AI-200 Exam Guide: Azure AI Cloud Developer Associate Explained

AI-200, Developing AI Cloud Solutions on Azure, is the exam for the Azure AI Cloud Developer Associate certification. It replaced AZ-204, which retired on 31 July 2026. It tests the back end of AI apps: containers, vector search in Cosmos DB and PostgreSQL, Redis, messaging, Functions, Key Vault, OpenTelemetry and KQL. Python is the expected language, and 700 passes.

At a glance

• Exam: AI-200, Developing AI Cloud Solutions on Azure
• Certification: Azure AI Cloud Developer Associate, intermediate level, renewed free every year
• Replaces: AZ-204 and the Azure Developer Associate certification, retired 31 July 2026
• Available since: 9 July 2026, according to Microsoft's Partner Center announcements
• Skills measured: as of 7 October 2026, in four areas
• Language: Python
• Exam time: 100 minutes, or 120 if the exam contains labs
• Pass mark: 700

What happened to AZ-204

Microsoft retired AZ-204 on 31 July 2026, together with the Azure Developer Associate certification and its renewal assessment. Microsoft's July 2026 Partner Center announcements name AI-200 as the replacement, available from 9 July 2026. The official course changed too: AZ-204T00 retired on 29 May 2026 and was replaced by AI-200T00, Develop AI cloud solutions on Microsoft Azure.

If you already hold Azure Developer Associate, the credential retirement page says it stays in the Active section of your transcript until it expires, then moves to history. You cannot renew it. For employers that are Microsoft partners, specialization requirements now accept Azure Developer Associate or Azure AI Cloud Developer Associate, so AI-200 is the one to plan for. If you are heading for DevOps Engineer Expert, check that certification's page for the current prerequisite list before you book.

Who AI-200 is for

Microsoft's audience profile describes a developer who contributes to every phase of implementing AI solutions on Azure, "with an emphasis on back-end services and components", from requirements gathering to monitoring. You should be proficient with Azure SDKs, Azure data services, monitoring, messaging and eventing, vector databases, containers and Python.

In practice that is the engineer who makes an AI feature work in production: storing embeddings, scaling the API, queueing document ingestion, caching answers, protecting secrets and tracing a slow request. The model and agent work itself sits mostly in AI-103.

AI-200, AZ-204 and AI-103 compared

Compared with AZ-204, the AI-200 skills list no longer includes API Management, Blob Storage, Microsoft Graph, user sign-in with the Microsoft identity platform, shared access signatures, deployment slots, Container Instances, Event Hubs or Queue Storage. It adds Azure Kubernetes Service, KEDA scaling in Container Apps, Container Registry Tasks, PostgreSQL with pgvector, Azure Managed Redis, vector similarity search, retrieval-augmented generation (RAG) with metadata filters, OpenTelemetry and KQL.

Which first? Developers who build services, APIs and data stores should start with AI-200. Developers who build prompts, agents and model evaluations in Microsoft Foundry should start with AI-103. Together they cover a complete AI application.

The four skill areas and what to practise

The AI-200 study guide is the authority. Here is how to turn each area into lab work.

Develop containerized solutions on Azure (20 to 25%)

Build, version and manage images in Azure Container Registry, including with ACR Tasks. Deploy containers to App Service with environment variables and secrets. Deploy to Container Apps with environment configuration, revisions and KEDA event-driven scaling. Deploy to AKS with manifest files, and troubleshoot through logs, events and connectivity checks.

Practise: build an image with an ACR Task, deploy it to Container Apps, then add a KEDA rule that scales on queue length.

Develop AI solutions by using Azure data management services (30 to 35%)

The heaviest area. One trap: the section heading inside the study guide still shows 25 to 30%, but the skills at a glance list and the change log show that its weight increased on 7 October 2026. Plan for 30 to 35%.

Cosmos DB for NoSQL: SDK queries, Request Unit tuning with indexing policies and consistency levels, embeddings with vector similarity search, and the change feed processor. A vector query in the pattern Microsoft's Python quickstart uses:

results = container.query_items(
    query=(
        "SELECT TOP 5 c.title, VectorDistance(c.embedding, @embedding) AS score "
        "FROM c ORDER BY VectorDistance(c.embedding, @embedding)"
    ),
    parameters=[{"name": "@embedding", "value": query_embedding}],
    enable_cross_partition_query=True,
)

Azure Database for PostgreSQL: schema and index design, pgvector tuning, compute sizing for vector workloads, RAG with metadata filters and connection optimisation. A classic gotcha from the pgvector documentation: you must add the extension to the server's azure.extensions allowlist before CREATE EXTENSION, and its name is vector, not pgvector. You also create it separately in each database.

SHOW azure.extensions;
CREATE EXTENSION IF NOT EXISTS vector;

SELECT id, title
FROM policy_chunks
WHERE department = 'finance'
ORDER BY embedding <=> '[0.12, 0.03, 0.88]'::vector
LIMIT 5;

Azure Managed Redis: caching, expiration and invalidation, plus vector indexing for similarity search.

Connect to and consume Azure services (20 to 25%)

Service Bus queues, topics, subscriptions and dead-letter handling; Event Grid filters, custom events and retries; Azure Functions triggers, bindings, configuration and deployment.

Practise: send a document-ingestion message to a Service Bus queue, process it in a Function, then deliberately send a malformed message and watch it land in the dead-letter queue.

Secure, monitor and troubleshoot Azure solutions (20 to 25%)

Key Vault secrets with rotation and retrieval, Azure App Configuration, distributed tracing with OpenTelemetry SDKs and KQL over logs and metrics. Get fluent enough in KQL to answer "which operation fails most?" in a minute:

requests
| where timestamp > ago(1d)
| summarize failures = countif(success == false), total = count() by name
| order by failures desc

Build one project in the Singapore region

The fastest way to cover the outline is one small RAG back end, deployed in Azure's Southeast Asia region (southeastasia), which is physically in Singapore and paired with East Asia:

1. A Function triggered from a Service Bus queue chunks documents and writes embeddings to Cosmos DB or PostgreSQL.
2. An API in Container Apps runs vector search and caches frequent answers in Redis with an expiry.
3. Secrets live in Key Vault, settings in App Configuration.
4. OpenTelemetry sends traces to Application Insights, and a KQL query shows failures and latency.

Check that each service and feature you need is offered in Southeast Asia before you start, because some are limited to certain regions. Use public or synthetic documents. If you practise on work data that contains personal data, the PDPA still applies; our PDPA checklist for AI chatbots covers the basics. If your vector work lives mainly in SQL Server or Azure SQL instead, our DP-800 guide covers that route.

An eight-week study plan

The free practice assessment is on Microsoft's AI Skills Navigator, and you must sign in to launch it. During the exam you can open Microsoft Learn in a split screen, but the timer keeps running, so use it to check a parameter, not to learn a topic. You can schedule up to 90 days ahead through Pearson VUE, and if the exam isn't available in your preferred language you can request an extra 30 minutes. Check the exam page for the current price in Singapore.

How CertTulen Academy can help

CertTulen Academy is a Microsoft Training Services Partner running live, MCT-led classes with hands-on labs. Azure AI Cloud Developer Associate (AI-200) follows this outline and includes an exam voucher, mock practice tests and lab access (see the course page for details). If you will build the model and agent side too, pair it with Azure AI Apps and Agents Developer Associate (AI-103). Contact us or WhatsApp +60 16-241 8686 for a batch that suits your time zone.