AI-200 vs AI-103: Which Azure AI Developer Exam Should You Take?

Take AI-103 if you build with models, and AI-200 if you build the platform underneath them. The two exams launched in the same wave, share the AI prefix, and both earn an Azure developer associate certification, which is why people book the wrong one. The fastest way to separate them is to count a single word in each official skills outline. "Agent" appears 21 times in AI-103's and zero times in AI-200's. "Container" appears 11 times in AI-200's and zero times in AI-103's. They are peers aimed at different jobs, not a beginner and an advanced version of the same thing.

This is not a hypothetical mix-up. AI-200's name is Developing AI Cloud Solutions on Azure and AI-103's is Developing AI Apps and Agents on Azure. Read those two titles cold and tell me which one covers prompt engineering. Most people guess AI-200 because the number is higher, and they guess wrong.

The one table that settles it

Every figure below is a count from Microsoft's own published skills outline for each exam, taken in July 2026.

What AI-103 actually tests

AI-103 is the exam for the person writing the code that talks to a model. Its largest domain, at 30 to 35 percent, is "Implement generative AI and agentic solutions," and the bullets under it are exactly what you would hope: deploy and consume LLMs, small models and multimodal models, implement retrieval-augmented generation in an application, design tool-augmented flows and multistep reasoning pipelines, and evaluate apps for fabrications, relevance, quality and safety.

Then it goes further into agent territory than any Microsoft exam has before. You are expected to define agent roles, goals, conversation-tracking approaches and tool schemas. Build agents that integrate retrieval, function calling and conversation memory. Implement orchestrated multi-agent solutions. Build autonomous or semiautonomous workflows with safeguards and approval flow controls. If your job involves shipping something that decides what to do next on a user's behalf, this is your exam.

The second domain, at 25 to 30 percent, is the governance around all that: choosing models, managing quotas and cost footprints, monitoring drift and grounding quality, configuring safety filters and guardrails, and governing agent behavior with oversight modes and tool-access controls. The remaining three domains, computer vision, text analysis and information extraction, are 10 to 15 percent each. Those are the classic Cognitive Services topics that used to make up most of AI-102, and they are now a minority of the paper.

What AI-200 actually tests

AI-200 sounds like the AI exam and is not. Its biggest domain is Azure data management services, and what it means by that is Cosmos DB with vector similarity search, PostgreSQL sized for a pgvector workload, and Redis vector indexing. Around that sit containers (Container Registry, App Service, Container Apps, AKS), messaging (Service Bus, Event Grid, Functions), and observability (Key Vault, OpenTelemetry, KQL queries).

Read that list again. It is AZ-204's skill set with a vector database bolted on, which makes complete sense once you remember AI-200 is the exam replacing AZ-204. Microsoft did not invent a new AI developer role here. It took the existing Azure developer role and updated it for a world where the app you are deploying happens to have a model in it.

The practical implication is a pleasant one if you already hold AZ-204: your containers, Functions, Service Bus, Event Grid, Cosmos DB and Key Vault knowledge transfers directly. What you need to add is the vector layer and Python, which Microsoft names explicitly in the recommended proficiency. What you do not need is a single hour of prompt engineering study, because there is not one prompt question on the exam.

How to choose in ten seconds

Ask yourself what you would be doing on a Tuesday afternoon.

• Writing a system prompt, wiring a tool schema, checking whether the model hallucinated, deciding where a human approval step goes: AI-103.
• Sizing a pgvector index, debugging a Container Apps revision, tracing a request through Service Bus with OpenTelemetry: AI-200.
• Both, because you are the only engineer: start with AI-200 if you hold AZ-204 already, because most of it is a delta on what you know. Then take AI-103.

One thing to be clear about: neither is a prerequisite for the other, and AI-200 is not "AI-103 but harder." Microsoft lists no formal prerequisite for either. The numbering implies a ladder that does not exist.

Which one your employer probably means

If a job posting or a manager says "get the Azure AI certification," they almost always mean the one that used to be AI-102, the Azure AI Engineer Associate. That certification retired on June 30, 2026, and its successor is AI-103. If a posting still lists AI-102 as a requirement, it was written before the retirement and has not been updated. Point them at AI-103.

If instead the ask is "get the Azure developer certification," that was AZ-204, which retires on July 31, 2026, and its successor is AI-200. The AI in the name is the confusing part, because nothing about that exam requires you to understand how a language model works.

Neither exam has an official question bank

This is the part worth planning around. Microsoft states on both certification pages that the practice assessment is not currently available, and that practice assessments usually appear within 8 weeks of an exam leaving beta and becoming generally available. So for both exams, right now, there is no official bank from anyone at any price.

Microsoft also publishes no per-exam question count and no pass rate. It never has, for any exam. Its only public guidance is that most exams "typically contain between 40-60 questions." When a site advertises "AI-103: 65 questions, 82% pass rate," both numbers were invented, and the specificity is the sales pitch.

There is a second, sneakier problem with the recycled banks that do exist for AI-103. They are recycled from AI-102, and AI-102 is the one exam whose vocabulary Microsoft just rewrote. The phrase "Azure OpenAI" appeared five times in AI-102's outline and appears zero times in AI-103's, because model work now runs through Microsoft Foundry. Drill an AI-102 bank and you will learn the right concepts under the wrong names, which is exactly the failure mode a multiple-choice exam punishes.

The workaround that actually holds up for a new exam is to generate practice questions from the current source material rather than waiting for a bank. Microsoft publishes the full outline and the documentation behind every bullet in it, and that material is current by definition. Upload it, along with your own architecture notes, and turn your own documentation into exam questions with an answer key. If your team runs a customer-facing assistant in production, the same trick works on your internal runbooks, and the notes you keep for a chatbot that trains on your own content tend to make unusually good exam material, because they are written in the operational language the exam now uses.

The wider retirement wave

AI-103 and AI-200 are two of six certifications changing hands in Microsoft's Cloud and AI rebrand. Three of the retirements have already happened.

If you already hold one of the retired certifications, it stays valid until its existing expiration date. What you lose is the renewal, because the renewal assessment retires alongside the exam.

Start practicing

Both exams are new enough that the honest answer to "what will be on it" is: exactly what Microsoft published, and nothing else has been verified by anyone. Build your practice from that.

For the agents exam, the AI-103 practice test generator covers all five domains and the Foundry vocabulary shift. For the backend exam, the AI-200 practice test generator lays out the container and vector material. If you are still deciding whether to sit AZ-204 before it dies, the AZ-204 practice test generator has a what-to-do-by-situation table. And if you are coming up from the fundamentals level, note that AI-900 is gone too: the AI-901 practice test generator covers its replacement.