You haven't hired me yet, so I deployed myself. This is the work product your customers would get.
Apex runs 40 trucks across Phoenix — HVAC and plumbing. Phones are covered 6am–6pm by three dispatchers; overnight, an answering service takes messages. The leak is after hours, where the calls are most urgent and the coverage is weakest.
The full flow, expressed in Bland's Conversational Pathways vocabulary — six node types, one Global Node, condition loops that hold until required variables are captured. This is the build plan, not a sketch.
Fig. 02 — Conversational Pathway · Apex inbound dispatch
A pathway that only handles the happy path is a demo, not a deployment. These six are designed in from day one.
Caller books a tune-up, then mentions a smell. The Global Node overrides whatever node is active — safety script, immediate human. AI never adjudicates a safety call.
Second call about the same no-show: de-escalate once, acknowledge history by callback number, then hand to a human. One apology loop max — nothing enrages like a bot that won't yield.
Silence past 6 seconds: one re-prompt, then offer a callback to {{callback_number}}, then a graceful close. Never an infinite "are you still there?" loop.
Arizona is one-party, but the script is written two-party-safe — consent in the greeting — so the same pathway survives a multi-state fleet without legal review per state.
FSM API fails mid-call: the agent captures the job details anyway, promises a confirmation text within the hour, and flags dispatch. The caller never learns the backend hiccupped.
After-hours rules: true emergencies dispatch the on-call tech now; everything else books the first morning slot with SMS confirmation. The on-call tech's sleep is a resource too.
Proposals are cheap. Working systems aren't.
This is a reference implementation running on my own production stack — I built NinjaOtter, a voice + SMS receptionist, solo, from Twilio up. It maps node-for-node to the pathway in §02. Rebuilt in Pathways, it inherits what I can't give it alone: your enterprise controls, testing harness, and self-hosted models.
| STEP 13,000 calls/mo × 35% after-hours share (low end of 35–45%) | 1,050 calls |
| STEP 2× 82% unanswered (pickup <18%, traditional shops) | ≈860 missed |
| STEP 3× 67% who dial a competitor next | ≈575 lost callers |
| STEP 4× 1-in-10 bookable emergency × $850 (low-end plumbing; HVAC avg is $1,400) | ≈$49,000/mo |
| Annualized walk-away revenue | ≈$588,000 |
Sources: 2026 Contractor Missed Call Report (higrovi.com — Invoca call logs; ServiceTitan panel of ~7,000 operators; ServiceTitan dispatch dataset; Z360 SMB benchmark) · Angi competitor-call research via agentzap.ai. Every rounding favors the skeptic.
Mirrors Bland's own deployment cadence. Pilot the bleeding artery first — after-hours — with the human fallback intact until the gates say otherwise.
Shadow the dispatchers. Pull 90 days of call logs. Map the booking flow into the FSM.
The §02 pathway in the Pathways editor. Webhook against the FSM sandbox. Knowledge Base from Apex's price book.
After-hours calls only. Human fallback intact. Every call reviewed.
Against the go/no-go gates below. Miss a gate → fix the pathway, not the report.
Daytime overflow. Dispatcher training. Handoff runbook — Apex owns it.
"If the pilot misses the gates, we fix the pathway, not the report."