AI-900: Azure AI Fundamentals Before Its Retirement

Microsoft Exam AI-900 was retired on June 30, 2026. It once provided a fundamentals route into artificial intelligence workloads on Azure, but it is no longer an examination new candidates can book. Microsoft now lists AI-901 as the required exam for its Azure AI Fundamentals certification. Keeping the AI-900 page useful therefore means explaining its historical subject matter and the transition accurately rather than continuing to advertise an obsolete testing path.

The older exam introduced AI workloads, fundamental machine learning concepts, computer vision, natural language processing and generative AI. It was designed for broad understanding, not for proving professional-level proficiency in training models or deploying complex applications. These concepts remain relevant in an AI project conversation, but their presentation, product names and relative exam weight have changed. Candidates preparing for a current certificate should use AI-901's own study guide instead of assuming that five older AI-900 domains still define the active syllabus.

Recognize the AI workloads AI-900 introduced

An organization considering AI needs to know which problem it actually has. Classification assigns a category to an observation, such as identifying the type of a support request. Regression predicts a numeric value, such as expected equipment demand. Clustering looks for groups without relying on predefined class labels. Language processing extracts or generates meaning from text, while computer vision works with visual signals. Generative AI may create prose, code or multimedia from instructions and context. These tasks carry different data requirements and accuracy expectations.

It is tempting to label every automation requirement an AI problem. A fixed rule might be more transparent and reliable for a known threshold, while a predictive model may be justified when relationships are too complex for stable rules. AI can also sit within a larger conventional application: a model suggests an invoice category, while a deterministic accounting service validates tax codes and posts transactions. Understanding these boundaries was part of the value of fundamentals study, and it remains essential regardless of which Azure product currently exposes the capability.

The transition from AI-900 to contemporary Azure AI study provides historical context, but old blog material should be interpreted alongside Microsoft’s official retirement notice. AI-900 was not a prerequisite to advanced Azure AI credentials. It was a conceptual foundation that could help teams communicate about model capabilities, expected errors and appropriate human oversight before a specialist implemented a production system.

Understand learning types and model evaluation

Supervised learning uses examples with known desired outcomes. A regression model might estimate delivery time from historical journeys, while a classification model predicts whether a transaction belongs to a fraud-review category. Unsupervised learning looks for patterns without supplied labels; clustering can segment customer behavior for analysis, but a discovered group does not automatically correspond to a valid business category. Training data quality, representative samples and a clear problem formulation matter more than the brand of algorithm.

A model's evaluation must align with the real error costs. Accuracy alone can be misleading when fraud is rare. Precision describes the proportion of predicted positives that are correct, while recall describes how many actual positives are found. A fraud investigator may prefer one trade-off; an automatic medical decision would require much stronger safety review. For regression, errors may need to be analyzed by amount, direction and the consequences of severe misses, rather than a single aggregate number.

Data leakage invalidates a seemingly impressive score. If a training set includes information that becomes available only after the event being predicted, a model can appear excellent during testing and fail in production. Splitting data according to time or entity relationships may be more realistic than random splitting for some workloads. A fundamentals learner should understand the principle without needing to implement a complete MLOps pipeline. Model and data drift extends this discussion into the separate operational challenge of detecting when production conditions have changed.

Explain computer vision and document processing

Computer vision workloads differ by the required output. Image classification identifies an overall category; object detection locates individual items; optical character recognition extracts visible text. An application that scans receipts needs accurate field values and layout association, not just a statement that the image contains a receipt. A visual accessibility assistant may instead need a concise natural-language description. Choosing a service depends on whether the output will guide a human, populate a record or trigger an automated action.

Quality varies with the source. Low resolution, glare, handwriting, language, camera angle and crowded tables can reduce extraction accuracy. A system should preserve uncertainty and route exceptions when confidence or validation rules indicate a problem. An extracted price of 18.00 and 180.00 can look superficially similar but create a serious accounting error. Sensitive images may also contain identities, location information or confidential text that requires retention and access controls.

Generative visual systems introduce additional questions about provenance and appropriateness. Producing a plausible image is not the same as documenting what an original photograph shows. For an insurance or compliance case, maintain the distinction between observed evidence, inferred descriptions and synthetically produced media. This is a broader lesson than any one retired exam's product naming: choose AI behavior in light of the truth claims the application must make.

