AI-900 vs AI-901: What Changed in Azure AI Fundamentals

Exam AI-900 retired on 30 June 2026 and AI-901 took its place. The credential itself — Microsoft Certified: Azure AI Fundamentals — did not retire, so if you already passed AI-900 your certification stays valid on your transcript and there is nothing to redo. If you have not sat it yet, AI-901 is the only route.

The important part is that this is not a version bump. AI-900 was organised around machine learning principles and a tour of one Azure service per workload. AI-901 keeps a concepts half and replaces the rest with implementation on Microsoft Foundry, which now carries the majority of the marks. A study plan built for the old exam spends most of its time in the wrong place.

Underneath that sits a naming problem that catches almost everyone: Microsoft renamed the products the exam is about. Below is what changed domain by domain, which names are now wrong, and how to tell whether a course or practice test predates the switch. Gnoseed does not sell, proctor or guarantee this exam — it is a free study companion. Confirm every exam detail on Microsoft Learn before you book, since Microsoft updates the skills measured periodically. If you are weighing this exam against another vendor, see AI-901 vs AWS AI Practitioner; if you are looking at the associate tier above it, the same replacement story played out very differently for AI-102 and AI-103.

AI-900 spread its marks across a familiar fundamentals shape: AI workloads and considerations, machine learning principles on Azure, computer vision, natural language processing, and generative AI. It rewarded recognising which Azure service matched which workload.

AI-901 publishes two skills-measured domains: Identify AI concepts and capabilities (40–45%) and Implement AI solutions by using Microsoft Foundry (55–60%). The concepts domain covers the six responsible AI principles, how generative models are configured and chosen, and which workload family a business problem belongs to. The implementation domain covers building generative apps and agents in Foundry, text and speech, computer vision and image generation, and information extraction with Content Understanding.

The practical consequence: more than half the exam is now implementation. Classification-versus-regression style questions and the old service-matching drills no longer carry the paper. If your revision is mostly "which Azure service does X", you are studying the retired exam.

Microsoft consolidated its AI platform at the same time, and the old names are still everywhere in tutorials, videos and practice tests. Azure AI Studio and Azure AI Foundry are now Microsoft Foundry. Azure AI Services — and before that, Cognitive Services — are now Foundry Tools. The Assistants API became the Responses API, and its vocabulary of threads, messages, runs and assistants became conversations, items, responses and agent versions.

The resource model changed with it. Where you previously combined a hub, an Azure OpenAI resource and an Azure AI Services resource, there is now a single Foundry resource containing projects. Hub-based projects still exist, but in the Foundry (classic) portal, and new investment goes to the new one.

This is worth more than trivia on exam day, because the wrong names make good distractors. A question that offers "Azure AI Foundry" alongside "Microsoft Foundry" is testing exactly whether you learned the platform this year or last year.

Open any course, video or practice test you are considering and look for three tells. Does it list five or six skill areas rather than two? Does it say Azure AI Foundry or Azure AI Services in the present tense? Does it teach regression, classification and clustering as exam content? Any one of those means it targets AI-900.

The reverse tell is just as useful: current material talks about prompt agents and hosted agents, the Foundry model catalog, Content Understanding analyzers, and guardrails with their intervention points. None of those existed as exam content before April 2026.

Microsoft states the skills measured as of a date on the study guide itself — at the time of writing, 15 April 2026. Check that date first; it is the fastest way to tell whether what you are reading has been revised.

Start with the responsible AI principles. They are six named ideas, they appear across both domains, and they are the cheapest marks on the paper because they need no Azure access to learn.

Then take the concepts half properly: how generative models are configured, what the deployment types mean for data residency and cost, and which workload family — generative and agentic, text analysis, speech, vision, information extraction — a given scenario belongs to. This is where "which tool" questions now live.

Leave Foundry implementation for last but give it the most time, because it is the majority of the paper and the part you cannot bluff. Create a free Foundry project, deploy a model, and build one prompt agent. An afternoon in the portal is worth more than a week of reading, and the exam assumes you have seen the SDK and the portal rather than only read about them.