Historical status as of October 2026: Microsoft retired the AI-900 examination on June 30, 2026. This article explains its historical Azure AI Fundamentals subject matter and what remains worth learning; the retired exam cannot be booked. Anyone pursuing the active certification should consult Microsoft Azure AI Fundamentals (AI-901) and its current requirements.
The transition matters because AI-900 emphasized recognition of workload categories and Azure service purposes, while AI-901 includes more practical implementation with Microsoft Foundry. Old explanations of regression, computer vision or responsible AI can still be useful, but they are not a substitute for the current Python, resource, prompting and lightweight application skills.
The retired AI-900 syllabus and its continuing value
Microsoft’s historical AI-900 examination objectives described workloads and considerations, machine learning, computer vision, language processing and generative AI. Those were the old examination domains; they should not be presented as topics candidates will encounter in a new AI-900 sitting. The former assessment tested recognition of concepts rather than independent delivery of a production application.
Regression estimates a numerical value, classification predicts among categories and clustering finds groups without preassigned labels. Neural networks can support several kinds of tasks, but their presence does not guarantee accurate predictions. These distinctions survive a change in exam code because they explain what a model is supposed to do and how its output might be evaluated.
Older Azure examples distinguished generic image analysis from structured document extraction, speech-to-text from text-to-speech, and predictive machine learning from generation. For an invoice, the necessary output might be dates, amounts and supplier names rather than a descriptive image label. That kind of workload reasoning remains useful even if a service has been renamed.
Historical lesson screenshots often show older Azure AI services, portals, or configuration controls. Interface names, SDK versions and restrictions evolve. Retain an old diagram only when it supports a concept accurately; use current Microsoft documentation for any executable instruction or product-availability claim.
Old exam durations, scoring advice and booking instructions have no current registration value. The active AI-901 examination has its own versioned guide and delivery policies. Reused AI-900 practice questions should be explicitly identified as historical concept exercises, never promised as a valid prediction of current exam content.
Current AI-901 focuses on implementation
Microsoft’s AI-901 assessed skills, effective April 15, 2026, divide the examination into identifying AI concepts and capabilities (40–45%) and implementing AI solutions with Microsoft Foundry (55–60%). This differs from the old AI-900 blueprint. Learners should understand both workload selection and the steps needed to build and test a small solution.
The active audience includes people new to AI solution development with foundational Python syntax and Azure resource knowledge. This does not mean a beginner must be an experienced developer. It does mean that candidates should practise authenticating a client, constructing an input, interpreting a response and handling a failed request instead of relying entirely on theory videos.
Current implementation areas include deploying and interacting with a model in Foundry, writing system and user prompts, building a lightweight chat client with the Foundry SDK and testing a single-agent solution. The outline also includes examples involving text, speech, vision, image generation and multimodal extraction. These are public domains, not guarantees about a particular item format.
An introductory exercise might access a model from a nonproduction project, submit a reviewed prompt and compare the answer with an expected result. Change the instruction to see how the output differs. Then test an invalid request or unavailable endpoint so the client handles error conditions deliberately rather than displaying a misleading success.
The exact portal and API experience may change. Use Microsoft’s currently published objectives, not a stale course screenshot, before registration. A learner may retain useful old AI-900 explanations but needs practical current implementation exercises to address the new Foundry emphasis.
Model evaluation and generative AI fundamentals
A reliable workflow begins with a clear task. Regression might predict demand, classification could distinguish document categories, clustering might reveal patterns and generative systems can produce text or imagery. A model should be chosen for the required input and output, not merely because one approach is more fashionable. Simpler methods are often easier to evaluate.
Representative data matters. Inconsistent labels, duplicate examples, missing records and information leakage into the evaluation set can make performance look better than it is. State what failures would be costly, select useful metrics and compare the system against a baseline rather than trusting a single successful demonstration.
Generative models sometimes produce persuasive but unsupported statements. A chat client should make it possible to compare an answer against trusted material and identify missing information. An elaborate prompt does not eliminate hallucination or substitute for validation of results.
