Plan and manage an Azure AI solution
Foundry Infrastructure, Deployments, and CI/CD
CoreLay out the Foundry resource, projects, connected resources, and agent setup for an AI app; pick deployment types that meet residency, throughput, and cost needs; configure model and agent deployments; and automate them through CI/CD.
Aligned to the AI-103 study guide with skills measured as of April 16, 2026; guide and product behavior verified October 12, 2026.
Why this matters
Infrastructure choices made on day one decide where prompts are processed, how capacity is billed, whether agent data stays in your own storage, and whether releases can be repeated safely. A deployment type chosen without checking data residency, or an agent created by hand in the portal, becomes a compliance or reliability problem that is expensive to unwind.
Must Know
- A Foundry resource is an Azure resource of kind AIServices, and Foundry projects are its child resources. Connected resources such as Azure Storage, Azure Key Vault, and Azure AI Search remain separate Azure resources with their own networking and access settings.
- Agent Service basic setup stores agent state in Microsoft-managed storage. Standard setup brings your own Azure Storage, Azure AI Search, and Azure Cosmos DB so that conversation and file data stay in resources you control and are isolated per project.
- Deployment types for models sold by Azure fall into standard (pay per token), provisioned (reserved provisioned throughput units), and batch (asynchronous, about 50% cheaper, with a 24-hour target turnaround). A Developer type exists only for evaluating fine-tuned models and has no SLA.
- Global types may process prompts in any Azure region, Data Zone types keep processing inside a Microsoft-defined data zone, and Standard or Regional Provisioned keeps processing within the chosen Azure geography. Data at rest stays in the resource's geography for all types.
- Start with Global Standard for most workloads: new models arrive there first and it has the highest default quota. Move to Data Zone or regional types for residency, to provisioned types for predictable latency at steady high volume, and to batch for large jobs that are not time-sensitive.
- Each model deployment receives a tokens-per-minute allocation from the subscription's regional quota. Prompt agents are defined by a model deployment, instructions, and tools, and every change creates a new agent version that you can compare or roll back.
- Foundry has no automatic cross-region failover; multiregion resilience means deploying separate resources and routing traffic at the application layer.
- Bicep templates can create the Foundry resource and its projects, and an existing resource's template can be exported from the Azure portal. Runtime role assignments for the project managed identity are a separate step.
- Agents and evaluations can be defined code-first in pipelines. The AI agent evaluation GitHub Action (preview) runs evaluations against agent versions and can compare a candidate with a baseline, authenticating to Azure with OpenID Connect.
Compare and Distinguish
- Basic versus standard agent setup: Microsoft-managed agent storage versus your own Storage, AI Search, and Cosmos DB resources.
- Global versus Data Zone versus regional processing: widest capacity and earliest models versus processing inside a data zone versus processing inside one Azure geography.
- Standard versus provisioned versus batch: pay-per-token best effort versus reserved capacity with lower latency variance versus discounted asynchronous jobs.
- Co-located versus isolated topology: several projects sharing one Foundry resource and its quota versus one Foundry resource per production workload.
Scenario examples
- Scenario: A European insurer requires prompts to be processed only inside the EU but wants the broadest capacity available within that boundary. Think: a Data Zone Standard deployment in an EU region.
- Scenario: A retailer must summarize ten million product reviews overnight at the lowest cost. Think: a Global Batch or Data Zone Batch deployment with a JSONL batch file.
- Scenario: Security requires agent conversation history and uploaded files to stay in the company's own Cosmos DB and Storage accounts. Think: standard agent setup with bring-your-own resources.
Exam traps
- Global Standard is not a data residency control; it can process prompts in any region.
- Provisioned capacity is billed hourly whether or not it is used, so it is a poor fit for idle or bursty workloads.
- Having provisioned quota does not guarantee that capacity is available in a region when you create the deployment.
- A Developer deployment is for evaluating fine-tuned models, not for production traffic.
- Exporting a template captures resources, not the data-plane role assignments the project identity needs at runtime.
Key takeaways
- Choose the deployment type from residency first, then traffic shape, then latency and cost.
- Use standard agent setup when agent data must live in your own resources.
- Treat agents, deployments, and evaluations as code that a pipeline creates and verifies.
How it works
- New models typically launch on Global types, then Data Zone, then geography-based types, so the newest models may not yet exist in a regional deployment.
- Batch jobs draw on a separate enqueued-token quota, so they do not consume the tokens-per-minute quota of online deployments.
Objects and administrative surfaces
- Foundry portal Deployments and Model quota pages.
- Bicep resources Microsoft.CognitiveServices/accounts and accounts/projects.
- Azure Developer CLI (azd) templates for hosted agent projects.
- GitHub Actions or Azure Pipelines running the Foundry SDK and evaluation tasks.
When to use it
- Use one Foundry resource per production workload when blast radius and quota isolation matter.
- Use Data Zone or regional deployments when regulation limits where inference can happen.
Security and governance implications
- Azure Policy can deny specific deployment SKUs, such as Global types, across subscriptions.
- Pipelines should authenticate with workload identity federation rather than stored secrets.
Troubleshooting signals
- If a new model is missing from a regional deployment list, check whether it has launched only on Global or Data Zone types so far.
- If a pipeline-created project cannot call models, check the data-plane role assignment for the calling identity.
More detail
- Design Foundry resources, projects, and connected resources for apps and agents.
- Choose Global, Data Zone, regional, provisioned, batch, or Developer deployments.
- Configure model deployments and versioned prompt agents.
- Automate resources, agents, and evaluations with Bicep and CI/CD pipelines.
Ready for the quiz?
- Which deployment type keeps processing inside the EU while still pooling capacity across several regions?
- What changes when an agent project uses standard setup instead of basic setup?
- Why is batch a better fit than provisioned throughput for a nightly bulk job?
- What must a pipeline do after a template creates a Foundry project before the app can call a model?
Related objectives
- D1.2.S1 — Design Azure infrastructure for AI apps and agent-based solutions
- D1.2.S2 — Choose appropriate deployment options
- D1.2.S3 — Configure model and agent deployments
- D1.2.S4 — Integrate Foundry projects with continuous integration and continuous deployment (CI/CD) pipelines