Distinguish natural language processing from generative AI

Natural language processing can extract entities, key phrases or sentiment, translate text and classify intents. A text-analysis pipeline used for customer complaints might find product names and route messages; a translation service attempts to preserve meaning across languages; a summary model compresses longer material into selected information. These outputs are not equally verifiable. Extracted names can sometimes be checked against the source, but a generated explanation may require additional human judgment and grounding.

Speech services convert voice to text, text to speech or one language to another. Acoustic quality, accents, background noise and domain terminology affect real-world performance. A meeting transcription is useful even when imperfect, but an automated call-center action needs a more explicit confirmation step before changing a customer's account. A learner should be able to state when the system is recognizing a signal, predicting a class or generating content based on probabilities.

Generative AI can build coherent responses from a prompt and context, but plausibility is not proof. Retrieval-augmented generation may supply supporting evidence, yet the search index and access rules still need validation. Trusted data grounding for enterprise agents explains the difference between having a fluent model and building an answer service that can support its claims. The old AI-900 concepts provide vocabulary for this distinction; AI-901 uses a newer Foundry-oriented skills map.

For example, consider an automated customer-service classifier trained primarily on short English messages. A launch review should include longer notes, different dialects, messages with multiple requests and users whose accessibility needs change the structure of the input. It should then analyze whether the model misroutes certain customers more often and whether a human can correct those decisions promptly. The intended business outcome is correct service, not maximizing a headline accuracy statistic that hides systematic disadvantages. This example gives fairness and reliability a measurable operational meaning.

Transparency also has limits. Telling customers that a model is involved is useful but insufficient when there is no route to contest a high-impact decision. Accountability requires an owner who can investigate the evidence, repair the dataset or revise the workflow. These duties persist even when the examination used to introduce their vocabulary has retired.

Apply responsible AI principles to practical decisions

Responsible AI was never only a topic list to memorize. Fairness asks whether error rates and consequences fall unequally across populations; reliability examines failures under real operating conditions; privacy limits information collection and disclosure; transparency helps people understand what a system does and cannot establish; and accountability assigns responsibility for decisions and remediation. These concerns depend on the use case. A creative-writing assistant and an automated benefit-eligibility tool should not be governed by the same level of risk tolerance.

Testing requires examples that reflect actual deployment. A language model can be evaluated against ambiguous prompts, dialects, unusual file layouts, malicious inputs and contradictory sources. Track the performance of versions over time. Human review matters when a recommendation may alter an employee's status, a customer's entitlement or a patient's treatment. Policy labels are not enough; controls must influence the system's real behavior. The operation of responsible AI in production makes this principle concrete through monitoring and corrective action.

Security is separate from truthfulness. A model might accurately summarize a sensitive document that the caller should not have been able to access. Identity permissions, data classification, tool authorization and logging must be built into the surrounding application. Conversely, a model may be authorized to read a public document yet generate an unsupported conclusion. Responsible operation considers both access safety and output reliability rather than treating them as one test score.

Choose the current route without erasing the history

The current AI-901 exam emphasizes AI concepts and capabilities alongside implementing AI solutions with Microsoft Foundry. That is more hands-on and differently organized than the former AI-900's five-domain overview. Candidates should check the current exam's expectation of basic Python familiarity and Azure resources, even though both credentials use the Azure AI Fundamentals name. Studying only AI-900 flashcards may not cover the operational examples now expected.

Students aiming beyond fundamentals have distinct options. AI-103 centers on developing apps and agents, while AI-300 centers on operating machine learning and generative systems. Neither becomes a substitute for the foundations of evaluating models, data and responsible use. A sensible practice progression is to classify a few sample use cases, build a limited Foundry experiment with mock data and document how retrieval, permissions and evaluation change the outcome.

Microsoft's official retired-exams record settles the status of AI-900 as of June 30, 2026. Any site still offering AI-900 registration or claiming it is the active prerequisite for another certification must be treated skeptically. Historical courses may still teach useful ideas, but the exam title, effective outline and availability should be verified against Microsoft Learn for the candidate's planned sitting. The point of retaining this page is accurate continuity of knowledge, not preserving an obsolete booking call to action.