AI-200: Developing AI Cloud Solutions on Azure is the exam for Microsoft Certified: Azure AI Cloud Developer Associate. It replaced AZ-204 (Azure Developer Associate), which retired on 31 July 2026. Think of it as AZ-204 rebuilt for AI back ends. The containers, Functions, messaging, Key Vault and monitoring topics are still there. New is a big block (25–30%) on vector search and RAG data stores: Cosmos DB, PostgreSQL with pgvector, and Azure Managed Redis. The exam expects Python.
This guide follows Microsoft's study guide: every skill area, the code patterns to know, a 7-week plan, labs and 20 original scenario questions.
Quick facts (Microsoft Learn, November 2026)
| Item | Detail |
|---|---|
| Exam | AI-200: Developing AI Cloud Solutions on Azure |
| Certification | Microsoft Certified: Azure AI Cloud Developer Associate |
| Time | 120 minutes; proctored; may include interactive components |
| Passing score | 700 |
| Renewal | Every 12 months, free online assessment |
| Language for code | Python |
| Price | Set by the country where you take the exam |
Read the official AI-200 study guide before booking. For context, see Microsoft certifications retired in 2026.
Who should take AI-200?
- Back-end developers (.NET, Java, Python) building on Azure. This is now the Microsoft developer credential.
- Engineers who build the data and serving layer behind RAG and agent apps.
- People who were preparing for AZ-204.
Microsoft expects you to be proficient in Azure SDKs, data services, monitoring, messaging, vector databases, Python and containerised apps. AI-200 pairs well with AI-103: AI-103 covers models and agents, AI-200 covers the cloud back end around them.
Skills measured and weights
| Skill area | Weight |
|---|---|
| Develop containerized solutions on Azure | 20–25% |
| Develop AI solutions by using Azure data management services | 25–30% |
| Connect to and consume Azure services | 20–25% |
| Secure, monitor, and troubleshoot Azure solutions | 20–25% |
1. Containerized solutions (20–25%)
- Hosting:
- build, version and manage images in Azure Container Registry, including ACR Tasks;
- deploy containers to App Service with environment variables and secrets.
- Orchestration:
- Container Apps: environments, revisions and KEDA event-driven scaling;
- AKS deployments from manifest files;
- troubleshooting AKS and Container Apps from logs, events and end-to-end connectivity.
See Kubernetes for AI applications.
2. AI solutions with Azure data services (25–30%)
This is the biggest change from AZ-204.
Azure Cosmos DB for NoSQL:
- connect and query with the SDK;
- tune indexing policies, consistency levels and request units (RUs);
- store embeddings and run vector similarity search;
- process changes with the change feed.
-- Cosmos DB for NoSQL: top 5 most similar items to a query embedding
SELECT TOP 5 c.id, c.title, VectorDistance(c.embedding, @queryVector) AS score
FROM c
ORDER BY VectorDistance(c.embedding, @queryVector)Azure Database for PostgreSQL:
- connect with SDKs;
- design schemas, data types and indexes;
- reduce pgvector compute overhead and query latency;
- size compute, memory and storage for vector workloads;
- run RAG with metadata filters;
- optimise connections.
-- PostgreSQL + pgvector: nearest chunks for one tenant, cosine distance
SELECT id, content
FROM doc_chunks
WHERE tenant_id = %(tenant)s
ORDER BY embedding <=> %(query_vec)s
LIMIT 5;A full walkthrough: the pgvector RAG tutorial.
Azure Managed Redis:
- caching, expiry and invalidation;
- vector indexing for similarity search, useful for semantic caching of LLM answers.
Background: vector databases explained and embeddings explained.
3. Connect to and consume Azure services (20–25%)
- Messaging:
- Service Bus queues, topics and subscriptions, including dead-letter handling;
- Event Grid with filters, custom events and retries.
- Azure Functions: serverless APIs with triggers and bindings; configuring and deploying function apps.
4. Secure, monitor and troubleshoot (20–25%)
- Security: Key Vault secrets, including rotation and retrieval; App Configuration for settings.
- Monitoring: distributed tracing with OpenTelemetry; KQL queries over logs and metrics.
