Back to index
OtherSelf-paced; 50 to 80 hours for a working Azure developer, a lot more if containers and Cosmos DB are new to you·Training free on Microsoft Learn; associate exams are priced by country (about $165 in the US); free annual renewal

Microsoft AI-200: Developing AI Cloud Solutions on Azure (Azure AI Cloud Developer Associate)

4.1

Despite the AI in the name, this is a back-end cloud developer exam. Roughly a quarter is about vector storage and RAG plumbing, the rest is containers, messaging, serverless and observability. That is a fair picture of what building AI products on Azure actually involves, and I think it is a better exam than AZ-204 was.

What We Liked

  • The data section is genuinely useful: Cosmos DB vector search, pgvector index tuning, Redis vector indexing and RAG with metadata filters
  • Swaps out stale AZ-204 material for Container Apps, KEDA scaling and AKS manifests
  • OpenTelemetry tracing and KQL are in scope, which most AI courses never mention
  • Python only, which matches where most AI back-end work happens

What Could Be Better

  • Very little about models themselves; anyone expecting prompt engineering or agents should look at AI-103 instead
  • Broad: you need working knowledge of around fifteen Azure services, and the exam punishes gaps
  • No practice assessment yet, and Microsoft Learn paths were still catching up when the exam launched
  • AZ-204 holders get no bridge exam, so a still-valid credential quietly becomes a dead end

Detailed review

Microsoft retired AZ-204 and its renewal assessment on 31 July 2026, and AI-200 is what replaced it. Search interest shows people noticed: the exam code went from around 100 searches a month at the start of the year to 880 in July and August. The exam is called Developing AI Cloud Solutions on Azure and passing it earns Microsoft Certified: Azure AI Cloud Developer Associate. It is 120 minutes, proctored, 700 to pass, with the usual free annual renewal on Microsoft Learn. The skills outline has four parts.

Containerized solutions (20 to 25%) covers Azure Container Registry and ACR Tasks, App Service containers, Container Apps with revisions and KEDA autoscaling, and AKS deployments from manifests. Data (25 to 30%) is where the AI label comes from: Cosmos DB for NoSQL including embeddings, vector similarity search and change feed processors; Azure Database for PostgreSQL with pgvector, index tuning and RAG patterns using metadata filters; and Azure Managed Redis for caching and vector indexing. Connecting services (20 to 25%) is Service Bus with dead-letter handling, Event Grid and Azure Functions. Security and monitoring (20 to 25%) covers Key Vault rotation, App Configuration, OpenTelemetry tracing and KQL. My honest take is that the name oversells the AI part and undersells the rest.

There is nothing about calling a model, choosing one, prompting it or evaluating its output. What you get is the plumbing around a retrieval system: where the vectors live, how much it costs in request units, how to keep the index fresh when data changes, how to queue slow work and how to trace a request through it all. Having watched plenty of RAG prototypes fall over in production for exactly these reasons, I think that is the more valuable half to certify. The pgvector objectives on reducing compute overhead and sizing memory for vector workloads are the kind of thing people usually learn the hard way. The weak spots are breadth and timing.

The exam assumes hands-on use of a lot of services, and the Learn paths were still being filled in when it went live. There was no practice assessment at the time of writing. If you already hold AZ-204 it stays valid until it expires, but you cannot renew it, so you will be sitting this eventually. If you want the model and agent side, the matching exam is AI-103 (Azure AI Apps and Agents Developer). For most Azure back-end developers I would take AI-200 first, since it is closer to the work you already do.

[ final ]

The verdict.

For an Azure developer, this is the default associate cert to take now that AZ-204 is gone. Just know what you are buying: it proves you can build the infrastructure around an AI feature, not the AI feature itself. Pair it with AI-103 if you want both halves.