Enterprise Capstone: Design a Regulated Azure Landing Zone and AI Workload

This capstone designs a regulated enterprise platform and onboards Northstar as a private AI workload. The deliverable is an evidence-backed architecture, code plan, threat model, policy set, operations model, cost forecast, and tested proof of concept. Do not deploy enterprise-wide controls into a live tenant without the required authority and change process.

Method note: The capstone is an architecture and approval exercise spanning many services, not an interchangeable Portal-versus-code procedure. Use the implementation alternatives in the governance, private networking, recovery, and AI service chapters to build the authorized proof of concept.

Scenario

A global organization requires:

• Production and non-production subscriptions.
• Primary and recovery Azure regions with stated residency.
• ExpressRoute primary and VPN backup.
• No public access to storage, databases, search, model endpoints, or management APIs beyond approved edge services.
• Entra identity, PIM, workload federation/managed identity, and central emergency access.
• Central Defender/Sentinel and controlled logs.
• Governed RAG with synthetic regulated-document patterns.
• RTO 4 hours and RPO 30 minutes for the critical journey.
• Per-workload budgets and chargeback.

What architecture must you produce?

1. Management-group/subscription hierarchy.
2. Identity/admin/support access model.
3. Hub-spoke or Virtual WAN route and DNS diagrams.
4. Ingress/egress, Firewall, WAF, DDoS, and private endpoint design.
5. Management/monitoring/Sentinel/Defender data flows.
6. Workload compute, storage, database, AI Search, Foundry, and Key Vault.
7. Backup, region recovery, traffic, and data consistency design.
8. CI/CD/GitOps and artifact flow.
9. Data classification, lineage, retention, deletion, and AI evaluation.
10. Cost allocation and capacity/quota model.

What policy initiative should you design?

At minimum, audit or enforce by scoped rollout:

• Allowed regions and approved resource types/SKUs.
• Required workload/owner/data-classification tags.
• Public network access disabled for protected PaaS.
• Secure transfer/TLS baseline.
• Managed identity and approved authentication.
• Diagnostic settings to central destinations.
• Defender plans and vulnerability controls.
• Encryption/key requirements for the regulated class.
• Backup protection and private DNS patterns.

Include exemption owner, justification, compensating control, and expiry.

How do you prove the platform boundaries?

Run tests in a canary subscription:

• Workload deployer cannot change management-group policy or hub firewall.
• Platform network operator cannot read application data.
• Security responder can investigate without permanent workload Owner.
• App identity can reach only approved search/storage/vault/model endpoints.
• Unapproved public endpoint creation is denied.
• Hybrid DNS resolves private names; internet clients do not.
• Central logs detect a denied policy change and unusual secret access.

How do you prove recovery?

Deploy the secondary region baseline and reserve/confirm capacity for critical services. Execute a controlled failover exercise:

1. Declare simulated regional loss.
2. Validate replicated/restored authoritative data.
3. Deploy/scale secondary compute and regional AI dependencies.
4. Verify identities, DNS, certificates, private routes, and quotas.
5. Switch synthetic traffic.
6. Run the critical user journey and access-leak tests.
7. Measure RTO/RPO.
8. Reconcile and fail back.

If Microsoft Foundry model availability differs by region, use an already evaluated fallback and record the quality difference.

What evidence goes to the approval board?

• Architecture decision records and diagrams.
• Threat model/privacy/data-impact assessment.
• IaC plans and policy compliance report.
• RBAC/PIM and emergency-access test.
• Penetration/red-team and AI evaluation summaries.
• Load/capacity/quota and regional failover results.
• Backup restore result.
• Monitoring/incident runbooks and alert tests.
• Cost forecast, unit economics, and commitments plan.
• Residual risk register with owners/expiry.

Scoring rubric

• 20% identity/security/data protection.
• 20% network/DNS/hybrid correctness.
• 15% AI governance and evaluation.
• 15% reliability/recovery evidence.
• 10% policy/landing-zone repeatability.
• 10% operations/observability/incident response.
• 10% cost and organizational operating model.

The capstone passes only if critical access-leak, public-exposure, restore, or regional-capacity risks are resolved or explicitly block launch.