Does Claude train on your data? Training, retention, zero data retention and where data lives

No, not on business plans by default. For Claude, powered by Anthropic, the Commercial Terms state that Anthropic may not train models on customer content from Team, Enterprise or API use. There is one practical exception: if a user submits feedback with the thumbs up or down button, that conversation can be used for training, and admins can switch that button off. Consumer accounts (Free, Pro, Max) work differently: training there is an opt-out setting.

Which Claude plans use your data for training?

The answer depends on which terms govern the account, not on the price. Team, Enterprise and API accounts sit under the Commercial Terms of Service, where Anthropic acts as a processor and your company is the controller of the data your users submit. Free, Pro and Max accounts sit under the Consumer Terms, even when an employee pays for Pro with a company card. That distinction is the single most important fact in this article.

The Commercial Terms also settle ownership. Your company keeps its rights to inputs and owns the outputs, and Anthropic assigns to you whatever rights it has in those outputs. Anthropic says it processes Team and Enterprise data only on your instructions to provide the service. A Data Processing Addendum, including Standard Contractual Clauses, is part of the Commercial Terms automatically when you accept them.

What is the feedback exception, and how do you close it?

When a user clicks thumbs up or thumbs down and submits a report, Anthropic stores the whole related conversation for up to five years. That includes content, custom styles, conversation preferences and model settings. Anthropic says it de-links feedback from user and customer IDs before use and does not combine it with your other conversations. It may then use that feedback for research and to train models. Raw content pulled through connectors such as Google Drive or MCP servers is not included unless someone pasted it into the chat.

If your policy says no company data leaves for any training purpose, close the gap in settings rather than in a training slide. A Primary Owner or Owner on Team or Enterprise can turn off the Rate chats setting under Organization settings, then Data and Privacy. With it off, members cannot submit feedback through the thumbs buttons. On HIPAA-ready Enterprise organizations with the HIPAA configuration applied, feedback, bug reports and ratings are already unavailable in Claude Code and Cowork.

How long does Anthropic keep your data?

Not training on data is different from not storing it. Claude keeps chat history so people can return to their work, and Anthropic keeps some data longer for safety and legal reasons. These are the periods Anthropic publishes for commercial accounts.

Custom retention on Enterprise needs care. Owners set it under Organization settings, then Data and Privacy. Shortening the period schedules everything older for permanent deletion as soon as you save, and the data cannot be recovered. Retention settings do not reach standalone Claude Design, Claude Tag, Claude Managed Agents or Claude Code running in the cloud. Changes to retention and the resulting deletions are recorded in the audit log, which helps when legal asks you to prove a policy was in force.

What is zero data retention, and who can get it?

Zero data retention (ZDR) is an arrangement, approved by Anthropic per organization, under which Anthropic does not store inputs or outputs except where the law or misuse prevention requires it. Anthropic still keeps safety classifier results. ZDR applies only to eligible Anthropic APIs, products that use your commercial API key (including Claude Code through the API), and Claude Code on Enterprise plans. It does not apply to chat in the Claude apps. If you run several organizations, each one needs its own approval. API customers can confirm ZDR under Settings, then Privacy Controls, then Data retention period.

Where does Claude process and store your data?

Anthropic says customer data is stored in the US. By default, it may route traffic for processing to select countries in the US, Europe, Asia and Australia, unless you agree otherwise or instruct it to. Anthropic may also process data for internal work such as safety review, support or incident response in countries where it or its affiliates operate. Its subprocessor list names the cloud providers involved.

Usage-based Enterprise plans have a US-only inference setting under Organization settings, then Data and privacy. When it is on, all model processing for the organization runs on US servers, including background work such as conversation titles and memory. Usage on Claude Opus 4.6, Claude Sonnet 4.6 and later models is then billed at 1.1 times standard API rates. The setting does not control where connected services such as Slack or Google Drive process data. It also does not change where your stored data sits. API customers have separate data residency options in the developer documentation.

If you reach Claude through Amazon Bedrock, Google Cloud or Microsoft Foundry, Anthropic says the platform's terms govern that use. For Covered Models on Bedrock and Google Cloud, Anthropic says retained data stays in that provider's environment. That route suits companies that want AI spend and data controls inside an existing cloud agreement.

What should an admin configure in the first week?

• Decide whether thumbs up or down feedback is allowed, and set Rate chats to match.
• On Enterprise, set a custom retention period that matches your records schedule. Remember projects default to indefinite.
• If you have US data-location requirements, turn on US-only inference and budget for the 1.1x usage rate.
• If you need customer-managed encryption keys, provision a key in your own cloud account. Anthropic lists this as an Enterprise feature.
• Turn on audit logs and the Compliance API so security can see activity and export chats for legal hold.
• Review which connectors are allowed. Data sent to Slack, Google or another service is processed under that service's terms.
• On Enterprise, plan domain claiming: once SSO is enforced and JIT or SCIM provisioning is on, admins can migrate personal Free, Pro and Max accounts on the company domain into the organization within a 30-day window. Turning on domain capture cannot be reversed.

Where do companies get this wrong?

• Tolerating personal Pro accounts. An employee paying for Pro is under consumer terms, where training is a user setting your company does not control. Team plans can block new personal accounts on a verified domain but cannot claim existing ones; Enterprise can.
• Treating deletion as retroactive. Deleting a chat does not pull back a feedback report already submitted, which is kept for five years.
• Assuming ZDR covers everything. It does not cover chat in the Claude apps, and it blocks Covered Models.
• Shortening retention without checking legal holds. Shortened periods delete older data permanently, and a daily job removes it within days.
• Forgetting connector data flows. US-only inference and the Anthropic DPA do not change how a third-party service handles data Claude sends it.