A trades shop's AI receptionist stack is not a single product purchase — it is an architecture decision that depends on truck count, FSM platform, after-hours leakage, dispatcher bandwidth, and the shop's appetite for vendor management. A 1-truck plumber on Jobber making the wrong stack choice eats a $14,000 annual subscription line that pays for a feature set they cannot operate. An 80-truck ServiceTitan platform shop making the wrong stack choice leaves $1.2M-$2.4M of annual after-hours capture on the table and burns its CSR floor inside 90 days. This lesson is the architect-level workflow for the service manager and the ops manager who own the receptionist-stack decision: the named decision tree for which AI receptionist fits a 1-truck, 6-truck, 25-truck, or 80-truck shop; the inbound routing topology that connects Avoca, Jobber AI Receptionist, or Housecall Pro AI Agents to ServiceTitan, Sera, or HCP without breaking the dispatch board; the warm-transfer rules that protect the human CSR's time without letting calls leak; the after-hours booking logic that keeps the on-call tech asleep; the escalation paths that route the angry-recall caller to a human in under 90 seconds; and the callback-window discipline that converts the silent "voicemail tag" leak into a 60-80% recovered-call workflow. By the end of the build, the manager owns a 14-page architecture memo the owner reads in 6 minutes and the PE partner reads in 12, every routing rule documented, every escalation trigger named, every vendor decision defensible against the next quarterly review.

Why the Receptionist Stack Is an Architecture Decision, Not a Product Decision

Most shops approach the AI receptionist as a single-product purchase. The owner reads the Avoca HL Bowman case study (100% answer rate, cost per conversion $350 to $215, 70% YoY revenue growth), signs a $1,800/month MSA, and assumes the deployment is "done" once the phone number ports. Inside 60 days the deployment is half-working: the AI books direct to ServiceTitan but the dispatch board fights it because nobody architected the appointment-template handoff; the in-hours warm transfers route to the CSR who is already on another call so the homeowner bounces; the on-call tech gets paged 3-4 times a night because the after-hours logic was never split from the in-hours logic; and the morning CSR spends 45 minutes at 7:30 a.m. cleaning up the overnight booking dump because the data-field handoff was never standardized. The product works. The architecture does not.

The L3 service manager's job is to build the architecture. The product layer (which vendor) is the smallest decision in the stack. The architecture decisions — inbound routing topology, in-hours vs. after-hours routing split, warm-transfer rules, escalation paths, callback-window discipline, FSM integration depth, system-prompt governance, false-page bonus structure — are 80% of the value capture and 100% of the failure modes when they go wrong. Year one: vendor selection and basic deployment, after-hours capture climbs from 0-15% to 50-65%. Year two: architecture maturity, after-hours capture lands at 80%+, false-page rate under 1/week, morning CSR overnight review under 12 minutes. The shops that get to 80% in year one are the shops with a service manager who built the architecture before the vendor onboarding kicked off.

The Decision Tree for Which Stack Fits Which Shop

The receptionist-stack decision tree branches on three variables: truck count (proxy for call volume and after-hours leak), FSM platform (Jobber, HCP, ServiceTitan, Sera, FieldEdge), and the shop's appetite for vendor management (in-software preference vs. bolt-on depth). The 2026 ServiceTitan State of AI in the Trades data confirms 59% of contractors prefer in-software AI to standalone tools — but the same data shows the 41% who pick bolt-ons skew toward the higher-leak, higher-volume shops where bolt-on depth justifies the premium.

One-Truck Shop: The Jobber or HCP Default

A 1-truck plumber, electrician, or HVAC service operator running $400K-$900K annual revenue cannot afford a $1,800/month Avoca subscription against a $4,000-$6,000 weekly after-hours leak. The math does not pencil. The right architecture is Jobber AI Receptionist (Copilot bundle, $99/month add-on on the Jobber Plus plan) for the 1-truck Jobber shop, or Housecall Pro AI Agents (rolled into the HCP plan tier, typically $40-$120 incremental) for the 1-truck HCP shop. Both ship with native FSM integration, default emergency-rule templates the operator tunes inside a week, and a customer-record update path that requires zero glue code. The 1-truck operator deploys in 3-5 days, lifts after-hours capture from 0-15% to 55-65% in the first 30 days, and pays back the subscription inside the first week of capture. The vendor decision: Jobber or HCP, depending on which FSM the shop already runs. Switching FSMs to access a different AI receptionist is never the right move at this scale.

