Claustel · AI you own

You don't own the AI your organisation now runs on.

Someone else does — and they can change the price, the rules, or whether you have it at all. This is how that happened, and what we build instead.

Request a first meeting Go to the details

We are looking for our first pilot organisations. No system of ours is installed anywhere yet. That means unusual attention from the two people who build it — and a risk you should weigh before the first meeting rather than after.

Scroll

Act one

It arrived, and it was remarkable.

Nobody had to be persuaded. Your people started using AI because it worked. Contracts summarised in seconds. Reports drafted before lunch. Questions answered that used to cost an afternoon of searching. Within a year it stopped being a novelty and became part of how the work gets done. Nobody wants to go back — and that is exactly the problem.

Act two

Nobody asked where it went.

Every one of those questions left the building. Not a summary of it — the actual text. The contract, the patient note, the drawing, travelling across the internet to a computer owned by a company you have never met, in a country whose laws you did not choose. It comes back so fast that it feels local. It isn't.

Act three

Then, one Tuesday, the terms change.

Example · Pricing updateEffective 1 March, pricing for your tier increases by 40%.

Example · DeprecationThe model version you rely on will be retired on 30 June.

Example · Policy changeWe are updating how long submitted content is retained.

Example · AccountYour organisation's access has been suspended pending review.

We wrote these notices as examples. The situations are real: price rises, retired models, changed terms and suspended accounts are the ordinary business of renting something. What has changed is what you are renting: not a tool any more, but something your people now need in order to do their jobs. Most AI is rented, and what is rented can be taken back.

Act four

So we asked a simpler question. What if it never left?

Not a stricter contract. Not a promise about deletion. No journey at all. A computer — a tower under a desk, or at most a filing cabinet — goes into your building. The AI lives on it. It reads your own documents, and we teach it to write the way you write — and the line around your premises, the one that was never really closed, closes.

Act five

Three things have to be true at once.

Useful

It has to actually help

An AI that has never read your work gives generic answers. Ours reads your documents first — and shows your staff which document each answer came from, so anyone can check it.

Private

Nobody sees it but the people you let in

Not a promise buried in a contract that you have to trust us to keep. Questions and answers are handled on the machine in your building, and no document is sent to an outside AI service. The one way in from outside is the maintenance access you grant us — and can close at any moment.

Yours

Nobody can take it away

You own the machine and the AI on it. No licence to renew, no account to suspend, and nobody upstream — us included — who can switch it off.

Act six

The moment you see for yourself.

On the last day of the installation we do something that sounds reckless. We disconnect your building from the internet. Then we ask your people to carry on working.

InternetDisconnected

Your AI systemStill answering

Questions answeredAll of them

Your AI keeps answering, at the same speed. That is when you understand what you have actually bought — not a subscription, but a machine that is yours.

Act seven

And then we stay.

Machines need looking after. Parts fail, your documents change, the work changes, and the person who understood the whole setup takes another job. We monitor it, put faults right or call in the machine's own warranty, add your new documents as they come, and keep it current — using access you grant us and can withdraw at any moment, without anything breaking.

Two ways on

You can hear the whole argument, or check the whole machine.

Talk to us about your own situation — 90 minutes, free, no obligation. Or read the detailed version first: the hardware, the process step by step, exactly what you own at handover, and how we keep it running.

Request a first meeting Read it in detail

Built for you · Installed in your building · Yours to keep

Your own AI. In your own building.

Nearly every AI tool in use today runs on computers owned by a handful of very large companies. Every question your staff ask, and every document they show it, is sent to those companies. We do the opposite. We build a complete AI system, install it in your building, and give it your own documents to read, so its answers are about your work. We adapt it to your terminology and the way you write. Nothing is ever sent to an outside AI service. Nobody can change the rules, retire the model or switch it off. And we stay on to look after it.

Request a first meeting See how it works

Where it lives: In your building
Your documents: Stay on your machine
Who controls it: You do
Who looks after it: We do

In plain terms

What we actually install.

Imagine hiring someone who has read every document your organisation has ever produced — every contract, report, procedure and email thread — and who can answer questions about all of it in seconds, at any hour, without ever getting tired.

That is roughly what modern AI can do for an organisation. The catch is where that assistant sits. If you use one of the well-known AI services, it sits in a data centre on another continent, owned by a company you have no real relationship with. Every time you ask it something, you have to send your documents there for it to read.

