Blog · Sovereign
Sovereign AI what your customer conversations reveal, and who can access them
GDPR, CLOUD Act, FISA 702, Data Act, SecNumCloud: what the law says about your customer conversations and the questions to ask an AI agent provider.
- Text
- The QLOA team
- Published
- Reading time
- 12 min · 12 chapters
A customer conversation contains what no form collects: a purchase intent, a hesitation, an objection, an address, sometimes a health problem. As soon as an AI agent runs these exchanges, the question is no longer just where the data is stored. You also need to know who runs the model, who trains it, which law applies to the provider and how to leave.
A conversation is data of a particular density
A customer who writes “I’m looking for a mattress, I have a herniated disc” has just shared health data without being asked for it. The GDPR places data concerning health among the special categories of Article 9, along with political opinions, religious beliefs, trade union membership, genetic and biometric data or sexual orientation. Processing them is prohibited in principle, except in narrowly listed cases1.
That is what sets conversational data apart. A form collects fields chosen by the company. A conversation collects whatever the customer decides to say, in their own words. It contains identity data, transaction data, and a layer that other channels rarely capture: the customer’s reasoning. Why they hesitate, what they compare, what makes them give up.
A conversation history therefore describes a market’s expectations, the perceived weaknesses of an offer and price sensitivity. It is strategic information for the company, and personal information for each customer.
What the GDPR requires of an AI agent
The GDPR applies to any conversation that makes it possible to identify a person, directly or indirectly. For an AI agent, four requirements shape the project.
The legal basis. Article 6 provides for six: consent, performance of a contract, legal obligation, protection of vital interests, public interest task, legitimate interest1. Answering a customer about their order generally falls under performance of a contract. Reusing the same conversation to improve a model is a different purpose, which calls for its own analysis.
Minimization. Data must be “adequate, relevant and limited to what is necessary in relation to the purposes” (Article 5(1)(c))1. An agent that asks for an order number has no reason to keep the rest of the purchase history in its logs.
Retention period. Data may be kept in an identifiable form only for as long as necessary for the purposes (Article 5(1)(e))1. An “unlimited by default” period for conversation transcripts does not meet this requirement.
Transfers outside the Union: Chapter V
Chapter V of the GDPR (Articles 44 to 50) governs any transfer of personal data to a third country1. There are three main routes: an adequacy decision by the European Commission (Article 45), appropriate safeguards such as standard contractual clauses (Article 46), or derogations for specific situations (Article 49).
Article 48 deserves particular attention. A decision of a court or administrative authority of a third country requiring the disclosure of personal data may be recognized or enforceable only if it is based on an international agreement, such as a mutual legal assistance treaty1. This provision is at the heart of the conflict of laws described below.
The EU–U.S. Data Privacy Framework
On July 10, 2023, the European Commission adopted Implementing Decision (EU) 2023/1795, which finds that the United States ensures an adequate level of protection for data transferred to organizations participating in the Data Privacy Framework2. This framework succeeds two earlier decisions invalidated by the Court of Justice in the Schrems I and Schrems II judgments3. The Schrems II judgment, delivered on July 16, 2020, had invalidated the Privacy Shield4.
On September 3, 2025, the General Court of the European Union dismissed the action for annulment brought by Philippe Latombe against this decision (Case T-553/23)3. The applicant challenged the independence of the American Data Protection Review Court and the framework governing bulk collection by intelligence agencies. The Court rejected both pleas. It notes in particular that the Schrems II judgment does not require prior authorization by an independent authority, but at a minimum a judicial review after the fact3. It states that the Commission must continuously monitor developments in the American legal framework and may, if necessary, suspend, amend or repeal its decision3.
On October 31, 2025, Philippe Latombe lodged an appeal before the Court of Justice (Case C-703/25 P)5. As of the publication date of this article, the case is pending and the Data Privacy Framework remains in force.
CLOUD Act and FISA 702: a question of access, not storage
Hosting data in Europe does not, on its own, settle the question of access.
The CLOUD Act (2018)
The CLOUD Act was enacted on March 23, 2018 (Public Law 115-141, Division V)6. It introduced into the U.S. federal code a provision (18 U.S.C. § 2713) under which providers of electronic communication services and remote computing services must comply with their preservation and disclosure obligations for data in their possession, custody or control, whether that data is located inside or outside the United States6.
Under the most widespread reading of this text, the relevant criterion is therefore the jurisdiction the provider is subject to, not the location of the server. A provider subject to U.S. law may receive an order covering data hosted in a European data center. It can then find itself facing two legal systems: the U.S. order on one side, Article 48 of the GDPR on the other1.
