Data sovereignty in AI use — the board's decision on what goes out to someone else's models

Three routes as a portfolio, not a choice of one

A company has three routes available: public tools, the enterprise cloud under a contract with a provider, and a private AI instance, that is, a model run in infrastructure controlled by the company. A common mistake is to look for one answer for the whole organisation.

In practice companies use several routes at once. Public and non-sensitive content can go to public tools. Most office work on internal data can run in the enterprise cloud, within the limits of the contract. A private instance makes sense where data should not leave the company: contracts, cost estimates, customer documentation and know-how. A private AI instance does not replace the cloud; it works alongside it. The choice depends on the data, not on fashion.

Model interchangeability as a condition of sovereignty

Sovereignty does not end with the place of processing. A company whose processes are described in the language of a single tool depends on its provider even when the data sit on its own servers. A change in price, licence terms or the legal status of a model can then mean losing the work put into describing processes.

We wrote about this in the analyses the US and China in the AI race and on the EU code of practice and the dispute with large providers: the answer to market instability is organisation on the company’s side. A process description that is not tied to a specific model turns the model into an interchangeable component rather than a foundation. This is the essence of semantic transformation, that is, organising a company’s names, processes and data so that people, language models and board reports understand them the same way.

The European framework sets the rules of the game; it is not a barrier

Companies in the Union operate within a framework set by, among others, the GDPR and the AI Act — Regulation (EU) 2024/1689, as amended by Regulation (EU) 2026/1744, known as the AI Omnibus. This framework does not rule out the use of models from outside Europe in advance, but it sets conditions, among others for transferring personal data outside the European Economic Area (EEA); nor does it require a company to build its own infrastructure. What it does require is that the company knows what data it processes, where, on what basis and in what role it acts in relation to the AI system. For a board that has these answers, the framework is a predictable set of rules. For a board that does not, it becomes a risk.

The specific scope of obligations depends on the use and on the company’s role; it is determined by the company’s legal adviser on the basis of a description of the process.

Checklist for the board

1. Data classification. What categories of data the company holds and which of them may not leave the approved environment.
2. List of uses. Which AI tools are in use in the company, including those employees use without a formal decision, and which processes they serve.
3. Place of processing. For each use: where the data are processed and stored.
4. Contracts. Whether contracts with providers set out ownership of outputs, confidentiality, use of the company’s data to train models, and the terms for exiting the service.
5. Exit plan. Whether the company can change model or provider without losing its process descriptions and data, and how long that would take.

The service in which we take a company through these steps is described on the page Data sovereignty and private AI instance.