Microsoft AI-900: Retired Azure AI Fundamentals and the AI-901 Path

AI-900: Microsoft Azure AI Fundamentals retired on June 30, 2026. This is now a historical certification-exam page, not a live scheduling destination. The exam gave beginners a way to understand AI workloads, responsible AI, machine learning, computer vision, natural language processing and generative AI through the Azure service portfolio. Those concepts continue to matter even though Microsoft's current fundamentals assessment is AI-901, with a stronger orientation toward Microsoft Foundry and practical interaction with AI services.

The version distinction matters because exam code and skill set are not interchangeable. A learner reading an old AI-900 preparation course may encounter a service name or weighting that has been revised since its publication. That material can still explain the difference between classification and generation, but Microsoft’s live AI-901 study guide—not an old practice-question collection—determines the current testing objectives. A person who already earned a credential should verify its transcript and expiration details through Microsoft rather than infer them from an exam article.

Understand what the original fundamentals assessment taught

AI-900 introduced AI as a collection of workloads rather than one universal model. Predicting a numeric value from historical observations is a regression task; assigning a category is classification; identifying unusual records can require anomaly detection. Conversational question answering and text generation are different again. The exam asked learners to identify which general method might suit a business problem before they explored the implementation details of a particular Azure product.

Consider an insurer examining claims. Detecting whether a claim resembles known fraud patterns is a predictive analytics problem that needs labeled examples and carefully measured error costs. Writing a concise explanation of policy language is a generative problem whose output must be grounded in approved documents. Reading an invoice image is an extraction problem. The application might combine all three, but a reliable design gives each component its own data sources, confidence thresholds and review procedures. A single chatbot should not be allowed to silently make financial decisions.

Fundamentals knowledge also involves recognizing where AI adds little value. A policy that says a reimbursement may not exceed a fixed limit should be enforced through a deterministic rule. A model may help interpret a free-text request, but the final threshold should remain testable and auditable. This principle continues to be useful when comparing expensive model deployments with simpler automation, and it is a helpful starting point for someone building practical skills beyond a certification.

Separate machine learning from generative AI

Traditional machine learning learns relationships from data to predict labels, values or groupings. Supervised learning needs examples with suitable targets; unsupervised methods find structures without those targets. Data quality, sampling, bias and leakage determine whether evaluation results mean anything. A fraud-detection example trained on records that inadvertently contain later investigation outcomes can appear highly accurate while failing on real-time claims. Learners should understand why an honest train/test split and representative data are essential.

Generative AI produces new content conditioned on a prompt and context. It may write a summary, propose code or answer a question based on retrieved sources. Such content can be fluent and still be factually wrong. Quality depends on prompt structure, model capabilities, grounding, input data and application checks. Use independent validation when output feeds a consequential transaction. Predictive systems infer from learned statistical patterns; generative and agentic systems add content creation and, in some cases, tool-assisted actions that require separate authorization and evaluation.

Learners also need to distinguish training, fine-tuning, retrieval and prompting. Updating a model's weights is not the same thing as making relevant documents available during inference. A knowledge assistant can often answer company-policy questions with controlled retrieval rather than costly custom training. However, retrieval requires versioned documents, permission checks and a strategy for expired content. Selecting the correct method involves balancing task specificity, evidence requirements, maintainability and cost.

Recognize vision and language workloads without oversimplifying them

Computer vision includes image classification, detecting objects, reading text and describing visual content. These are related but not identical. A system that labels an image "warehouse" does not know how many pallets are present or whether a safety sign is legible. Object detection identifies locations of items, while OCR extracts characters from imagery. Image quality, language, lighting and perspective affect the outcome, so sample demonstrations should use representative production conditions rather than perfectly framed pictures alone.

Natural language processing covers language identification, entity recognition, sentiment estimation, key-phrase extraction and speech capabilities. A sentiment score should not be equated with a customer's actual legal or commercial intent. Entity extraction might identify a name but cannot reliably resolve which of several customers shares it without an authoritative database. Speech transcription must be evaluated with real accents, ambient noise and specialized vocabulary. Beginners benefit from learning these limitations before building automation around the outputs.

