Claude's Security and Compliance in Business: GDPR, AI Act and Governance
This article was updated on September 27, 2026, based on Anthropic's official documentation (Privacy Center, platform documentation, help center), Regulation (EU) 2026/1744, and the European Commission's Q&A on AI governance, reviewed page by page. The first version was dated August 16, 2026.
Every enterprise AI deployment eventually comes up against the same meeting: the one with the CISO, the DPO, and the legal department. This is good news. In the organizations we support, including those in sectors with stringent requirements such as aeronautics and space or defenseThis meeting is precisely what allows things to move quickly afterwards. The question is never whether Claude can be used in a company, but under what conditions, within what framework, and who is responsible for it.
Here are the five questions that consistently come up, what official documents from Anthropic and regulators say about them, and for each one, the specific artifact that unlocks the project.
1. Where does the data go?
This is the question that most often blocks projects, frequently due to confusion between offers. Claude exists in four very different contractual forms, and the answer changes on each line of the table below.
| It offers | Training models on your data | Conservation | Management by the company |
|---|---|---|---|
| Free, Pro, Max individual accounts | Yes, if the user allows it in their privacy settings (this choice can be changed at any time; Incognito mode is always excluded). | Up to 5 years in training pipelines if model improvement is enabled; a deleted conversation disappears from the servers within 30 days | None: it's a contract between Anthropic and the person, not with you. |
| Claude Team and Claude Enterprise | Not by default; only explicit feedback (thumbs up or down) can be used, and the Enterprise administrator can disable this sharing. | According to the organization's policy; on Enterprise, the administrator sets the duration from 1 day to unlimited | SAML or OIDC SSO, SCIM provisioning, connector whitelisting, audit logs, compliance API (Enterprise, excluding the public sector), activation of Cowork by the organization |
| Anthropic API (Claude Platform) | No | Input and output data is deleted after 30 days by default; zero retention (ZDR) is available upon request for eligible uses, although some recent models require a 30-day retention period. | Data Processing Agreement (DPA) with standard contractual clauses integrated into the commercial terms; inference option limited to the United States |
| Claude via AWS Bedrock, Google Vertex AI or Microsoft Foundry | No (partner cloud conditions) | Partner Cloud Terms and Conditions | Bedrock and Vertex AI offer European regions; on Bedrock, the model prefix eu. guarantees that prompts and replies do not leave the Union. According to Foundry, Microsoft had not yet announced a European data zone for Claude as of the date this article was updated. |
Three points that are repeated in all legal journals. First, Anthropic indicates that the data from its direct offers is stored in the United States, with inference potentially routed to several regions (United States, Europe, Asia, Australia) unless otherwise instructed; the only strict residency offered on the API is inference_geo: "us"There is no direct "Europe" option. For European residents, the process therefore involves Bedrock or Vertex AI. Furthermore, even with zero retention, content flagged by automated trust and security systems can be retained for up to two years. Finally, commercial products are covered by SOC 2 Type I and II, ISO/IEC 27001:2022, and ISO/IEC 42001:2023 (AI systems management) certifications, with HIPAA-compliant configuration available upon request.
The artifact that unlocks: A contractual validation note, signed by the legal department, specifies the exact offering selected (Enterprise, direct API, or API via a partner cloud), the inference region, the configured retention period, the applicable clauses, and their verification date. One page is sufficient. This document authorizes deployment, not a personal opinion or a blog post, including this one: the policies above date from September 2026, and Anthropic updates them several times a year.
2. What data can be entered into it?
Most incidents stem not from the tool itself, but from its use: a colleague copying a sensitive document into a personal account. The previous table illustrates why this action makes all the difference: on a Free, Pro, or Max account, the conversation can be used to train models if the option is enabled, whereas it cannot on Teams, Enterprise, or via the API. This shadow AI already exists within your organization, whether you've deployed an official tool or not; a controlled deployment is precisely what eliminates it.
The artifact that unlocks: A three-level classification matrix, fitting on a single page, with illustrative examples for each level. Green level: non-sensitive public or internal data, freely usable within the approved environment. Orange level: confidential data, authorized use only within the corporate environment, never on a personal account. Red level: regulated or strategic data, which must remain within designated systems. Three lines that everyone remembers are worth more than a fifteen-page charter that no one reads.
