Azure AI-200: What Changed From AZ-204 (and Who Should Take It)

If you were halfway through AZ-204 prep this summer, you're studying for a retired exam.

Microsoft's Partner Center announcements for July 2026 put it in one line:

"Azure Developer Associate (AZ-204) retires July 31, 2026. It's replaced by Azure AI Cloud Developer Associate (AI-200), available July 9, 2026."

The new exam is AI-200: Developing AI Cloud Solutions on Azure, and it leads to the Microsoft Certified: Azure AI Cloud Developer Associate certification. It's still a developer exam, but it isn't AZ-204 with a new name. More than half of the outline is new.

Here's what changed, what carried over, and who should take it.

The exam at a glance

From the official certification page and study guide, checked on September 16, 2026:

• 120 minutes. The page says "You may have interactive components to complete as part of this exam."
• 700 to pass, on Microsoft's usual scaled score.
• No published question count. Be wary of any course that claims to know the exact number.
• US$165 in the United States. The price depends on where you take the exam.
• 13 languages, including English, Japanese, Chinese, Spanish and Portuguese (Brazil).
• No official practice assessment yet. Microsoft says these usually arrive within about eight weeks of an exam leaving beta.
• Annual renewal with a free online assessment, like every associate certification.

The four domains

AZ-204 had five domains, and compute was the largest (25-30%). In AI-200, data services are the largest domain, and "data services" now means vector search.

What was added

• Vector search in three places. Cosmos DB for NoSQL (embeddings and vector similarity search), Azure Database for PostgreSQL with pgvector (indexing, compute and memory sizing, RAG with metadata filters), and Azure Managed Redis (vector indexing).
• PostgreSQL as a whole topic. Connecting with SDKs, schema and data type choices, index strategy, and connection optimization for throughput.
• Azure Managed Redis. Caching, expiration and invalidation are now listed on their own, next to vector indexing.
• Container orchestration. ACR Tasks, Container Apps revision management, KEDA event-driven scaling, and deploying and troubleshooting AKS with manifest files. Troubleshooting means reading logs and events and testing end-to-end connectivity.
• Change feed processor for Cosmos DB, now worded as a skill you implement.
• OpenTelemetry SDKs for distributed tracing.
• KQL for analyzing logs and metrics, as its own bullet.
• Python. The audience profile names Python programming, and it's the only language the study guide mentions.

What was removed

Compared with the last AZ-204 skills outline (January 14, 2026), these are gone:

• The App Service web apps topic: deployment slots, autoscaling, TLS and API settings. (Deploying containers to App Service is still there, along with configuring environment variables and secrets.)
• Blob Storage: properties, metadata, lifecycle policies.
• User authentication and authorization: the Microsoft identity platform, Microsoft Entra ID, shared access signatures, Microsoft Graph.
• Managed identities as a separate objective.
• API Management.
• Event Hubs and Queue Storage.
• Azure Container Instances.
• Application Insights availability tests.

You'll still use managed identities and Application Insights, because Key Vault and OpenTelemetry depend on them. The difference is that AI-200 treats them as tools you use along the way, not as objectives.

What carried over

If you already studied for AZ-204, this part isn't wasted:

• Azure Container Registry and Container Apps
• Azure Functions triggers and bindings, plus configuring and deploying function apps
• Cosmos DB SDK operations, consistency levels, and change feed
• Service Bus (now with dead-letter queues, topics and subscriptions spelled out)
• Event Grid (now with filters, custom events and retries spelled out)
• Key Vault and App Configuration (Key Vault rotation is new)

That's a little over a third of the new outline. Start with the added list above, because that's where your gaps are.

Who should take it

AI-200 is a good fit if you:

• Build back ends for AI features: retrieval-augmented generation, semantic caching, asynchronous inference jobs.
• Planned to take AZ-204. This is the same developer associate slot in Microsoft's catalog.
• Write Python and want an Azure cert that matches how you work.
• Run containers, whether that's App Service, Container Apps or AKS.

It's a poor first choice if you want to train models or tune prompts. It barely touches either. It's about the systems around the model: where embeddings are stored, how work is queued, how secrets are kept out of code, and how you trace a slow request.

If you're new to Azure, it's worth learning the basics first. If you already hold AZ-204, AI-200 is the current version of that developer credential, and the new material is concentrated in data services and observability.

How I'd study for it

Plan for eight to twelve weeks alongside a full-time job. The most efficient approach I know is to build one small RAG service that touches every domain:

1. Containerize it. Build the image with ACR Tasks and run it on Container Apps. Deploy a second revision and split traffic between them. Then run the same image on AKS from hand-written YAML and break the Service selector on purpose.
2. Store embeddings twice. Once in Cosmos DB with a vector index, once in PostgreSQL with pgvector. Run EXPLAIN on your pgvector query, and make sure the index's operator class matches the distance operator you query with. If they don't match, the planner skips the index and falls back to a sequential scan.
3. Add a cache. Use Azure Managed Redis for cache-aside first, then add a vector index for semantic caching. Redis modules are chosen when the cache is created, so select RediSearch at that point.
4. Make it asynchronous. Put ingestion behind a Service Bus queue, trigger a function from it, and dead-letter a message on purpose so you have to recover it.
5. Secure and observe it. Keep secrets in Key Vault, settings in App Configuration, and send traces through the Azure Monitor OpenTelemetry distro. Then write the KQL that finds your slowest dependency.

When the build is done, work through our AI-200 practice questions in timed mode. They're scenario questions written to the new outline, and every wrong answer has an explanation. The AI-200 study guide page has the domain notes and a week-by-week plan.

A few more pages that help:

• If you use the Kubernetes parts less often, the Docker and Kubernetes cheat sheet covers the kubectl commands the AKS objectives assume.
• Our LLM inference system design walkthrough covers the architecture around the services this exam tests.
• If you still have AZ-204 notes, the AZ-204 guide is still available for comparison. Use it only for the topics that carried over.
• If you're aiming for architect next, AZ-305 is the usual follow-on.

The short version

AZ-204 is retired. AI-200 replaced it on July 9, 2026. It's still a developer associate exam, but most of its weight is now on vector search, containers and event-driven pipelines. If you know Python and have shipped something with a vector database, you're closer than an old AZ-204 course would suggest. If you haven't, build something first.