If you were preparing for AI-102 (Azure AI Engineer Associate), stop and read this first. Microsoft retired AI-102 on June 30, 2026. Its replacement is AI-103: Developing AI Apps and Agents on Azure, which leads to the new Azure AI Apps and Agents Developer Associate certification.
Quick answers
- Can I still take AI-102? No. It retired on June 30, 2026.
- Is my AI-102 certification still valid? Yes. Microsoft says retirement does not revoke certifications already earned. They stay valid until they expire.
- What should I take now? AI-103. It is the direct replacement.
- Passing score? 700 out of 1000, like other Microsoft role based exams.
What changed from AI-102 to AI-103
AI-102 was a broad exam about Azure AI services: vision, language, speech, search and a small amount of generative AI. AI-103 is built around generative AI and agents on Microsoft Foundry. The skills outline (as of April 16, 2026) is weighted like this:
| Domain | Weight |
|---|---|
| Plan and manage an Azure AI solution | 25 to 30% |
| Implement generative AI and agentic solutions | 30 to 35% |
| Implement computer vision solutions | 10 to 15% |
| Implement text analysis solutions | 10 to 15% |
| Implement information extraction solutions | 10 to 15% |
The biggest shifts:
- Agents are now central. Expect questions on agent roles, tool schemas, function calling, multi agent orchestration, and approval flows for risky actions.
- Security is keyless. Managed identity, private networking and role based access come up far more than API keys.
- Vision is generative. Image and video generation, mask based editing, alt text and multimodal understanding replace much of the old custom vision content.
- Information extraction runs through Content Understanding and retrieval pipelines for RAG, not only classic Document Intelligence models.
- Python is the expected language.
Try three AI-103 style questions
A web app must call a Foundry model deployment without storing any API keys, including in Key Vault. What should you do?
Correct answer: B
A managed identity lets the app get Microsoft Entra ID tokens at runtime, so no secret exists to store or leak.
An agent can issue refunds through a tool. Compliance requires a human to approve each refund before it runs. What should you implement?
Correct answer: D
Tool approval is an enforced human in the loop control. Instructions alone are not.
Your RAG answers sometimes contain claims that are not in the retrieved documents. Which evaluator measures this?
Correct answer: A
Groundedness checks whether the answer is supported by the provided context.
How to prepare
- Read the official AI-103 study guide on Microsoft Learn and note every sub skill.
- Build hands on: a RAG app with citations, an agent with a function tool and an approval step, and a Content Understanding analyzer.
- Practice with scenario questions and read every explanation, not only the answer.
- Take Microsoft’s official practice assessment in the final week.
Try the free practice questions on our AI-103 page, then get the complete guide: original questions mapped to the current outline, every option explained, a four week study plan and a cheat sheet.