For a lot of work, that is fine. For a law firm, a notary, a fiduciary, a clinic, a bank or a manufacturer with drawings worth protecting, it raises real questions: professional secrecy, medical confidentiality, the duty of discretion, trade secrets, what you owe your clients. Using an outside AI service is not necessarily forbidden, depending on your profession and the kind of data; where it is allowed, it takes careful contracts, checks on the provider and sometimes your clients' consent.

So we build the same thing, in your building instead. A computer, from a tower that fits under a desk to something the size of a filing cabinet, goes into your office or server room. The AI lives on that computer. We put your own documents where it can read them, teach it your vocabulary and the way you write, and connect it to the systems your staff already use. Then it answers their questions from those documents, naming the one each answer came from — without any of it being sent to an outside AI service.

That is the whole idea. Everything below is detail about how we do it, what it costs, and what you end up owning.

The difference

It comes down to one arrow.

Here is the same everyday task — someone asking a question about a company document — drawn twice. Once the way it works today, and once the way it works with a system of your own.

In A the line around your building is drawn broken, because it is: every question sends a piece of your files outside, and what happens to them out there — who reads them, how long they are kept, what it costs next year, whether you still have access at all — is somebody else's decision. In B the line is closed. Two things cross it, both travelling inward and both on your terms: an update from us, and a maintenance session you open and close again. Your documents and your questions stay on your own machine. Unplug the connection and everything carries on working.

Our three goals

Useful, private, and yours. In that order.

Most suppliers can give you one of these. All three have to be true at once, because a system nobody uses is a waste of money — and a system everybody depends on, that somebody else can take away, is a much bigger problem than the one you started with.

Goal 01 · Utility

It has to be genuinely useful

An AI that has never seen your work can only give you generic answers. Ours reads your material first — your documents, your terminology, the particular way your organisation does things — so its answers are about your work rather than about the world in general.

And when it tells your staff something, it shows them which of your own documents it took that from, so anybody can check it. An assistant you cannot check is not much use in serious work.

Before we build anything, we agree with you how we will test whether it actually helps. If it doesn't beat the way you do the job today, you don't have to accept it.

Goal 02 · Privacy

Private in a way you can check

Your documents and the questions your staff ask are never sent to an outside AI service — not to us, not to whoever published the model, not to a cloud. They are read and answered on the machine in your building.

We are an outside company, and we do not ask you to take that on trust. If you keep us on to look after the system, our only way in is a door you open — and each part of our privacy promise comes with the way you check it, set out in full just below.

You open the door, you can read every visit, and you can close it for good.

Goal 03 · Sovereignty

Nobody can take it away from you

Anything you rent can be taken back. Prices rise, terms change, useful features get retired, accounts get suspended. Once an AI system is genuinely woven into how your organisation works, none of those are small inconveniences.

What we build is not rented. You own the computer and you own the AI running on it. There is no licence to renew, no account that can be closed, and no company — including ours — able to switch it off.

We prove this at handover: we disconnect your building from the internet and show you that everything still works.

The privacy promise

Don't take our word for it.

A privacy promise is worth what you can check for yourself. Ours has three parts, and each one comes with its check.

Promise 01 · Where your work goes

Never sent to an outside AI service

Your documents and the questions your staff ask are read and answered on the machine in your building — not sent to us, not to whoever published the model, not to a cloud. And because they are never sent to such a service, they cannot end up training somebody else's AI.

How you check: At handover we unplug your building from the internet and the system keeps answering. That shows it needs nobody outside to do its work; it does not show what happens once the line is back — so your IT provider is welcome to watch the machine's network traffic, that day and afterwards.

Promise 02 · Who can get in

One way in for us, and you hold the door

If you keep us on to look after the system, our only way to reach it is a door that you open: closed by default, open only to named people, and every session logged in a record you can read.

How you check: Read the log whenever you like, and close the door for good whenever you like. The system keeps running without it; what ends is our remote maintenance.

Promise 03 · What can be taken back

Nothing to renew, so nothing anyone can take away

There is no licence fee and nothing to renew — not to us, not to anybody else. The model comes under a perpetual licence that cannot be revoked for the version you receive, and its full text is in the written plan. Security updates are included in the service contract, which is optional, and are never sold as a separate licence. If you end that contract, the system works the next day exactly as it did the day before; what stops is our work on it, updates included.

