There is a bargain almost everyone has quietly accepted: an assistant can be useful, or it can be private, and you choose.
Useful means it knows your situation, which means your situation has to go somewhere. Private means you keep your material to yourself, which means the assistant stays a stranger and gives you the same generic answer it gives everyone else. Most people take a position somewhere in the middle and feel slightly uneasy about it.
That bargain is not a law of nature. It is a consequence of where the computer is.
Why the trade exists at all
To answer questions about your world, a system has to have read your world. There is no way around that part. The question is where the reading happens.
When the system lives in a data centre, your material has to travel to it. Once it has travelled, you are relying on terms of service, retention policies and the good behaviour of a company you do not control. Those may all be perfectly sound. But the guarantee you have is a promise, and promises are the kind of thing that changes when a company is acquired, changes policy, or is compelled by someone else.
That is why sensible people keep their most valuable material out. And it is exactly the material that would have made the assistant worth having.
Move the computer, not the material
The alternative is unglamorous and effective: put the machine where the records already are.
A small appliance sits in your building. Your material is loaded onto it. The training happens there. The model that results lives there. Nothing has to travel, because the thing that needed to read your material is in the same room as your material.
What changes is the nature of the guarantee. It stops being a promise about what a company will refrain from doing, and becomes a fact about where the data physically is. Those are different kinds of assurance, and only one of them survives a change of ownership at the other company.
What is trained, and for whom
This is the part worth being precise about, because the industry has made the word “training” frightening for good reasons.
When people object to their data being used for training, what they object to is their material improving a model that someone else owns and everybody else uses. That objection is correct, and it does not apply here.
Your material trains your model, on your machine, for you alone. Nothing you provide is pooled with anyone else’s, shared, or used to improve any other model. There is no common corpus for it to join. The model is an asset of your household or your practice in the same way your filing cabinet is, and it is about as portable.
The honest limits
Three things are true and worth saying before anyone sets expectations too high.
A local model is not the largest model in the world. For general knowledge — a question about history, or law in the abstract, or a language you do not speak — the big hosted models remain better. A sensible system lets you send a question out when you decide it is worth it, and tells you plainly that it is doing so.
Local is not automatically private. A machine in your building that quietly sends telemetry somewhere is not meaningfully more private than a hosted one. The property that matters is whether anything leaves without your say-so, and that is a design decision, not a location.
Personalisation still depends on the material. An assistant trained on a disorganised archive gives disorganised answers. Where the training happens solves the privacy problem. It does not solve the preparation problem.
What it looks like when it works
You ask what the roof work cost, and get a number, the year, and the note that the same leak was fixed twice. You ask who has not replied, and get four names and the drafts. You ask what changed in the new contract, and get the clause.
None of those questions could be asked of a system that has not read your papers. All of them are ordinary. That is the point: the impressive thing about an assistant that knows your world is how unremarkable it makes the rest of the week.
How local privacy actually works →
Find out what your own material supports →
Local AI. Private data. Local training.
$8,995 the first year — everything included. $4,995 each year after. The first year costs more because it contains the Jetson Orin placed and configured, your archive loaded and the first training run; every year after is the service running.
Or add the Archive Assessment first — $1,950, credited in full
You order and pay at checkout; we write back within days — a decline returns every dollar, and until our letter confirms the year you may withdraw in writing. From that letter the year is final, and it arrives on the date the letter names. Prefer to write to us first?