Custom AI · sovereign
Custom AI, trained on your data.
A model that learns your business, and improves alongside your teams.
Generic AIs know a little about everything and nothing about your company. We do the opposite: we specialise an open model on your data. It runs where you decide, even on your own premises. Buy it, or take it on subscription, retraining included.
| A generic AI | Your custom AI | |
|---|---|---|
| What it knows | A little of everything | Your documents, your vocabulary, your exceptions |
| Your data | Handed to a third-party service | Trains only your model |
| When it gets it wrong | The same mistake tomorrow | The correction is used to retrain it |
| Where it runs | At the vendor’s | Cloud, sovereign cloud or SABER, your choice |
| What it costs | Per user or by volume | A fixed subscription, or buying the model |
| Who owns it | Never you | You, if you buy it |
The method
From your documents to your AI, in four steps.
- 1Your real examples
Delivery notes, orders, invoices, emails: we start from what you already handle, anonymised where needed.
- 2An open model, fine-tuned
A compact model is fine-tuned on your examples, and connected to your documents so it can cite its sources.
- 3In production, under control
It plugs into your tools. Uncertain cases go through a person, never forced through.
- 4It learns from your corrections
Your teams’ corrections are used to retrain it: it makes fewer mistakes week after week.
Two levers
Search your documents, or learn your business.
The AI finds the right passage and cites its source. Ideal for answering from your procedures, contracts and catalogues.
The AI learns your formats, your vocabulary and your rules. Ideal for reading, sorting and extracting at high volume.
Most often, we combine the two: one for memory, the other for know-how.
Proof
Trained on real documents, not on a demo.
Sovereignty
Your data trains only your model.
Where your AI runs, and where it is trained: you decide.
Your data trains no other model: not another client’s, not a public model.
Cloud, sovereign cloud for your tools as well as for the AI, or a SABER server on your premises.
On your premises on SABER, or in a sovereign or standard cloud, depending on your requirements, your budget and your security level. Off your premises: under agreement, then your data is deleted.
Use cases
An AI that knows your business.
Delivery notes, CMRs, invoices, orders: read, extract and check at high volume.
Answer with your procedures, your products and your tone, not generalities.
Quotes, tender responses and emails, drafted from your catalogue.
An assistant that knows your standards, your contracts and your history.
Go further
Questions
Your questions, answered.
Do we need a lot of data?
No. We start with what you have: search across your documents works from setup, and the model’s fine-tuning is refined afterwards.
Is my data used to train other models?
Never. It trains only your model. And if training takes place off your premises, your data is deleted afterwards, with a certificate to prove it.
Who owns the model?
Your choice. Buy it, and the trained model is yours. Or take it on subscription: it stays ours, and we maintain and retrain it for you. Either way, your data stays yours.
Where does training happen?
Depending on your requirements: on your SABER server, in a sovereign cloud or in a standard cloud. We choose with you, based on how sensitive your data is, your security constraints and your budget. Off your premises, training runs under a data-processing agreement, and your data is deleted afterwards.
How is this different from ChatGPT?
ChatGPT is general-purpose and runs on a third party’s infrastructure. We deliver a model specialised on your data, which runs where you decide: in the cloud, in a sovereign cloud or on your premises.
Which models do you use?
Open-weight models, chosen for your case. We are not tied to any vendor: if a better model comes out, your AI can benefit from it.
Does it make up answers?
Search across your documents grounds the answers and cites the source. And on uncertain cases, a person on your team approves.
How much does it cost?
Two options. On subscription: a short setup, then a fixed monthly amount that covers retraining, supervision and updates. To buy: you pay for the training once, and the model is yours. The amount is in your proposal, after the free demo.