AI-901 Microsoft Azure AI Fundamentals Roadmap

Build the conceptual foundation, but connect every concept to a design or evaluation decision.

• Explain fairness and why aggregate quality can hide subgroup disparities
• Apply reliability and safety through boundaries, testing, safe failure, monitoring, and escalation
• Connect privacy to purpose, minimization, retention, deletion, and appropriate access
• Connect security to identities, credentials, least privilege, data, models, tools, and infrastructure
• Explain inclusiveness, transparency, accountability, accessible recourse, and named ownership
• Describe tokenized context, learned parameters, probabilistic generation, and hallucination risk
• Select models by capability, modality, latency, safety, availability, deployment, and cost
• Recognize deployment choices and parameters without memorizing one static model list
• Distinguish generative and agentic AI, text analysis, speech, vision, and information extraction
• Match keyword extraction, entities, sentiment, summarization, recognition, synthesis, and multimodal extraction to scenarios

Move from recognition to a working Foundry project. Keep the first lab small, synthetic, and disposable.

• Create or enter a Microsoft Foundry project and identify models, deployments, endpoints, and access
• Compare candidate models with the same representative prompts and acceptance criteria
• Deploy a supported model and interact with it in the Foundry portal
• Write system and user prompts with purpose, relevant context, constraints, output, and fallback behavior
• Understand temperature and output limits as controls, not guarantees of truth
• Use Python configuration, JSON requests and responses, and basic exception handling
• Follow the current official Foundry SDK or REST quickstart to create a lightweight chat client
• Use Microsoft Entra-compatible credentials instead of placing keys in prompts or source code
• Handle timeouts, transient failures, throttling, and safe user-facing errors
• Record model, deployment, prompt, evaluation, latency, token, and cost assumptions

Create one bounded assistant before considering broad tools or autonomy. An agent's model and instructions do not replace access control.

• Explain an agent as a model plus instructions plus tools or knowledge
• Create and test a single agent in the Foundry portal
• Define scope, evidence use, uncertainty, refusal, and human-escalation behavior
• Connect only approved read-only synthetic knowledge for the fundamentals lab
• Create a lightweight client for the agent using the current supported development pattern
• Test normal, ambiguous, unsupported, harmful, and adversarial prompts
• Evaluate relevance, groundedness or source support, task success, safety, fairness, and refusal quality
• Keep identity, authorization, business rules, and consequential approvals outside model instructions
• Monitor errors, latency, throttling, token use, quality trends, and estimated cost
• Document limitations, incident ownership, budget thresholds, and a complete cleanup plan

Practice lightweight implementations and choose capabilities by workload rather than product-name memorization.

• Build a small text-analysis flow for keywords, entities, sentiment, or summarization
• Evaluate supported languages, confidence, false positives, false negatives, privacy, and fairness
• Recognize speech-to-text and text-to-speech scenarios and current Azure Speech capabilities in Foundry Tools
• Build a lightweight speech application and handle approved audio securely
• Respond to spoken prompts with a deployed multimodal model where supported
• Interpret images in prompts by using a deployed multimodal model
• Build a lightweight vision application and preserve uncertainty about visual claims
• Create visual output with a generative model and apply safety and review controls
• Test accessibility, missing objects, invented objects, harmful content, and visual text attacks
• Compare quality, latency, availability, and cost across specialized and multimodal approaches

Use Content Understanding to connect documents, images, audio, and video to one measurable workflow, then consolidate the complete blueprint.

• Define a schema and ground truth before configuring an analyzer
• Extract information from documents and forms with Content Understanding in Foundry Tools
• Extract source-linked fields and descriptions from images
• Extract structured information and time evidence from audio and video
• Build a lightweight application with content submission, status handling, result retrieval, and schema validation
• Route low-confidence, missing-evidence, inconsistent, or policy-sensitive output to human review
• Evaluate field quality by modality rather than relying only on one aggregate score
• Apply least privilege, short retention, safe logs, monitoring, budget controls, and complete deletion
• Complete both synthetic portfolio projects and explain each responsible AI and cost decision
• Use original questions and flashcards, review every distractor, recheck official docs, and avoid AI-900-only study plans