AI-200 Exam Domains Explained

The short version

AI-200 has four domains. Data management with vector search is the largest at 30-35%, and the other three each carry 20-25%. The skill names come from Microsoft’s study guide, “Skills measured as of October 7, 2026”. The traps and tips below are our own general advice, not Microsoft statements. The study guide itself says its bullets illustrate how a skill is assessed and that related topics may appear.

Domain 1: Develop containerized solutions on Azure (20-25%)

What Microsoft lists. Build, store, version, and manage images in Azure Container Registry. Build and run images with Registry Tasks. Deploy containers to App Service, including environment variables and secrets. Deploy to Azure Container Apps, including environment configuration and revision management. Use KEDA for event-driven scaling in Container Apps. Deploy and manage applications on AKS by using manifest files. Monitor and troubleshoot AKS and Container Apps by inspecting logs, events, and end-to-end connectivity.

In practice. You should know how an image gets from source to registry (a local build and push, or a build that runs in Azure), how tags and versions behave, and how a container app rolls out a new revision. For AKS, you should be able to read a manifest and say what it will create: a Deployment, a Service, a ConfigMap, a Secret. For scaling, know that KEDA scales on events such as queue length instead of CPU.

Typical traps.

  • Mixing up which Azure service fits a workload: App Service for a simple web container, Container Apps for revisions and event-driven scale, AKS when you need Kubernetes itself.
  • Confusing configuration with secrets, and not knowing which one a question’s constraint (for example “no secrets in plain settings”) rules out.
  • Reading a Kubernetes problem as an Azure problem. Often the answer is in the pod events or logs.

What to practice. Run one Registry Task, make two revisions of a container app and split traffic, add a queue-based scale rule, and break an AKS deployment on purpose so you can find the cause from the events.

Domain 2: Develop AI solutions by using Azure data management services (30-35%)

What Microsoft lists. Three services, each with its own set of skills.

  • Cosmos DB for NoSQL: connect with the SDK and run queries, tune indexing policies and consistency levels to control performance and request units, store embeddings and run vector similarity search, and implement a change feed processor.
  • Azure Database for PostgreSQL: connect and query with SDKs, design tables and data types, choose indexing strategies, reduce pgvector compute overhead, size compute, memory, and storage for vector workloads, run similarity search and retrieval-augmented generation (RAG) patterns with metadata filters, and tune connections for throughput and latency.
  • Azure Managed Redis: caching with expiration and invalidation, and vector indexing for similarity search.

In practice. Vector search appears three times, so expect to compare the three services. You need to know how each stores an embedding, how it builds an index, and what you give up for speed. A RAG pattern means: embed the question, find the nearest stored items, filter by metadata such as tenant or date, and pass the results to a model.

Typical traps.

  • Approximate indexes trade some recall for speed. A question that needs exact results wants a different index than one that wants low latency at large scale.
  • Index choice depends on the service. Names you know from one product are not always options in another.
  • Consistency levels in Cosmos DB trade freshness against cost and latency; read what the scenario needs to be true.
  • Caching questions hide the real issue in invalidation: when does stale data become wrong?

What to practice. Create a vector index in each of the three services and query it. Write one query that combines a vector search with a metadata filter. Change an indexing policy and watch the request unit charge.

Domain 3: Connect to and consume Azure services (20-25%)

What Microsoft lists. Queue and process back-end work with Azure Service Bus, including dead-letter handling, messages, topics, and subscriptions. Build event-driven workflows with Azure Event Grid, including filters, custom events, and retries. Build serverless APIs with Azure Functions, including triggers and bindings. Configure and deploy function apps.

In practice. The core skill is choosing between messages and events. A message is a command or work item that someone must process, so you want queues, ordering, and dead-lettering. An event is a notification that something happened, so you want filters, fan-out, and retries. Functions then tie these together through triggers and bindings.

Typical traps.

  • Picking Event Grid when the scenario needs ordered, durable processing, or Service Bus when it only needs a lightweight notification.
  • Forgetting what happens to a message that keeps failing. Know where it ends up and how to read it.
  • Treating a function’s trigger and its configuration as the same thing. Settings, hosting plan, and deployment method are separate decisions.

What to practice. Send a message that fails until it dead-letters, then read it from the dead-letter subqueue. Create an Event Grid subscription with a filter. Write a function with a queue trigger and an output binding, and deploy it.

Domain 4: Secure, monitor, and troubleshoot Azure solutions (20-25%)

What Microsoft lists. Secure secrets in Azure Key Vault, including rotation and retrieval. Store and retrieve configuration in Azure App Configuration. Trace distributed systems with OpenTelemetry SDKs. Write KQL queries to analyze logs and metrics.

In practice. Think “no passwords in code”. An app uses a managed identity, reads configuration from App Configuration, and fetches secrets from Key Vault with the least privilege needed. When something is slow, you follow a request across services with traces and then query the logs.

Typical traps.

  • Mixing up role names. Reading a value and managing a value are different permissions, and the wrong role fails quietly in the real world.
  • Putting a secret where plain configuration belongs, or the reverse.
  • KQL details: grouping by time with `bin()`, counting distinct values with `dcount()`, and choosing the right join kind when some rows have no match.

What to practice. Give an app a managed identity and read a secret without any stored credentials. Add OpenTelemetry to a small service and find a trace in Application Insights. Write ten small KQL queries until `summarize`, `bin`, `join`, and `where` feel automatic.

A note on the change

Microsoft’s change log says the data management domain grew on October 7, 2026. If your study material predates that, it probably underweights vector search.

More guides for AI-200

Next step

Put the domains into a schedule with the study plan, or see how they look as questions on the sample questions page. Our AI-200 practice tests include sets that cover all four domains.

Frequently asked questions

Which domain should I study first?

Start with containers if they are new to you, because the later domains assume you can deploy something. If you are strong there, begin with data management, since it carries the most weight.

Do I need to know every CLI command?

The study guide does not list commands. Know what a command or setting does and when to use it, but memorizing flags matters less than understanding the service.

Are preview features tested?

Microsoft says most questions cover generally available features, and the exam may include preview features that are commonly used.

Try the AI-200 practice exams