Enterprise Deployment Controls¶

Level: Intermediate · Time: ~40 min · Exam relevance: CCAR-P D5 (5.4), CCAR-P D7 (7.1), CCAO-F D6 (6.2, 6.3)

In plain English¶

Everything earlier in this module is about controls you build into your own application. This page is about the controls Anthropic's products give administrators: who can sign in, how accounts are provisioned and removed, how long data is kept, where inference runs, which networks can reach the service, and which settings users can or cannot change.

The aim is to separate what is actually offered from what people assume is offered. Every feature listed here was checked against Anthropic's official documentation on 2026-10-24. Availability depends on plan and platform and changes over time, so confirm against the linked pages and your contract before you design around a control.

Why it matters¶

Security and procurement reviews ask for these controls by name: SSO, SCIM, audit logs, retention, encryption keys, residency, IP restrictions.
Choosing the wrong surface (for example a consumer plan for regulated data) can make an otherwise good design non-compliant.
CCAR-P-7.1 covers configuring Claude tools for teams, and CCAR-P-5.4 covers regulatory compliance.

How it works¶

Map controls to surfaces¶

Surface	Who administers	Main control points
Claude apps on Team and Enterprise plans (claude.ai, desktop, mobile)	Organisation admins	Identity (SSO, JIT, SCIM), roles, retention, audit logs, Compliance API, connectors, network restrictions
Claude API (Claude Console)	Organisation and workspace admins	Workspaces, API keys, spend and rate limits, data-retention arrangements, Admin API, workload identity
Claude Code	Admins via managed settings, plus Console or plan admin	Managed settings and permissions, managed MCP configuration, sandboxing, monitoring
Claude on cloud platforms (Amazon Bedrock, Google Cloud, Microsoft Foundry) and Claude Platform on AWS	Your cloud account administrators	The cloud's IAM, networking, logging and regional controls; data processing terms depend on the platform

Claude Enterprise plan controls¶

Anthropic's help centre article What is the Enterprise plan? and the Enterprise security and compliance collection list these administrative controls (as of 2026-10-24):

Control	What it gives you
SSO, just-in-time provisioning and SCIM	Central authentication through your identity provider; automated provisioning and de-provisioning (documented in the help centre's identity management collection; confirm plan availability)
Audit logs	Records of user actions, system events and data access
Custom data retention controls	Organisation-defined retention for chats, files and projects
Compliance API	Programmatic access to activity logs, chat histories and file content, filterable by user and time range, for compliance tooling
Analytics API	Aggregated engagement and adoption metrics
Customer-managed encryption keys (CMEK)	Use a key in your own cloud provider to protect chats, projects and files
US-only inference	Keep the organisation's inference in the United States
Spend limits	Organisation and per-user limits
HIPAA-ready configuration	Eligible organisations can enable it and accept a business associate agreement
IP allowlisting	Restrict access to Claude to approved IP ranges
Tenant Restrictions	Network-level control to stop users signing in to organisations other than yours from your network
Connectors administration	Control workplace connectors such as Google Drive, Gmail, GitHub, Microsoft 365 and Slack

The last three items come from article titles in the Enterprise security and compliance collection ("Restrict access to Claude with IP allowlisting", "Enforce network-level access control with Tenant Restrictions"); read those articles for configuration details and plan requirements.

Claude API controls¶

Control	Details (verified 2026-10-24)
Workspaces	Separate API keys, members, and spend and rate limits per workspace; organisation-wide limits still apply
Spend limits	Per-tier monthly caps, plus your own lower limits for the organisation and for workspaces
Rate limits	RPM, input and output tokens per minute per model class; readable via the Rate Limits API
Admin API	Manage members, invites, workspaces, API keys, service accounts and federation rules programmatically with an admin credential
Workload identity federation	Exchange your platform's identity token for short-lived credentials instead of long-lived API keys
Inference geography	The `inference_geo` request parameter (supported on Opus 4.6 / Sonnet 4.6 and later) controls where inference runs, with `"us"` or `"global"` as the only values as of 2026-10-24; `usage.inference_geo` reports it. For other regions, use a cloud platform's regional endpoint
Usage and cost reporting	Usage and Cost Admin API; Console usage export by key and model
OpenTelemetry	SDK and API tracing (beta) to your backend, metadata-only by default

