Quick answer: sovereignty does not depend first on the nationality of the model, but on what leaves your perimeter and on who can access it. Choosing a French model is not enough: if the request contains a name, an IBAN or a diagnosis, the data is exposed whatever the model. Conversely, data detected and masked before sending stays protected, even with the best American models. That is what lets an SME or a mid-sized company stay competitive without giving up control. This guide sets out the definition, the three levels, what US law and EU law can do, then six criteria to write into the contract.
What is a sovereign AI?
Quick answer: a sovereign AI is a use of AI in which your organisation keeps the decision over its data and its providers. It is neither a matter of the flag on the model nor merely a matter of hosting: it is a matter of control, defined below in five points.
Sovereign AI: a use of AI in which the organisation controls where its data is processed, the law that applies to it, who can access it, what is done with it (including the training of models) and the ability to change provider without disruption.
Digital sovereignty is this same decision-making capacity applied to all digital tools; AI is its most sensitive case, because requests carry entire business contents.
What is the difference with a French AI? A French AI describes the origin of the vendor. A sovereign AI describes your ability to decide where your data goes, who can access it, on what terms, and to change provider without disruption. A French AI can be used without sovereignty, from a consumer account for instance, and American models can be part of a sovereign set-up if the detected data is masked before reaching them.
The central confusion fits in one sentence: the place of processing and the applicable law are two different things. The place of processing is the country where the servers run. The applicable law is the law that the company holding your data must obey, wherever those servers are. The matrix below crosses the two.
| European model vendor | American model vendor | |
|---|---|---|
| Platform hosted in France, European operator | Infrastructure and vendor under European law: the CLOUD Act does not apply, unless one of them depends on an American parent company. | The infrastructure is protected; the request sent to the model, however, falls under US law. |
| Platform hosted in the United States, or with a host under US law | The vendor is European, but the host holds the data and falls under the CLOUD Act. | Both links fall under US law. |
Three movements drive the subject. Shadow AI first: files pasted into consumer tools, these unmanaged uses expose data that no contract covers. Vendors' terms next, which change often: an “improve the model” box ticked by default, retention that evolves, a sub-processor added. The instability of the transfer framework finally: the Data Privacy Framework that governs transfers to the United States is being challenged before the Court of Justice, and the US surveillance law FISA 702 has expired while remaining active through its current authorisations, as the section on US law details.
Sovereignty is decided first by the data.
Digital sovereignty: the three levels, infrastructure, model, data
Quick answer: the sovereignty of an AI use is decided at three distinct levels: the infrastructure that hosts the platform, the model that processes the request, and the data the request contains. The first two limit exposure. The third acts on the data itself, before it leaves. Each level reads with the same grid: the question to ask, what protects, what does not protect.
Level 1: the infrastructure, an AI hosted in France
The question to ask: where does the platform itself run, with the accounts, histories, uploaded documents, audit logs and the table that links a pseudonym to real data, and under which law is the operator that holds them placed?
What protects: an AI hosted in France, with a European operator not subject to US law, places these elements under European law. That is why the host is chosen as much for its jurisdiction as for its country.
What does not protect: this level says nothing about what happens when the request leaves for the model.
Level 2: the model and its endpoint
The question to ask: who publishes the model, under which contract is it called, and through which endpoint? The endpoint is the technical address to which the platform sends the request; its location sets the place of processing, not the law applicable to the vendor.
What protects: business access under a contract, with a no-training clause and a written retention period, offers a very different framework from a consumer version of the same model.
What does not protect: an endpoint located in Europe does not take an American vendor out of the CLOUD Act. This is mainly where the CLOUD Act weighs, as the next section explains.
Level 3: the data, what actually leaves
The question to ask: what does the request contain at the moment it leaves your perimeter? What leaves for the model is everything the employee wrote, the client's name, the amount, the address, sometimes a diagnosis.
