Microsoft confirms that the retired Microsoft AI-900 ended on June 30, 2026. The current Azure AI Fundamentals examination is Microsoft AI-901, whose English skills guide took effect April 15, 2026. This distinction matters if you saved an old AI-900 course or study plan: its AI vocabulary is still useful, but a historic exam outline is not a booking guide. Check the current Microsoft exam listing for the language, accommodations and delivery choices applicable to your attempt.
AI-900 was a product of an earlier stage of Azure artificial intelligence. It introduced machine learning workloads, computer vision, natural language processing, conversational AI and responsible AI, mostly by asking candidates to recognize capabilities and match them to business problems. Those subjects have not disappeared. What changed is the expected depth and the platform context. Microsoft’s April 15, 2026 AI-901 objective guide emphasizes implementing lightweight applications with Microsoft Foundry, not merely identifying service names from a description.
If you already invested time in AI-900 materials, do not discard everything. Separate the durable concepts from retired exam objectives, legacy product terminology and instructions for an obsolete portal. The result is a transition path that retains what you learned while directing your practical work toward the exam that now exists.
What AI-900 taught, and why the distinction still matters
AI-900 introduced a useful vocabulary for turning business requests into technical requirements. When a manager wants to forecast demand, identify anomalies, understand customer sentiment or interpret a photograph, the first question is not which Azure screen to open. It is whether the task involves predicting a category, estimating a number, finding patterns, extracting facts or generating new content. This way of classifying the problem remains valuable regardless of the service packaging.
Its machine learning lessons emphasized the difference between supervised learning, which uses labeled examples, and unsupervised learning, which seeks patterns without those labels. Classification produces classes such as fraud or non-fraud; regression estimates a continuous quantity such as delivery time. Clustering groups similar cases when a predefined label is unavailable. These are distinct operations. A generative model can describe their results in natural language, but it does not eliminate the need to select the right operation.
Vision coverage presented image classification, object detection, optical character recognition and analysis of visual content. Language coverage addressed sentiment, entity recognition, key phrases and conversational experiences. These categories were often taught as separate Azure services. The important retained lesson is that outputs differ: a document field, a transcription, a numerical confidence score and an image description cannot be evaluated as though they were interchangeable.
Responsible AI also had a prominent place in the historic syllabus. Microsoft teaches fairness, reliability and safety, privacy and security, inclusiveness, transparency and accountability. Treating those principles as a memorized list was never enough. In practice, an inclusive interface requires accessibility testing; a private document assistant needs permission boundaries; and an accountable automated decision needs an identifiable owner and an investigation path. The dated AI-901 skills guide includes responsible AI, while its implementation objectives ask candidates to apply these principles to actual inputs, outputs and data access. Study the difference between naming a principle and proving that a particular control enforces it.
What changed when Microsoft moved to AI-901
Microsoft groups AI-901 into two weighted areas: Identify AI concepts and capabilities (40–45%) and Implement AI solutions by using Microsoft Foundry (55–60%). These are the issuer’s skills measured as of April 15, 2026, not estimates from an archived course. The implementation area expects a candidate to understand lightweight chat and agent clients, text and speech workflows, vision, and information extraction. Build a modest example for each material capability while recognizing that not every AI service or preview feature is available in every region. The balance is not an invitation to ignore responsible AI: those controls belong throughout both areas.
| Dimension | Historic AI-900 emphasis | AI-901 focus in archived outline |
|---|---|---|
| Primary learning mode | Recognize AI workload types and corresponding Azure capabilities | Recognize capabilities and implement small Foundry-backed solutions |
| Generative AI | Understand generative AI concepts and responsible use | Work with prompts, model choices, deployment and a lightweight chat client |
| Agents | Conversational AI was often discussed in a chatbot context | Create and test a single-agent solution and call it from a small client |
| Information extraction | Recognize OCR, forms and language processing tasks | Use Azure Content Understanding with documents, images, audio and video |
| Developer background | Primarily conceptual foundations | Basic Python, Azure resources and familiarity with REST, SDKs and CLIs |
The platform names also require care. Older tutorials may describe Azure Cognitive Services, Azure AI Studio or particular endpoint patterns as though they were the current default workflow. Some of these services or endpoints still exist, but naming alone is not proof that an old tutorial matches the current Foundry project model. For current learning, follow Microsoft’s active documentation and confirm whether an example uses a classic resource, a Foundry project, or a tool-specific endpoint.
