AI that writes, figures that stay exact, a decision that stays yours
Detailed example: the Suite QVCT assistant, on NOVALIA, a fictional company of 250 employees · sample data
In our tools, AI replaces neither calculation nor judgement. It reads your data, writes what it is asked for (a summary, a diagnosis, an action idea) and shows what it read so that anyone can check. It runs at your place, offline, with no third-party service.
For whom: management, HR and team leaders who want to save time on writing and reading, without losing control of their data.
The three core ideas
The Suite QVCT example
In Suite QVCT, the assistant steps in at five specific places, and elsewhere only if you ask it to.
- Dashboard: four analysis leads that cross-check results instead of repeating what is already displayed.
- Team sheet: a diagnosis written from the team's indicators and testimonials.
- Employee voice: a two-sentence summary of the testimonials, with the recurring themes.
- Action plan: a suggested next step, taking into account the actions already open.
- Tailored analysis and conversation: you choose the team, the topic and the sources, or ask the question in your own words; the assistant produces a summary for management, meeting questions or a message to employees.
What the assistant produces can be exported as a PDF or a PowerPoint presentation for management, with what it read attached as an appendix.
A question, an answer, and the proof
This demo replays a real question put to the assistant, on NOVALIA's sample data. Watch the answer being written, then follow how we check that it is correct.
Demo summary. Question: what do the workers say about workload? The assistant asks which teams to treat as “workers”, then reads the barometer, three testimonials, one pain point and one idea. It answers that the pace is too intense on some days, especially at month end, that urgent interventions are announced too late, and that the pace is generally bearable but hard to manage during peaks. Each sentence is matched to its sources, the figures are recalculated separately, and a faulty draft is caught by the safeguard.
Real answer from the “balanced” assistant, captured on 5 October 2026 on NOVALIA's sample data. Four successive runs of the same question gave four different wordings, all consistent with the data: the demo shows one of them. It proves that this answer is correct and that the check works, not that AI never gets it wrong: that is why the text stays “to be reviewed” and the sources are always displayed. The assistant's answer is shown in French, as produced.
The online Suite QVCT demo shows the interface with sample data. The assistant itself works in the installed version, because it relies on an engine that runs on your machine.
Our rules
A language model writes well, but it can get things wrong. The software is built so that this is visible and harmless.
Why a button rather than an automatic run?
Because AI uses computing power and you should stay in control of the timing. A text already written comes back from memory without recalculation, and a short waiting time separates two requests so that runs are not chained.
What happens if the model gets it wrong?
The software checks texts before displaying them: figures, direction of a comparison, repetition, off-topic content. A minor flaw is fixed in place, without rewriting everything. A text stopped midway is never kept. And since every text is presented as “to be reviewed” with its sources, a mistake remains easy to spot.
Where it runs
No subscription to an AI service, no key, no data sent out: the engine is delivered with the software.
Three assistant levels to choose from
Fast, the lightest on resources, for modest machines.
Balanced, offered by default: clear and concise, a good compromise between quality and waiting time.
Precise, slower, for analyses where finesse matters.
Open models, licences in plain sight
The models used are under open licences that allow commercial use. Their licence texts are delivered with the software and can be read in the “About and licences” section.
The model is never presented as developed by Centauri: we integrate it, check it and frame it.
Its limits
Better to know them before you decide.
- It assists, it does not decide: an action idea or a diagnosis remains a proposal to discuss with the teams.
- It is slower without a graphics card: a few seconds for a short text, around a minute for a detailed analysis on an ordinary computer.
- It only knows what the tool contains: without the data, it says so instead of making it up.
- It writes in French: the assistant is not translated, and the English version of the interface does not offer it.
In a tool other than Suite QVCT?
The same foundation (local model, figures calculated by the code, visible sources, anonymity) can be reused when AI genuinely adds something. We never add it by default: it is decided together during scoping, tool by tool.
Frequently asked questions
Is AI mandatory in a Centauri tool?
No. It is offered where it saves time, never imposed. A tool works fully without it.
Is my data sent to a third-party service?
No. The engine runs on your computer or your server. Nothing is sent to an external AI service, and none is added without your explicit agreement and documented scoping.
Do we need a very powerful computer?
A recent office computer is enough for the “fast” or “balanced” assistant. A dedicated graphics card speeds up answers noticeably but is not required. Larger models are best placed on a central server. We check this with your IT department during scoping.
Who is responsible for the text produced?
You are, as with any text you approve: the assistant prepares, you review and decide. That is why every text is presented as “to be reviewed”, with its sources.
Can it be connected to our own data?
In principle, yes, as long as the data is in the tool. How to link it to your existing software depends on each one and is checked during scoping.