Security & Governance

Built for firms that take their fiduciary duty seriously.

Implementing AI inside a firm handling HNW client data is a high-stakes decision. This page sets out where Anthropic's controls do meaningful work, where the firm's IT and compliance functions retain ownership, and what we will and won't ask for.

01 Commercial Terms

Claude Enterprise does not train on your prompts, files, or outputs.

This is a contractual commitment under Anthropic's Commercial Terms — not a toggleable setting. Customer Content stays out of model training, full stop. Data is encrypted in transit (TLS 1.2+) and at rest (AES-256).

SOC 2 Type II

Independent audit of controls over time — not just on paper.

ISO/IEC 27001:2022

Information Security Management Systems.

ISO/IEC 42001:2023

AI Management Systems — the international standard for AI governance.

Verify at trust.anthropic.com. Detailed SOC 2 reports available under NDA on request.


02 Enterprise Controls

The controls IT and Compliance need to do their jobs.

Audit Logs

Continuous record of user actions, project activity, data access, file uploads, authentication. Exportable as CSV by Org Owners.

Compliance API

Programmatic access to activity logs, chat histories, file content by user and time range. Forensic retrieval on demand.

Custom Data Retention

Configurable per chat and per project. Zero Data Retention available via contractual addendum for the most sensitive workloads.

SAML SSO & SCIM

Centralised identity. Automated provisioning and de-provisioning when staff leave or change roles.

IP Allowlisting + RBAC

Network-level access restrictions and role-based access controls.


03 The Four Locks

Claude does not ingest your documents. It queries them on demand.

Files remain inside your firm's tenant. The integration is governed by four mechanisms, each controlled by your IT function:

Delegated Permissions

Claude inherits each user's existing M365 permissions. If a junior administrator cannot see a folder today, Claude cannot either.

Read-Only

The connector cannot edit, delete, send, or create. All outbound actions remain human.

Per-User Authorisation Gate

In Microsoft Entra ID, IT sets Assignment required = Yes. Only users IT assigns can authenticate the connector.

Selective Scope Revocation

IT can switch off any permission category — SharePoint, Outlook, Teams, OneDrive — or restrict to specific sites.


04 Residual Risk

No control set replaces firm discipline.

Anthropic's controls are real and well-designed, but they don't substitute for ongoing operational discipline. We name the residual risks here so the firm, IT, and operations all see them clearly.

Adversarial Content

If a CDD pack or piece of correspondence contains content engineered to redirect Claude's behaviour, human review of every consequential output is the mitigation.

The Discipline of Review

Every control assumes humans actually read AI output before signing or sending. If review becomes rubber-stamping, the controls become theatre.

Scope Drift

Permissions set on day one are not the permissions you will have in year three unless someone owns reviewing them. Quarterly access review by IT is the answer.

Plausible Wrongness

Claude is usually right — which is what makes the wrong moments dangerous. Train staff to verify rather than trust, especially on figures, names, dates, and jurisdictional detail.


05 Who Owns What

Your firm retains every meaningful access lever.

All configurations, projects, prompts, and rule sets reside in your firm's own tenant. If the engagement ends, the firm retains everything. We take nothing with us.

Domain Owner
Platform compliance, security architecture, Entra ID scoping, audit log review Your IT function
Retention policy, user provisioning approvals, connector enablement Your firm's owners
Regulatory, signatory, banking, and compliance decisions on client matters Your Compliance & Directors
Build quality, automation logic, prompt and project configuration, staff training ctrlroom.ai
Model behaviour, platform certifications, contractual data handling Anthropic

06 Out of Scope by Design

The shorter this list is, the more comfortable everyone gets.

No Write Access

Deployments are read-only by design. Outbound actions remain human.

No Banking Permissions

No signatory authority. No payment instructions. No call-back verification.

No Admin of Your Tenant

Your IT function holds every admin role. We operate as a configured user under your governance.


07 Due Diligence Starting Points

What to ask for before work begins.

A serious vendor risk assessment is the right first step, not a checkbox cleared once work has started. We recommend your committee and IT function request the following from Anthropic before the build queue begins:

SOC 2 Type II Report

Available under standard NDA via trust.anthropic.com.

Data Processing Addendum

The DPA covers SCCs for international transfers and establishes the customer as data controller.

Zero Data Retention

Available via contractual addendum for the most sensitive workloads.

We can request expedited engagement from Anthropic's compliance team to support your due diligence, not to bypass it.


Questions for IT or Compliance before you go further?