What protects: Guard detects personal and sensitive data, replaces it with placeholders before sending, then restores the real values in the answer. A placeholder is a tag such as “[PERSON_1]” or “[AMOUNT_1]”; the table that links each placeholder to its real value stays in the infrastructure hosted in France. This mechanism is called reversible pseudonymisation, in the sense of the GDPR. The model works on placeholders, never on the detected values. What still leaves is the non-identifying business context: the type of document, the task, the expected tone. This level settles the essential for personal data.
What does not protect: masking protects what identifies a person or a transaction: a name, an IBAN, an address, an amount, a reference. It does not remove the meaning of strategic content: an acquisition project, a roadmap or a commercial strategy remain understandable, even without proper names. For those uses, Crosstensor provides a model hosted in France on its own servers: the organisation chooses the uses directed to it, and the request does not leave the country.
Sovereignty is decided first by the data.
What US law can do: CLOUD Act, FISA 702, Data Privacy Framework
Quick answer: US law can force access to the data held by a provider subject to that law, wherever it is stored. The CLOUD Act allows it for investigations, FISA 702 for intelligence. Hosting in France is therefore not enough if one link remains under US law; a request whose detected data was masked before sending, on the other hand, does not expose the real values.
The Clarifying Lawful Overseas Use of Data Act (CLOUD Act, 2018) requires providers under US jurisdiction to comply with the legal obligations to preserve and disclose data within their possession, custody or control, regardless of whether that data is located inside or outside the United States.1 An endpoint located in Europe does not change the law that applies to the vendor: a model called in Europe from an American vendor remains subject to the CLOUD Act.
Separately, section 702 of the Foreign Intelligence Surveillance Act (FISA) allows the collection of foreign intelligence targeting non-US persons reasonably believed to be outside the United States, with the assistance of electronic communication service providers. As of the update date of this guide, this provision expired on 12 June 2026 under its sunset clause, after the House of Representatives rejected a short extension the day before, and no reauthorisation has been adopted; transitional provisions nevertheless keep earlier orders, authorisations and directives in force until they expire, and the authorisations approved in March 2026 allow the covered activities to continue until March 2027.2
The Court of Justice of the European Union invalidated the Privacy Shield in 2020 because of these surveillance programmes, notably FISA 702.3 The Data Privacy Framework has governed transfers since 2023 without making those laws disappear; upheld by the General Court of the European Union in September 2025, it is now under appeal before the Court of Justice.4
Masking changes what these laws can reach, because a request concerns what the provider holds. The real values escape the CLOUD Act on the model vendor's side, because they were never transmitted to it. The CLOUD Act can only reach what the vendor keeps: the retention period at the model vendor is checked in the same way as no-training. The Court of Justice also ruled, in September 2025, that pseudonymised data may not constitute personal data for a recipient that has no reasonable means of re-identifying the persons concerned.5 That is the situation a model vendor that only receives placeholders may be in. For the organisation and for Crosstensor, which holds the mapping table, the data remains personal: the GDPR obligations remain in full.
| Real values | Placeholders | Mapping table | |
|---|---|---|---|
| Crosstensor and OVHcloud (France) | Yes, in the infrastructure hosted in France | Yes | Yes: it never leaves this infrastructure |
| The model vendor | No: never transmitted | Yes | No |
| A US authority, through the CLOUD Act | No: the vendor does not hold them | Yes, what the vendor keeps | No |
Is a European host that is a subsidiary of an American group concerned? Yes, as soon as the American parent company can access the data or order its disclosure: the criterion is the control exercised over the data, not the address of the head office.
Is a European model called through an American cloud concerned? Yes, the host remains subject to US law.
Sovereignty is decided first by the data.
What EU law requires: GDPR, HDS, SecNumCloud
Quick answer: EU law does not require a European model. It requires processors to be bound by contract, the data processed to be limited to what is necessary and every transfer outside the European Union to be justified; for health data, it has required storage within the EU or the European Economic Area since September 2026; and it offers, with SecNumCloud, a benchmark for judging a host.