A practical difference shows up in troubleshooting. In a recognition-only question, knowing that an image model can interpret a picture may be sufficient. In a current lightweight application, you also need to know which resource and model deployment you are calling, whether the application’s identity has access, what input shape is accepted, and how you will inspect an error or evaluate the returned answer. You do not need to design a vast production system for a fundamentals exam, but a plausible implementation now matters.
Keep the core AI ideas, not the old product checklist
The most productive way to reuse AI-900 notes is to extract the reasoning behind them. If an exercise explains why sentiment analysis differs from text generation, keep that concept. If a screenshot demonstrates a menu that has since moved, treat the screenshot as historical and find the new operation in the current portal. If a practice question assumes the existence of a specific retired exam skill, do not interpret its answer key as current AI-901 guidance.
For machine learning, retain the distinction between training and inference, overfitting and generalization, and classification versus regression. These concepts explain where a predictive model belongs even if their depth differs between exam versions. A business might use a classifier to route a support request, then a generative model to draft an explanation. The two stages solve different problems, need different evaluation criteria and should not be treated as interchangeable merely because both use AI.
For language, retain the difference between extraction and generation. Extracting a customer’s account number from a document should preserve the exact identifier; generating a helpful reply can tolerate varied wording but must stay grounded in authorized data. Speech recognition converts spoken input into text, while speech synthesis produces audio from text. Both may appear in the same experience, yet latency, speaker identity and sensitive data handling give each distinct failure modes.
For vision, distinguish describing a scene, reading text and extracting structured fields. A model might correctly describe a photo of an invoice and still miss the precise amount due. That is why a controlled extraction workflow needs field validation, provenance and a way to reject uncertain results. The newer computer vision work in AI-901 makes more sense when these older conceptual boundaries are already clear.
Finally, retain the habit of choosing a technology from a user requirement instead of from a brand name. Consider cost, quality, privacy, language support and operational constraints before choosing a model. An AI assistant that can call external tools introduces more authority and risk than one that only provides information. An image generator is not a replacement for a deterministic data extractor. Those practical distinctions survive every change in the Azure product catalog.
How to rebuild your knowledge around Foundry
The implementation work should start with a small environment. Choose a Microsoft Foundry project, identify its project endpoint, deploy a supported model, and practice a harmless chat request. In the current SDK model, Python applications commonly authenticate with Microsoft Entra credentials and use a project client to obtain the appropriate model interface. This is different from copying a key into an old notebook and assuming every Azure AI service uses the same endpoint.
A useful evidence exercise is to start with a synthetic returns policy, ask a deployed model to answer a question, and then change one rule in the policy. The correct outcome is not a beautifully phrased response; it is a response that respects the updated rule, identifies missing facts and cannot access a source the signed-in identity is forbidden to read. Keep a compact test record with the deployment name, SDK version, input, expected decision and actual output if you perform the exercise. The test is a proposed learning procedure here, not a claim that this site executed a Foundry tenant or validated a specific deployment.
In your first exercise, write a short system instruction explaining the assistant’s role, supply a user message and examine the response. Then change a model configuration deliberately, repeat the request and compare the results. Record the deployment name, prompt version and observed behavior. That exercise connects historical generative AI concepts to the real task of operating a small application, without needing advanced orchestration.
Next, build a simple agent in Foundry. Give it a narrow responsibility, such as answering questions from approved help-desk policy material, and test several requests in the portal. A single-agent experiment is enough to learn the difference between instructions, retrieved data and an action tool. The agent and AI workflow concepts help explain why a chatbot is not automatically an agent and why authority must come from the application rather than the model’s confidence.
