Microsoft Azure AI Foundry: Enterprise Deployment Guide for Regulated Industries

Core Components

• Model Catalog: Access to OpenAI models (GPT-4o, GPT-4), open models (Llama, Mistral, Phi), and Microsoft models — all deployable within your Azure tenant with your compliance controls.
• Prompt Flow: Visual and code-based orchestration for building AI applications. Version-controlled, testable, deployable to managed endpoints.
• Content Safety: Built-in content filtering with configurable severity thresholds. Critical for regulated use cases where output must be controlled.
• Responsible AI Dashboard: Model evaluation, fairness metrics, and explainability tools. Required for EU AI Act high-risk system documentation.
• Managed Endpoints: Deploy models as APIs within your VNet, with private endpoints, managed identity, and Azure RBAC. No public internet exposure required.

Network Isolation

• Private endpoints: All AI Foundry resources should be deployed with private endpoints in your VNet. No public IP exposure.
• Managed VNet: Use AI Foundry's managed VNet option to ensure compute resources (training, inference) never communicate over public internet.
• Data exfiltration protection: Enable Azure Storage firewall rules and service endpoints. Prevent models from reaching external endpoints during inference.
• DNS configuration: Private DNS zones for all AI Foundry endpoints. Ensure all resolution stays within your network boundary.

Identity & Access

• Managed Identity: Use system-assigned managed identities for all AI Foundry resources. No API keys stored in configuration.
• RBAC granularity: Separate roles for model deployment (ML Engineer), data access (Data Scientist), and configuration (Platform Admin). Principle of least privilege.
• Conditional Access: Require MFA, compliant devices, and named locations for AI Foundry access. AI model management is a privileged operation.
• Just-in-Time access: Use PIM (Privileged Identity Management) for model deployment and configuration changes. Time-bound, approval-required access.

Catalogue Controls

• Approved model registry: Don't give all teams access to the full model catalogue. Create an approved list of models that have passed your security and compliance review.
• Version pinning: Pin model versions in production. Never use "latest" — you need reproducibility for audit and incident investigation.
• Model evaluation gates: Before any model reaches production, it must pass evaluation benchmarks specific to your domain. Accuracy, fairness, safety, and latency thresholds.
• Deprecation policy: When Microsoft deprecates a model version, you need lead time. Monitor deprecation announcements and maintain a 90-day upgrade runway.
• Fine-tuned model lifecycle: Fine-tuned models are YOUR responsibility. Version them, test them, and have rollback capability. Store training data lineage for audit.

Azure AI Foundry → DORA Articles

• Article 6 (ICT risk management): Document AI Foundry in your ICT risk framework. Identify threats specific to AI systems (model poisoning, prompt injection, data leakage).
• Article 9 (Protection and prevention): Network isolation, encryption at rest and in transit, access controls. Azure AI Foundry supports all — but you must configure them.
• Article 11 (Response and recovery): What's your plan if Azure AI Foundry is unavailable? Multi-region deployment, failover to alternative providers, or graceful degradation?
• Article 28 (Third-party risk): Microsoft is your ICT third-party provider. Document the relationship, assess concentration risk (especially if you use Azure for compute + AI + identity).

Recommended Topology

1. Hub-spoke network: AI Foundry in a dedicated spoke VNet, peered to your hub. Controlled egress through Azure Firewall.
2. Dev/staging/prod separation: Separate AI Foundry projects per environment. Promote models through environments with approval gates (Azure DevOps/GitHub Actions).
3. Monitoring: Azure Monitor + Log Analytics for all AI Foundry operations. Custom alerts on model latency degradation, error rate spikes, and content safety triggers.
4. Cost management: Azure Cost Management tags on all AI resources. Chargeback to business units based on token consumption. Set budget alerts at 80% threshold.
5. Backup & DR: Export Prompt Flow definitions to git. Model weights stored in Azure ML registry (geo-redundant). Configuration as code via Bicep/Terraform.