Artificial intelligence (AI) is software that imitates human abilities such as seeing, understanding language, making predictions and creating content. On AI-900 you are not expected to build AI from scratch. You are expected to recognize what kind of problem AI can solve, which Azure service fits, and what responsible use looks like. Almost everything in modern AI rests on one idea: a model that has learned patterns from data.
Traditional software is rule-based. A developer writes explicit logic, such as 'if the order total is over 100, apply free shipping'. That works when the rules are known and stable. It breaks down for tasks where nobody can write the rules, such as telling a cat from a dog in a photo or deciding whether a review sounds angry. There are too many variations to list.
Machine learning (ML) turns this around. Instead of writing the rules, you give an algorithm many examples of inputs and, usually, the correct outputs. Training adjusts the model's internal values until its outputs match the examples as closely as possible. The result is a model: a function that takes new input and returns a prediction. Using the trained model on new data is called inferencing (or inference).
Because a model learned from examples, its output is a probability-based best guess, not a guaranteed truth. A vision model might say 'dog, 94% confidence'. A language model might write a fluent paragraph that contains a mistake. This is why AI solutions need testing, human oversight and the responsible AI principles you will study later in this domain. It is also why the quality and representativeness of training data matter so much: a model can only learn patterns that are present in its data, including unwanted ones such as bias.
Most AI you will meet on the exam is delivered as a service. Azure AI services (Vision, Language, Speech, Translator, Document Intelligence, Content Safety and others) expose pretrained models through an API, so you send an image or text and get results back. Azure Machine Learning is for training your own models on your own data. Microsoft Foundry brings together generative AI models, agents and AI services in one portal. Knowing which of these to pick for a scenario is a large part of AI-900.
Key terms
- Model
- A function learned from data that takes new input and returns a prediction, label, score or generated content.
- Training
- The process of adjusting a model's internal values using example data until its outputs match the examples well.
- Inference
- Using a trained model to make predictions on new data; also called inferencing or scoring.
- Rule-based software
- Software whose behavior comes from explicit logic written by a developer rather than from patterns learned from data.
A spam filter written with rules might block any email containing the word 'prize'. Spammers quickly change the wording. A machine learning spam filter is trained on thousands of emails already marked spam or not spam, learns hundreds of subtle signals, and keeps improving when it is retrained on newly reported messages.
Check yourself
What is the difference between training and inference?
Training builds the model from example data; inference uses the trained model to make predictions on new data.
Why can two similar inputs to an AI model give a wrong answer for one of them?
A model gives a probability-based prediction learned from its training data, not a guaranteed result, so unusual or poorly represented inputs can be misclassified.