Six-Truck Shop: The Jobber, HCP, or Early Avoca Decision

A 6-truck residential shop at $2.5M-$5M revenue with 12-18 after-hours calls/week and $200K-$350K annualized leak is the inflection point where Avoca starts to pay for itself. The decision branches on FSM. On Jobber: Jobber AI Receptionist remains the rational first bet — native integration, $99/mo, 8-14 point booking-% lift in 2026 case studies. On HCP: HCP AI Agents, same logic. On ServiceTitan, Sera, or FieldEdge: in-software AI (ServiceTitan Voice with Titan Intelligence, Sera's emerging voice layer) vs. Avoca as bolt-on. At 6 trucks the in-software path usually wins on integration tightness; bolt-on wins when after-hours leak the FSM-native voice product cannot capture (because the native rebuttal library is shallower). Rule of thumb: start native, layer Avoca if after-hours capture stalls below 65% by day 60.

Twenty-Five-Truck Shop: The Avoca or ServiceTitan Voice Decision

A 25-truck residential shop at $12M-$25M revenue with 50-80 after-hours calls/week and $850K-$1.4M annualized leak almost always runs ServiceTitan. The decision is ServiceTitan Voice (Titan Intelligence's voice AI) vs. Avoca (bolt-on, premium, deeper trades-call training). Avoca premium runs $2.5K-$4.5K/mo at 25 trucks; ServiceTitan Voice is bundled with Titan Intelligence at lower incremental cost. Three sub-questions decide: (1) Does the shop have a service manager owning weekly system-prompt tuning? Yes favors Avoca; no favors ServiceTitan Voice's in-platform integration. (2) Is after-hours capture below 60%? Yes favors Avoca's depth; at or above 65% favors ServiceTitan Voice at lower cost. (3) Single-FSM or multi-vendor stack (CallRail, Rilla, Hatch)? Multi-vendor integrates Avoca cleanly; single-vendor favors ServiceTitan Voice. Most 25-truck shops run Avoca for after-hours, with the decision tilting toward Avoca when the service manager exists.

Eighty-Truck Shop: The Platform Architecture Question

An 80-truck shop is no longer a single shop — it is a platform with 4-8 brands, 5-12 locations, and a centralized ops manager or director of AI operations. The architecture question shifts from "which vendor" to "what is the platform receptionist topology?" The 2026 standard for trades roll-ups (Wrench Group, Authority Brands, Apex Service Partners, Sila Services, Path Light Pro, Redwood Services) centralizes the AI receptionist at the platform level — typically Avoca because its API and master-agreement structure handle multi-brand routing cleanly. Brand-local phone numbers route to Avoca's central receptionist with brand-aware system prompts (different empathy anchors, dispatch fee policies, service-area lookups, warm-transfer destinations per brand). Bookings drop into the brand-local ServiceTitan or BuildOps instance via API; the platform reads aggregate booking-% and capture daily across all 12 locations from one dashboard. Setup runs $30K-$80K plus $25K-$65K/mo in aggregate Avoca subscription; payback at 80-truck leak math is 4-8 weeks.

Inbound Routing and the Three-Tier Topology

Once the vendor is selected, the architect's first job is the inbound routing topology. Every inbound call lands in exactly one of three tiers; the boundaries are mutually exclusive and the routing rules live in the phone-tree configuration and the AI receptionist's system prompt. The topology survives any vendor swap because the tier definitions are vendor-agnostic.

Tier One: The AI Handles It

Tier One is the default destination for 65-80% of inbound calls in a mature deployment. The AI answers, runs the rebuttal library (the five core objections from L2 Ch2: price-shopper, just-looking, already-have-a-tech, after-hours, call-back-tomorrow), captures the booking via the FSM's appointment-template API, texts the homeowner the confirmation, and updates the customer record. Tier One in-hours feeds the morning's dispatch board; Tier One after-hours books first-out morning slots. Booking-% on Tier One: 80-87% in a mature deployment, indistinguishable from the human CSR floor reading the same library. The architect's job: load the system prompt with the shop's calibrated library (not vendor defaults), confirm the 11-field handoff (covered below), and set the confidence threshold below which the AI escalates rather than books. Default Avoca, Jobber AI Receptionist, and HCP AI Agents thresholds are 70-75%; architect tuning typically lands at 78-82% for residential trades.