AI services can be accessed through SDKs, REST APIs or managed tools, but the fundamentals lesson is the task-to-capability mapping. Consider a support center that needs to transcribe calls, summarize issues and check product entitlements. Transcription, summarization and entitlement checking belong to different components. The final eligibility rule should remain in the organization's transactional system. The operational architecture can protect customers more reliably than attempting to solve every step through unrestricted natural-language instructions.

Make responsible AI an operational requirement

Microsoft's responsible AI principles emphasize fairness, reliability and safety, privacy and security, inclusiveness, transparency and accountability. In an AI-900 setting, candidates learned to recognize those principles in straightforward scenarios. A hiring classifier trained on unrepresentative data can introduce unfair outcomes; a system relying on unauthorized private records violates privacy; a voice assistant that excludes some users fails inclusion even if it performs well for a narrow test group.

Convert principles into tests. Fairness requires exploring whether important groups experience unequal errors. Reliability means testing normal and exceptional inputs and having a recovery plan. Privacy involves limiting who can access data in source systems, prompts and logs. Transparency can involve showing when content is generated and which sources support claims. Accountability requires a named owner and a path to contest or correct harmful output. These controls belong in the business workflow, not only in an ethics statement attached to an AI project.

A useful beginner exercise builds a small question-answering service and deliberately introduces contradictory documents. The learner should report which source is authoritative, what happens when no source is sufficient and how a user can request review. Even without advanced code, that exercise teaches why an attractive answer is not necessarily an authorized or reliable one. The same reasoning underpins more advanced examinations and operational roles.

Compare retired AI-900 with the current AI-901 assessment

AI-901 covers conceptual knowledge but introduces a more hands-on Foundry emphasis. Microsoft's April 15, 2026 outline measures AI concepts and capabilities alongside implementing AI solutions with Microsoft Foundry. Candidates are expected to recognize basic Python syntax and be familiar with Azure resources, REST APIs, SDKs and command-line interfaces. That does not turn it into an expert development certification, but it moves beyond the entirely conceptual assumptions often associated with introductory AI material.

Prepare with a small Foundry project: inspect a model deployment, construct system and user prompts, interact through the portal or SDK, and explore how an agent can call a restricted tool. Then test text analysis, vision or information extraction against representative examples. Record what the service does when the input is ambiguous or an instruction conflicts with organizational policy. Beginners gain practical intuition from seeing controlled failure cases, not merely selecting product names from a list.

Use AI-901 Azure AI Fundamentals as the approved internal destination for the current exam. Do not tell readers that an old AI-900 voucher, completed course or booking will automatically convert to the new examination. Scheduling rules depend on Microsoft's current registration system. Historical AI-900 content remains valuable as a concept reference, but the two test codes must be labeled according to their actual lifecycle.

An additional comparison helps readers see why the historical fundamentals program was valuable. Imagine a recruitment team asking for an automated shortlist. A regression model is not a candidate-ranking policy; a language model can summarize résumés but may reproduce irrelevant or biased impressions; and a deterministic workflow can screen only explicitly permitted qualifications. Before any automation, the organization should clarify the legal and fairness constraints, decide which steps need human judgment, and document the right to correct source information.

The best technical option may be a small extraction service combined with auditable business rules, rather than an unrestricted generative system. This scenario demonstrates that responsible AI is not a final layer pasted onto a completed solution. Risk and accountability should influence the choice of workload, data model, validation and user-facing explanation from the outset.

Preserve useful technical foundations while updating a study plan

Start with the distinction between input, model behavior, output and business action. For each workload, note the source of data, expected structure, measurable error and person accountable for final use. Compare a classifier with a generative assistant and a rule-based workflow on the same task. This reinforces the idea that AI capability selection is a design decision, not a popularity contest between products.

Next, check the Microsoft Learn AI-901 objectives against what the old course covers. Keep valid explanations of regression, classification, computer vision, natural language processing and responsible AI. Replace outdated directions about service portals, model deployment and exam scheduling with the live program's instructions. An effective transition sheet marks transferred knowledge, newly tested Foundry operations and historical-only references so candidates can allocate time efficiently.

The clear editorial conclusion is that AI-900 deserves to exist as a historical exam and learning resource, but only the current AI-901 program should be presented as the live fundamentals route in October 2026. A reader should finish understanding both the concepts that continue to matter and the practical steps that have changed. That preserves value without inventing ongoing availability for an exam Microsoft retired on June 30.