Data retention, ZDR and HIPAA on the API¶

Anthropic's API and data retention page (checked 2026-10-24) states:

Only what is technically necessary for each feature is retained, and conversation content (prompts and outputs) is not retained by default. The exception is Covered Models, which require 30-day retention. Standard commercial retention policies outside the ZDR and HIPAA arrangements are documented separately on Anthropic's privacy pages. Claude Fable 5.1, Claude Mythos 5.1, Claude Fable 5 and Claude Mythos 5 are listed as requiring 30-day retention and are not available under ZDR unless expressly authorised.
Zero data retention (ZDR) is arranged with Anthropic's sales team and enabled per organisation. Under ZDR, Anthropic does not store prompts or responses at rest after the response is returned, for eligible features of the Messages and Token Counting APIs listed in the page's feature-eligibility table. Some features that use /v1/messages, such as code execution, are not covered.
ZDR applies to Claude Code used with API keys from a commercial organisation (metrics logging, if enabled, is exempt). Claude Code on Claude Enterprise has a separate ZDR offering for qualified accounts.
ZDR does not cover the Claude Console (including the playground), Claude Managed Agents (session transcripts persist until deleted), consumer plans, the Team and Enterprise product interfaces, third-party integrations, or flagged content and legal holds. CORS (direct browser calls) is not supported for ZDR organisations.
HIPAA readiness for the API is available with a signed BAA and a HIPAA-enabled organisation, for eligible features.
Who processes your data depends on the route. On Amazon Bedrock, and for Claude models on Google Cloud that Google manages, the cloud provider is the processor, so its retention and compliance terms apply. On Google Cloud offerings that Anthropic manages, on Claude Platform on AWS and on Microsoft Foundry, Anthropic is the processor (checked 2026-10-24 against Anthropic's API and data retention page).

Design implication: build a feature-by-feature data-flow table. A pipeline that is ZDR-eligible for plain Messages calls may lose that property the moment you add a stateful or file-based feature.

Claude Code organisational controls¶

Per Claude Code security (checked 2026-10-24):

Managed settings enforce organisational standards and permission rules that users cannot override.
Permission modes: organisations can turn auto mode off; explicit allow and deny rules apply in every mode.
Managed MCP configuration restricts which MCP servers can run.
Sandboxing isolates bash commands at the filesystem and network level; the whole process can run in a sandbox runtime or dev container.
Monitoring via OpenTelemetry metrics, and ConfigChange hooks to audit or block settings changes mid-session.
Cloud sessions run in isolated VMs with limited network access by default, server-side GitHub credentials, push restrictions and audit logging.

See Claude Code team rollout for a rollout plan.

Network controls you own¶

Route all API traffic from servers through an egress proxy with an allow-list for Anthropic endpoints, so you can monitor volume and block unapproved destinations.
Never call the API from browsers or mobile apps with embedded keys; use a backend.
On cloud platforms, use the cloud's private networking and IAM features as documented by that provider.
Combine IP allowlisting (for the Claude apps) with your corporate network controls and Tenant Restrictions to stop data going to personal accounts.

Example¶

A deployment-readiness check that turns this page into a repeatable gate. Fill the YAML-free dictionary from your configuration management database and fail the pipeline on gaps.

