You searched “AI-104” because AI-102 retired and you assumed the next exam would be the next number up. Reasonable guess. It’s also wrong.
There is no AI-104. Microsoft did not renumber AI-102 by adding one.
The Exam That Replaces AI-102 Is AI-103
AI-102 retired June 30, 2026, at 11:59 PM CST. The credential it earned, “Microsoft Certified: Azure AI Engineer Associate”, retired the same day.
The replacement is AI-103: “Developing AI Apps and Agents on Azure.” Pass it and you earn a differently named credential: “Microsoft Certified: Azure AI Apps and Agents Developer Associate.” Passing score is 700 out of 1000, same as AI-102.
So the number did jump by one. But it’s AI-103, not AI-104, and the name of both the exam and the certification changed. If you’re checking Microsoft Learn for “AI-104,” you’ll find nothing, because nothing with that number exists in the Azure AI catalog.
Why the Number Confusion Happens
Microsoft’s numbering isn’t sequential in the way people expect.
When AZ-103 was replaced, the new exam was AZ-104, one higher. That pattern trains people to add one. But Azure AI went a different route. The predecessor to AI-102 was AI-100, not AI-101. Microsoft skips numbers, reuses ranges, and picks a number when a new exam ships rather than incrementing the old one.
There’s a second reason “AI-104” feels plausible: AI-900, the fundamentals exam, also retired June 30, 2026, and was replaced by AI-901. Two Azure AI exams retiring on the same day, both with successors, makes it easy to assume a tidy numbering scheme. There isn’t one.
What AI-103 Actually Tests
This is not AI-102 with a new sticker. The content moved.
AI-103 has five domains:
| Domain | Weight |
|---|---|
| Plan and manage an Azure AI solution | 25–30% |
| Implement generative AI and agentic solutions | 30–35% |
| Implement computer vision solutions | 10–15% |
| Implement text analysis solutions | 10–15% |
| Implement information extraction solutions | 10–15% |
The generative AI and agentic domain is 30–35% of the exam, the largest single block. In AI-102, agents were a 5–10% corner. Now they’re the center of the exam. Microsoft’s audience profile is explicit: you build and deploy agents in Microsoft Foundry, and you need hands-on Python experience. Being able to describe what an agent is won’t get you a passing score. You need to have built one.
If your day job already involves deploying models, wiring up RAG, or orchestrating multi-agent workflows in Foundry, the exam maps to what you’re doing. If it doesn’t, that 30–35% is where you’ll spend most of your prep.
Where to Start
Read Microsoft’s official AI-103 study guide first. It has the full objective list, and the “skills measured” section is the exact blueprint the exam is built from.
If you were mid-prep for AI-102 when it retired, don’t assume your notes carry over cleanly. The plan-and-manage and computer vision material overlaps, but generative and agentic work is now the largest single domain, with agents rising from a 5–10% corner to the center of it, and NLP as a standalone topic is gone, folded into “text analysis” at 10–15%. We broke down that shift in detail in the AI-102 to AI-103 transition post.
Pass-IT has AI-103 practice questions grounded in the current exam guide, with per-domain tracking so you can see whether that 30–35% agentic block is where you’re losing points before you book. The AI-103 cert page has the full domain breakdown. Free to start, no card for your first session.