Agents introduce actions beyond content generation. An agent’s external tools should verify the user’s authorization and use the least privileges necessary for a narrow task. Prompt instructions cannot grant access to unrelated records. Test both normal and refused tool actions, and record any difference between what the model claimed and what the system actually did.
A useful learner experiment involves a small set of reviewed questions, several prompt approaches and a record of inaccurate or incomplete outputs. That exercise shows how model behavior changes with context and why quality assurance needs negative test cases. It is more valuable than merely memorizing the word generative.
Vision, language, speech and multimodal information
Computer vision can classify an image, locate objects or help interpret visual input. Structured document processing answers a more specific question, such as extracting a date, amount or invoice identifier. Design the evaluation around the fields the workflow actually needs, and do not treat a plausible description as proof that the extracted values are correct.
Some Azure Face capabilities changed since older study materials were created. Microsoft’s Azure Face attribute restrictions identify emotion and gender inference as retired and limit access to several other characteristics, including age, smile, facial hair, hair and makeup. Historical exam examples must not imply that these analyses are freely available today.
Language tasks include summarization, translation, entity recognition, key-phrase extraction and sentiment analysis. Speech workflows recognize spoken words or produce synthesized speech. Practical accuracy can vary with language, background noise, regional pronunciations and specialized vocabulary. Test representative cases rather than relying on a single easy sample.
AI-901 information-extraction topics also involve documents, images, audio and video through Foundry capabilities such as Content Understanding. A meaningful exercise compares required structured fields with manually reviewed results and records missing, low-confidence or inconsistent values. A syntactically correct JSON response can still be factually wrong.
Try contrasting a clean invoice with a blurred scan or an audio recording containing an unfamiliar name. Describe the expected data before running the model. When a field is wrong, identify what downstream application behavior would be safe, including flagging it for human review instead of confidently accepting an error.
Responsible AI controls in practice
The old AI-900 syllabus introduced fairness, reliability, privacy, inclusiveness, transparency and accountability. These concepts remain relevant, but they have practical implications. Fairness requires appropriate evaluation, privacy calls for data boundaries, and reliability depends on failure handling rather than simply producing fluent answers.
Document what information enters a model, whether it appears in logs, who can read an output and whether a tool has permission to alter external state. Authorization should be enforced outside model text. An agent should not expose personal data because a prompt requests it.
Transparency should help users understand that content was generated, what evidence supports it and how a mistake may be challenged. For consequential decisions, define appropriate human oversight and an escalation path instead of treating a model’s answer as final authority.
Evaluation should include bad cases: hallucinated summaries, privacy-sensitive requests, false extracted amounts, unexpected language variations and prohibited tool use. Record each issue, apply a specific control and rerun the test to confirm the behavior changed as intended.
A beginner can test access boundaries with synthetic records and a narrowly scoped tool. Verify that authorized reads succeed and unauthorized ones fail; inspect the application’s actual response rather than accepting a model’s explanation of what it did. That builds a distinction between persuasive output and controlled system behavior.
An AI-901 preparation plan grounded in current tools
Start from the active AI-901 objectives. For concepts, explain workload choice, model limitations and responsible use. For implementation, practise basic Python, Azure resource setup, prompting and a simple Foundry client. Use synthetic or nonproduction data rather than exposing real customer content.
Build a small chat client that makes a request, displays its result and handles an error. Run a few reviewed questions, vary the prompt and track where the response becomes incomplete or unsupported. A useful test record contains expected and actual behavior, not only screenshots.
Next examine a narrow single-agent workflow whose tools have explicit permissions. Simulate a failed tool call and an out-of-scope instruction, then confirm the application rejects unauthorized actions and communicates errors appropriately. Complexity is not the goal; clear boundaries and verifiable outcomes are.
For vision or information extraction, define fields that a model should return and evaluate them against reviewed examples. Speech exercises should include ambiguous or noisy input. Record uncertainty and missing values, then decide when to require a person to check the output rather than accepting plausible but incorrect data.
Finally, consult Microsoft for current AI-901 exam registration, delivery, language and support rules. Preserve AI-900 historical lessons only where they accurately explain durable concepts. This page should help learners understand the transition and plan realistic introductory AI practice, not imply that the retired AI-900 exam remains obtainable.