Sovereign AI: a practical guide for decision-makers (CEO, CIO, head of legal) — what to demand, what to avoid

In brief.
Everyone uses the term “sovereign AI”, and it does not mean the same thing from one vendor to the next. Sovereignty plays out on three levels (data, processing, models) and across three deployment modes. What you need to demand comes down to the chain of contractual responsibility.

Your head of legal has just sent you a draft contract with an AI provider. Forty-two pages. In the middle, a clause you probably did not read carefully: “Data submitted to the platform may be used to improve our models.”

That is where everything is decided.

In 2026, “sovereign AI” is on everyone’s lips. Vendors use it. Consultancies sell it. Public tenders require it. But behind the word, the realities are radically different, and so are the consequences for your company, your clients and your regulatory exposure.

This guide is not an academic definition. It is a decision tool: what you must demand before signing, what you must avoid at all costs, and the concrete questions to ask any AI provider that claims to operate in sovereign mode.

What “sovereign AI” really means

Let’s start by clearing up the confusion.

AI sovereignty is not a characteristic of the model. GPT-4, Claude, Mistral: the model itself is neither sovereign nor non-sovereign. What is, or is not, sovereign is the infrastructure it runs on, and the conditions under which your data passes through it.

For French and European companies, sovereign AI rests on three inseparable pillars:

1. Data hosted on European territory, under the jurisdiction of European law, which by design rules out any provider subject to the US Cloud Act, even if it has data centres in France.

2. Strict isolation of data between clients. Your documents do not pass through a shared environment. Your queries do not “contaminate” the training data of a shared model. What you upload stays within your perimeter, and only within your perimeter.

3. Full traceability of processing. At any moment, you can find out which data was processed, by which component, and at what time. Not a marketing promise: an auditable technical reality.

If any one of these three pillars is missing, the sovereignty claimed is partial, and potentially misleading.

The three deployment models: what they really imply

Standard shared cloud

This is the default model for most consumer AI tools. You log in to a SaaS platform, submit your documents and get answers. Fast, accessible, inexpensive.

The problem: your data passes through an environment shared with thousands of other clients. Depending on the terms and conditions (which nobody reads), it may be used to improve the model. Hosting is often in the United States or Ireland, under a jurisdiction that is ambiguous with regard to GDPR. And in the event of a leak, it is practically impossible to determine what was exposed.

This model is unacceptable for: contractual data, due diligence, certification questionnaires, sensitive client information, and any document covered by professional secrecy or by your clients’ confidentiality policies.

A dedicated instance on a European public cloud

This is the intermediate model. You have an isolated instance (your own environment, physically separated from other clients) hosted by a European hyperscaler (OVHcloud, Scaleway, etc.) or by AWS/Azure/GCP with data residency guarantees.

It is better. But the nuances matter.

The technical isolation is real: your data does not mix with that of other clients. But legal dependence on the cloud provider remains a risk, depending on the context. A US hyperscaler hosting your data in France remains potentially subject to the PATRIOT Act and the Cloud Act for certain categories of data.

This model is suitable for: most enterprise use cases, provided the contract explicitly guarantees that the data will not be used for training and that it will reside strictly on European territory.

The dedicated private instance, the sovereign standard

This is the highest level of sovereignty. An environment entirely dedicated to your organisation, hosted on European infrastructure, with contractual and technical guarantees on isolation, traceability and non-reuse of data.

In practice: your documents never leave your perimeter. The AI model operates within your space, not the other way round. Every answer generated is sourced from your own documents, with the exact reference (document, page, date) for every statement produced.

This is the model Optivalue.ai has chosen for all its clients: one private instance per organisation, hosted in France, with no reuse of data for training and full logging of processing.

This model is essential for: data under NDA, M&A due diligence, certification data (ISO 27001, NIS2), documents covered by professional secrecy, and level 2 and 3 client information.

What to demand: the decision-maker’s checklist

Before signing with any AI provider that processes your internal data, here are the 8 non-negotiable questions to ask, and the satisfactory answers to obtain.

1. Where is my data hosted, physically?

Satisfactory answer: a data centre in France or the European Union, with a contractual commitment on data residency.

