Three principles
that reinforce each other.
An AI deployed at a client's site rests on three things: it has to run on your side, all of its traffic has to pass through a single door, and your teams' judgement has to stay inside the company. Kept apart, these three ideas dilute. Together, they are the difference between renting an AI and owning one.
1 · Sovereignty & compliance
Sovereignty isn't the point.
It's the consequence.
As long as your AI “lives” with a provider, you're renting it: their terms, their servers, their memory. The day it runs on your infrastructure, the question of sovereignty no longer arises — it's settled by design.
It runs on your side
The system is self-hosted: on your own server, your VPS, or a European sovereign cloud in your name (OVH, Scaleway, Outscale). Nothing depends on an account someone else can shut down.
Your own keys
You connect your provider keys. You choose which models you call, when, and you can switch without rebuilding anything. No intermediary sits between you and the model.
Your data stays put
Your documents, your memory, your trade vocabulary never leave your infrastructure. Only the requests you choose to send go out — never your assets, and never to train a third-party model.
How far you want to go.
You choose based on your sector.
Data sovereignty is the foundation: it comes with owning your AI. Model sovereignty is a further step, for the trades that cannot let a single request leave.
Data sovereignty
Your data never leaves your infrastructure. Hosting and the knowledge base are under your control; only the requests to the models you choose go out, and none of them serve to train a third-party model.
- Self-hosted on your server, your VPS or an EU sovereign cloud (OVH, Scaleway, Outscale)
- Hosting in your company's name, within your perimeter
- Your own provider keys — you decide which models you call
- GDPR by default; HDS (health data) compliance on request
- No third-party model trained on your data
- Full extraction of your data and your assets at any time
Model sovereignty
For sensitive sectors (defence, health, legal, finance): the model itself runs on your side. Zero external API calls — not a single request crosses your perimeter.
- Open-source LLM running on your infrastructure
- Fine-tuned on your vocabulary and your trade documents
- On-premise or dedicated sovereign cloud deployment
- No calls to an external LLM provider
- Full audit of the AI pipeline (inputs / outputs / logs)
Compliant by design
Humans
keep control.
We automate the data entry, not the decision. Because the AI we deploy touches sensitive functions, it's designed from the outset to stay on the human's side — not to tick a regulatory box after the fact. These guardrails are written into the specification you sign off before any development starts.
See how deployment runsHuman sign-off everywhere
Never an automatic decision about a person. The AI proposes, the team decides. We automate the data entry, not the decision.
0% on payroll & payments
No automation on the most sensitive flows. These are guardrails written into the specification you sign off, not options.
Logged, hosted in the EU
Every action traced, the data on your infrastructure — a method compliant with the AI Act by design, with no anxiety-inducing countdown.
Only the requests you choose to send
cross your threshold.
2 · The gateway
One single door.
All your AI passes through it.
Once the AI is deployed, all of its traffic needs a door. Without one, usage scatters: shared keys, subscriptions taken out on the side, data slipping into consumer-grade tools. The gateway is the single door through which everything passes — to see, to govern and to capitalise.
AI is everywhere in the company.
And under control nowhere.
Nobody built a door. Every tool, every team, every colleague calls the AI on their own side — and both the spend and the data slip through your fingers.
Shared keys
A single API key gets passed around between several people or tools. There's no way to know who is consuming what, nor to cut off one usage without breaking everything.
Individual subscriptions
Everyone takes out their own subscription on the side. The spend fragments into dozens of line items that nobody ever consolidates.
Shadow AI
Staff paste company data into consumer-grade tools, outside of any policy. And nobody sees it happen.
An unreadable bill
At the end of the month, the total lands with no breakdown. Which team, which task, which model cost what? The question goes unanswered.
Five functions,
one single door.
As soon as all the traffic passes through one point, you can finally observe it, route it and get value from it.
See
Who is using AI, for what, and at what cost. Every call is tracked and tied to a team. It's the end of shadow AI.
Route
The right model for every task. A simple task sent to a frontier model costs 20 to 50 times too much; by routing each task to the right model, you reduce the bill sharply, without loss of quality.
Capture
The “delta” — the human correction applied to a response — is recorded in passing, without asking anything more of the team member.
Learn
De-escalation: feed the corrections back in, specialise a model on your data, remove AI wherever a simple rule is now enough.
Reveal
The map of the tasks where teams struggle and correct the most. A skills diagnostic obtained along the way, with no questionnaire.
Why routing changes everything.
Three figures are enough to understand where the money goes — and why a single door gets it back.
Compatible with your tools,
with no development.
The gateway speaks the OpenAI protocol. Any tool that accepts an API URL connects by changing a single URL — no code to write, no integration to develop. Behind that door, four mechanisms work continuously.
Compute cache
A calculation already done is never paid for twice. The same request returns the cached result — instant, and free.
Trimmed context
Only what's necessary is sent to the model. The re-sent context accounts for two-thirds of an agent's bill; compressing it is the most direct saving of all.