FISA, Section 702: a regime in limbo
Section 702 of the Foreign Intelligence Surveillance Act authorized intelligence collection targeting non-U.S. persons located outside the United States, through directives issued to service providers8. It had been reauthorized by the Reforming Intelligence and Securing America Act, enacted on April 20, 2024 (Public Law 118-49)7.
According to the Congressional Research Service, Section 702 lapsed by operation of law on June 12, 2026, for lack of reauthorization by Congress8. Orders, authorizations and directives already in force remain applicable until they expire, within a limit of one year after issuance8. According to the Brennan Center, the Foreign Intelligence Surveillance Court approved the most recent certifications in March 2026, which allows collection to continue until March 17, 20279. In its October 16, 2026 update, the Brennan Center reports no reauthorization by Congress9.
For a European company, the lesson goes beyond this timeline: a foreign access framework can be changed by a parliamentary vote outside the Union, and an architecture that depends on it remains exposed to its variations.
Data Act: reversibility becomes a right
Regulation (EU) 2023/2854, known as the Data Act, has applied since September 12, 202510. Its Chapter VI governs switching between data processing service providers, a notion that covers cloud services.
The contract must provide for a maximum notice period of two months to initiate the switch, followed by a maximum transition period of thirty calendar days during which the original provider assists the customer and maintains service continuity (Article 25)10. The contract must also specify all categories of exportable data and digital assets. Since January 11, 2024, switching charges may not exceed the costs directly linked to the operation. From January 12, 2027, they will be prohibited (Article 29)10.
Article 32 further requires providers to take technical, organizational and legal measures to prevent third-country public authorities from accessing non-personal data held in the Union, where such access would conflict with Union or Member State law10. For personal data, the GDPR applies.
For an AI agent, reversibility is not limited to exporting transcripts. It also covers business rules, configurations, knowledge bases and, where applicable, models fine-tuned on your data. The question to ask is simple: if you leave, what do you get back, in what format, and how quickly?
SecNumCloud: explicit criteria against non-European law
The ANSSI’s SecNumCloud framework, in its version 3.2 of March 8, 2022, formalizes criteria for protection against non-European law11.
Its chapter 19.2 requires that customer data and technical data be stored and processed within the European Union, and that the service be administered and supervised from within the Union11. Its chapter 19.6 goes further11:
- the provider’s registered office, central administration and main place of business must be established in a Member State of the Union;
- share capital and voting rights may not be held, directly or indirectly, by entities established outside the Union at more than 24% individually and more than 39% collectively, and those entities may neither hold a veto right nor appoint the majority of the governing bodies;
- a subcontractor established or controlled outside the Union must not have the technical ability to obtain the data operated through the service;
- the provider must inform the customer, within one month, of any change that may affect its compliance with these requirements.
Our reading: protection against extraterritorial laws cannot be decreed in a contractual clause. It results from the provider’s ownership structure, its legal attachment and the technical impossibility of access for a non-European third party. This framework remains useful for a project that does not aim for certification.
Model training: where your conversations go after the conversation
The most sensitive question is often reuse. A conversation can serve to answer the customer, then to improve the model that answered them. If that model belongs to a third party and serves other customers, part of your business knowledge may contribute to it.
The CNIL’s first practical guidance on developing AI systems, published on April 8, 2024, covers in particular defining a purpose, qualifying the actors, the legal basis and the impact assessment12. On June 19, 2025, the CNIL clarified that legitimate interest is a possible legal basis for developing AI systems, subject to safeguards13. It gives an example directly applicable to customer relations: reusing the future conversations of a conversational agent’s users to improve the model can rely on legitimate interest “provided that certain strong safeguards are put in place”, such as informing individuals, a discretionary right to object and limiting use to pseudonymized or anonymized data13.
On July 22, 2025, the CNIL finalized this cycle of recommendations14. It draws on Opinion 28/2024 of the European Data Protection Board, adopted in December 202415, according to which the GDPR applies “in many cases” to AI models trained on personal data, because of their memorization capabilities14. It recommends analyzing and documenting whether a model falls under the GDPR14. Practical consequence: a model trained on your conversations may carry part of the data that fed it.