3. What information do the agents have access to?
As soon as Claude is connected to the information system, via connectors or the MCP protocol, and even more so with Cowork When an agent performs tasks on entire files, the question changes in nature. It is no longer "what is put in the prompt" but "what does the agent access, on whose behalf, with what permissions, and who authorized it?" The CNIL (French Data Protection Authority) and the AI and Digital Council formalized this in an exploratory note dated July 20, 2026: agentic systems create "a change of scale" by accessing large volumes of data from the multiple sources to which they are connected, with a risk of loss of control for the user and a more difficult distribution of responsibilities to establish.
On the product side, Claude Enterprise allows the administrator to maintain a whitelist of connectors and extensions, and to choose who can activate them. Coworkand to export the activity (conversations, files, audit events) via the compliance API, which also covers Cowork and Claude Code on the workstation. Sessions processed by Bedrock or Vertex AI do not appear in this API: if you mix channels, traceability is built on each side.
The artifact that unlocks: A connector register, maintained like any application integration register, with five columns: connector, data scope, authorized populations, access level (read by default, write by exception), and review date. The guiding principle is that of least privilege: an agent inherits exactly the rights it is granted, no more and no less, and this is what makes the process manageable.
4. And what about the GDPR?
The steps are the same as for any personal data processing, without any exoticism: list the processing activities in the register, identify the legal basis, apply data minimization by only disclosing the necessary data, inform the individuals involved, and verify whether a Data Protection Impact Assessment (DPIA) is required given the risk. Two points are specific to Claude. The subcontractor is based in the United States: the transfer is based on standard contractual clauses, automatically integrated into the Data Processing Agreement (DPA) of Anthropic's commercial terms, or on the partner cloud contract if you use Bedrock, Vertex AI, or Foundry. And the list of Anthropic's subsequent subcontractors is publicly available on its Trust Center: this is what your DPO must attach to the processing record, not a screenshot.
The CNIL has published a series of practical guides on AI and the GDPR (purpose, legal basis, data minimization, informing individuals, security, rights) and, more recently, the note on agentic AI mentioned above. These provide an up-to-date starting point for your DPO.
The artifact that unlocks: the standard processing sheet for AI uses, prepared once with the DPO and then adapted by use case, instead of being rediscovered for each project.
5. What does the AI Act require?
Since February 2, 2025, Article 4 of the AI Act has required companies that provide or use AI systems to take measures to develop their employees' AI proficiency, taking into account their role and the context of use. Omnibus Regulation (EU) 2026/1744, published in the Official Journal on July 24, 2026, and in force since July 27, 2026, has made this an obligation of means ("fostering the development of AI proficiency"): it no longer requires guaranteeing a "sufficient level" for each individual. Enforcement has been the responsibility of national authorities since August 2, 2026, without the specific penalties stipulated in Article 4.
The European Commission's Q&A, updated on July 27, 2026, sets out the expected level: no certification is required, and an internal register of training and awareness-raising activities is sufficient; the content must be tailored to the roles of the individuals involved and the risks of the deployed systems, with hallucinations being explicitly mentioned in the context of generative AI; simply asking teams to read the instructions is not enough; and the obligation covers external service providers and contractors who use your systems on your behalf. The Commission and Member States, for their part, must support and facilitate the efforts of businesses, particularly SMEs. Training teams is therefore no longer just a good adoption practice, but a documented compliance requirement.
The remaining obligations depend on the risk level of the systems used. For office and intellectual production use, the transparency rules of Article 50 have applied since August 2, 2026: they primarily concern the software publisher (labeling of generated content, with a deadline of December 2, 2026 for systems already on the market before August 2), and then you as soon as you expose a chatbot to individuals or publish generated content intended to inform the public. The AI omnibus bill, however, has postponed the obligations for high-risk systems: December 2, 2027 for cases in Annex III (recruitment, employee management, access to credit, etc.) and August 2, 2028 for security components of products in Annex I. Have your legal counsel confirm the exact deadlines applicable to your situation.