Fallback provider
If a provider goes down or slows, traffic switches automatically to another. The AI doesn't stop in the middle of the work.
Per-team budgets
A spending cap per team or per use case. Beyond it, the alert goes out before the bill — no more end-of-month surprises.
3 · Your AI assets
The AI you rent goes back to your providers.
The assets belong to you.
Every AI operation produces two things: a fleeting deliverable — which we keep — and a lasting piece of data, your trade's judgement — which we throw away. The assets are the system that harvests that second one, the one everybody lets slip.
What's worth its weight in gold
is the delta.
What matters isn't the interaction with the machine: it's the gap between what it proposed and what the human validated. An interaction without a correction is worth almost nothing — the machine was already right. An interaction with a clear correction is worth its weight in gold: that's where your trade's judgement expresses itself, in black and white.
This delta is captured without asking anything of your people. They have nothing to document, nothing extra to enter: the correction they already make, in their normal work, is the signal. An invisible by-product of the trade at work, harvested along the way.
The trap would be to keep everything blindly — the naive dataset, where noise drowns out the signal. We don't capture everything: we sort by the verdict. Validated, corrected, abandoned. It's that human judgement that decides what enters the assets and what we set aside.
The assets take shape
as three concrete objects.
Captured judgement doesn't stay abstract. It refines into three usable objects that document themselves as they happen.
Validated corrections
The collection of retained deltas: what the machine proposed, what the trade corrected, and the verdict. The raw material to specialise your models on your own decisions.
Procedures, references, context
The company's trade context — procedures, references, cases handled — made searchable through semantic search (RAG). The AI draws on what your trade already knows, instead of reinventing it.
Codified trade know-how
Chains of decisions codified and replayable. Know-how documents itself as it happens: a navigable map is generated through use, with no dedicated writing workshop.
Software depreciates.
A memory grows.
Software loses value the moment it's installed: it ages, a better version comes out elsewhere, it has to be replaced. AI assets do the opposite. Each additional correction enriches them — they grow without losing value, month after month.
It appreciates instead of depreciating
Every validated use adds to the capital. The more the company works, the more its memory is worth — instead of going out of date.
It stays when people leave
Your teams' judgement stops living only in their heads. A departure no longer takes the knowledge with it: it has been capitalised.
It belongs to you
It's your data, your judgement, your capital. Not a vendor's. No one can take it back or monetise it in your place.
The tool can be copied.
The accumulated knowledge, never.
That's where the real defensive moat lies: not the tool, which is replicated within a few months, but divergence over time. Two companies start from the same software; the one that capitalises its judgement pulls a little further ahead every year. Eventually, the gap can no longer be closed.
That reversal also changes what makes a person valuable: when knowledge becomes shared assets, it's no longer seniority that counts, but the quality of their judgement. A junior plugged into the assets becomes productive immediately — they inherit the best of what the company has validated, and each of their corrections in turn feeds the shared capital.
Apart, they're three concepts.
Together, they're a system.
Sovereignty makes the capture legitimate: you only capitalise what stays on your side. The gateway makes the capture possible: without a single door, judgement scatters. And the assets are what makes sovereignty worth having: owning an AI that learns nothing would be pointless.
None of this gets decided before the numbers are in. We always start by measuring what your manual processes cost — the level of sovereignty, the scope and the roadmap are decided in the light of the audit.
Frequently asked questions
The technical questions about the AI once it's deployed. Pricing, the audit and the delivery process are detailed on The offer page.
Your AI runs within your perimeter, in your name, with your keys. Sovereignty isn't an option you bolt on: it's the consequence of owning your system rather than renting it. As long as your AI lives with a third party, you depend on their terms; the day it runs on your side, the question is settled by design.
Data sovereignty (Level 1) is the foundation: hosting, knowledge base and assets stay on your side; only the requests to the models you choose go out. Model sovereignty (Level 2) goes further: the model itself runs on your infrastructure, with no external API calls at all — for the sectors that cannot let a single request leave.
GDPR is built in: because the data stays within your perimeter, under your control, you keep full command of the processing. HDS (health-data hosting) compliance is available on request. On the AI Act, compliance comes from the way it's built — human sign-off in the flow, action logging, EU hosting — not from a box ticked after the fact.
No. It sits in front of them: any software that accepts an API URL connects by changing a single URL, with no code. Your tools keep working — they simply pass through the single gateway.
Your people have nothing to document or enter extra: the correction they already make, in their normal work, is the signal. We harvest it at the gateway, then sort by the verdict — validated, corrected, abandoned. Only what a human has settled enters the assets.
Because you connect your own keys, you choose which models you call and you can switch without rebuilding anything. And you can extract all of your data and your assets at any time — you're not locked in with anyone.
Start by putting a number on it.
30 minutes, free, to spot where your manual processes cost you the most. If there's nothing to gain, we'll tell you.