It is the question that comes up most often. A business owner wants to use Claude to save time, but hesitates to feed it internal data: client files, contracts, accounting records, source code. The worry is simple. Will this data train the AI? Can it leak? Is it stored, and where?
The short answer is that there is no single answer. Everything depends on the plan you use and the setup you put in place. The same tool can be GDPR-compliant for one company and entirely unsuitable for another. This article sorts it out, with verified data, without taking Anthropic's messaging at face value.
Key takeaway. On consumer accounts (Free, Pro, Max), your data can train Claude since August 2025, unless you turn the option off. On business accounts (API, Team, Enterprise), Anthropic does not train on your data. But in every case, data is hosted in the United States, which raises a transfer question for European personal data.
The golden rule: consumer or business account
This is the most important distinction, and the one most people miss. Anthropic applies two radically different privacy regimes.
| Criterion | Consumer (Free, Pro, Max) | Business (API, Team, Enterprise) |
|---|---|---|
| Training on your data | Yes, by default (since Aug 2025) | No |
| Opt-out possible | Yes, manual opt-out | N/A, never trains |
| Retention period | Up to 5 years if training is on | 30 days standard |
| Zero Data Retention | No | Yes, on eligible Enterprise |
| Data Processing Agreement (DPA) | No | Yes |
The lesson is clear. A Claude Pro account at 20 euros a month is not a professional tool in data-protection terms. It is fine for personal use or non-sensitive content. The moment you touch client or confidential data, you need a business account with a signed processing agreement.
The August 2025 change many missed
Until 2025, Anthropic held a simple position on consumer accounts: no training on conversations, deletion within 30 days. It was a selling point against laxer competitors.
On August 28, 2025, Anthropic changed that policy. On Free, Pro and Max plans, conversations now train the models unless the user explicitly refuses. Retention rises to five years when training is on. New users choose at signup. Existing users were given a window to decide.
What to do on a Pro or Max account. Go to your account privacy settings and check the training option. If it is on and you work on sensitive topics, turn it off. That brings retention back to 30 days and excludes your data from training.
This change is not a scandal in itself, many vendors do the same. But it makes a key point: a privacy policy can change overnight. For a small business, relying on a contractual commitment (the DPA of a business account) is firmer than relying on the default settings of a consumer account.
The real blocker for an accounting or law firm: data residency
Here is the case we are asked about most. An accounting firm wants to use Claude to analyse balance sheets, draft letters or sort accounting documents. Can it feed it its clients' data?
The honest answer is: not without serious precautions, and not on just any plan.
The main problem is not training (a business account solves that). The problem is the location of the data. Anthropic hosts its entire infrastructure in the United States. There is, to date, no EU data residency option on claude.ai or the direct Anthropic API. All requests are processed on US servers.
For European personal data, this triggers GDPR Articles 44 to 46 on transfers outside the European Union. Anthropic does not currently hold a certification under the EU-US Data Privacy Framework. The transfer therefore relies on Standard Contractual Clauses and supplementary measures, whose analysis falls on the using company.
For an accounting or law firm bound by professional secrecy, putting identifiable client data into Claude through a standard account poses a real compliance risk. Professional secrecy does not sit well with an uncontrolled transfer to a US subprocessor.
A solution exists, but it requires a precise setup. To get EU data residency, you must deploy Claude not directly, but through AWS Bedrock (European regional profiles) or Google Vertex AI (European endpoints). These platforms keep data inside the Union, with a suitable processing agreement. It is heavier to set up, but it is the clean path for regulated sectors.
Anthropic's actual certifications
Anthropic is not a lightweight on security. Beyond the residency question, the company holds several verifiable certifications.
| Certification | Coverage |
|---|---|
| SOC 2 Type I and II | Operational security controls |
| ISO 27001:2022 | Information security management |
| ISO/IEC 42001:2023 | AI management systems |
| HIPAA (BAA available) | Health data, on API, Enterprise, Claude Code (ZDR) |
These certifications are serious and cover good operational security practices. They do not settle the European residency question, which is a separate matter. A company can be fully certified and still host its data in the United States. The two questions must be handled separately.