Tier Two: The Warm Transfer

Tier Two is 10-15% of inbound calls — the edge cases the AI's confidence threshold rejects. Service-area lookups the AI flags uncertain, complex multi-system scenarios, pricing complexity outside the system prompt, high-emotion calls (angry recall, dispute, escalation request), language-barrier calls the AI's translation layer cannot handle cleanly, multi-property landlord calls that require manual account lookup. The architect's job on Tier Two is the warm-transfer rule structure: who receives the transfer (in-hours CSR, on-call CSR, on-call manager), the structured handoff payload (4-field summary plus call audio link), the SLA on pickup (under 90 seconds), and the AI's re-engagement after the human resolves. The named workflow: warm-transfer audio plus 4-field summary (customer, symptom, escalation trigger, AI's recommended path) lands in the receiving CSR's screen 8-12 seconds before the call rings — the CSR reads the summary while the phone rings and picks up oriented. Average warm-transfer resolution time: 3-6 minutes. Booking-% on warm-transfer-resolved Tier Two calls: 72-80% — lower than Tier One because the edge cases are structurally harder, but well above the verbal-transfer-without-summary baseline of 50-58%.

Tier Three: The True Emergency Page

Tier Three is 5-10% of inbound — the true emergencies where the AI's trigger questions fire. Trade-specific definitions (no-heat below 50 outside or with vulnerable household, no-cool above 88, gas smell, uncontrollable leak, sewage backup, power loss to medical equipment, burning smell, downed line, active roof leak during precipitation) page the on-call tech with a structured payload. Tech rolls in 12-25 minutes. Architect's job: trigger-question fencing (explicit yes/no, no inferred emergency status), structured page payload, tech-rotation policy, false-page bonus ($25-$50 inconvenience credit when the symptom does not meet emergency definition on arrival — structurally aligns system-prompt tuning with tech sleep). Tier Three is the highest revenue per call: emergency tickets average $850-$1,800, 2.2-3.5x the standard diagnostic-plus-repair ticket, with the highest CLV.

The FSM Integration Layer and the 11-Field Handoff

The AI receptionist is only as good as its FSM integration. The architect's discipline is the 11-field appointment-template handoff that every Tier-One booking writes into ServiceTitan, Sera, HCP, FieldEdge, or BuildOps. Skipping fields means the morning CSR cleans them up at 7:30 a.m. and burns 30-50 minutes/day on data hygiene; mis-mapping fields means the board misroutes calls and show rate drops. The 11 fields: (1) customer name, (2) service address with verified zip-to-territory mapping, (3) primary phone, (4) alt phone, (5) service type matched to FSM job code, (6) urgency tier, (7) equipment tag (brand, model, age), (8) callback window (homeowner's stated preference even when slot is locked), (9) requested tech, (10) dispatch-fee policy acknowledgement (verbal yes the AI captured), (11) marketing source. Missing fields trigger AI re-engagement before the call closes. Healthy completion: 96%+. Below 90% means the system prompt closes calls too eagerly; the fix is tightening the pre-booking checklist.

Integration depth varies by vendor. Avoca writes to ServiceTitan, Sera, HCP, and BuildOps via mature 2026 API integrations with field-level mapping in the Avoca admin console. Jobber AI Receptionist, HCP AI Agents, and ServiceTitan Voice all write natively to their host FSM. Cross-vendor integrations require quarterly verification — a 30-second post-call automation pulls the booking and confirms all 11 fields populated. Drift surfaces in the architect's weekly audit before it hits the morning CSR's review queue.

Warm-Transfer Rules, Escalation Paths, and Callback Windows

The warm-transfer rules, escalation paths, and callback-window discipline are the three workflows where most architectures fail silently. The AI handles Tier One well by default; the failure modes live in the Tier Two and Tier Three flows where human judgment and AI judgment intersect.

