Sovereign AI in practice: training teams on models and data you control

What does sovereign AI mean in practice?

For the people who approve training environments, sovereignty comes down to four questions, each with a clear answer:

• Where does the model run? On a provider's cloud, on an approved in-country cloud, or on your own infrastructure.
• Where does the data sit? Prompts, uploaded files, outputs and logs, at rest and while being processed. This is data residency for AI, and it covers more than the documents people upload.
• Who can see it? The model provider's staff, the training provider, other tenants, or only your own administrators.
• What is kept, and for what? Retention periods, and whether anything you send can be used to train someone else's model.

At Cader Group we put it in one line: sovereign by design, your models, your data, in-country. Every other decision about tools, vendors and training follows from those four answers.

Why does this matter for training now?

Because use is already running ahead of the rules. In a study of more than 48,000 people in 47 countries, 48% of employees said they had uploaded company information into public AI tools, and only 40% said their workplace had policies or guidance on generative AI (University of Melbourne and KPMG, 2025; Indicative). A separate survey found that 78% of AI users bring their own AI tools to work (Microsoft and LinkedIn, Work Trend Index 2024; Vendor data).

The cause is rarely bad intent. People reach for public tools because no approved alternative has been set up, or because no one has shown them how to use it on real work. A training program run on those same public tools makes the problem worse: it signals that the tools are acceptable and builds workflows that cannot be deployed.

Hosted or open-weight models: what are the trade-offs?

Most organisations end up using both, matched to the sensitivity of the work.

Three points tend to settle the choice. First, classify the work before the model: the data a workflow touches decides where it may run. Second, an open-weight model only delivers sovereignty if it is hosted where your rules require; downloading the weights is the start, not the finish. Third, test output quality on your own tasks, in Arabic and English, before committing a workflow to either route.

How should teams be trained on approved tools?

On the same set-up they will use the following week. That means three things.

The approved model, in the approved environment. If the organisation has an enterprise agreement with a hosted provider, training runs inside that tenant. If sensitive work must stay in-country, training runs on the open-weight model hosted there. Participants learn the tool they will be held to, including its limits.

Sample data that mirrors real records. Real personal or confidential data does not belong in a training room. But generic examples teach little. The answer is sample data built to the same structure as your records: the same fields, formats, document types and mix of Arabic and English, with no real individuals behind it. Teams then build prompts, workflows and agents that work on the real thing once access is granted.

Approval points built in from the start. Every workflow built in training includes a defined point where a named person signs off before anything leaves the team. Governance taught as a closing module does not shape what people build.

What should you ask a training provider about data handling?

Ask these before you sign, and ask for written answers:

1. Which models will participants use, and where is each one hosted?
2. Where do prompts, files and outputs go during the program, and who can see them?
3. What is retained after the program, for how long, and is any of it used to train a model?
4. Will any of our real data be used? If not, how is the sample data produced?
5. Can the program run inside our own tenant or in-country environment?
6. Who owns the prompts, workflows and agents participants build, and how are they handed over?
7. How are participant accounts closed when the program ends?

A provider that answers with "our platform is secure" has not answered. For the wider set of questions to ask before signing, see how to choose an AI training provider.

Measure where your own team stands: take the free Readiness Snapshot →