How you check: It is written in the plan you receive and, if you take one, in the service contract — before you sign anything.

What we do not promise

We do not promise that nothing can ever go wrong. The machine is as safe as the office around it — its door, its network, its passwords — and we go through that with you, and with your IT provider if you have one, before anything is installed. If your work involves client matters or personal data, keeping them away from outside AI services makes the questions of professional secrecy and data protection simpler to answer; the answer itself stays with your own adviser.

We are an outside company. What makes that acceptable is not our word — it is the door you hold, the log you can read, and a system that keeps working whether or not we are still around.

How we do it

Built around your work, not around a box.

We don't hand you a sealed box. You see what goes into the system, you get a written list of every piece of software in it and the terms it comes on, and each stage produces something you keep — each one ending at a point where you can stop.

Working out what you need

We spend a working session with the people who would actually use this, and work out what would genuinely help — which jobs currently take too long, which documents matter, and what that work costs you in hours today. You leave with it in writing: what the system would do, what it would run on, an honest price, and the test we will judge it by. If our conclusion is that you shouldn't do this yet, we will say so, and what we wrote is yours to keep either way.

Specifying the equipment

We work out exactly what equipment you need — no more than necessary — and where it will physically stand, including power, noise and cooling. You receive a named list of parts, so the machine can be bought in your own name, from your own supplier or your own IT contractor, with the warranty in your name too — and the people who already look after your computers go on doing it. Your own IT and security people are part of this rather than presented with it afterwards. Where you would rather we handled the ordering and the assembly ourselves, we can do that instead.

Installing it

Once the machine is in your building we install the system on it and connect it to your network, together with whoever looks after your IT. Everything is set up in a way your own staff can read and follow — no sealed boxes, and nothing we can't explain to them while standing next to it.

Fitting it to your work

This is the stage that makes it yours rather than generic. We gather and tidy up your documents and set the system up to read them, so that every answer can point back to the document it came from. We also adapt its vocabulary and the way it writes to your own. Then we test it with you against the standard we agreed in stage one, on real tasks out of your own files — because the problems that really matter only show up in real work.

Handing it over

We train your team, hand over the written instructions, and run the agreed tests together. Then we disconnect the building from the internet to show you it still works — and give you every password and key.

What you own

What you have when we walk out.

"You own it" is easy to say. This is the list it actually comes down to — written into the contract before any work begins.

01The AI itself — the actual thing, installed on your own machine, yours to keep and to go on using. Nobody can take that back, and there is nothing to renew.
02Your documents as we prepared them, the searchable index built from them, and the settings and tools we used to adapt it — kept on your own hardware, so that somebody else could rebuild all of it without us.
03All of the software we wrote for you, kept somewhere you control.
04Plans of the physical setup — what is installed where, how it is wired, how much power it draws.
05Written instructions for everything routine: shutting it down and starting it up, backing it up, restoring it, adding new documents, adapting it again, replacing a broken part.
06Every password and administrator account — and ownership of the equipment itself.
07A written list of anything the system still needs from the outside world. Our aim is for it to be one line long, and for that line to say "nothing".

About our own access

If you keep us on to look after the system, we work on it using access that you give us, through your own systems, recorded where you can see it.

You can withdraw that access at any moment, without warning, and nothing breaks. A supplier you can remove cleanly is the only kind worth having on something this important.

Looking after it

Building it is the short part.

A system like this runs for years. Parts wear out, your documents change, the work changes, and the person who understood how it all fitted together takes another job. Looking after all of that is our second service.

What we cover, how quickly we respond and how often we are on site are agreed with you, around your system and your team — there is no standard package. We go through it together when you book a demo.

Request a first meeting

Compared with the alternatives

Three ways to have AI, and what each one costs you.

The middle column is the one worth reading slowly. A "private" version hosted by the same companies really does keep your data more separate from other customers — and changes nothing at all about who your organisation depends on.

	Public AI services	A "private" hosted version	Your own system, on site
Where the AI actually runs	Their data centre	Their data centre	Your building
Where your documents go	Sent over the internet	Sent over the internet	Nowhere outside — they stay on your machine
Works if your internet is down	No	No	Yes
Who owns the AI	They do	They do	You do
Can the price rise without your agreement	Yes	At renewal	No — you already own it
Can it be discontinued	Yes, on their timetable	Yes, on their timetable	Only if you decide
Can your access be cut off	Yes	Yes	There is nobody to cut it off
What each question costs	Charged per use — grows as staff use it more	Charged per use, or paid for in advance	The electricity it uses
What it costs to leave	Start again from scratch	Start again from scratch	Nothing — you already hold everything

Common questions

The questions you are probably asking first.

