate to counsel on contractor-board, CFPB, state-AG, EPA, TCPA, FTC, catastrophic field event.
• The phone is the source-of-truth port — Wisetack contractor support, manufacturer warranty desk, supplier catalog, building inspector, licensed individual, regulator contact line. 5-10 minutes per call saves 5-day-to-5-hour recovery. Re-prompting AI is a different fluent fabrication with the same source-of-truth gap.
• Build rules in week three — Lesson 1 installs verify (week 1), Lesson 2 builds the board (week 2), Lesson 3 defines decision rules (week 3). Without the rules, week-1 discipline becomes universal friction or universal exposure.