7-week study plan
| Week | Focus | Lab |
|---|---|---|
| 1 | ACR, ACR Tasks, App Service containers | Build an image with az acr build, run it on App Service with Key Vault references |
| 2 | Container Apps (revisions, KEDA), AKS manifests | Queue-driven Container App scaled by KEDA |
| 3 | Cosmos DB: SDK, indexing, consistency, RUs, change feed | Change-feed processor that embeds new documents |
| 4 | Cosmos DB vectors and PostgreSQL pgvector | Same RAG retrieval on both stores; compare latency |
| 5 | Azure Managed Redis caching and vector index; Service Bus, Event Grid | Semantic cache in front of an LLM; dead-letter handling |
| 6 | Azure Functions; Key Vault and App Configuration | HTTP-triggered function API with secrets from Key Vault |
| 7 | OpenTelemetry, KQL, troubleshooting; Practice Assessment twice | Trace one request end to end; book the exam |
20 scenario questions with answers
These are original scenarios written for this guide, not exam questions.
1. Images must build in the cloud on every commit without a local Docker install. What do you use? ACR Tasks, for example az acr build or a commit-triggered task.
2. A Container App must scale with the number of Service Bus messages. What do you configure? A KEDA scale rule on Service Bus queue length.
3. You need to test a new version on 10% of traffic. What does Container Apps provide? Revisions with traffic splitting.
4. A Cosmos DB query costs too many RUs. What do you tune first? The indexing policy (include or exclude paths, composite indexes), the query shape and the partition key.
5. A RAG app needs semantic search in Cosmos DB for NoSQL. What do you add? An embedding property with a vector index, queried with VectorDistance ordering.
6. New documents must be embedded automatically when they're written to Cosmos DB. What do you use? The change feed processor (or a Functions trigger on the change feed).
7. Retrieval must return only the current tenant's chunks. How, in PostgreSQL? A metadata filter such as WHERE tenant_id = ... combined with pgvector ordering, with suitable indexes.
8. pgvector queries are slow on a large table. What helps? An approximate vector index such as HNSW or IVFFlat, the right distance operator, enough memory, and filtering early.
9. The same LLM questions repeat and cost money. What do you add? A semantic cache: Azure Managed Redis with a vector index, returning cached answers to similar questions.
10. Messages that fail processing three times must be set aside for review. Where do they go? The Service Bus dead-letter queue.
11. Several services must react when a blob is uploaded. Which service? Event Grid, with event subscriptions and filters.
12. You need a serverless HTTP API that writes to a queue. What do you use? Azure Functions with an HTTP trigger and a queue or Service Bus output binding.
13. Database passwords must not appear in app settings or code. What do you use? Key Vault (or Key Vault references), accessed with a managed identity.
14. Feature flags and settings must change without a redeploy. What do you use? Azure App Configuration.
15. A request crosses API → Function → Cosmos DB and is slow. How do you find the slow hop? Distributed tracing with OpenTelemetry, then analyse the spans.
16. Count failed requests per operation over the last hour. What do you write? A KQL query in Log Analytics or Application Insights.
17. A pod on AKS keeps restarting. Where do you look? kubectl describe events, container logs, resource limits and liveness and readiness probes.
18. Reads in another region must never see stale data. Which Cosmos DB consistency level? Strong consistency, accepting the extra latency and cost.
19. A secret was rotated and the app must pick it up. What pattern? Read secrets at runtime from Key Vault with caching and refresh (or Key Vault references with refresh), not baked into the image.
20. An App Service container needs environment-specific settings. How? App settings (environment variables) and Key Vault references, configured per deployment slot or environment.
After AI-200
- Add AI-103 for the model and agent side, or AZ-104 for infrastructure.
- Interview practice: vector database interview questions and Azure AI services interview questions.
- See live cloud and AI jobs.
Cloud Soft Solutions in Ameerpet, Hyderabad trains .NET, Java and Python developers on Azure with real-time projects. See the .NET full-stack career path and all courses.
📞 Call +91 96660 19191 · WhatsApp · Flat No. 513, 5th Floor, Nilagiri Block, Aditya Enclave, Ameerpet, Hyderabad
Frequently Asked Questions
What replaced AZ-204?
AI-200: Developing AI Cloud Solutions on Azure, for the Azure AI Cloud Developer Associate certification. AZ-204 retired on 31 July 2026.
How long is the AI-200 exam?
Microsoft lists 120 minutes; proctored, and it may include interactive components.
What is new in AI-200 compared with AZ-204?
A 25–30% area on AI data services: vector search in Cosmos DB for NoSQL, PostgreSQL with pgvector, and Azure Managed Redis.
Which language does AI-200 use?
Python.
Is AI-200 a prerequisite for AZ-400?
Microsoft Learn currently lists Azure Administrator Associate or Azure Developer Associate as the AZ-400 prerequisite; check the AZ-400 page for updates.
Should I take AI-200 or AI-103?
AI-200 for back-end and data engineering around AI apps; AI-103 for building the AI apps and agents themselves. Many developers take both.