Warning sign: “Our servers are in Europe”, with no country or jurisdiction specified.

2. Is my data used to train or improve your model?

Satisfactory answer: No, contractually and technically. A written commitment in the terms of use and the DPA.

Warning sign: “Your data is anonymised before use”. Anonymisation is not a guarantee of non-use.

3. Is my instance shared with other clients?

Satisfactory answer: No. A dedicated instance, with guaranteed and auditable technical isolation.

Warning sign: “Secure multi-tenant environment”. Multi-tenant means shared infrastructure.

4. Who can access my data on the provider’s side?

Satisfactory answer: Strictly limited, logged access, subject to NDA. A list of named individuals available on request.

Warning sign: No precise answer, or a referral to a general privacy policy.

5. Do you hold a security certification: ISO 27001, SOC 2, HDS (the French health data hosting certification)?

Satisfactory answer: A valid, auditable certification, with a scope that covers the service you are buying.

Warning sign: “We are in the process of certification”, with no validity date and no defined scope.

6. What are your obligations in the event of a data breach?

Satisfactory answer: Notification within 72 hours, a documented procedure, a GDPR-compliant DPA.

Warning sign: A referral to the terms of use, with no formal incident procedure.

7. Are the generated answers sourced: document, page, date?

Satisfactory answer: Yes, every statement is traced back to its source in your documents.

Warning sign: Answers generated with no reference, impossible to verify or defend in an audit.

8. Can I have your infrastructure audited by a third party of my choice?

Satisfactory answer: Yes, a right to audit provided for in the contract.

Warning sign: “We carry out regular internal audits”. An internal audit is no substitute for a right to external audit.

What to avoid: the 5 classic traps

The label trap. Some providers display certifications (ISO, SOC 2) whose scope does not cover the service you are buying. Always ask for the original certificate and check that its scope explicitly includes the AI platform.

The trap of terms that can be changed unilaterally. A sovereignty clause in a contract is only valid if it cannot be changed without your consent. Check that the terms of use do not contain a clause allowing the provider to change its data processing practices with mere notice.

The anonymisation trap. “Your data is anonymised before any processing” is a reassuring formula that guarantees nothing about reuse. Anonymisation can be reversible, partial, or defined differently from what GDPR means by it. The only acceptable guarantee is contractual non-use, not technical anonymisation.

The invisible subcontractor trap. Your AI provider may itself subcontract to a third party (OpenAI API, Azure Cognitive Services, AWS Bedrock). In that case, it is the subcontractor that processes your data, under its own terms, which are often American. Demand the full list of subcontractors and the terms that apply to them.

The demo mode trap. Some providers apply different sovereignty conditions during the trial or POC phase. Your demo data is no less sensitive than your production data. Demand the same guarantees from the very first query.

Sovereignty and performance: a false dilemma

When sovereign AI comes up, one decision-maker in two immediately thinks of a trade-off on performance. Private instance = weaker AI = less relevant results.

It is an understandable reflex, and an inaccurate one.

The quality of a document AI does not depend on whether or not the model is shared with other users. It depends on the quality of the documents you provide, the precision of the question asked, and the robustness of the information retrieval architecture (RAG) that links the two.

A well-designed private instance produces better results than a shared model, precisely because it is trained on your own document corpus, with your own conventions, your own terminology and your own formats.

Among Optivalue.ai clients, moving to a dedicated instance has consistently led to better-quality answers: more relevant, better sourced, and more consistent with the organisation’s internal standards.

Where to start

AI sovereignty is not a transformation project. It is a contractual requirement you can set today, before anything is signed.

Three concrete actions for this week:

Audit your current AI tools. For every AI tool your teams use, ask the 8 questions above. You will be surprised how many answers are missing or unsatisfactory.

Reread your DPAs. Every AI provider’s Data Processing Agreement should contain explicit commitments on data residency, non-reuse for training, and subcontractors. If it does not, that is a contractual anomaly.

Test with your real data. A provider that refuses to guarantee sovereignty during the trial phase will not guarantee it any better in production. How it behaves in a demo is revealing.

Optivalue.ai operates as a dedicated private instance, hosted in France, with no reuse of your data. Every answer is sourced back to the original document. Request a personalised demonstration →