Claude and the GDPR: Free, Pro and Max compared with Team, Enterprise and the API

How do Claude Free, Pro and Max differ from Team, Enterprise and the API?

Anthropic applies different rules for training, retention and contract to consumer and to commercial products. Free, Pro and Max are consumer products, where each user makes a training choice. Team and Enterprise (Claude for Work) and the API are commercial products, to which that choice does not apply.^[1]

The retention row for commercial products is deliberately left open: the pages we checked give no fixed period that we could state as fact here. Ask the provider to confirm the retention rules for your plan and record them in your documentation. For a comparison with ChatGPT, Copilot and Gemini, see the privacy comparison of the four assistants. This guide covers the EU GDPR; UK teams should check the UK GDPR and the ICO’s position separately.

What changed on 28 August 2025?

Under the consumer terms introduced from 28 August 2025, users of Free, Pro and Max, including Claude Code on those accounts, choose whether their chats and coding sessions may be used to train the models. Existing users had to decide by 8 October 2025. The choice applies to new and resumed chats; deleted conversations are not used.^[1]

The decision has a consequence for retention: if a user agrees, Anthropic keeps their data for up to five years; if not, the previous period of 30 days stays.^[1] This matters for business data because one click by one employee extends the storage period without your organisation noticing.

Anthropic now names three cases in which consumer chats can be used for training: the model improvement setting is on, chats are flagged for safety review, or the user takes part in explicit programmes. Incognito chats are not used for improvement. When a user rates a response with a thumbs up or down, the whole conversation is stored for up to five years, de-linked from the account, and may be used for training.^[2]

What applies to Claude Team, Enterprise and the API?

For commercial products Anthropic writes that, by default, it does not use inputs or outputs. The exception is feedback that is submitted. Owners of Team or Enterprise can switch off member feedback (“Rate chats”).^[3] The consumer setting of 28 August 2025 does not apply to these products: Anthropic names Claude for Work (Team, Enterprise), Claude Gov, Claude for Education and the API, including via Bedrock and Vertex, as not covered.^[1]

On the contract: the Data Processing Addendum covers commercial products and, according to Anthropic, is automatically incorporated, including the Standard Contractual Clauses, into the Commercial Terms. You do not have to sign it separately; consumer products are excluded. If you reach Claude through a third-party platform, Anthropic says that platform’s terms govern your use.^[4] That is convenient but has a practical consequence: nobody in your organisation signs anything, so there is no moment at which data protection reads the contract. File the DPA version in force yourself and check that it covers the points of Article 28 GDPR, such as processing only on documented instructions.^[9]

Where is the data processed?

Anthropic’s regional compliance page lists Europe as available through AWS Bedrock and Google Vertex; Microsoft Foundry appears there as “coming soon”. The page names no EU legal entity and no native EU region for Anthropic’s own apps and API.^[5] If keeping data in Europe is mandatory for you, the route is therefore likely to be access through a hyperscaler; whether that is enough for your case, clarify with the provider and your data protection officer. The information can change, so check the page on the day of your decision.

If you use Claude through Microsoft Copilot the position is different: Microsoft currently excludes models from Anthropic as a subprocessor from the EU Data Boundary.^[10] See Microsoft Copilot and the GDPR.

Where does Claude data go, and how much of that can you control?

Once a prompt has been sent to Claude, you cannot observe where it is processed or stored. What you have is Anthropic’s documentation, and on commercial plans a contract. The documentation shows where the commitments are narrower than they sound. European residency is described through AWS Bedrock and Google Vertex, and the regional page names no native EU region for Anthropic’s own apps and API^[5]. That means a further company in the chain, which is permitted under Article 28 GDPR but is one more party you cannot inspect^[9]. On consumer plans retention depends on a choice each user makes: up to five years if training is allowed, and a thumbs rating stores the conversation for up to five years^[1][2]. On commercial plans Anthropic says it does not use inputs by default, with submitted feedback as the exception^[3].

None of this suggests that Anthropic does not keep its word. It marks where you rely on what is written rather than on what you can check. Switching a setting or deleting a chat acts on the future; what has been sent cannot be recalled. So the control you hold is the input itself: send only what the task needs and replace names, numbers and addresses before sending. See where your data goes when you use an AI assistant for the same question across providers.