Claude Code and Cowork: extra risks to know
Anthropic's agentic tools, Claude Code and Claude Cowork, deserve special attention because they access your files directly.
Claude Code is the command-line interface for development. Important verified point: every file Claude Code reads is sent in full to Anthropic for analysis, over an encrypted connection. Execution stays local, but file content travels. On a consumer account, that code can train the model. On a business account, it cannot. For a firm working on proprietary code or sensitive data, the Enterprise account with Zero Data Retention is the only acceptable setup.
Claude Cowork is an agent that acts on your local files: it reads, writes, runs commands, organises documents. The risk here is not only confidentiality. It is also autonomous action. A reasoning error can lead to file modification or deletion without human validation. On top of that, these actions run locally and often escape the company's security monitoring tools. Cowork demands a strict usage framework and human oversight on sensitive actions.
How to use Claude securely, by use case
Here is a simple grid to decide what you can entrust to Claude, depending on your situation.
Marketing content, drafts, public topics. A Pro or Max account is enough, with the training option off. No sensitive data is involved.
Non-personal internal data (processes, documentation, non-critical code). A Team or Enterprise account with a processing agreement. No training, controlled retention.
Client data, personal data, professional secrecy. An Enterprise account with Zero Data Retention, deployed through AWS Bedrock or Vertex AI in a European region to guarantee data residency. Documented transfer analysis. This is the level required for an accounting firm, a law firm or a healthcare professional.
Health data. On top of the above, sign a BAA (Business Associate Agreement) with Anthropic, available on eligible plans.
The reflex to adopt. Before pasting anything into Claude, ask a simple question. If this content leaked tomorrow, would it cause a legal or trust problem? If yes, it does not go into a consumer account. It goes into a configured business environment, or it is anonymised before processing.
An often forgotten point: Fable 5 and mandatory retention
The release of Claude Fable 5 in June 2026 introduced a nuance that sensitive businesses should know. For Mythos-class models, which Fable 5 belongs to, Anthropic enforces a 30-day data retention, with no Zero Data Retention option on that specific traffic. The stated reason is safety, as these models are more powerful.
This constraint has a concrete consequence. According to reports published in June 2026, Microsoft restricted its employees' access to Claude Fable 5 precisely because of this mandatory retention. For a small business, this means you must distinguish models. Usage on Claude Sonnet or Haiku through Enterprise can benefit from Zero Data Retention. The same usage on Fable 5 cannot, at least for now. Model choice is part of the compliance decision. Our analysis of Claude Fable 5 details this model's capabilities.
Frequently asked questions
Can an accounting firm put client data into a Claude project?
Not on a consumer account, and not without precautions on a standard account. For identifiable client data under professional secrecy, you need an Enterprise account with Zero Data Retention, deployed through AWS Bedrock or Google Vertex AI in a European region to guarantee data residency in the Union, plus a documented transfer analysis. Otherwise, data must be anonymised before any processing.
Do my Claude conversations train the AI?
On Free, Pro and Max accounts, yes by default since August 28, 2025, unless you turn the option off in settings. On business accounts (API, Team, Enterprise), no, Anthropic never trains on your data.
Is Claude GDPR-compliant?
Claude can be used in a compliant way, but compliance depends on your setup. The sensitive point is data residency, hosted in the United States by default. For European personal data, a deployment through a European region of AWS or Google and a transfer analysis are necessary.
Does Claude Code send my source code to Anthropic?
Yes. Every file Claude Code reads is transmitted in full to Anthropic for analysis, over an encrypted connection. On a consumer account, that content can train the model. On a business account with Zero Data Retention, it is neither stored nor used for training.
Should you trust Anthropic's privacy promises?
The certifications (SOC 2, ISO 27001, ISO 42001) are verifiable and serious. But a privacy policy can change, as the August 2025 update showed. The firmest protection for a company remains a signed processing agreement on a business plan, not the default settings of a consumer account.