If you were heading for AZ-204 and watched it retire out from under you on July 31, AI-200 is where that path continues. Microsoft’s course retirement list names AI-200T00 as the replacement for AZ-204T00 outright. But it is not a straight rename. Half of it is the developer exam you were already studying for; the other half is new.
Knowing which half is which is the whole point of preparing efficiently.
What the exam is
The credential is Microsoft Certified: Azure AI Cloud Developer Associate, earned by passing Exam AI-200: Developing AI Cloud Solutions on Azure. It is aimed at back-end developers who build AI solutions, with an emphasis on services and components rather than front-end work. Microsoft expects you to be comfortable with Python, Azure and third-party SDKs, containerized applications, messaging and eventing, and vector databases before you sit it.
A passing score is 700 out of 1000. The certification renews the same way the rest of the Azure associate track does: a free online assessment on Microsoft Learn once a year.
Check one thing before you book: AI-200 launched as a beta exam. Beta exams score on a delay, ship in English first, and can carry a rougher question set than a settled exam. If you sit it early, budget for the wait on your result.
The four skill areas
Microsoft publishes the weights as ranges, not fixed numbers, so treat these as the shape of the exam rather than a promise of exact question counts:
| Skill area | Weight |
|---|---|
| Develop AI solutions using Azure data management services | 25–30% |
| Develop containerized solutions on Azure | 20–25% |
| Connect to and consume Azure services | 20–25% |
| Secure, monitor, and troubleshoot Azure solutions | 20–25% |
The data services area is the heaviest, and it is also the one with the least overlap with AZ-204. Those two facts line up for a reason: the data half is where Microsoft added the weight that makes this an AI exam rather than a developer one.
Where AZ-204 study still counts
Three of the four areas are the AZ-204 spine.
Containers: Azure Container Registry for building and versioning images, Azure Container Apps with KEDA-driven event scaling, and AKS deployments from manifest files. Connecting to services: Service Bus queues and topics with dead-letter handling, Event Grid workflows, and Azure Functions triggers and bindings. Security and operations: Key Vault for secrets and rotation, App Configuration, distributed tracing with OpenTelemetry, and KQL for reading logs and metrics.
If you prepared for AZ-204, you have seen all of this. The service names and the patterns carry straight over. That is roughly three-quarters of the exam where your existing study is not wasted.
Where the exam actually changed
The data half is the new material, and it is all built around one idea: storing and searching vectors.
You are expected to work with Azure Cosmos DB for NoSQL, tuning Request Units and consistency levels, and running vector similarity search for semantic retrieval. With Azure Database for PostgreSQL, the exam moves into pgvector: schema and indexing choices, vector similarity search, and implementing retrieval-augmented generation with metadata filtering. Azure Managed Redis appears as a third vector store, for caching and vector indexing.
Generic blob storage, the safe default answer on AZ-204, is rarely the right answer here. The data question on AI-200 is almost always “which store, and how do you search it by meaning.” If you have never stored an embedding or written a RAG retrieval step, that is the gap to close first, because it is both the newest material and the heaviest-weighted.
How to prepare
Start from what you already know. If AZ-204 was on your radar, spend your first pass confirming the container, messaging, Functions, and security material still holds, because it does, and you can move through it quickly.
Then put your real hours into the data half. Stand up a Cosmos DB for NoSQL container, store some embeddings, and run a similarity search. Do the same with pgvector in PostgreSQL and build a small RAG retrieval that filters on metadata. The exam is testing whether you can reach for the right vector store and query it correctly, and that is a thing you learn by building it once, not by reading about it.
Practice questions are worth more here than on a settled exam, precisely because AI-200 is new and the community notes are thin. Working AI-200 practice questions grounded in the current skills outline tells you fast whether the vector-search material has actually landed, or whether you only recognize the service names. On a beta exam with a delayed score, finding that out before you sit it is the difference that matters.
If you were an AZ-204 candidate, none of your work was lost. AI-200 kept the developer spine and added the part that makes it an AI exam. Study the new half like it is new, and let the old half carry the rest.