Where do your contracts go when you use AI? Hosting and confidentiality

When a user sends a contract to an AI assistant, several separate operations may take place. They do not necessarily happen in the same place, or with the same party.

• Storage. The document is kept in the application or in a dedicated space.
• Processing by the model. The text, or part of it, is sent to the model that produces the answer.
• Temporary retention. Prompts and answers may be kept for history, support or abuse detection.
• Ancillary services. Application monitoring, email delivery or backups involve other providers.

To assess a tool, you therefore need to follow the full path of the data, not just look for a country name on the home page.

Not necessarily. Hosting describes where data is stored. The location where the model processes it is a separate piece of information, which depends on the architecture chosen by the provider.

Microsoft’s documentation on models sold by Azure, including Azure OpenAI, dated May 18, 2026, illustrates this difference well. It states that, for a “Global” deployment, prompts and responses may be processed in any geography where the model is deployed. For a “DataZone” deployment created in the European Union, processing may take place in any EU member state. Stored data, for its part, stays in the geography designated by the customer.

The same document shows that a no-training commitment does not settle everything. Microsoft states that prompts and responses are not used to train foundation models without the customer’s permission or instruction. It also describes an abuse detection system that may lead to retaining certain exchanges for human review, unless the customer has been granted an exemption.

Trade secrets are defined in Article L. 151-1 of the French Commercial Code, introduced by Law No. 2018-670, which transposes Directive (EU) 2016/943. As the French Constitutional Council (Conseil constitutionnel) noted in its decision No. 2018-768 DC, information is protected if it meets three criteria: it is not generally known or readily accessible to persons familiar with this type of information, it has commercial value because it is secret, and its holder takes reasonable measures to keep it secret.

This last criterion is directly relevant to teams that use AI. Pricing terms, negotiation strategy or a partner’s data may fall under this regime. How this information flows through tools is part of the broader question of the protective measures in place. The text gives no AI-specific answer: the assessment is made case by case.

There is also a contractual issue. The directive considers unlawful, without the holder’s consent, use or disclosure by a person acting in breach of a confidentiality agreement or of a duty to limit the use of the secret. Before submitting a partner’s contract to a tool, it is therefore useful to check what your own non-disclosure agreement with that partner provides, in particular regarding authorized third parties and providers. This point is a matter for your counsel’s analysis.

A contract often contains names, job titles, contact details or information about employees. The GDPR may therefore apply to their processing by an AI tool. Three points structure the analysis.

• Roles. The CNIL (the French data protection authority) FAQ of July 18, 2024 asks organizations to define their own role and that of the provider, and to question the provider about the compliance of its system.
• The data processing agreement. Under Article 28, the controller uses only processors that provide sufficient guarantees. The agreement covers, in particular, instructions, confidentiality, security, what happens to the data at the end of the service and audits. Engaging another processor requires prior written authorization, specific or general.
• Transfers outside the EU. Articles 44 to 46 allow a transfer only within the framework provided: an adequacy decision or appropriate safeguards, such as the standard contractual clauses adopted by the Commission.

For the United States, the Commission adopted an adequacy decision on the EU-U.S. Data Privacy Framework on July 10, 2023. The General Court of the EU dismissed the action brought against it in case T-553/23, and an appeal was lodged on October 31, 2025 (C-703/25 P). The status of this procedure and the exact legal basis relied on by each provider should be checked. For details, read our Article 28 checklist.

The delivery model changes the level of control. The CNIL FAQ distinguishes several situations. On-premise deployment is presented as preferable for personal or sensitive data, because it limits the risk of extraction by a third party. The cloud requires a contract specifying responsibilities and access. With an API, control lies almost entirely with the provider, which calls for particular attention to contractual terms and transfers.

For services available to the general public, the CNIL considers non-confidential use with safeguards, such as dedicated work addresses. ANSSI (the French national cybersecurity agency), in its April 2024 guide, recommends ruling out generative AI tools available on the internet for professional use involving sensitive data. These rules belong in your AI use policy.

These questions apply to a general-purpose assistant, a legal tech tool or a CLM. A written, dated answer is more useful than a marketing statement.

• Hosting. Where are documents, indexes and backups stored?
• Processing location. Where does the model process prompts? What type of deployment is used, and can it change without notice?
• Model. Who supplies the model, and does data pass through that supplier’s services?
• Training. Is our data used to train or improve a model? Where is this commitment written?
• Retention. Which prompts and answers are kept, for how long, and for what purpose?
• Human access. Who can view our data, at your company and at your subprocessors, and under what circumstances?
• Subprocessors. What is the current list, with the role and location of each? How are we notified of a change?
• Transfers. Does any data leave the European Union, and on what legal basis?
• Requests from authorities. Are you subject to a foreign law that allows access to data to be demanded? What is your procedure in that case?
• Security. What encryption, what key management, what penetration testing? Which documents are available, and on what terms?
• Certifications. What is the exact scope: the provider, its hosting provider or the product purchased?
• End of contract. How do we retrieve our data, and how is its deletion confirmed?

The platform’s contracts and data are hosted in France, on Microsoft Azure. The security page lists the subprocessors, including Azure OpenAI (Microsoft) for the assistant’s artificial intelligence, a subprocessor marked as optional, with “Microsoft Azure” as its location. Some subprocessors process data outside France, in particular for sending transactional emails, which is operated from the United States.

Data is encrypted in transit with TLS 1.2+ and at rest with AES-256. Docform acts as a processor for its customers, and a DPA is available on request. The questions above also apply to Docform: ask us during a demo.