Support
Can open
- Care plans
- Booking rules
- 2 outside this role’s access
Nothing this role can open answers that.
Behind the answer2 of 4 documents searched. The document that answers it is outside this role’s access, so it was never searched.
Private AI
We build AI that works on knowledge you can’t hand to a public tool.
That material usually comes with rules: who may read it, where it may be processed, what happens when an answer is wrong. We look at those conditions first, then say what is worth building, or whether anything is.
Where this starts
Three things usually bring this up.
Most work needs none of them. When one is true, the ordinary answer stops fitting and the question turns into an investigation.
-
Knowledge that can’t leave
The answers live in material you are not free to upload: signed client work, personnel records, medical or financial files, contracts with confidentiality terms.
What we do firstList what the material actually is, which parts each role may read, and who in your company decides that.
-
A model that has to be adapted
The work has its own vocabulary, its own format, or a way of being right that a general model can’t infer from the words alone. It gets close and stops there.
What we do firstLook for examples of the finished result, enough of them, and a way to tell a good answer from one that is nearly right.
-
A constraint on where it runs
A contract, a regulator or one of your own clients has written down where this data may be processed, and by whom.
What we do firstRead the constraint as it is written and work out which arrangements satisfy it, including the ones that cost less than you expect.
If none of this is true of your work, an assistant on the tools you already use, or an agent doing one job, is usually faster and cheaper. AI Integration AI Agents
The same question, from three roles.
Built by Version2
Four of our own documents and three roles with different access to them. A document a role may not open is removed before the search runs, so nothing in it reaches the answer, not even its wording. Put your own documents and your own roles in place of ours.
- 01 Care plans After launch
- 02 Quote and payment rules Before we start
- 03 Booking rules The first call
- 04 Offer rules Discounts
Question
Project lead
Can open
- Care plans
- Quote and payment rules
- Booking rules
- 1 outside this role’s access
A quote up to €800 is paid in full. Above that a 20% deposit reserves the work, never less than €200, and the quote stands for fourteen days.
A quote up to €800 is paid in full. Above that, a 20% deposit reserves the work, and never less than €200. A quote stands for fourteen days.
Behind the answer3 of 4 documents searched.
Finance
Can open
- Quote and payment rules
- Offer rules
- 2 outside this role’s access
A quote up to €800 is paid in full. Above that a 20% deposit reserves the work, never less than €200, and the quote stands for fourteen days.
A quote up to €800 is paid in full. Above that, a 20% deposit reserves the work, and never less than €200. A quote stands for fourteen days.
Behind the answer2 of 4 documents searched.
Every answer here is worked out in your browser from the four documents above. Nothing is sent anywhere.
How much is due before a project starts?
The approaches
Three approaches, and what each one leaves where it was.
Most projects need one of them, some need two, and the third is usually a constraint rather than a choice.
ApproachWhat it isFits whenWhat it leaves where it was
01Retrieval over approved documents
The system searches only the documents you have approved, and answers with the passage it used.
The knowledge is already written down and it changes often.
Documents that are out of date or contradict each other. An answer is only as current as its source.
02Adapting a model
A model is trained further on your own examples so it follows your format or your vocabulary.
You have enough good examples of the finished result, and the task repeats.
Facts that change. An adapted model still has to be checked against a source, and you need test material that shows it improved.
03Where it runs
With a model provider under agreed terms, in a private cloud, or on hardware you control.
A contract, a regulator or the sensitivity of the material limits where it may be processed.
Accuracy. This decides who can reach the material, not what the answers say.
Running a system privately does not make it correct. Where it runs decides who can reach the material. What it answers is decided by the documents, by the retrieval and by the checks you can run against it.
The build
What a build is made of.
Five parts, in the order they get built. The first three decide whether the thing can be trusted, and none of them is the model.
-
Connections
The material is read where it already lives, with the same login and permissions a person has.
-
A memory of it
Documents are indexed so every answer names the passage it came from.