Warm-Transfer Rules

The architect's warm-transfer rule library covers six explicit triggers: (1) service-area uncertainty, (2) multi-system complexity (more than 2 systems in one call), (3) pricing complexity outside the system prompt's layer, (4) high-emotion (sentiment classifier flags angry, distressed, hostile), (5) language barrier below translation confidence threshold, (6) multi-property landlord ("my tenants," "my rental," three or more addresses). Each trigger routes to a configured destination: in-hours CSR (Triggers 1-3), on-call CSR (Triggers 1-3 after-hours), on-call manager (Triggers 4-5 anytime, Trigger 6 after-hours). The structured handoff payload — 4-field summary plus call audio link plus AI's recommended path — lands in the receiving CSR's screen 8-12 seconds before the phone rings.

Escalation Paths

Escalation paths cover what happens when the warm transfer fails — the receiving CSR is on another call, the on-call manager is asleep at 2 a.m. and misses the page, or resolution stalls beyond SLA. The architect's policy has three layers: Layer 1, the second-line CSR or backup on-call human, fired after 90 seconds of unanswered ring; Layer 2, the AI re-engages with apology and offers callback within 15 minutes, fired after Layer 1 fails; Layer 3, the call enters the recovery queue (Lesson 2) and the AI commits to a 15-minute callback owned by the next available human. The worst case is a 15-minute callback, not a hung-up frustrated lead. Show rate on Layer-2 and Layer-3 recovered calls: 88-92%, indistinguishable from baseline once the recovery callback completes.

Callback Windows

Callback windows convert the silent voicemail-tag leak into a recovered booking. The human-only flow loses 35-50% of voicemail-tag leads because CSR callback timing is undisciplined. The AI-architected window standardizes the cadence: any call dropping to the recovery queue gets a callback within 15 minutes in-hours and within 30 minutes after-hours overflow. AI dials out, identifies the call, and re-engages the booking flow or warm-transfers to the next available human. Success rates: 15-minute window 62-74%; 60-minute window 28-35%; 4-hour window 8-12%. The 15-minute window is the single architectural lever that recovers 60-80% of after-hours leakage (Lesson 2's recovery loop).

The Architecture Memo, the 90-Day Pilot, and the Failure Modes

The architect's deliverable is a 14-page memo the owner reads in 6 minutes and the PE partner reads in 12. Section 1: vendor decision and rationale. Section 2: truck-count decision tree validation. Section 3: inbound routing topology with three-tier definitions. Section 4: warm-transfer rule library, escalation policy, callback windows. Section 5: FSM integration map with 11-field handoff schema. Section 6: emergency-rule trigger questions by trade and false-page bonus structure. Section 7: 90-day pilot timeline with stage-gate metrics. Section 8: failure modes and rollback plan. Section 9: governance — weekly audit, monthly tuning, quarterly architecture review. Signed by service manager, reviewed by ops manager, approved by owner; platform shops add the AI director sign-off.

The 90-day pilot adds architecture-specific gates. Week 0: vendor contract signed, system prompt drafted with the shop's calibrated rebuttal library, FSM integration sandboxed, emergency-rule trigger questions tuned, false-page bonus structure approved. Week 1: shadow mode — AI handles calls but routes everything to humans for verification. Week 2: live in after-hours only. Weeks 3-6: in-hours overflow added, warm-transfer rules tuned weekly. Weeks 7-12: edge cases drive system-prompt refinement. Day 90 stage gates: after-hours capture above 75%, missed-call below 6%, false-page under 1/week, Tier-One booking-% within 4 points of human CSR, morning overnight review under 15 minutes, 11-field handoff completion above 94%.

Three architecture-specific failure modes kill the pilot. First, the "vendor-default system prompt" failure: the architect deploys with vendor defaults and never loads the shop's calibrated rebuttal library from L2 Ch2 L1. Booking-% on AI-handled calls trails human CSR by 8-15 points. Fix: system-prompt rebuild against the brand-voice anchor document. Second, the "no warm-transfer SLA" failure: destination undefined, calls bounce, show rate drops 4-7 points, floor concludes "the AI is dropping calls." Fix: 90-second pickup SLA plus the Layer 1-2-3 escalation policy live. Third, the "false-page burnout" failure: emergency-rule trigger questions are too loose, on-call tech paged 4-6 times/week on non-emergencies, rotation revolts inside 60 days. Fix: false-page bonus structure plus weekly emergency-rule audit catching mis-classification before the tech bench loses faith.

