What is sovereign AI? What it means for confidential work

“Sovereign AI” usually comes up when governments talk about national data centres and home-grown models. For a law firm, a clinic, a town hall or a field unit, it means something more concrete: you decide where your prompts, documents and models live, and which rules apply to them.

Why many organisations still hold back

In 2025, one in five EU companies with ten or more staff used any AI technology at all, according to Eurostat. Among the companies that had considered AI but weren’t using it, 70% said they lacked the expertise, 54% weren’t sure of the legal consequences and 53% were worried about data protection and privacy.

Holding back doesn’t keep AI out of the building, though. In KPMG and the University of Melbourne’s 2025 global study of trust in AI, 48% of employees who use AI at work said they had put company information, such as financial, sales or customer data, into public AI tools.

So the data leaves anyway, just without anyone deciding that it should. People end up paying for AI twice: once with their subscription, and again with their data.

Four questions that decide who is in control

Whatever a vendor calls its product, four questions decide whether an AI system leaves you in control:

• Where does the data go? Prompts, files and answers stay on your devices and your servers, rather than being uploaded to someone else’s.
• Who can see it? No third party stores, reads or processes it, and it is never used to train a model for anyone else.
• Whose rules apply? The data protection, professional secrecy and national rules that already govern your work, not a provider’s terms of service.
• Does it keep working without the provider? Offline, at a predictable cost, and without a change in someone else’s prices or terms deciding what you can do.

Cloud AI can answer some of these with the right contracts, regions and settings. Running the AI where your data already is answers them by design.

Why it’s possible now

Not long ago, a useful language model needed a data centre. Three things have changed:

• The hardware is ready. New laptops and phones ship with processors built for AI, and one server can serve a whole office.
• Smaller models have caught up. Built on established foundation models and trained for one job, a model small enough for that hardware can take on focused professional work.
• Specialising beats size. A specialist trained on one profession’s rules, cases and vocabulary, and grounded in its sources, doesn’t need to know everything. It needs to know its field and show where its answers come from.

Sovereign doesn’t mean on your own

Keeping AI in-house doesn’t remove the need for judgement. A specialist proposes and a professional decides, and when a case is unclear, “needs review” beats a guess. Your backups, access rules and records stay as they are.

It also doesn’t mean every task needs it. Drafting a public announcement can go anywhere. A client’s file, a patient’s record or a citizen’s request is different.

How Inovia builds it

Inovia builds specialist AI that runs on your own devices and premises. One sovereign AI core powers three product lines:

• Office and industry, for law firms, accountants, doctors, banks, insurers and engineering teams.
• Public administration, for municipalities, tax offices, courts, hospitals and schools.
• Defence and security forces, for field units, interpreters, technicians and command posts.

Kimiko, our private AI note taker for Mac and Windows, is live today. It transcribes, summarises and dictates on your computer, and it’s free to try, no credit card required.

If you have work you can’t send to the cloud, talk to us.

Sources

• Eurostat, Artificial intelligence by size class of enterprise (isoc_eb_ai), 2025: EU enterprises with 10 or more persons employed; reasons for not using AI among enterprises that had considered it.
• KPMG and the University of Melbourne, Trust, attitudes and use of artificial intelligence: A global study 2025, p. 75: employees who use AI at work.