AI-901 Course Notes

AI-901 starts with recognizing the workload before choosing a service. Common Azure AI workloads include machine learning prediction, anomaly detection, natural language processing, speech, computer vision, document extraction, content safety, and generative AI grounded with retrieval. Azure AI Services provide prebuilt capabilities through APIs, while Azure Machine Learning is the platform for training, evaluating, deploying, and managing custom models.

• Choosing Azure Machine Learning for every AI problem instead of recognizing prebuilt Azure AI Services, Stream Analytics, or Azure AI Search workloads.

• Identify the workload first: language, speech, document, safety, search, ML, anomaly detection, or generation.

Responsible AI review checks whether an AI system is fair, reliable, safe, private, secure, inclusive, transparent, and accountable. Fairness depends on representative data and evaluation across affected groups, not just overall accuracy. Transparency means users and reviewers can understand what the system is for, what data it uses, and where its limitations are.

• Treating rule-based systems and machine learning as interchangeable even though rules need explicit logic and ML learns patterns from data.

• Representative data and subgroup evaluation are fairness clues.

Machine learning fundamentals start with task type. Classification predicts a category, regression predicts a numeric value, clustering finds groups in unlabeled data, and reinforcement learning learns actions through rewards. AI-901 questions often describe the output first: a yes/no label, a product category, a price, a customer segment, or an action policy.

• Using classification when the target is a number, or regression when the target is a category.

• Classification predicts categories; regression predicts numbers; clustering finds unlabeled groups.

Azure AI workloads are easiest to choose by input type and required output. Vision analyzes images and video frames, OCR extracts printed or handwritten text, Face detects or analyzes human faces, and object detection returns locations with bounding boxes. Image classification labels an entire image, while object detection answers where specific objects appear.

• Choosing general Vision when the scenario specifically requires OCR, Face, or Document Intelligence.

• OCR extracts text; Document Intelligence extracts structured document fields.

Large language models generate text by processing prompts, system prompts, chat history, retrieved context, and other input as tokens. A system prompt sets behavior and constraints, while user and assistant messages form the conversation. Context windows limit how much input and output the model can consider at one time, so long documents usually require chunking, retrieval, or summarization.

• Fine-tuning a model when the problem is missing or changing source knowledge that RAG should retrieve.

• System prompts set behavior and constraints; user prompts provide the task.

AI governance starts with access control. Azure RBAC assigns permissions to users, groups, and managed identities, while Key Vault protects secrets, keys, and credentials used by AI services and applications. Network isolation limits which networks can reach machine learning workspaces or AI resources. Encryption protects data at rest or in transit, but permissions and network design still decide who can use the resource.

• Using encryption as the answer for access control when the scenario really asks for RBAC or Key Vault.

• RBAC controls who can access resources; Key Vault protects secrets and keys.

The Responsible AI framework provides a structured approach to managing AI risks, prioritizing proactive risk mitigation and preventing harm. This framework centers on establishing clear ownership for AI system outcomes and integrating technical controls with robust governance. Administrators utilize this framework to identify and address potential biases within AI systems, ensuring equitable outcomes and fostering trust. Successfully implementing this framework requires continuous monitoring and adaptation to evolving risks and regulatory requirements.

• Memorizing Responsible AI terms without tying them to a concrete design or review action.

• Fairness, reliability and safety, privacy and security, inclusiveness, transparency, and accountability are the core Responsible AI ideas.