Rolling Out Claude Safely: Plugins, Skills and Governance for Enterprises

One of the first questions we hear from our customers once Claude is running productively in their team: “Who actually controls what the AI does?” Shortly after, IT asks: “How do we make sure no confidential data leaks out?”

These are exactly the right questions. For finance teams, governance isn’t a feature you bolt on later – it’s the prerequisite for AI to enter production at all. The good news: Claude Enterprise was built for these requirements from the ground up.

Which Subscription Do You Need?

Claude comes in several tiers – but for governance features, you need at least the Team or Enterprise plan:

The Admin Panel: Your Control Center

Since February 2026, the “Customize” system in Claude Cowork provides a dedicated admin layer with four core functions:

Data Privacy: Four Layers of Protection

Data privacy determines whether your compliance team will approve AI usage at all. Claude Enterprise offers four layers:

Cost Management: Full Transparency at Fixed Costs

Claude Enterprise works with a fixed price per user. Usage is still tracked in detail – so you can see exactly which teams and plugins deliver the most value.

Shadow AI: The Post-it Problem

You know the story of the password on a Post-it stuck to the monitor? When password policies get too complicated – 16 characters, special symbols, monthly rotation – employees just stick the password to their screen. The security problem doesn’t come from bad technology, but from rules that miss everyday reality.

With AI, we see exactly the same thing. If your official Claude setup is too restrictive or too cumbersome, employees will use their own tools – uncontrolled, without logging, without data protection. That’s Shadow AI, and it’s the biggest governance risk of all.

The best countermeasure: Make the official path the easiest path. Well-configured plugins that deliver real value drastically reduce the incentive for workarounds.

Real-World Examples

These governance features are already being used productively in regulated industries:

• Allianz (January 2026): Claude for all employees, AI agents for claims processing and document intake – with full logging of all AI interactions
• PwC (February 2026): The “AI Native Finance” initiative brings Claude into regulated environments where auditability is mandatory
• Intapp (February 2026): Claude at audit firms and investment banking firms – with Ethical Walls and Conflicts Management as an integrated governance layer
• Thomson Reuters (March 2026): Integration partner for AI-powered research tasks in regulated areas
• Infosys (February 2026): AI agents for telecommunications and financial services – including compliance reporting

Rollout in 5 Steps

The following section is intended for IT teams and admins who are technically setting up and securing Claude Enterprise. Feel free to forward this part to your IT department.

Compliance API: Audit-Ready Logging

The Compliance API provides programmatic access to everything auditors want to see: chat histories, file contents, tool calls – filterable by user and time range. Audit logs are retained for 30 days and can be exported as JSON/CSV or pushed directly to SIEM systems like Splunk, Datadog or Elastic.

Core features:

• Complete activity logs of all AI interactions
• Automated flagging of potential policy violations
• Selective deletion for targeted data retention
• Direct SIEM integration for existing compliance dashboards

Managing MCP Servers Securely

The Model Context Protocol (MCP) is the interface through which Claude accesses external data sources. There are currently over 12 official connections (Google Calendar, Drive, Gmail, DocuSign, FactSet, Salesforce and more). The key security levers:

• Central Configuration: Via a managed-mcp.json, admins deploy a fixed set of connections that users cannot modify
• Allowlist/Denylist: Precisely define which MCP servers are permitted
• OAuth 2.1 with PKCE: Standard authentication for all MCP connections (no static API keys)
• Least Privilege: Each server only receives the permissions needed for its function
• Versioning: Make changes traceable, roll back when issues arise
• MCP Gateway: For larger deployments, a central control point for authentication, rate limiting and audit logging

Skill and Plugin Versioning

Skills and plugins evolve over time. Without versioning, you lose track of which version is running in production. Recommended: Semantic Versioning (MAJOR.MINOR.PATCH) with Git tags and staged rollouts (test group first, then everyone).

Regulatory Context

The EU AI Act becomes enforceable for high-risk AI from August 2026. BaFin published guidance on ICT risks related to AI in late 2025. FINRA dedicates a GenAI section for the first time in its 2026 Oversight Report. Document your configuration carefully – it will be relevant at the next audit.

Conclusion

AI governance is not a one-time project but an ongoing process. Claude Enterprise provides the tools – from the Compliance API to plugin management to real-time cost transparency. Start with the 5-step plan, work through the checklist and establish a monthly review cycle. This way you build a framework that grows with your requirements.

Sources