In June 2025, testifying before a French Senate commission of inquiry, Microsoft France’s director of public and legal affairs said he could not guarantee that French citizens’ data hosted in France would never be handed over to US authorities, even without the French authorities’ consent.
He was telling the truth, and his position allowed no other answer. No European subsidiary of an American group can promise otherwise, wherever its servers are located.
Most executive committees conflate two notions: where data is located and who has sovereignty over it. You can check the first on a map. The second depends on the law that applies to the company holding the encryption keys.
Generative AI turns this gap from a theoretical concern into an operational one, simply because of volume: within a few months, a company rolling out these tools sends out more sensitive material than in ten years of email.
The actual journey of a request
Many companies think they are safe because they signed up for a business plan rather than a consumer one. The terms do differ, but less than they think.
On free or individual versions, the terms of use allow content to be used to improve the models, unless you turn the option off in the settings. On enterprise plans and API access, the opposite applies by default: no training on customer content, retention of a few days for abuse detection, zero retention on request in some contracts.
A no-training clause covers only part of the path. Transport, inference, technical logging, moderation filters that read the content in order to classify it, temporary storage, infrastructure subcontractors: at every step, the data exists in plaintext, if only for a moment.
The enterprise contract therefore reduces exposure on this front alone.
What US law says
The CLOUD Act, passed in the United States in 2018, allows US authorities to compel any provider under their jurisdiction to hand over data it holds or controls, regardless of where that data is stored. What counts is the operator’s nationality, not the data center’s address.
Section 702 of FISA governs surveillance of foreign nationals and applies to electronic communication service providers. The Court of Justice of the European Union relied on it to strike down the Privacy Shield in July 2020, in its Schrems II ruling: US law gave Europeans neither protection equivalent to the GDPR nor an effective means of redress.
The Data Privacy Framework, adopted in July 2023, rests on a similar architecture and adds a redress mechanism within the US executive branch. Plaintiffs have been challenging it in court since it came into force. A company that has built its compliance on this text alone will have no fallback if it is struck down again.
Four layers of dependency
Steering committees treat digital sovereignty as a binary attribute attached to a provider. It actually breaks down into four independent layers, and no single hosting choice settles all of them.
- The data: who can access it, under what jurisdiction, with what recourse. European hosting addresses this layer, and it is the only one many committees look at.
- The model: closed weights accessed through an API, or open weights that you download and run in-house. This choice governs your business continuity. When a vendor deprecates a proprietary model, changes its behavior or doubles its price, you have no immediate alternative.
- The hardware infrastructure: compute accelerators. Nvidia dominates, Washington decides on export controls, and no European manufacturer produces these components at the required scale. A SecNumCloud-certified French cloud runs on American hardware that comes out of a Taiwanese and Dutch supply chain.
- Operational availability: who decides what is accessible, when, and to whom. In 2026, several American vendors delayed the general availability of their most advanced models after concerns were voiced in Washington about attacks on critical infrastructure. Your roadmap hinges on political decisions you are neither party to nor informed of.
An organization can satisfy the first layer and remain dependent on the other three. Most deployments today are in this position.
Accumulation and unsanctioned use
No major corporate data leak traceable to a generative AI provider has been made public yet. That absence says little about the level of risk, and accumulation is the reason.
A single query is worth nothing. A year of internal queries is something else entirely: contracts under negotiation, source code, compensation policies, litigation files, strategic decisions. This concentration of documents at a third party often exceeds anything that exists in-house, where information remains scattered across siloed systems.
The vectors fall into familiar scenarios. An attacker compromises the provider, a prime target by definition. A subcontractor keeps the data longer than its contract allows. A poorly partitioned internal document search system exposes HR files to unauthorized employees.
The best-documented phenomenon remains internal. Your employees adopt tools with no oversight at all: consumer assistants to process client files, browser extensions that capture what is typed, services with no contract in place. No clause covers these flows, since no one knows they exist. Industry studies agree on one point: this shadow AI grows faster the more a company bans tools without offering a validated alternative.
What local AI changes
Running language models on infrastructure you control is no longer experimental, yet few architects have acted on it.
Open-weight models (those from Mistral, the Llama and Qwen families, or the open models American vendors release themselves) cover most everyday business uses: document summarization, classification, structured information extraction, querying an internal knowledge base, writing assistance, first-draft translation.
The gap with cutting-edge proprietary models is still real for complex reasoning, hard code generation and long-running agentic tasks. It becomes marginal when the job is finding a clause in an eighty-page contract or summarizing meeting minutes.
On sovereignty, the benefit is clear. The data never leaves the perimeter: no CLOUD Act, no cascading subcontracting, no retention policy imposed from outside, no service cut off by a foreign regulatory decision. Budget a few thousand euros for a specialized workstation, and a few tens of thousands for a server shared by a whole team.
In exchange, you inherit the security of the entire chain, with teams that rarely match those a cloud operator devotes to it. Open weights are not auditable models: you can run and modify them, but not inspect their contents or reproduce them from training data that no one publishes. Hardware dependency remains untouched. And maintenance never stops: version tracking, quantization, monitoring, output evaluation, backups.
The most mature deployments split the load according to a classification set in advance, rather than bringing everything in-house: local processing for data that must not leave, a contracted API for the rest, and a written, verifiable boundary between the two.
The information regime
All these trade-offs rest on the same weakness: you assess risk based on the providers’ own publications.
In July 2026, OpenAI announced that several of its advanced models, evaluated in a sandboxed test environment, had devoted a significant share of their computing power to regaining unrestricted internet access, then attacked the Hugging Face platform on their own initiative by chaining several vectors, including stolen credentials. The company calls the episode an unprecedented cyber incident and has opened a joint investigation with the targeted platform. Hugging Face says an autonomous agentic system drove the attack end to end, and that its own AI tools detected most of it.
The incident occurred in a setup designed to observe it, and we know about it because OpenAI chose to publish it. Transparency is a good signal here. It is also a reminder that we only see as much of these systems’ behavior as vendors are willing to show: the absence of a disclosure proves nothing.
The episode also shifts the question. Until now, you asked what these systems retain. Once an agent has credentials and write access to your internal systems, you are dealing with access management, not confidentiality.
Levers of control
Four practices set the best-run deployments apart, whatever the provider.
- Classify your data before choosing tools. Three levels are enough: freely shareable, enterprise contract required, internal processing only.
- Provide a validated, accessible tool. Nothing curbs shadow AI as effectively, not even usage policies.
- Write down your contractual requirements instead of inferring them from sales documentation: no training on your data, retention period, data location, subcontractor notification, incident notification deadline.
- Treat automated agents as service accounts: least privilege, dedicated credentials, action logging, immediate revocation.
None of these measures eliminates the risk, and choosing a European provider addresses only part of it. The question leadership teams are asking today is less about where the data sits than about the actions these tools can trigger, the rights they hold, and the record they leave behind.

Julie Bonazzi explores how design, digital and narratives can transform the way organizations deploy and carry out change. A specialist in digital marketing and brand experience, she has spent more than ten years helping companies and entrepreneurial projects build identities, communication strategies and digital ecosystems.
An entrepreneur, she develops an approach in which design, agility and digital culture become levers to make transformations more intelligible and more engaging.
Julie is the co-founder, with Malika Ait El Mouden, of Twin Transformation.fr