Finish with an information-extraction exercise. Take a sample invoice or receipt, run an appropriate Content Understanding analyzer and inspect whether the returned fields match the source. Repeat with a non-document modality such as an image or short audio sample. The important result is not a screenshot that says the request succeeded. It is evidence that the right fields were extracted, uncertain fields were handled properly, and the application would not silently invent missing data.
How to adapt legacy study material without wasting time
Begin by labeling each source as conceptually useful, implementation-current, or historically interesting. An AI-900 question about the difference between classification and regression belongs in the first category. A current Foundry SDK quickstart belongs in the second. A retired portal walkthrough belongs in the third unless you are studying how the platform evolved. Mixing the categories creates confusion because a correct historical statement may not describe the workflow you need to perform today.
Use the official AI-901 objective wording as your map. For every topic, ask what decision you need to make and, where the objective calls for it, what small application you could build. If the objective asks for a lightweight text-analysis application, opening a glossary is not enough: you should know the input, resource, request, output and one useful validation check. If it asks you to create a single-agent solution, know how to test the agent’s instructions and inspect a conversation.
It also helps to keep version-sensitive facts out of memorized notes. Model availability, deployment options, SDK versions and regional support can change. Record where the official configuration is documented and check it when building the lab. A page or workbook containing a URL does not establish that a service remains current, and an old code sample may use a different client library from the current Foundry SDK. Understanding why the call works is more durable than memorizing its exact historical method names.
Do not carry forward claims that passing AI-900 is a required step before AI-901. The two are different exam versions in the Azure AI Fundamentals area, not a progression requiring the earlier retired exam. If your immediate goal is the current credential, prioritize its active blueprint and practice environment. If your goal is to understand the field more broadly, the historical material can still provide useful perspective.
A transition checklist for someone who studied AI-900
| Review session | Keep from AI-900 | Add for AI-901 | Evidence you understand it |
|---|---|---|---|
| Concepts | Prediction, classification, text, vision and responsible AI | GenAI workloads, agents and multimodal inputs | Match six realistic business cases to appropriate capabilities |
| Models | Training versus inference and workload matching | Foundry deployment and model settings | Explain the input, model choice and trade-off for a deployed model |
| Applications | What language and vision services can do | Basic Python SDK, chat client and agent | Send a request and inspect an error or tool boundary |
| Extraction | OCR and form recognition concepts | Content Understanding across document, image, audio and video | Validate an extracted field against its source |
This sequence is a bridge, not a fixed number of weeks or a promise of exam readiness. Someone comfortable with Python and Azure may move quickly through the application exercises. Someone new to cloud resources may need extra time for authentication, deploying a supported model and understanding endpoint conventions. Allocate effort according to the gaps that appear in your own tests rather than following a schedule because it looks tidy.
When you practice questions, use them to diagnose the reasoning failure. If you picked text generation for a request that needs exact field extraction, revisit the output requirements. If you selected an agent for a simple lookup, ask whether autonomy solves a real problem. If your sample application fails, first check resource scope, identity and deployment name before assuming the model is broken. These habits are useful in work as well as in the exam.
What an AI-900 article can—and cannot—do now
A historical syllabus is useful when it makes the fundamentals clearer and shows which assumptions changed. It is not a substitute for Microsoft’s current scheduling page, objective list or hands-on Foundry instructions. This article intentionally preserves the distinction. AI-900 remains the name of the retired exam; AI-901 is the code to use when preparing for today’s Azure AI Fundamentals assessment.
The most valuable transfer is conceptual: knowing which task you are solving, why a model or service fits, how to recognize risk and how to verify an output. Build the newer practical skills on top of those foundations. That approach respects what the old exam taught without pretending that its legacy practice questions, screenshots or service names constitute an up-to-date study plan.
Continue exploring the technologies and certification concepts that matter most to you.