"""enterprise_controls_check.py - check a deployment's control posture against its data class."""
CONTROL_REQUIREMENTS = {
    "public":   {"sso"},
    "internal": {"sso", "scim", "audit_logs", "workspace_isolation", "spend_limits"},
    "personal": {"sso", "scim", "audit_logs", "workspace_isolation", "spend_limits",
                 "retention_policy_set", "egress_proxy", "no_client_side_keys"},
    "phi":      {"sso", "scim", "audit_logs", "workspace_isolation", "spend_limits",
                 "retention_policy_set", "egress_proxy", "no_client_side_keys",
                 "hipaa_ready_org", "baa_signed", "features_checked_against_eligibility_table"},
}

def check(deployment: dict) -> list[str]:
    required = CONTROL_REQUIREMENTS[deployment["highest_data_class"]]
    missing = sorted(required - set(deployment["controls"]))
    problems = [f"missing control: {c}" for c in missing]
    if deployment.get("requires_zdr") and deployment.get("uses_stateful_features"):
        problems.append("ZDR required but deployment uses features outside ZDR scope; review eligibility table")
    return problems

if __name__ == "__main__":
    print(check({
        "name": "clinical-notes-summariser",
        "highest_data_class": "phi",
        "controls": ["sso", "scim", "audit_logs", "workspace_isolation", "spend_limits",
                     "retention_policy_set", "egress_proxy", "no_client_side_keys", "hipaa_ready_org"],
        "requires_zdr": True,
        "uses_stateful_features": True,
    }))

Enterprise use case¶

Contoso Health deploys Claude in three ways:

Claude Enterprise for 4,000 staff: SSO with SCIM from its identity provider, IP allowlisting to corporate egress addresses, Tenant Restrictions on the corporate proxy, custom retention of 90 days, audit logs and the Compliance API feeding the security information and event management (SIEM) platform, HIPAA-ready configuration with a BAA, and connectors limited to the corporate Microsoft 365 tenant.
Claude API for a clinical-notes summariser: a dedicated workspace with its own spend and rate limits, HIPAA readiness for eligible features, workload identity federation instead of static keys, an egress proxy, and a design review that confirmed every feature used is in the eligibility table.
Claude Code for 120 engineers: managed settings with deny rules for secret files, sandboxing on, managed MCP allow-list, auto mode disabled for repositories with production access, and OpenTelemetry metrics to the engineering observability stack.

Common mistakes¶

Assuming "enterprise" means every feature is ZDR-eligible.
Letting staff use personal consumer accounts for work data because no network control blocks them.
Using one Console workspace and API key for every application.
Enabling connectors organisation-wide without reviewing what data they expose.
Designing around a feature without checking plan, platform and regional availability.

Security considerations¶

Retention settings and ZDR do not remove your own obligations for logs and copies you keep.
Review admin roles regularly: Console organisation admins and Claude Enterprise owners hold powerful rights.
Treat Compliance API access as highly privileged; it can read conversation content.
When using cloud platforms, the cloud provider's controls and terms apply; validate feature parity for the controls you need.

Troubleshooting¶

Symptom	Likely cause	Fix
Browser app calls fail after ZDR enablement	CORS not supported for ZDR organisations	Route calls through your backend
A model returns a 400 for a ZDR organisation	The model requires 30-day retention (Covered Model)	Use a model available under your arrangement, or discuss with Anthropic
De-provisioned staff still have access	SCIM not configured or not syncing	Configure SCIM and monitor sync errors
Usage appears in the wrong cost centre	Shared keys across teams	One workspace and key per application or team

Check your understanding¶

Name four administrative controls available on the Claude Enterprise plan.
Which Anthropic surfaces does the API ZDR arrangement explicitly not cover?
Why must browser-based applications in a ZDR organisation call Claude through a backend?
Which Claude Code feature lets an organisation stop users from overriding permission rules?

Answers

Any four of: SSO/JIT/SCIM, audit logs, custom data retention, Compliance API, Analytics API, CMEK, US-only inference, spend limits, HIPAA-ready configuration, IP allowlisting, Tenant Restrictions.
The Claude Console (including the playground), Claude Managed Agents, consumer plans, the Team and Enterprise product interfaces (except the separate Claude Code on Enterprise offering), third-party integrations, and flagged content or legal holds.
CORS is not supported for ZDR organisations, and in any case API keys must never be exposed client-side.
Managed settings.