The artifact that unlocks: A structured training plan, including target groups, dates, content for each role, and certifications. This is both the answer to the requirement for AI proficiency and the best way to accelerate adoption.
Who does what
| Actor | Main responsibility |
|---|---|
| CIO and CISO | Choice of offer and inference region, configuration of retention and SSO, connector whitelist, least privilege, periodic review |
| DPO and legal | Contractual validation (DPA, standard clauses, subcontractors), processing register, impact assessments, regulatory monitoring AI Act |
| Career paths | Classification applied to use cases, designation of priority cases, quality control of deliverables |
| HR and training | Skills development plan outlined under Article 4, integration into the career path of new arrivals and service providers |
The 30-day framing
This comprehensive framework can be established in a month, and we regularly see it implemented in less: week 1, selection of the offering, contract validation, and role assignments; week 2, classification matrix and usage rules on a single page; week 3, connector registry, space configuration (SSO, data retention, authorized connectors), and pilot scope; week 4, training of pilot users and launch. Organizations that deploy smoothly are not necessarily those with the fewest constraints: they are those that have transformed their constraints into simple, universally understood, and regularly reviewed tools.
FAQ
Does Claude use my company's data to train his models?
No, not on commercial offerings (Team, Enterprise, API, partner clouds), unless a user sends explicit feedback, which the Enterprise administrator can disable. Yes, potentially, on an individual Free, Pro, or Max account where the user has left the model enhancement option enabled. This is the primary reason why an enterprise deployment must use an enterprise offering.
Can Claude's data be kept in Europe?
Not currently available on Anthropic's direct offerings: the data is stored in the United States, and the only strict residency option available on the API is US. A European residency is possible by consuming Claude models via AWS Bedrock European regions (using inference profiles). eu.) or Google Vertex AI, under the contractual conditions of these clouds.
Is Claude compliant with the GDPR?
GDPR compliance applies to the processing itself, not to the tools used. Anthropic provides the essential building blocks expected of a data processor: Data Processing Agreement (DPA) with standard contractual clauses integrated into the terms and conditions, a public list of processors, ISO 27001 and SOC 2 certifications, and no training on commercial data. The remaining aspects (purpose, legal basis, data minimization, information, and Data Protection Impact Assessment - DPA) are the responsibility of your data processing register and your Data Protection Officer (DPO).
Is training your teams in AI really mandatory?
Yes. Article 4 of the AI Act, as amended by Omnibus Regulation (EU) 2026/1744, which came into force on July 27, 2026, requires that measures be taken to develop the AI skills of staff and contractors using your systems, without requiring certification or a guaranteed level of proficiency. An internal record of actions taken is sufficient to demonstrate this; a complete lack of action is no longer justifiable.
Is an impact analysis required before deploying Claude?
Not always. A Data Protection Impact Assessment (DPIA) is required when the processing is likely to result in a high risk to individuals, for example, if Claude processes employee, patient, or customer data on a large scale, or if his output informs decisions concerning them. For office use involving data classified as green or orange, a well-maintained processing record is usually sufficient; the Data Protection Officer (DPO) decides on a case-by-case basis.
Sources
- Anthropic Privacy Center: use of data from commercial products for training, use of individual account data (policy of March 16, 2026) organizations' retention periods, server location, certifications et Data Processing Agreement (DPA).
- Claude Platform, documentation: API data preservation and zero retention et data residence.
- Claude, Help Center and Academy: Compliance API et Claude Enterprise Administrator's Guide.
- European Union : Regulation (EU) 2024/1689 (AI Act) Regulation (EU) 2026/1744 (omnibus IA) and Questions and answers from the Commission on the control of AI (updated on July 27, 2026).
- CNIL: AI: How to comply with the GDPR et Exploratory note from CNIL and CIANum on agentic AI and personal data (July 20, 2026).
This article provides general guidelines and does not constitute legal advice. Have your analysis reviewed by your legal department and your Data Protection Officer (DPO).
To learn more: discover how Generative AI is disrupting the GDPR, Why AI transparency It remains a gamble on trust and how Claude CoworkCopilot Cowork and ChatGPT Work They compare themselves on governance.