Legislative movement is under way. On November 19, 2025, the Commission presented a proposal for a “digital omnibus” regulation (COM(2025) 837) that would, among other things, amend the definition of personal data and the use of legitimate interest for training AI systems16. According to the European Parliament, the proposal had not been adopted as of September 20, 2026: the Council had not set its negotiating mandate and examination was continuing in committee in Parliament16. This strand is separate from the digital omnibus on AI, which amends the AI Act and was adopted on July 8, 2026 (Regulation (EU) 2026/1744) without amending the GDPR17. The GDPR applies today as currently worded.
Sovereignty is not a flag, it is control
Here is our analysis. The sovereignty of conversational data comes down neither to a logo nor to a country of hosting. It rests on four questions of control.
Where the model runs. Inference, meaning the moment when the model reads the conversation and produces the answer, is a processing operation in its own right. If inference is entrusted to a provider located outside the Union, conversational data is in principle transferred outside the Union, which brings Chapter V of the GDPR into play1.
Who trains it, and on what. A model fine-tuned on your data and reserved for your use does not have the same status as a shared model enriched by the conversations of all of a provider’s customers.
Who can compel it. The answer depends on the jurisdiction of the provider and that of its own subprocessors, as the CLOUD Act and the SecNumCloud criteria illustrate611.
How to leave. The Data Act sets a contractual framework for switching providers10. The useful assets (rules, knowledge, models) must still be truly portable.
Our hypothesis: these four questions will become ordinary purchasing criteria, just like security. As AI agents take on conversations that commit the company (advice, sales, booking, resolution), the value of those conversations increases, and with it the cost of losing control over them.
What this means for your organization
The first step is a mapping, channel by channel: categories of data, possible presence of sensitive data, retention period of transcripts, location of inference and location of storage.
The second step is documentary: one legal basis per purpose, distinguishing the handling of the request from the improvement of the model. If reuse for training is considered, the safeguards described by the CNIL serve as a reference: information, objection, pseudonymization13. A data protection impact assessment (Article 35 of the GDPR) should be considered as soon as sensitive data may appear in the exchanges112.
The third step is contractual, with a grid to put to any conversational AI provider.
| Topic | Question to ask | Why |
|---|---|---|
| Location | Where are conversations, logs and backups stored? | Article 5 and Chapter V of the GDPR1 |
| Inference | Where does the model that reads and produces the answers run? | Chapter V of the GDPR if inference takes place outside the EU1 |
| Subprocessors | Which subprocessors are involved, in which countries, under which jurisdiction? | Article 28 of the GDPR1 |
| Jurisdiction | Is the provider or its parent company subject to non-European law that allows a disclosure order? | CLOUD Act6, SecNumCloud 19.6 criteria11 |
| Transfers | If a transfer exists, what is it based on (adequacy, standard clauses)? | Articles 45 and 46 of the GDPR1 |
| Training | Are our conversations used to train a model shared with other customers? Can this be excluded by contract? | CNIL recommendations13 |
| Fine-tuned model | Who owns a model fine-tuned on our data, and what happens to it at the end of the contract? | A model can fall under the GDPR1415 |
| Retention | What default period, configurable by data type? | Article 5(1)(e)1 |
| Sensitive data | What mechanisms detect and filter health data or other special categories? | Article 9 of the GDPR1 |
| Reversibility | Which data and assets are exportable, in what format, within what time? | Articles 25 and 29 of the Data Act10 |
| Foreign access | What measures prevent access by a third-country authority, and will we be informed of a request? | Article 48 of the GDPR1, Article 32 of the Data Act10 |
| Changes | Will we be told of a change of shareholders or hosting provider? | SecNumCloud requirement 19.6 f11 |
A vague answer flags a point to document before signing, so that your teams keep control over what is done with your conversations.
Where QLOA stands
QLOA is an AI agent platform for customer experience, hosted and operated in Europe. Its architecture rests on a Router, an Orchestration layer, an AI Analyzer, Skills and Templates. Model fine-tuning goes through QLOA Forge, offered in five levels depending on the degree of control sought: Shared Forge (pooled), Forge Cloud (managed private model), Forge Cloud Dedicated (dedicated GPU), Forge On-Prem (on-site installation) and Forge Fleet (federated multi-site). Agent rules and tone are written in plain language, and every conversation is tracked. Agents operate on the website, on messaging apps including WhatsApp, and by voice.
Frequently asked questions
What is a sovereign AI?
A sovereign AI is an AI whose conditions of use the company controls: where the model runs, where the data is stored, the jurisdiction of the provider and its subprocessors, the use of data for training and the ability to switch providers. European hosting of the AI is one component, not the whole picture. ANSSI’s SecNumCloud criteria give a precise example of what independence from non-European laws covers.