Is it as good as the well-known AI services?

For general questions — trivia, travel advice, writing a poem — no, and we will not pretend otherwise. For your work, it has read your documents without them being sent to an outside service. We settle this by testing rather than arguing: before we build, we agree with you on a test made of real tasks from your work. If the system does not beat the way you work today, you do not have to accept it.

What does it cost?

Two things, in Swiss francs. First, the system itself: the machine it runs on, and our work to specify it, install it and adapt it to your work. However the machine is paid for, it ends up yours — we do not rent it to you. Second, if you want us to look after it, a service contract written around your system. No licence fee and no charge per question or per document, so heavier everyday use does not change what you pay us. The bill changes only if you add capacity or change the service you want from us. We don't publish a price list. After the first meeting you have a first estimate of size and price for your case. If you want to compare it with what you spend today, we do that calculation with your own figures, and we will tell you if the honest answer is "not yet", or if the case rests on confidentiality rather than cost.

Do we need to hire AI experts?

No. That is exactly what the support agreement is for. If you would rather build that skill inside your own team, tell us early and we will arrange the work so that we train your people and hand over completely. That is a good outcome, not a lost customer.

What if we grow?

You move up to a bigger machine rather than starting again: your documents, the index, the adaptation and the software all move with you. We tell you in advance where the next step up sits and what changes when you do, so growing is a decision you make rather than a surprise you get.

Can it work with no internet at all?

Yes, if you need it. Some organisations require it; updates then arrive on physical media rather than over a connection. It is slower and costs more to run, so we will tell you frankly if your situation does not call for it.

Whose AI is it, and where does it come from?

The part that does the thinking is a model somebody else has published openly, and we install a copy of it on your machine. It is not ours and it is not rented from anybody: once it is installed, it never contacts whoever made it. We only install models whose published terms cannot later be withdrawn — that is what makes the rest of this page true, because no company, ours included, can then raise a price, change the rules or switch it off. Which model we propose depends on the work and on the machine it has to run on; we name it, and our reasons, in the written plan before you commit to anything, and you receive it with its licence text. What makes the system yours is what goes on top: your own documents for it to read, the reference it gives for every answer, and the terminology and style we adapt to your work. If a better model appears later, it can be swapped without starting again.

What if your company disappears?

Your system carries on working, and somebody else can look after it. You already hold the AI, the documents, the software and the instructions. Building it so that our disappearance doesn't hurt you is the entire point — a supplier you cannot survive losing isn't really a supplier.

First meeting

Start with your own work.

A first meeting takes about 90 minutes. We look at the first task you would want help with, what your documents are like, and the rules on where your data may be kept. You leave with a first estimate of size and price for your case — and sometimes with our advice to wait. The build, the timetable and the service contract are worked out for you, and we give you the details afterwards.

There is no charge for it and nothing expected of you afterwards.

Useful to bring, if you have it

One job you would want help with first · roughly how many people would use it · what your documents look like and where they are kept · any rules about where your data is allowed to be · what that work costs you today.

Your organisation

What kind of work you do

The first thing you would want it to help with

Roughly how many people would use it

How sensitive is the material

When are you thinking of doing this

Your name

Your email

Your request is sent to us and answered by e-mail. Only your name and a way to reach you are needed, and we reply within two working days. The text is shown to you before it goes, and if sending fails you can send it yourself from your own mailbox.

Your request — ready to send Open in my email

Send it to

Who we are

The founders.

We are two mathematicians, trained at both EPFL and ETH Zurich, based in Geneva and Vaud. We work in French and English. Claustel is new, and we are looking for our first pilot organisations — so you would be early, and treated accordingly.

Co-founder

Ismail Bouhaj

He has done research on robust machine learning at EPFL, and produced expert training data to improve large language models.

Co-founder

Antoine du Fresne von Hohenesche

At ETH Zurich he builds machine-checked proofs of how algorithms behave, and he has evaluated the mathematical reasoning of large AI models.

Request a first meeting