The GDPR requires the controller to choose processors offering sufficient guarantees, to bind them with a data processing agreement, to keep a record of processing activities and to limit the data processed to what is necessary; it governs transfers outside the European Union through its Chapter V.6 A pseudonymised request sent to an American provider remains a transfer from the controller's point of view: the transfer framework, Data Privacy Framework or standard contractual clauses, remains necessary, and masking greatly reduces its risk.
For health data, French decree no. 2026-209 of 24 March 2026 now requires that their hosting, where it involves storage, take place exclusively on the territory of a member State of the European Union or a State party to the European Economic Area, with contractual transparency from the host about its exposure to third-country laws; these provisions have applied since 26 September 2026.7 They add to the French HDS certification (Hébergeur de Données de Santé, health data host), which is mandatory for hosting personal health data.8 It is, to date, the most concrete sovereignty requirement in French law.
SecNumCloud, the qualification issued by ANSSI, the French cybersecurity agency, is a useful benchmark for judging a host without being a prerequisite for an SME; it may be required for sensitive data of the French State and its operators.9
Sovereignty is decided first by the data.
Sovereign model or the best models on the market: do you have to choose?
Quick answer: no, if protection is ensured at the data level. The dilemma “sovereign AI model or performance” only exists when the sole protective measure is the choice of vendor. As soon as detected data is masked before sending, your teams use Claude, GPT, Gemini or Mistral depending on the task, and you keep control.
Models progress in waves, each vendor takes the lead in turn, and an organisation locked into a single model (a lock-in with a single provider, vendor lock-in, chosen this time) sees its teams go back to consumer tools as soon as the gap becomes visible. The European origin of a vendor such as Mistral simplifies the contract and reduces exposure to extraterritorial laws; it does not protect the content of the request, because a French model that receives an IBAN sees the IBAN.
Can ChatGPT or Claude be used in a sovereign way? Yes, under three conditions: business access under a contract with a no-training commitment, documented endpoints, and masking of personal and sensitive data before sending. Together, they protect personal data better than a European model without masking. For strategic content, the model hosted in France completes the set-up.
Sovereignty is decided first by the data.
The six criteria for choosing a sovereign AI
Quick answer: six criteria make it possible to assess any offer described as sovereign: infrastructure hosting, model endpoints, CLOUD Act exposure, contractual no-training and retention, masking before sending and reversibility. Each is checked with a precise question to the provider, whose answer must be in the contract.
| Criterion | What to check | Question to ask the provider | At Crosstensor |
|---|---|---|---|
| Infrastructure hosting | European host not subject to US law for accounts, histories, logs and the mapping table; encryption at rest and two-factor authentication. | Who operates the infrastructure, in which country, under which law, and with which account security measures? | Platform hosted in France at OVHcloud, a European operator not subject to US law, on HDS servers. Database encrypted at rest, two-factor authentication configurable by the organisation, including for SSO accounts when the SSO does not enforce it. Histories, documents, audit logs and the mapping table never leave this infrastructure. |
| Model endpoints | Endpoints documented by the provider, within a professional contractual framework. | Through which endpoints is each model called, and can you prove it? | Models are called through endpoints located in the European Union, under professional contracts. A European endpoint sets the place of processing, not the law applicable to the vendor: masking therefore applies whatever the model. A model hosted in France by Crosstensor is available for uses that must not leave the country. |
| CLOUD Act exposure | The provider can say which links fall under US law, and what limits their reach. | Which links in the chain are subject to the CLOUD Act, and what data do those links actually see? | Anthropic, OpenAI and Google fall under US law, and therefore under the CLOUD Act. That is why they only receive requests in which the detected identifying data has been masked. For uses that must not leave the country, the model hosted in France is available. |
| Contractual no-training and retention | Written commitment, in the data processing agreement, excluding the use of your exchanges to train the models, with the retention period stated. | Are no-training and the retention period in the signed contract, for every model provider used? | No-training is guaranteed contractually with each model provider and carried into the data processing agreement you sign with Crosstensor. On Crosstensor's side, your data is kept for the duration of the contract, then deleted two months after the end of the subscription. |
| Masking before sending | Personal and sensitive data detected and replaced before leaving your perimeter, then restored in the answer. | What exactly does the request contain at the moment it reaches the model? | Guard detects personal and sensitive data, replaces it with placeholders before sending, then restores the real values in the answer. Depending on the organisation's settings, the user can see what has been masked before sending. |
| Reversibility | Change of model or provider without re-integration; data and audit log retrievable on exit. | What actually happens if you change model, or if you leave the platform? | Model change from the back office, without touching the uses. Access to the main models (Claude, GPT, Gemini, Mistral) behind a single gateway, and an exportable audit log. |
No-training is often confused with a tick box in a consumer tool. In a company, it must take the form of a written contractual commitment, in the data processing agreement (DPA) signed with the platform, for each model provider used, with the retention period.
