Private data Claude without risk

Configuring Private Knowledge Bases in Claude Enterprise Without Data Retention Risk

Item	Details
Purpose	Enable teams to build a secure, department-specific Claude knowledge base (Projects) while minimizing exposure of sensitive customer, financial, or HR data to unnecessary retention.
Issuing Authority	Configured by the organization's Claude Enterprise account Owner/Admin through Anthropic's Organization Settings; governed by Anthropic's data handling and retention policies.
Average Processing Time	Initial hardening and retention configuration typically takes 1–2 weeks; full knowledge base rollout across departments (Sales, HR, Finance) generally spans 2–4 weeks with a pilot phase.
Prerequisites	Claude Enterprise plan with Owner-level admin access, SSO/SCIM identity provider integration, a data classification policy, and audited source documents ready for upload.

What Is a Zero-Retention Knowledge Base and When Is It Mandatory?

A zero-retention configuration ensures that prompts and responses sent to Claude are processed and then deleted rather than stored for model training or indefinite logging. For Claude Enterprise, retention is admin-defined, with a minimum retention window of 30 days or an option to disable history entirely, unlike the indefinite default on lower-tier plans.

This configuration becomes mandatory, not optional, when a knowledge base will contain regulated or sensitive data such as customer information, financial records, health data, or proprietary methods. Organizations bound by HIPAA, PCI DSS, or similar frameworks should treat zero-retention as a compliance requirement rather than a preference, since Anthropic separately offers zero-data-retention (ZDR) and HIPAA-eligible API access for exactly this use case.

Step-by-Step Configuration: Admin Settings vs. Department Rollout

Organization-level (Admin) configuration

• Enforce SSO through your identity provider and enable SCIM for automated user provisioning before any documents are uploaded.
• Restrict allowed email domains to your corporate domain(s) only to prevent unauthorized external accounts from joining workspaces.
• Navigate to Organization Settings > Data and Privacy and set a custom retention period; note that the minimum is 30 days, and retention timers reset from the last message in a chat or the last update to a Project's knowledge base.
• Disable public Projects, disable chat sharing through connectors, and disable location metadata organization-wide to limit unnecessary data exposure.
• Set audit log retention to permanent and configure a defined connector approval process so every data source feeding the knowledge base is reviewed and justified.

Department-level (Knowledge Base) rollout

• Audit existing document repositories (Drive, Notion, Confluence, OneDrive) and flag which files are current versus outdated before uploading anything.
• Classify each document as Public, Internal, or Sensitive before upload, since only Public and reviewed Internal content is appropriate for standard Projects, while Sensitive data requires the Enterprise zero-retention path.
• Standardize file formats (Markdown or text-layer PDF) and use a clear naming convention, such as department-document type-version, to improve retrieval accuracy.
• Create separate Projects per department, such as Sales-KB, Finance-KB, and HR-KB, rather than a single combined knowledge base, to keep retrieval scoped and context windows manageable.
• Write detailed custom instructions for each Project covering Claude's role, output format, and how to handle missing information, and pilot the configuration with a small group of users before a full rollout.

3 Common Configuration Failures and How to Fix Them

1. Uploading sensitive data into standard Projects instead of zero-retention mode

Teams often default to Claude Projects for convenience without checking data classification, exposing customer or financial records to longer retention than necessary. Fix this by requiring every uploader to answer a mandatory data-classification question before upload, routing anything marked "Sensitive" to the Enterprise zero-retention configuration instead.

2. Leaving retention settings at default instead of minimizing them

Organizations frequently skip the Data and Privacy settings entirely, leaving retention at longer defaults than their compliance posture requires. Fix this by explicitly setting the retention period to the practical minimum under Organization Settings and documenting the change with an approval artifact and review date.

3. Failing to review connectors and access scope over time

A knowledge base configured securely at launch can drift as new connectors, integrations, or admin accounts are added without re-review. Fix this by establishing a recurring review cycle that revalidates active connectors, role-based access mappings, and admin account access on a scheduled basis.

Best practice: Treat data classification as a gate, not a suggestion. Before any document enters a Claude Project, require an explicit tag — Public, Internal, or Sensitive — and route Sensitive material exclusively through Claude Enterprise's zero-retention configuration. This single control point prevents the most common enterprise mistake: sensitive data landing in a standard workspace simply because it was faster to upload.

Frequently Asked Questions

Does Claude Enterprise train its models on our uploaded documents?

No. Anthropic does not use enterprise prompts, responses, or uploaded knowledge base content for model training unless the organization explicitly opts in, and Enterprise plans give admins direct control over retention periods.

What is the shortest retention period we can configure for our Claude Enterprise workspace?

The minimum configurable retention period is 30 days, with retention timers resetting based on the last chat message or the last modification to a Project's knowledge base.

Should HR and Finance departments share one knowledge base or separate ones?

Separate Projects per department, such as HR-KB and Finance-KB, are recommended because they keep context windows focused and prevent unrelated teams from retrieving data outside their function, improving both accuracy and access control.