AfterHours · The night desk for service businesses - Vertical Agent Spec

One-line definition

An agent that answers inbound calls, screens and books leads, and delivers structured intake briefs for HVAC, plumbing, legal, and clinic businesses during hours when no staff is available.

The workflow it owns end-to-end

Call receipt: Inbound call arrives after business hours; the agent answers within two rings with a business-specific greeting, identifies itself as an automated assistant, and begins intake.
Urgency triage: Agent classifies the call as emergency (burst pipe, no heat in a Minnesota winter, possible medical situation) versus routine (scheduling, pricing inquiry, callback request) using a decision tree built on vertical-specific signals, not general reasoning.
Booking or message capture: For routine calls, the agent checks real-time calendar availability and books the appointment, confirming date, address, and contact info. For non-bookable calls, it captures a structured message and sets a callback expectation.
Brief generation: At call close, or at a configurable batch time like 6 AM, the agent compiles a plain-text brief per caller: summary, urgency tier, booked slot or open item, and a verbatim clip of anything ambiguous.
Delivery: Brief goes to the owner via SMS, email, or CRM note before the first technician or staff member arrives.

What it knows that a generic LLM doesn't

Triage vocabulary by vertical: "No heat" in January in Minnesota is a P1. "No heat" in July is a routine callback. The agent needs coded urgency rules per service category and geography, not improvised reasoning.
Dispatch eligibility logic: Whether a caller's zip code falls inside the service area, whether they hold a service agreement (priority scheduling), and what the after-hours surcharge is, all confirmed before a booking is completed.
Legal intake protocol: Conflict-of-interest screening questions must precede any case detail. The agent collects adverse party names first, stops there, and never crosses into anything resembling advice.
Clinic-specific deflection language: How to route a caller toward 911, a nurse line, or a next-morning appointment without making a clinical judgment. This is a rehearsed script with liability-aware language, not improvised output.
SMB owner brief format: These buyers skim on a phone at 6 AM. Three lines per caller, action item first, no narrative.
Seasonal urgency weighting: HVAC and plumbing call criticality shifts by month and region. A general model doesn't carry this without explicit injection, and injection at runtime is brittle.

What it explicitly declines

Clinical or medical advice of any kind: The agent cannot assess symptoms, recommend medications, or advise on whether a condition requires immediate care. It routes to 911 or a nurse line and ends the intake.
Legal guidance or case assessment: The agent collects intake information only. It does not evaluate case merit, estimate outcomes, or comment on deadlines.
Emergency dispatch decisions: If caller language meets a life-safety threshold, the agent states clearly that it cannot help and instructs the caller to dial 911. It does not attempt to manage or resolve the situation.
Irreversible financial commitments: The agent cannot quote a binding price, authorize a repair, or process payment. Booking a free estimate is within scope; committing to a job cost is not.

Tools and integrations required

Twilio Voice: Inbound call handling, call flow logic, and compliant call recording with required disclosure.
ServiceTitan or Jobber: Real-time schedule read and write for HVAC and plumbing booking.
Calendly or Acuity: Fallback scheduling for legal and clinic verticals that don't run field-service software.
HubSpot or GoHighLevel: Lead record creation, call log, and brief delivery via CRM activity note.
Twilio SMS or SendGrid: Morning brief delivery to the owner at a configurable time.
Deepgram or AWS Transcribe: Call transcription for brief generation and audit trail. For clinic clients, the transcription vendor must be a HIPAA-covered entity, which eliminates most low-cost options and is a real cost driver.

Trust escalation: when it pings a human

Life-safety keyword detection: Any call containing language associated with active flooding, fire, cardiac events, or injury triggers an immediate SMS to an on-call human and instructs the caller to dial 911. The agent does not continue the intake.
Caller distress signals: Crying, shouting, repeated escalation attempts, or explicit requests for a human pause the agent and route to a flagged voicemail in the morning brief.
Ambiguous booking data: If the caller's address falls outside the service area, or the requested slot conflicts with an existing booking, the agent captures the request without confirming and flags it for manual review rather than guessing.
Legal conflict-of-interest hit: If the adverse party name matches an existing client in the CRM lookup, the agent stops intake, tells the caller the firm may have a conflict, and asks them to call back during business hours.

Pricing model

The honest model is a monthly floor plus a per-booking fee: $99 per month for call answering and brief delivery covering up to 40 calls, then $8 per confirmed booking beyond that. At typical close rates for HVAC and plumbing, a single recovered after-hours booking covers two to three months of the floor. The $149-249 flat range in the GTM plan prices on perceived value, which works until a slow month triggers a cancellation. A usage-based floor protects retention because the bill drops when volume drops; flat SaaS does not. The clinic vertical warrants a separate tier at $299 minimum to cover HIPAA BAA overhead and compliant transcription costs, and that number should be stress-tested against what a small clinic will actually sign.

Differentiation from a generic LLM wrapper

A buyer who pastes their workflow into Claude will get something that handles the easy 80 percent of calls adequately and fails the hardest 20 percent in ways that are hard to anticipate before they happen. The value of a purpose-built agent is not the AI layer; it is what sits underneath: coded triage rules per vertical, live calendar read and write, CRM logging, compliant call recording, and a brief format calibrated to how an owner actually reviews their morning queue. Those integrations take months to build and tune against real call data, and the liability structure around emergency deflection and legal intake requires explicit design decisions that a generic wrapper leaves undefined. By 2026, "I'll just use Claude" is a real competitive threat only until the first emergency call goes wrong and there is no audit trail, no escalation log, and no documented liability framework in place.