Claude Cowork data residency and retention

TL;DR. Anthropic processes Cowork inference in US and EU regions; Enterprise customers can negotiate region pinning. Default retention is around 5 years on Pro and Max with a 30-day opt-out; Team and Enterprise are organisation-controlled with no training as the default and zero-retention available contractually. Customers whose compliance program excludes Anthropic-hosted inference entirely can deploy Cowork on Bedrock or Vertex AI — conversation data stays in their AWS or GCP tenant. Microsoft Foundry is not a residency solution; that route still terminates inference at Anthropic infrastructure.

The three questions buyers ask#

• Where is the data processed? US and EU on first-party. On Bedrock or Vertex AI, in the customer's chosen cloud region. On Foundry, still on Anthropic's infrastructure.
• How long is it kept? Plan-dependent on first-party. On Bedrock / Vertex AI, the cloud provider's posture applies — Anthropic does not store conversation data.
• Is it used to train the model? Default yes on Pro/Max with a per-user opt-out; default no on Team and Enterprise. On Bedrock / Vertex AI, structurally not applicable.

The rest of this page backs those three answers up.

Region and processing#

• Anthropic's primary first-party inference regions are US and EU. Verify against the current Trust Center page before you put it in a procurement document.
• Enterprise customers can negotiate region pinning at contract. This is the right move for any client with strict EU residency requirements.
• The desktop app processes some content locally before sending excerpts to inference — but assume the relevant content reaches Anthropic's region.
• For UK / EU residency-sensitive customers, the Enterprise plan is the standard answer. Pro and Max are not the right floor for tight residency requirements.
• For customers whose compliance program excludes Anthropic-hosted inference entirely, Cowork can run on a third-party inference provider — Amazon Bedrock or Google Cloud Vertex AI — keeping conversation data inside the customer's own cloud tenant. See Cowork on third-party platforms. Microsoft Foundry is not a residency solution; that route still terminates inference at Anthropic infrastructure — see the Foundry page for the gotcha.

Retention by plan#

Numbers verified against Anthropic's data privacy controls page in late April 2026. Re-verify on each monthly review of this page.

Training opt-out and how it actually works#

• The opt-out toggle prevents your prompts and outputs from being used to improve future Claude models.
• Opt-out applies forward. Data already used for training is not retroactively removed.
• For Team and Enterprise, the default is "do not train." No per-user toggle needed.
• Tinkso recommends enabling opt-out at the start of any Pro / Max pilot before any real data goes in. This is a one-click action that survives every subsequent procurement question.

Memory data#

• Memory entries are stored under the user's account.
• Same retention and training posture as the rest of the conversation data.
• Deletion is per-entry via the Memory UI. There is no bulk-export-and-delete tooling at GA, which is the one Memory limitation worth flagging in a security review.

MCP connector data#

• Each connector handles its own data flow according to its vendor's terms.
• Cowork does not store the connector's responses outside the conversation context unless told to.
• If a connector returns regulated data — PII, PHI, financial — the data is in the prompt; treat the inference call as in scope for that data class.

This is where most regulated-deployment design happens: which connectors return what data, and which inference calls accordingly inherit which compliance posture.

Cross-border considerations#

• US / UK mid-market customers should review the EU vs US processing question even if all employees are stateside. Supplier and customer data may pull GDPR scope into play.
• Enterprise customers with strict residency can negotiate region pinning at contract.
• Tinkso engagements include a 30-minute residency check during pre-flight for any client in regulated industries — see Pre-flight checklist.
• For customers who genuinely cannot send data to Anthropic at all, the 3P routes (Bedrock or Vertex AI) are the correct answer rather than a contractual workaround.

Tinkso's take#

Retention defaults trip up more procurement reviews than any other Cowork question. We brief every client to enable opt-out on day one of a Pro / Max pilot, and to upgrade to Team or Enterprise before scale. The cost delta is small; the security-review delta is enormous.

The other recurring trip-up: clients assume "do not train" and "zero retention" are the same thing. They are not. Zero retention is a contractual posture only available on Enterprise; do-not-train is widely available. Read your contract. And if the conversation drifts to "what if we go on Microsoft Foundry?", redirect immediately — Foundry is the right answer for Azure-led billing, not for residency.

Try this#

In account settings, find the Data Privacy Controls page. Turn on the "do not train" opt-out for every pilot user before they touch a real document. Take a screenshot. That screenshot is the artifact your security partner will ask for in week one.