How an AI receptionist, an ontology and an edge box fit together — written for owners, not engineers. Worked example: a Hill Country ranch resort.
Underneath every pattern in this guide sits the same rule: facts are typed the moment they arrive, carry who asserted them and how sure they were, and a model can propose — it never decides what's true.
Every service business has the same hidden leak: the phone. Calls after hours go to voicemail — and voicemail is where revenue goes to die. Calls during the rush stack up while your best person juggles a guest at the desk, a text about a late checkout, and a contractor asking for a gate code. None of this is a people problem. It is an information problem: the answers exist — in your PMS, your booking engine, your maintenance log, your policies — but no human can hold all of it at 6:00 PM on a Friday.
The AI industry’s answer has been a chatbot bolted to a website. Ours is different: give the machine the same thing you’d give a great new hire — a model of how your business actually works — then let it answer the phone with that model behind it.
Chapter 2An ontology is your business, written down as things and rules. Not documents — things: Guests. Bookings. Units. Trail rides. Releases. Work orders. And the relationships between them: a Guest holds a Booking; a Booking occupies a Unit; a trail ride requires a signed Release.
Once your operation exists in this form, a machine can answer questions the way your ten-year veteran does — by looking things up and following the rules, not by guessing. That distinction is the entire product. A language model guesses beautifully. An ontology doesn’t guess at all.
The edge box. A small appliance in your building runs everything: the voice agent, the ontology, local AI models, and the security stack. Your data never leaves. The cloud is our supply line — signed model updates travel in; your guest data doesn’t travel out.
The receptionist. Answers every call in under a second, speaks from the graph, books, reschedules, texts back — and hands to a human the moment a caller is frustrated or a rule says “this needs a person.”
The guardrails. Every answer is checked against the ontology before it’s spoken. Every action is logged. Anything the graph can’t verify, the agent won’t claim.
Worked exampleSilver Creek’s front desk handled 300+ calls a day in season. A third were the same twelve questions. Trail rides required signed releases — chased by hand. After 9 PM, calls went to a binder and a prayer.
Their ontology, at go-live:
| Entities | Links | Actions | Rules |
|---|---|---|---|
| Guest · Booking · Unit · Party | Guest holds Booking | book · reschedule · cancel | No trail ride without signed release |
| Activity · Release · Task · Channel | Booking occupies Unit | send release link by text | Quiet hours: no outbound after 9 PM |
| Rate · Policy · Staff · Vendor | Activity requires Release | open + route work orders | Pets: units yes, lodge no — always |
| (grows weekly from real questions) | Task assigned-to Staff | escalate to human on frustration | Refunds over $200: human approves |
A Tuesday, 7:02 AM. A guest calls: “Can my daughter ride this afternoon if my husband signed her release last year?” The graph checks: Release → signed 14 months ago → policy says annual renewal → the agent answers in one breath, texts a fresh release link, holds two saddles, and logs it. Zero staff seconds. The model never guessed — it looked it up.
By month two: 96% of calls answered inside one ring · the twelve repeat questions fully absorbed · release chase time to zero · an estimated $11k/month in recovered after-hours bookings. The ontology grew from 100 covered questions to 109 — because the corpus loop watches what real callers actually ask.
First principlesMost “AI companies” are a thin wrapper around someone else’s frontier model: your calls, your customers, your files — shipped out, marked up, handed back. Easy to build. Impossible to trust. Ontology is the hard road — modeling your business from first principles — and it’s the only one that ends with something you own.
Wrapping a frontier API takes a weekend. Modeling a business — its entities, links, actions and rules — takes real engineering, which is exactly why almost nobody does it. First principles, first things first: build the model of the business, then let AI speak from it. That ordering is the whole game.
And at scale, it only gets harder. Vibe-coded wrappers demo beautifully and then collapse under real call volume, real compliance, real attackers — and they leave your company exposed while they do it. This is the biggest thing we see in the field: businesses handing their alpha — their pricing logic, their playbooks, their customer relationships — to a wrapper, where it evaporates into the ether and trains someone else’s model.
An answering center runs 1,900 calls a day across 42 client brands. The ontology holds Brand · Script · Escalation · On-call — so the agent greets as the right company, follows that brand’s rules, and pages the right plumber at 2 AM. Agents now handle exceptions, not volume.
100% answered · brand never breaks character
A multi-provider clinic drowned every morning: reschedules, refills, “do you take my insurance?” The ontology gates everything — PHI never leaves the building, consent rules enforced, coverage checked from the graph. Recall campaigns text patients by rule, never after hours.
no-shows down · HIPAA-clean by architecture
Brokers call all day for check calls; drivers call about detention. The ontology binds Load · Lane · Stop · HOS-Clock to live ELD feeds — so status questions answer themselves and detention claims file on the clock, automatically. Dispatchers dispatch again.
HOS-aware always · detention captured, not forgiven
9 PM, golf-ball hail, 400 client households — 60 calls an hour. The ontology binds Policy · Coverage · Claim · Carrier, so every caller gets: claim opened, the carrier’s claims number, photos requested by text, adjuster expectations set — while the producers sleep. The graph knows who has comprehensive. Nobody’s coverage gets guessed.
renewals never slip · E&O-safe by architecture
Sixty techs, seasonal surge, and a 6 PM voicemail cliff. The ontology holds Equipment · Contract · Warranty · Priority rules — “no heat + infant in home” pages on-call now; “thermostat blank” books tomorrow 8 AM. Emergencies surface; everything else self-schedules.
after-hours revenue recovered · zero triage guesswork
We are a security company that built an AI platform, not the reverse. Every endpoint in your building — desk PCs, door controllers, cameras, the edge box itself — carries enterprise EDR and reports to a 24/7 SOC with 40-second containment playbooks. Your design files, guest lists and call recordings never leave your building. Staff get one governed way to use AI — so they stop pasting your business into public chatbots.
Chapter 6Everything above runs on hardware you can own, models you can buy outright, and an ontology that is yours — exportable, documented, portable. We feel like you’ll never leave us. But if you do: you’re welcome. That promise is on our pricing page, in writing, because a platform you can leave is the only kind worth staying on.
Next step: a 30-minute walkthrough scopes your exact stack. info@ontiqlabs.com →