Does the CLOUD Act apply to data hosted in Europe?
It can, if the provider falls under U.S. jurisdiction. The 2018 CLOUD Act provides that the disclosure obligations of the providers concerned cover data in their possession, custody or control, whether located in the United States or elsewhere. The criterion is the provider’s jurisdiction, not the server’s location. For its part, Article 48 of the GDPR limits the recognition of foreign orders that are not based on an international agreement.
Is the Data Privacy Framework still valid in 2026?
Yes. The adequacy decision of July 10, 2023 is in force. The General Court of the European Union dismissed the action for annulment on September 3, 2025 (Case T-553/23). An appeal was lodged before the Court of Justice on October 31, 2025 (Case C-703/25 P) and is still pending. The Commission must also monitor developments in U.S. law and may suspend or amend its decision.
Can an AI agent provider use my customer conversations to train its model?
Not without a legal basis and safeguards. According to the CNIL, reusing the conversations of a conversational agent to improve the model can rest on legitimate interest, provided there are strong safeguards: informing individuals, a discretionary right to object, limiting use to pseudonymized or anonymized data. As the controller, you can also frame or exclude this reuse in the processing contract.
What does the Data Act change for an AI agent contract?
Applicable since September 12, 2025, the Data Act requires data processing service providers to include switching clauses: notice of two months at most, a transition period of thirty days at most, a list of exportable data. Switching charges are limited to direct costs until January 12, 2027, then prohibited. It is worth checking at signing what will be exportable.
Sources
- 1
European Parliament and Council, “Regulation (EU) 2016/679 on the protection of natural persons with regard to the processing of personal data (GDPR)”, April 27, 2016. eur-lex.europa.eu
- 2
European Commission, “Implementing Decision (EU) 2023/1795 on the adequate level of protection of personal data under the EU-US Data Privacy Framework”, July 10, 2023. eur-lex.europa.eu
- 3
General Court of the European Union, press release No. 106/25, “Judgment in Case T-553/23, Latombe v Commission”, September 3, 2025. curia.europa.eu
- 4
Court of Justice of the European Union, judgment C-311/18, Data Protection Commissioner v Facebook Ireland and Schrems (“Schrems II”), July 16, 2020. curia.europa.eu
- 5
Court of Justice of the European Union, Case C-703/25 P, Latombe v Commission, appeal lodged on October 31, 2025 (notice published in the OJEU on December 22, 2025). curia.europa.eu
- 6
United States Code, “18 U.S.C. § 2713, Required preservation and disclosure of communications and records”, added by Public Law 115-141, div. V, March 23, 2018. law.cornell.edu
- 7
U.S. Government Publishing Office, “Public Law 118-49, Reforming Intelligence and Securing America Act”, April 20, 2024. govinfo.gov
- 8
Congressional Research Service, “The Impact of FISA Section 702’s Repeal”, Legal Sidebar LSB11444, June 22, 2026. everycrsreport.com
- 9
Brennan Center for Justice, “Section 702 of the Foreign Intelligence Surveillance Act, Explained”, updated October 16, 2026. brennancenter.org
- 10
European Parliament and Council, “Regulation (EU) 2023/2854 on harmonised rules on fair access to and use of data (Data Act)”, December 13, 2023. eur-lex.europa.eu
- 11
ANSSI, “Cloud computing service providers (SecNumCloud), requirements framework, version 3.2”, March 8, 2022. cyber.gouv.fr
- 12
CNIL, “AI: the CNIL publishes its first recommendations on the development of artificial intelligence systems”, April 8, 2024. cnil.fr
- 13
CNIL, “Development of AI systems: the CNIL publishes its recommendations on legitimate interest”, June 19, 2025. cnil.fr
- 14
CNIL, “AI: the CNIL finalizes its recommendations on the development of AI systems and announces its future work”, July 22, 2025. cnil.fr
- 15
European Data Protection Board, “Opinion 28/2024 on certain data protection aspects related to the processing of personal data in the context of AI models”, December 2024. edpb.europa.eu
- 16
European Parliament, Legislative Train Schedule, “Digital Omnibus Regulation Proposal” (COM(2025) 837, 2025/0360(COD)), updated September 20, 2026. europarl.europa.eu
- 17
European Parliament and Council, “Regulation (EU) 2026/1744 of July 8, 2026 amending Regulations (EU) 2024/1689, (EU) 2018/1139 and (EU) 2023/1230 (digital omnibus package on AI)”, OJEU L of July 24, 2026. eur-lex.europa.eu
