Phase 1: Approve Tool Matrix, Identify Enterprise, Choose the Owner

• Approved Tool Matrix: Define who can use what, for which use case, under which exception path.
• Enterprise Identity: Require SSO, Conditional Access, MFA, approved groups, and fast offboarding for every AI surface.
• Named Owner: Assign one accountable internal owner for integrations, exceptions, and escalation.
• Logging Baseline: Establish who used which tool, against what data, and when. Document any platform logging gaps.
• Shadow AI Cleanup: Move users out of personal accounts and into sanctioned, managed tenants.
• Claude Hardening: Disable Cowork, Desktop, Chrome, Code, connectors, plugins, skills, and write actions by default. Enable only approved capabilities and configure retention, sharing, feedback, web, network, and logging settings.

Phase 2: Decide the Plan, data Boundary and Integrations

• Plan: Choose Claude Business or Enterprise intentionally based on identity, provisioning, retention, compliance, audit, and group- control needs. Ask Atlas for the separate comparison sheet.
• Entry Points: Approve each Claude surface separately. Cowork now runs on web and mobile as well as desktop. Desktop is not required and remains disabled under Tier 0 unless approved.
• Data Scope: Name the exact libraries, mailboxes, applications, and data classes Claude may access. Review permissions and external sharing before connecting content.
• Read vs. Write: Treat read-only and write-capable access as different risk classes. Allowlist connectors, MCP servers, plugins, and skills; review scopes and data flows before approval.

Phase 3:  Pilot, Monitor and Expand

• Plan: Choose Claude Business or Enterprise intentionally based on identity, provisioning, retention, compliance, audit, and group- control needs. Ask Atlas for the separate comparison sheet.
• Entry Points: Approve each Claude surface separately. Cowork now runs on web and mobile as well as desktop. Desktop is not required and remains disabled under Tier 0 unless approved.
• Data Scope: Name the exact libraries, mailboxes, applications, and data classes Claude may access. Review permissions and external sharing before connecting content.
• Read vs. Write: Treat read-only and write-capable access as different risk classes. Allowlist connectors, MCP servers, plugins, and skills; review scopes and data flows before approval.

Phase 4: Decide the Plan, data Boundary and Integrations

• Pilot: Start with named users, curated data, low-risk workflows, and human approval. Exclude regulated, investment, legal, and client-facing decisions.
• Cowork: Recommend only enabling on sandboxed computers. If approved on non-sandboxed computers, recommend restricting to designated group of users, require per-task approval for write actions, configure network access controls, and minimize integrations. Insight into and control over CoWork is limited; see Atlas CoWork Security Brief.
• Monitor: Review usage, exceptions, incidents, support requests, connectors, and business value. Treat new models, features, connectors, and retention changes as change-control events.
• Expand: Add users, data, or capabilities only after the named owner approves the prior stage.
  

What Atlas Clients Need to Do?

• Choose Claude Business or Enterprise intentionally. Ask Atlas for the separate comparison sheet before purchasing. Complete Atlas Tier 0 before use.
• Grant Atlas the required admin privileges in Claude and the identity provider, so we can monitor configuration and make authorized client-requested changes.
• Provide the named owner, approved users, use cases, data sources, exception paths, and risk approvals. Keep Atlas involved as Claude capabilities and defaults change.