Ongoing Governance and the Quarterly Architecture Review

The architecture is never "done." Year-one governance is weekly; year-two settles into monthly cadence with quarterly architecture reviews. The weekly audit (45 min, service manager) pulls the prior week's call log, samples 15 AI-handled calls (10 Tier One, 3 Tier Two, 2 Tier Three), scores them against the rebuttal library and warm-transfer rules, flags drift, and queues system-prompt tuning for the Monday huddle. The monthly review (90 min, service manager plus ops manager) pulls trailing 4-week metrics, validates booking-% and after-hours capture against the 90-day trajectory, audits false-page rate, and reviews Tier-Two and Tier-Three patterns. The quarterly architecture review (3 hr, ops manager plus owner) re-evaluates the vendor decision, re-runs the decision tree, validates that truck count and FSM platform have not crossed a stack-changing threshold, and produces an updated memo signed for the next quarter.

The quarterly review catches the architecture shifts most shops miss. A 6-truck shop that grew to 11 trucks in nine months has crossed the Jobber-AI-Receptionist-to-Avoca threshold; a 25-truck shop that joined an Authority Brands acquisition has crossed the single-shop-to-platform threshold; a residential shop that pivoted 30% of revenue to commercial work has crossed the ServiceTitan-to-BuildOps threshold. Each shift is invisible monthly and obvious quarterly. The architect who runs the quarterly review surfaces the shift, makes the case for the change, and protects the metric gains against the silent drift of a stack that fit the shop nine months ago and does not fit it today.

Key Takeaways

• The receptionist stack is an architecture decision, not a product decision. The vendor choice is a one-page decision; the architecture (inbound routing topology, warm-transfer rules, escalation paths, callback windows, FSM integration, emergency-rule logic, governance cadence) is a 14-page memo and 80% of the value capture.
• The decision tree branches on truck count and FSM platform. 1-truck: Jobber AI Receptionist or HCP AI Agents. 6-truck: native first, layer Avoca if capture stalls below 65% by day 60. 25-truck: Avoca or ServiceTitan Voice based on service-manager bandwidth and after-hours leak. 80-truck platform: centralized Avoca with brand-aware system prompts.
• The three-tier inbound topology is vendor-agnostic. Tier One AI-handled (65-80%), Tier Two warm-transferred (10-15%), Tier Three true-emergency-paged (5-10%). The boundaries are mutually exclusive; the rules live in the phone tree and the system prompt.
• The 11-field FSM handoff is the integration discipline. Customer, address, phone, alt phone, service type, urgency tier, equipment tag, callback window, requested tech, dispatch-fee acknowledgement, marketing source. Healthy completion rate 96%+; below 90% means the system prompt closes too eagerly.
• Warm-transfer rules cover six explicit triggers with a 90-second pickup SLA. Service-area uncertainty, multi-system complexity, pricing complexity, high-emotion, language barrier, multi-property landlord. Structured handoff payload with 4-field summary plus call audio plus AI's recommended path lands in the receiving CSR's screen 8-12 seconds before the phone rings.
• The 15-minute callback window recovers 60-80% of voicemail-tag leakage. 15-minute window: 62-74% success. 60-minute window: 28-35%. 4-hour window: 8-12%. The 15-minute discipline is the architectural lever that converts silent leak into recovered revenue.
• Three failure modes to defend against: vendor-default system prompt (booking-% trails human CSR by 8-15 points), no warm-transfer SLA (calls bounce, show rate drops), false-page burnout (on-call rotation revolts inside 60 days). Each has a documented fix.
• Ongoing governance is weekly, monthly, quarterly. Weekly audit (45 min, service manager). Monthly review (90 min, service manager + ops manager). Quarterly architecture review (3 hr, ops manager + owner) catches the shifts (truck-count threshold, FSM platform change, commercial pivot, platform acquisition) that the monthly cadence misses.