Microsoft Azure AI Fundamentals is no longer examined through AI-900. Microsoft retired that exam on June 30, 2026; the current examination is AI-901. The certificate name remains Azure AI Fundamentals, but a current candidate should use the AI-901 objective list rather than an old AI-900 course outline. The change is substantive: the updated exam explicitly expects beginner implementation knowledge using Microsoft Foundry, not only recognition of AI concepts. It would be misleading to describe the new test as a pure vocabulary quiz.
A useful starting point is the AI-901 exam destination, alongside the AI-901 skills outline. Microsoft’s English-language blueprint identifies skills measured as of April 15, 2026. Always check the edition of any lesson or practice question. A tutorial demonstrating only retired AI-900 screens may still teach a useful concept, but it is not proof that it covers the new implementation requirements.
Know the two domains before choosing study resources
The current objectives assign 40–45% to identifying AI concepts and capabilities and 55–60% to implementing AI solutions through Microsoft Foundry. The first domain covers responsible AI, model components, deployment options and common workloads. The second goes further: prompt design, model deployment and testing, a lightweight Python client, a single-agent solution, text and speech, image input and generation, and extraction from documents and multimedia. Those percentages are ranges, not permission to ignore one domain because the other has more weight.
The old AI-900 examination emphasized a broader collection of conceptual categories. AI-901 retains important ideas such as computer vision and language processing but connects them to practical Foundry steps. The exam’s audience statement expects basic familiarity with Python coding syntax and Azure resources. A beginner does not need to operate a production-scale service, but should recognize credentials, a project endpoint, a model deployment name, and what an application actually sends to a model.
Learn what an AI workload does before naming a product
Start with a business problem rather than the service catalog. An airline needs to classify complaint themes, a hospital has handwritten intake forms, and a warehouse wants a searchable inventory assistant. Classification of complaints, document extraction, and retrieval-grounded question answering have different input assumptions and quality measures. A system may use more than one, but they are not interchangeable. To practice these decisions, compare the AI workload map with a real example from work or study.
A model that produces fluent text may struggle to identify a particular field in a scanned invoice reliably. A deterministic extractor may return a valid date while failing to understand a customer’s broader question. Select a technique based on the required output and evidence. When uncertain, a workflow can request clarification or hand a document to a reviewer rather than pretend a confident answer.
Understand the difference between a model, a deployment, and an agent
A model is the learned capability used for inference. A deployment makes a selected model accessible under an application-facing name and appropriate capacity or processing configuration. An agent adds an instructed task, optional tools and orchestration behavior around model interactions. A chat client can call a model without becoming an autonomous agent, and an agent should not acquire business permissions simply because it has a tool listed in its configuration.
Begin with the Foundry first-model walkthrough to see the resources and names involved. Then study the single-agent workflow as a separate concept. Keep the distinction clear on scenario questions: if the task is a one-off explanation, direct model inference may be enough; if a system must execute a bounded multi-step process, an agent may be appropriate, but it brings additional controls.
Build a beginner lab that produces evidence
Choose a fictional help-desk dataset with ten short requests and three expected categories. Write down what a correct answer looks like, which requests need human review, and the evidence source. First ask a model to classify them using a consistent prompt; compare the output against manually labeled expectations. Next change the prompt to request structured fields and see whether the error rate changes. The exercise is more valuable when mistakes are recorded than when screenshots only show a successful sample.
After that, try a small image or audio example with permitted nonsensitive material. Identify what the model actually observes, whether OCR or speech recognition is needed, and what happens if the input is ambiguous. The integrated beginner lab combines these concepts but keeps the task small enough to troubleshoot. Do not upload real customer records to a disposable learning environment.
Treat responsible AI as a design decision
Fairness means looking for different error rates across relevant groups, not announcing that a model is fair because it has a safety filter. Reliability includes handling unavailable resources and unclear inputs. Privacy requires minimizing collected information and governing its retention. Inclusiveness means testing for the people who will use the service. Transparency explains what is automated and what evidence the answer is based on; accountability assigns an owner and an escalation path. These principles influence whether an otherwise accurate AI feature can be safely used.
The responsible-AI failure scenarios connect Microsoft’s six principles to practical examples. For a help-desk classifier, an error that routes a normal request to the wrong queue may be annoying. A similar error in a health or benefits system can have a more consequential outcome. Decide the required oversight from the use case and consequences, not just the model’s apparent fluency.
Make a four-stage study schedule
First, learn the two-domain outline and build a short glossary: inference, model deployment, token, multimodal, agent, grounding, speech recognition, synthesis, OCR and extraction. Second, deploy or study a Foundry example and sketch a lightweight Python client that authenticates, passes a deployment name and handles a response. Third, test text, audio, vision and document scenarios; record when one approach is more appropriate than another. Finally, perform mixed scenario exercises, explain incorrect choices and use the official objective list to identify remaining weak areas.
Practice questions are useful only when their explanations match the current objective version and actual platform concepts. Avoid memorizing a vendor’s temporary portal button positions as though they are stable exam facts. Region, model availability, quota and preview status may change. Microsoft may include commonly used preview capabilities, but an examiner can also test sound design decisions that remain correct when the interface changes.
Verify administrative details at the issuer
Microsoft lists a passing score of 700 for AI-901. That is a scaled passing threshold, not a promise that exactly 70 percent of items will produce a pass. Exam pricing varies by location and can change; do not rely on a price copied from another country’s page. Check available language versions and accommodations in the official scheduling process. The certification page does not currently show a retirement date for AI-901, but lifecycle decisions can change and should be verified near your test appointment.
When choosing whether to proceed after fundamentals, separate credentials by role. AI-901 is an entry point; AI-103 applications and agents covers substantially deeper development responsibilities. A new learner benefits from recognizing this progression without pretending that passing AI-901 is a prerequisite for every advanced role. The goal of preparation is to explain a safe, working beginner AI solution and why its pieces fit the business task.