AI Receptionist Pilot Spec

Executive summary

Enderby handles a steady volume of inbound calls — delivery scheduling, tank monitoring, billing questions, emergency triage — all routed through a human dispatcher. That model works for complex situations and breaks down for everything else: after-hours calls go to voicemail, routine billing eats dispatcher time, and no-heat calls during cold snaps create handle-time spikes the team can't absorb.

The AI Receptionist is a voice-based frontline agent designed to answer calls 24/7, capture intent, resolve routine questions, and escalate urgent issues to the right human channel — without changing how Enderby fulfills service.

1 / 3

The business problem

Enderby's current call-handling model is fully human-driven. Effective for complex issues, but it creates five operational gaps that directly impact customer satisfaction, dispatcher capacity, and revenue:

• Missed calls. After-hours calls roll to voicemail. No-heat situations, tank outages, and delivery concerns wait until morning — creating delayed response and next-day call spikes that overload dispatch.
• Lost leads. New-customer inquiries that arrive after hours or during peak hold times go cold before a CSR can call back. In a low-switching-cost market, lead-capture velocity matters.
• Slow response time. During peak season, hold times climb. Long waits and repeated transfers create a perception of poor service even when internal effort is strong.
• Manual scheduling burden. Dispatchers spend hours per week on calls that require no judgment — confirming windows, answering "when is my next fill," capturing basic new-customer info.
• No consistent documentation. Call notes live in dispatcher heads, paper notepads, and inconsistent Cargas entries, so the next CSR starts over and patterns are invisible.

Pilot objectives

The pilot is designed to validate, not deploy. It must prove three things during the two-week window: that an AI agent can resolve at least 60% of routine calls without a human, escalate emergencies with at least 90% accuracy against human-reviewer judgment, and cut after-hours voicemail volume by at least half. If those benchmarks hold, the pilot becomes the foundation for a production deployment.

Workflow — six steps on every call

1. Customer calls. A standard greeting identifies the agent as AI and offers immediate human handoff.
2. AI answers. Twilio/Vapi receives the call and streams audio to the model, which asks an open-ended intent question.
3. AI qualifies intent — billing, delivery status, tank level, new customer, or emergency. One clarifying question, then escalate if still unclear.
4. AI books, routes, or summarizes using the NotebookLM knowledge base (sourced from Project 2's SOPs).
5. Business gets a clean follow-up — a structured summary to the dispatcher queue; emergencies trigger an immediate notification, not just a summary.
6. Caller gets confirmation of the next step. For escalations, the caller is told a human is being connected.

System architecture

Risks & guardrails — the part that matters most

The risk profile of a voice agent at a propane distributor is non-trivial; gas safety is the central concern. The spec includes four explicit guardrails:

• Emergency calls route to a human. Any gas odor, suspected leak, CO alarm, no-heat with a vulnerable resident, or critically low tank → immediate handoff. No diagnostics, no troubleshooting, no holding.
• Ambiguity escalates after one clarification rather than letting the agent keep probing. A needless escalation costs a few dispatcher minutes; mishandled ambiguity in this domain costs much more.
• Consent & recording disclosure in the opening greeting, reviewed against Texas telephone-consumer regulations; callers who decline route to a human.
• Human review of every call touching an emergency, vulnerable resident, or legal-risk statement within 24 hours — building the audit trail for a go/no-go decision.

Pilot cost (simulated)

All numbers are based on publicly available pricing and are simulated for portfolio purposes. The pilot is constrained to free or low-tier plans — estimated total cost under $500 across the two weeks.

Success metrics

Go / no-go decision logic

Appendix — a mock call

Simulated for demonstration: a no-heat call with no safety risk, where the agent handles intake, screens for danger, gathers structured information, and routes to dispatch.

Businesses that benefit from this pattern

A voice-agent pilot can make sense for businesses with high inbound volume, after-hours demand, repeatable questions, and a clear human escalation path. It is not a fit when calls are mostly high-risk, emotionally sensitive, or impossible to resolve without expert judgment.

Enderby Gas is a real company used with the owner's permission. All names, customer data, and operational figures are simulated for portfolio purposes; cost and impact figures are based on publicly available pricing and industry benchmarks. No proprietary records were used.