-
Permissions in the tools
A role that may not see a figure gets a tool that never returns it, so the model cannot leak what it was never given.
-
Evaluation cases
Real past work with the right answer already known. A change does not ship if the score drops.
-
Handover and care
Where it runs, who can pause it and what the log shows, handed over in your name.
A build takes8–14 weeksafter a discovery
One we built
The company agent for Tramontana Gastro runs on a dedicated server in Germany and reaches its model through an EU-region endpoint under a data processing agreement, with their data excluded from training. Its evaluation sets were built during the discovery week: 612 past orders, 180 supplier invoices and 90 complaints, each answer checked by the person who owns that work. Discovery to handover took twelve weeks.
How that agent worksWhere it starts
A week, and a written answer on whether to build.
We talk to the people who do the work, check the material and its access rules, and compare the approaches against your task.
What the week is
- Interviews with the people who do the work, not only the ones who commission it.
- A look at the material itself: what exists, what is missing, what contradicts what.
- The access rules as they stand today, and who is allowed to change them.
- Evaluation cases drafted from your own past work, so a build has something to pass.
A recommendation not to build is a valid result, and the most useful one when it is true.
What it costs
AI Discovery
€490
per project, fixedOne week
Credited against an AI build that follows.
Private AI Build
from €9,800
per project8–14 weeks
Scoped from the discovery’s recommendation.
Written recommendationOne week
- 01The task and who owns it
- 02The material we checked, and what is missing
- 03Who may see what
- 04The approaches compared
- 05Evaluation cases a build would have to pass
- 06Recommendation BuildNot yetDon’t build
- 07Scope, price and delivery window, if it is a build
- 08Questions that remain
What moves a build price: how much material there is and what state it is in, how many roles and rules the access needs, how many systems it has to reach, and how much checking the work has to survive before it can run on its own.
Calculate your priceThe limits
What a private system still will not fix.
-
A source nobody wrote down
If the answer lives only in someone’s head, or three documents disagree and nobody has said which one wins, there is nothing for the system to be right about. Writing it down is the work, and it is yours. A discovery usually finds several of these.
-
Permissions nobody has decided
If your company cannot say who may see what, we cannot encode it. That decision comes before any build, and it is often the part that takes longest.
Who decides what is correct
Your process owner decides what a correct answer looks like and what has to be approved before it leaves the company. We define the access rules, the test cases and what happens at the edges. Neither of those is a decision software makes.
Questions
What people ask before the first call.
Does our data leave the company?
That depends on the arrangement, and the engagement names it: which provider or which infrastructure, where the processing happens, under which terms. Keeping everything on hardware you control is possible, and it costs more and takes longer. We say which one your work actually needs.
Can you run a model on our own servers?
Yes, when the work needs it. It means hardware sized for the model, a model small enough to run on it, someone patching that machine, and a running cost you carry whether or not anyone asks a question that day. The discovery prices it beside the alternatives so the choice is made on numbers.
Will it stop making things up?
Answers come from your documents and name the passage they came from, which removes the invented-fact case. It does not remove a document that is wrong, one that is out of date, or a question your material does not answer. That is what the evaluation cases are for.
How do you know it will work on our documents?
Before the discovery, we don’t. That is what the week is for: your real material, drafted evaluation cases, and a written answer with the reasons in it.
What if you say don’t build?
You have a written recommendation with the reasons, the evaluation cases we drafted, and the access questions written down. The €490 is credited against an AI build, so if nothing is built it stays spent. That is the honest arrangement, and it is a week rather than a quarter.
Is our material used to train a model?
No. We choose providers whose terms exclude training on your data, and the exclusion is written into your agreement rather than assumed.
Who owns what you build?
You do. The code, the data and the accounts are handed over in your name at the end, and you can take them elsewhere.
Can we start smaller than this?
Often, yes, and that is a good outcome. If the material is already public, or the access rules are simple, an assistant on the tools you use is a smaller job. The discovery is also the place to find that out.