Lesson 1: Every Azure AI and ML Service on One Page

Why start with a map?

If you have run Azure workloads for a while, you will have seen the AI estate renamed several times. A Terraform module from 2022 says Cognitive Services. A blog from 2024 says Azure AI Studio. Today's portal says Microsoft Foundry. They are often the same service under a newer name, and sometimes they are not. Before you can choose a service, you need to know what exists, which group it belongs to, and whether it is still open for new work.

Think of this as a station map for a city that keeps renaming its stations. The track is mostly the same, but you need a lookup table of old and new names or you will end up at the wrong stop.

The six groups on one page

Gotcha: Foundry IQ is not simply a new name for Azure AI Search. It is a managed knowledge layer built on Azure AI Search, and you need a search resource to use it. When a pricing page says "Foundry IQ", the meter behind it is still Azure AI Search.

Reading old code: the resource-type decoder

Names in templates change more slowly than names in marketing. Most of the estate sits on just two Azure resource providers:

// 1) The modern Foundry resource: one account that holds projects, models, agents and tools.
//    Despite the name, this is the same provider that Cognitive Services used.
resource foundry 'Microsoft.CognitiveServices/accounts@2025-06-01' = {
  name: 'contoso-foundry'
  location: 'swedencentral'
  kind: 'AIServices'            // 'OpenAI' = an older standalone Azure OpenAI resource
  sku: { name: 'S0' }           // many single tools also offer F0 (free)
  identity: { type: 'SystemAssigned' }
  properties: {
    allowProjectManagement: true  // turns the account into a Foundry resource with projects
    customSubDomainName: 'contoso-foundry'
  }
}

// 2) The classic hub: a Machine Learning workspace with kind 'hub'.
//    Seeing this means you are reading a Foundry (classic) hub-based project.
// resource hub 'Microsoft.MachineLearningServices/workspaces@...' = { kind: 'Hub' ... }

In Terraform, the same split looks like this:

# Modern: a cognitive account of kind AIServices with project management enabled
resource "azurerm_cognitive_account" "foundry" {
  kind                       = "AIServices"
  project_management_enabled = true
  # ...
}

# Classic: azurerm_ai_foundry is the HUB (an ML workspace), not the new Foundry resource
# resource "azurerm_ai_foundry" "hub" { ... }
# azurerm_ai_services is deprecated in favour of azurerm_cognitive_account kind = "AIServices"

Pitfall: azurerm_ai_foundry sounds like the current product, but it creates a classic hub. Hub-based projects still work in the Foundry (classic) portal, but Microsoft puts new feature work into Foundry projects. Treat a hub in a new design as a warning sign.

Walkthrough: inventory what you already have

Run this against a subscription and sort what comes back into the six groups:

# Everything on the Cognitive Services provider: Foundry resources, old Azure OpenAI, single tools
az cognitiveservices account list \
  --query "[].{name:name, kind:kind, sku:sku.name}" -o table

# ML workspaces: plain Azure ML ('Default'), classic hubs ('Hub') and hub projects ('Project')
az resource list --resource-type Microsoft.MachineLearningServices/workspaces \
  --query "[].{name:name, kind:kind}" -o table

Sample output, with each line mapped to the table:

Name              Kind            Sku
----------------  --------------  -----
contoso-foundry   AIServices      S0     -> Generative platform (Microsoft Foundry)
legacy-aoai       OpenAI          S0     -> Azure OpenAI, can be upgraded to a Foundry resource
invoice-ocr       FormRecognizer  S0     -> Document Intelligence (old kind name kept)
faq-bot           QnAMaker        S0     -> RETIRED, migrate to Language question answering

Name              Kind
----------------  -------
churn-ml          Default        -> Azure Machine Learning
ai-hub-2024       Hub            -> Foundry (classic) hub

Notice that kind values such as FormRecognizer keep the old name forever. The resource type tells you what the service is, whatever the docs now call it.

Closed and retiring: never shortlist these

Best practice: add a "lifecycle" column to your own service map and check it against the Foundry Tools overview page every quarter. Retirements arrive with notice, but the endpoints really do go dark.

Key takeaways

• There are six groups: generative platforms and models, agents, search and retrieval, ready-made APIs, ML platforms and accelerators. Every Azure AI service fits in one of them.
• Azure AI Studio → Azure AI Foundry → Microsoft Foundry. Cognitive Services → Azure AI services → Foundry Tools. Azure Cognitive Search → Azure AI Search, which underpins Foundry IQ.
• A Microsoft.CognitiveServices/accounts resource of kind AIServices is the modern Foundry resource. A MachineLearningServices workspace of kind Hub is the classic hub.
• Old kind values such as FormRecognizer survive in templates, so decode by resource type, not by name.
• QnA Maker, LUIS, Metrics Advisor, Anomaly Detector and Personalizer are retired, and Content Moderator retires in March 2027. Keep all of them off your shortlist.