At Crosstensor, these contracts with the providers are managed by the platform and carried into the data processing agreement you sign: using Crosstensor means delegating that part.
Sovereignty is decided first by the data.
A sovereign AI platform without giving up the best models
Quick answer: a sovereign AI solution that stays competitive combines the three levels: an infrastructure hosted in France with a European operator, models called under no-training contracts, and automatic masking of detected data before every send. That is Crosstensor's architecture, detailed in the “At Crosstensor” column of the table above.
The masking mechanism is described step by step in the guide on how to mask data before sending. Here is what it gives on an ordinary request.
| Step | Content |
|---|---|
| What the employee types | “Draft a meeting proposal for Camille Roussel, chief financial officer of Verlaine Conseil, 12 rue de Rivoli in Paris, following our discussion about contract no. C-2026-087 for an amount of €24,600.” |
| What the model receives | “Draft a meeting proposal for [PERSON_1], [ROLE_1] of [ORGANISATION_1], [ADDRESS_1] in [PLACE_1], following our discussion about contract no. [REFERENCE_1] for an amount of [AMOUNT_1].” |
| What the employee reads | The complete proposal, with the name, role, company, address, reference and amount restored automatically. The model only worked on placeholders, never on the detected values. |
The employee changed nothing in the way they work, the model produced a complete answer, and the identifying data that was detected did not leave the infrastructure hosted in France. The exportable audit log feeds an inspection or your AI management system.
Who is sovereign AI indispensable for?
Quick answer: for any organisation that processes personal or confidential data with AI, which is to say almost all of them. It becomes indispensable in four situations: the SME or mid-sized company without a large IT department, healthcare, professions bound by secrecy, and the public sector.
Sovereign AI for SMEs and mid-sized companies. The same obligations as a large group, with a small IT team. Sovereignty must be obtained without a project: an official access, detected data protected by default, a single subscription, the scope of the Platform offer.
Healthcare. Health data falls under HDS-certified hosting and, since 26 September 2026, under mandatory storage within the European Union or the European Economic Area, with transparency from the host about its exposure to foreign laws.7 AI on patient reports is only conceivable with masking before sending; see the case of healthcare professionals.
Lawyers and professional secrecy. Pasting a file into a consumer tool may constitute a breach of professional secrecy. Masking names, references and amounts makes it possible to use AI on case preparation without exposing the client; see the case of lawyers.
Local authorities and the public sector. Citizens' data, transparency obligations, public procurement attentive to suppliers' jurisdictional guarantees: hosting in France with a European operator and full traceability of uses are prerequisites there, and SecNumCloud may be required for sensitive data of the French State and its operators.
Conclusion
Sovereignty is decided first by the data, not by the nationality of the model: a French model does not protect an IBAN present in the prompt, whereas data detected and masked before sending stays protected whatever the model. For strategic content, a model hosted in France completes the set-up. A sovereign AI is checked with six criteria written into the contract: hosting, endpoints, CLOUD Act, no-training and retention, masking, reversibility. That is the sovereign alternative to ChatGPT: the best models for your teams, visibility and proof for you.