AI-901 - Microsoft Azure AI Fundamentals Cheat Sheet

Updated 2026-08-06

Next Topic: Google Cloud Professional Data Engineer Cheat Sheet

🎯Take a practice test on this topic5 practice tests · 186 questions→

The Microsoft Certified: Azure AI Fundamentals certification (Exam AI-901) is in beta as of April 2026, with general availability expected in June 2026, and it replaces the retiring AI-900 with a syllabus rebuilt around Microsoft Foundry instead of the older standalone Cognitive Services portals. It targets people at the very start of an AI career, testing conceptual knowledge of AI solutions in Azure plus enough Python, REST, SDK, and CLI familiarity to deploy and call a model, not to build production systems. Two skill areas carry the exam: identifying AI concepts and capabilities such as responsible AI, model components, and workload types (40-45%), and implementing AI solutions with Microsoft Foundry across generative apps and agents, text and speech, vision and image generation, and information extraction (55-60%). The single most useful mental model for this exam is that almost every "implement" task routes through the same Foundry project, deployment, and playground flow, so learning that pattern once pays off across most of the second domain.

What This Cheat Sheet Covers

This topic spans 17 focused tables and 143 indexed concepts, 5 practice tests with 186 questions. Below is a complete table-by-table outline of this topic, spanning foundational concepts through advanced details.

Table 1: Responsible AI PrinciplesTable 2: How Generative AI Models Work and Model SelectionTable 3: Model Deployment Options and Configuration ParametersTable 4: AI Workload Scenarios and SelectionTable 5: Text Analysis TechniquesTable 6: Speech Recognition and SynthesisTable 7: Computer Vision and Image Generation CapabilitiesTable 8: Information Extraction TechniquesTable 9: Prompt Engineering for Generative AITable 10: Deploying and Testing Models in the Foundry PortalTable 11: Building Chat Client Apps with the Foundry SDKTable 12: Building Single-Agent Solutions in FoundryTable 13: Building Text Analysis Applications with FoundryTable 14: Building Speech Applications with FoundryTable 15: Implementing Vision and Image-Generation Solutions with FoundryTable 16: Content Understanding for Documents and ImagesTable 17: Content Understanding for Audio and Video, and App Building

Quick IndexSubscribe to unlock

A jump-to index of every table row in this cheat sheet.

Mind MapSubscribe to unlock

An interactive map of every table and concept in this topic.

Table 1: Responsible AI Principles

Describe principles of responsible AI: Microsoft's Responsible AI Standard and the six principles it operationalizes, fairness, reliability and safety, privacy and security, inclusiveness, transparency, and accountability, and what each one concretely requires of an AI solution.

Concept	Example	Description
Microsoft Responsible AI Standard	Goal A1: Impact Assessment reviewed before development starts; Goal F1: Quality of Service evaluated across demographic groups	Internal engineering framework that turns the six principles into numbered Goals and Requirements (Impact Assessment, Responsible Release Criteria) product teams must satisfy before shipping. • Not a marketing statement; each Goal has its own applies-to scope and evaluation checkpoints.
Fairness	A loan model recommends the same outcome for two applicants with matching income and credit history, regardless of gender	AI systems should treat all people fairly and give similar groups similar outcomes. • Requires identifying at-risk demographic groups AND ongoing evaluation after release, not a one-time check of training-data balance.
Reliability and Safety	A predictable false-negative case gets a documented fallback option and an estimated rollback time	AI systems must perform reliably and safely across the operational conditions and edge cases they'll actually meet. • Not the same as an uptime SLA: the Standard requires defined error rates, edge-case testing, and a documented remediation plan for predictable failures.
Privacy and Security	SmartNoise adds differential privacy so no single person's record can be reverse-engineered from a model's outputs	AI systems must comply with privacy laws on data collection and use, and be secured end to end. • Covers access restriction, network controls, vulnerability scanning, and compliance auditing together; encryption is only one part.
Inclusiveness	A speech app is checked against Microsoft's Accessibility Standards for screen-reader support and captions before release	AI systems should empower and engage everyone, including people with disabilities and different backgrounds. • A separate Standard goal tied to Accessibility Standards compliance, not to be confused with fairness's demographic quality-of-service goal.
Transparency	A loan app shows which factors drove a rejection; a support chatbot states upfront that it's an AI, not a person	People must understand how an AI decision was made and know when they're interacting with AI. • Interpretability (explaining model behavior) is only part of it; the Standard separately requires disclosing AI interaction so people are never fooled into thinking an AI is human.
Accountability	A hospital keeps a clinician in the loop so an AI's treatment suggestion is never the final word on care	People who design and deploy AI remain answerable for how it behaves. • Requires a documented Impact Assessment before release AND human oversight and control afterward, so the system is never the final authority on decisions affecting people's lives.

Table 2: How Generative AI Models Work and Model Selection

Covers the AI-901 "Identify AI model components and configurations" area: how a generative AI model actually turns a prompt into a reply (tokens, embeddings, next-token prediction, foundation models) and how to pick the right model from the Azure AI Foundry catalog for a given task.

