Foundry Resources, Projects and Access Control

What it is

A Microsoft Foundry resource is the top-level Azure resource for AI work. It holds model deployments, shared connections and the settings for networking, identity and security. A project is a child of that resource and is the working space where a team keeps its agents, files, evaluations and project level connections. Apps call a project endpoint.

Identity is controlled by Microsoft Entra ID and Azure RBAC. People and apps get the smallest role that does the job, and apps use managed identities rather than keys.

How it works

Resource and project layout

Foundry also still supports the older hub based projects, which are built on Azure Machine Learning workspaces. Know that both exist, but new designs start with the Foundry resource and its projects.

Create them with Bicep

param location string = resourceGroup().location
param foundryName string
param projectName string

resource foundry 'Microsoft.CognitiveServices/accounts@2025-06-01' = {
  name: foundryName
  location: location
  kind: 'AIServices'
  sku: { name: 'S0' }
  identity: { type: 'SystemAssigned' }
  properties: {
    allowProjectManagement: true
    customSubDomainName: foundryName
    disableLocalAuth: true
  }
}

resource project 'Microsoft.CognitiveServices/accounts/projects@2025-06-01' = {
  parent: foundry
  name: projectName
  location: location
  identity: { type: 'SystemAssigned' }
  properties: {
    displayName: projectName
    description: 'Customer support assistant'
  }
}

Two properties matter here. allowProjectManagement: true lets the resource hold projects. disableLocalAuth: true turns off API keys so only Entra ID tokens work.

Connect the project to other services

Add connections for services such as Azure AI Search, storage and Application Insights. Prefer connections that authenticate with Entra ID, and grant the project's managed identity the data role it needs on the target (for example Search Index Data Reader on Azure AI Search or Storage Blob Data Reader on storage).

Roles for people

Microsoft renamed these roles, so portals and exam items may show either name. The role IDs and permissions did not change. For Foundry work, do not assign roles whose names start with Cognitive Services, except Cognitive Services Usages Reader at subscription scope when users need to see quota.

Assign roles at the narrowest scope that works (project before resource before resource group), to Entra groups rather than individuals, and use Privileged Identity Management for elevated roles.

Managed identities for applications

An app hosted on Azure (App Service, Container Apps, Functions, AKS) gets a system assigned or user assigned managed identity. The code calls the endpoint with DefaultAzureCredential, and that identity holds a role on the Foundry resource or project.

from azure.identity import DefaultAzureCredential
from azure.ai.projects import AIProjectClient

client = AIProjectClient(
    endpoint="https://<resource>.services.ai.azure.com/api/projects/<project>",
    credential=DefaultAzureCredential(),
)

Use a user assigned identity when several services share one identity or when you need it to exist before the app does. Locally, DefaultAzureCredential uses the developer's own signed in account, so the same code runs everywhere with no keys.

A worked example

A company creates foundry-dev and foundry-prod resources with one project per app. Developers are in an Entra group with Foundry User on the dev project only. The production web app runs with a managed identity that holds Foundry User on foundry-prod, and key based authentication is disabled.

How the exam asks about this

• "Developers can run and evaluate in a project but cannot manage accounts." Foundry User (formerly Azure AI User) on the project.
• "The app must call the model with no secret in configuration." Managed identity plus a role assignment.
• "Isolate production deployments and quota from experiments." Separate Foundry resources per environment.