Every AI company has a demo video. This is a phone number. It rings a machine in Urbandale, Iowa that I built, and it will answer your question the same way it answers mine. Below is exactly what happens between you dialing and it speaking — including the part where it decides it doesn't know.
Reading about a phone system is a poor substitute for hearing one. It takes about ninety seconds, and the rest of this page will make more sense afterward.
Ask what it runs on. Ask how it works. Ask when I'm available. Ask something it can't possibly know — that's the interesting one.
Say anything else — or stay quiet — and you get the everyday line instead. That one only plays back answers I wrote and approved, and takes a message for anything else. Both lines are real; the difference is explained further down.
There's no mystery in it, and the ordering is the point. Each step is allowed to do one thing, and the step that could invent a sentence to a stranger doesn't exist on the everyday line.
In full, before anything is written down. It's the first thing that happens and it can't be skipped or shortened.
Speech becomes text on the machine itself. No transcription service, no outside API, no audio leaving the box.
The model's only job is to pick from a list — which of these approved answers fits? It's choosing, not writing.
Wrong category, suspicious input, or an answer already played this call — all of those disqualify a match before you hear it.
A match plays as a recording. No match takes the other branch, below, which is the more important half.
The audio was written by a human, recorded once, and approved by someone listening to the actual file. The model picked which one plays. It did not write a syllable of it, and it can't play the same one at you twice in one call.
Then it does the useful part: it drafts what it thinks the answer should have been and puts it in my queue with the original question attached. I approve or rewrite it, and the next caller who asks gets a real answer. The library grows from the questions people actually ask.
Most "AI won't hallucinate" claims are a promise about a prompt. A prompt is a suggestion. This one is built into the wiring instead, which is the only version that survives a bad day.
The part that turns text into speech refuses free text. Hand it a sentence nobody approved and it rejects it.
A rule in a document survives exactly as long as everyone remembers it. A machine that physically cannot render an unapproved sentence survives a bad config, a tired afternoon, and somebody new touching the settings.
So on the everyday line there is no code path that ends in a made-up sentence reaching a customer. Not one that's discouraged — one that doesn't exist.
The same system with one difference: where the everyday line would say "let me have Tyler follow up on that", the demo composes an answer instead, grounded in the approved material.
It runs on separate hardware, deliberately, so the two can't be confused. It exists so you can hear the ceiling of what's possible — then choose where your own line should sit. Most businesses want the strict one pointed at customers.
The system is only worth anything if the person running it never has to think about servers. Everything the machine wants from you shows up in one place, ordered by what needs you first.
Read a draft and send it, or fix the wording first — the system notices what you changed and drafts closer to your voice next time.
Write a new answer and hear it read back before you decide whether it goes live. Approval is against the recording you actually listened to, so nobody can quietly swap the file underneath it.
And see what happened while you were closed: what came in, what got answered, what's still waiting on you.
Nothing exotic and nothing rented. The interesting engineering is in what it refuses to do, not in the parts list.
A Ryzen 7 with 32GB of memory paired with a 16GB NVIDIA card. Roughly $1,700 of hardware, sitting in a closet drawing about as much power as a couple of light bulbs.
The model needs no internet connection and no outside service to answer a question. I've run one from a truck on a hotspot. No customer data goes to any AI company, mine included.
A carrier still carries the call, the same as every other call you take — that part doesn't change and I won't pretend it does. The encryption on that leg was verified on the wire, not assumed from a settings page.
Transcripts and message text are overwritten after the window your policy sets — mine is 30 days. There's a test that reads the raw database afterward and fails if the text is still findable.
The system opens no inbound ports on your network. Its security is the security of the network you put it on; it doesn't hand anyone a new door.
Automated checks run against the deployed system, not against the code that was supposed to be deployed. Where a claim on this page is testable, there's a test that fails if it stops being true.
Dory is one shape this takes. The same pieces — a model on your hardware, a library of things a human approved, a queue where you say yes — apply to whatever your version of "nobody has time for this" happens to be. The phone was just the one that cost me the most leads.
You've heard mine answer. A conversation about yours costs nothing, and if AI isn't the right answer for your situation I'll say so.
Start a discovery call