Concept	Example	Description
Tokens	A tokenizer splits "I heard a dog bark" into `I`, `heard`, `a`, `dog`, `bark`; GPT models instead use subword (Byte-Pair Encoding) tokenization.	The actual unit of text an LLM processes: a word, a partial word, or punctuation, not a fixed count of letters. • Not the same as characters or words 1:1. • Drives both context-window limits and per-token pricing.

The rest of this section is part of the full cheat sheet

Table 3: Model Deployment Options and Configuration Parameters

Under Identify AI model components and configurations, AI-901 candidates must know Microsoft Foundry's deployment types (Global Standard, Provisioned, Data Zone, Batch) and the request-level generation parameters (temperature, top P, max tokens, penalties, stop sequences, system message, streaming) that shape a model's output.

Concept	Example	Description
Global Standard	`SKU: GlobalStandard` Pay-per-token, routes to any Azure region, highest default quota	Dynamically routes traffic across Azure's global infrastructure with pay-per-token billing. • Highest default quota, no manual load balancing needed • Best for general workloads and getting started; high, steady volume may see latency variance (use Provisioned instead)

The rest of this section is part of the full cheat sheet

Table 4: AI Workload Scenarios and Selection

AI-901 domain "Identify AI workloads" (40-45% weight) opens with matching a business scenario to the right broad AI workload category, including the exam's newly expanded generative and agentic AI distinction, before drilling into any one category's techniques.

Workload	Example	Description
Generative AI	Prompt: "Write a 2 sentence product description for wireless earbuds" LLM outputs brand new marketing copy on demand	• Right fit when the ask is to CREATE new content (text, code, images) from a prompt • Powered by large language models responding to a single request at a time

The rest of this section is part of the full cheat sheet

Table 1: Responsible AI Principles

Concept	Example	Description
Microsoft Responsible AI Standard	Goal A1: Impact Assessment reviewed before development starts; Goal F1: Quality of Service evaluated across demographic groups	Internal engineering framework that turns the six principles into numbered Goals and Requirements (Impact Assessment, Responsible Release Criteria) product teams must satisfy before shipping. • Not a marketing statement; each Goal has its own applies-to scope and evaluation checkpoints.
Fairness	A loan model recommends the same outcome for two applicants with matching income and credit history, regardless of gender	AI systems should treat all people fairly and give similar groups similar outcomes. • Requires identifying at-risk demographic groups AND ongoing evaluation after release, not a one-time check of training-data balance.
Reliability and Safety	A predictable false-negative case gets a documented fallback option and an estimated rollback time	AI systems must perform reliably and safely across the operational conditions and edge cases they'll actually meet. • Not the same as an uptime SLA: the Standard requires defined error rates, edge-case testing, and a documented remediation plan for predictable failures.
Privacy and Security	SmartNoise adds differential privacy so no single person's record can be reverse-engineered from a model's outputs	AI systems must comply with privacy laws on data collection and use, and be secured end to end. • Covers access restriction, network controls, vulnerability scanning, and compliance auditing together; encryption is only one part.
Inclusiveness	A speech app is checked against Microsoft's Accessibility Standards for screen-reader support and captions before release	AI systems should empower and engage everyone, including people with disabilities and different backgrounds. • A separate Standard goal tied to Accessibility Standards compliance, not to be confused with fairness's demographic quality-of-service goal.
Transparency	A loan app shows which factors drove a rejection; a support chatbot states upfront that it's an AI, not a person	People must understand how an AI decision was made and know when they're interacting with AI. • Interpretability (explaining model behavior) is only part of it; the Standard separately requires disclosing AI interaction so people are never fooled into thinking an AI is human.
Accountability	A hospital keeps a clinician in the loop so an AI's treatment suggestion is never the final word on care	People who design and deploy AI remain answerable for how it behaves. • Requires a documented Impact Assessment before release AND human oversight and control afterward, so the system is never the final authority on decisions affecting people's lives.

Table 2: How Generative AI Models Work and Model Selection

Concept	Example	Description
Tokens	A tokenizer splits "I heard a dog bark" into `I`, `heard`, `a`, `dog`, `bark`; GPT models instead use subword (Byte-Pair Encoding) tokenization.	The actual unit of text an LLM processes: a word, a partial word, or punctuation, not a fixed count of letters. • Not the same as characters or words 1:1. • Drives both context-window limits and per-token pricing.

Table 3: Model Deployment Options and Configuration Parameters

Concept	Example	Description
Global Standard	`SKU: GlobalStandard` Pay-per-token, routes to any Azure region, highest default quota	Dynamically routes traffic across Azure's global infrastructure with pay-per-token billing. • Highest default quota, no manual load balancing needed • Best for general workloads and getting started; high, steady volume may see latency variance (use Provisioned instead)

Table 4: AI Workload Scenarios and Selection

Workload	Example	Description
Generative AI	Prompt: "Write a 2 sentence product description for wireless earbuds" LLM outputs brand new marketing copy on demand	• Right fit when the ask is to CREATE new content (text, code, images) from a prompt • Powered by large language models responding to a single request at a time