Microsoft Azure AI Fundamentals Training Course

AI-900 opens with a definitional domain, and it is tempting to skim it. Do not. Almost every service-selection question later in the exam is decided in the first five seconds, when you classify the scenario as a vision problem, a language problem, a forecasting problem, or a generative problem. If you classify it wrong, the four service names in the answer options will not save you. This lesson builds the frame everything else hangs on.

A working definition

Microsoft's framing of artificial intelligence is deliberately practical: software that imitates human capabilities and behaviors. That includes recognizing what is in an image, understanding and producing language, spotting something unusual in a stream of numbers, predicting a value that has not happened yet, and generating original content. Notice that the definition is about capability, not about technique. AI-900 does not ask you how a model is implemented. It asks you what a system does and which Azure service does that.

The counterpart to that definition is just as important. A program that applies rules a human wrote is not artificial intelligence, however sophisticated the rules are. A spreadsheet that flags any transaction over a fixed threshold is automation. A model that learned from historical transactions which patterns tend to precede a chargeback is machine learning. The difference is where the logic came from: a person, or the data.

Machine learning is the engine

Machine learning is the subset of AI in which a system derives its rules from examples rather than receiving them from a programmer. You supply historical data in which the outcome is already known, a training process finds the statistical relationships in it, and the result is a model — a function that takes new input and returns a prediction.

Practically every AI capability you meet on this exam is machine learning underneath. Image classification is a model trained on labeled images. Sentiment analysis is a model trained on text with known sentiment. Even the large language models behind generative AI are machine learning models, trained on very large text corpora. When a question asks what technique underpins a described capability, machine learning is very often the answer, and deep learning — machine learning using multi-layered neural networks — is the more specific answer for vision, speech, and language tasks.

Two consequences of the learned-from-data model matter for the rest of the course:

• A model is only as good as its training data. If the historical data underrepresents a group of people, the model will underperform for that group. This is the mechanism behind most fairness problems, and it is why the Responsible AI material sits in this same domain rather than in a separate appendix.
• A model produces a prediction, not a fact. Outputs carry confidence scores, and a confident answer can still be wrong. Systems that act on model output need a human review path when the stakes are high.

The three ways you can consume AI on Azure

Microsoft offers AI at three levels of effort, and recognizing which level a scenario needs is worth several marks across the exam.

• Prebuilt services. A model Microsoft already trained, called through an API. You send an image or a sentence and get a result. No training data of your own, no data science. Azure AI Vision reading text from a photograph is this level.
• Customizable services. A Microsoft model you adapt to your own domain with a modest amount of your own labeled data. Azure AI Custom Vision, custom named entity recognition, and custom speech all sit here. Choose this level when the prebuilt model does not know your categories — your specific product defects, your industry's jargon.
• Build your own. Azure Machine Learning, where you bring a dataset, choose or search for an algorithm, train, evaluate, and deploy. This is the right answer when the task is a numeric prediction or a segmentation specific to your business, and there is no prebuilt service for it.

The exam's preferred answer is nearly always the least effort that solves the stated problem. If a prebuilt service does the job, a scenario that proposes training a custom model is the wrong answer even though it would work.

Common AI terms you will see in stems

You do not need mathematics for AI-900, but you do need to read these words without hesitating. Training is the process that produces a model from data. Inferencing, sometimes called scoring or prediction, is using the finished model on new input. Features are the input values a model reads; a label is the known answer it learns to reproduce. A dataset is the collection of examples, usually split so that some of it is held back to test the model honestly.

One distinction is worth fixing now because it recurs throughout the exam. Narrow AI describes systems built for one specific task — every service in this course. General artificial intelligence, a system with broad human-equivalent competence across arbitrary tasks, does not exist as a product you can call. If an answer option describes a service that reasons like a person across any domain, it is a distractor.

What this buys you on exam day

When a stem opens with a business situation, do not start matching service names. Start by naming the workload in your own words: is the system being asked to see, to read or listen, to predict a number, to notice something unusual, to pull structure out of a document, to search a body of content, or to create something new? The next seven lessons take those categories one at a time, because that single act of classification is what the rest of the exam is testing.