Is a commercial plan enough for data protection at work?

It gives a firmer starting point, but it does not answer everything. Three things stay with you.

• Legal basis. The German supervisory authorities write that every processing step involving personal data in an AI application needs a legal basis.^[6] According to the Baden-Württemberg paper, provider and user each need their own.^[8]
• Information for data subjects. Where customer, applicant or employee data is involved, the use of the assistant and its recipients belong in your privacy notices.
• Professional secrecy. Where it applies, extra safeguards are needed; a processing agreement does not satisfy them by itself. In Germany, for example, this concerns lawyers and tax advisers under § 203 of the Criminal Code.^[11]

Data minimisation also applies. Even in a commercial plan, prompts do not need real names, account numbers or addresses if the task does not require them. How to pseudonymise prompts is described in anonymise data before ChatGPT; the method works the same way for Claude. Pseudonymised data remains personal data for you because you hold the key.

What should a company do now?

Work in this order: plan and contract first, then settings, then rules. The steps follow the supervisory authorities’ recommendations that employers provide work accounts and set accounts so that inputs are not used for training.^[6]

1. Choose a commercial plan. For personal data from your business use Team or Enterprise, or the API for your own applications, rather than Pro or Max on private accounts. Owner: management, IT.
2. File the contract. Save the Commercial Terms and the DPA in force, with the date, and check them against Article 28 GDPR. Owner: data protection.
3. Decide on feedback. If you do not want whole conversations to reach Anthropic through member feedback, switch off “Rate chats”.^[3] Owner: administrators in IT.
4. Rule out private accounts. The German authorities say that staff should not have to use private accounts and advise functional accounts without individual names.^[6] The Hamburg authority’s checklist likewise names business accounts without private use and without individual employees’ names.^[7] Owner: management.
5. Write the rules. What may go in, what may not, who checks results? A template is in the AI acceptable use policy template. Owner: management, data protection.
6. Train. Staff need examples of permitted and prohibited inputs; the Hamburg authority names training and checking of results as part of use.^[7] Owner: HR, data protection.
7. Document. Update the records of processing activities and privacy notices; where employee representatives exist, consider an agreement. Owner: data protection, management.

Which mistakes are most common in practice?

Most problems come not from the provider but from gaps between purchasing and daily work. Four patterns recur.

• The private subscription as a shortcut. A team lead books Pro on a personal card because the business plan is not yet approved. From the first customer email in a prompt, business personal data sits in a consumer account that is excluded from the DPA.^[4] Close the gap by providing the commercial plan quickly and writing the rule down.
• Training as a one-off decision. On consumer accounts the choice is a personal setting that can be changed. An employee’s assurance that “training is off” is therefore not evidence. On a commercial plan the risk shrinks, because Anthropic says training does not take place there by default.^[3]
• The thumbs-up button. Well meant, but on consumer accounts it stores the whole conversation.^[2] Where prompts contain business data, switching it off for members of Team and Enterprise is a simple measure.^[3]
• The contract nobody has read. Because the DPA applies automatically, it is easily overlooked. Review it once, record date and version, and decide who watches for future changes.

An example of a prompt where the type of account makes a difference: “Summarise the complaint from [CUSTOMER_1] of 3 October in three sentences and suggest a reply.” On a commercial plan the legal basis still has to be right; on a private account there is the further problem that no processor arrangement exists with the provider.

When does Claude stay only partly usable, even on a commercial plan?

When data needs special protection, such as health data, professionally privileged information or applicant data. There the plan is only one building block. You also need to check the legal basis, possibly carry out a data protection impact assessment (see DPIA for AI tools), and ask whether the details need to go into the tool at all. If a mistake has already happened, see what to do after pasting personal data into an AI chat; the same ChatGPT facts are in is ChatGPT GDPR compliant?.

A word on sources: the statements here come from Anthropic’s own pages and are provider statements, with no independent verification of the technical implementation. This text is general information, not legal advice. Read the linked pages on the day you decide, because provider terms change often.

This article is general information and not legal advice. It does not replace an assessment of your own case by your data protection officer or a lawyer. Last updated: 13 November 2026.