What's changed: Added per-section figures (cert-figure-retrofit). New AI-901 Chapter 1 (Domain 1 "AI concepts and responsibilities": workloads = generative/agentic/text analysis/speech/CV/information extraction, Microsoft Foundry overview, six Responsible AI principles plus genAI/agent-specific risks and mitigations)
1.1AI Workloads and Their Types
Understand what AI is, the relationship among AI/ML/deep learning, and the common AI workloads emphasized in AI-901—generative AI, agentic AI, text analysis, speech, computer vision, and information extraction—plus how Azure delivers them in a unified way on Microsoft Foundry.
AI (Artificial Intelligence) is an umbrella term for software that performs tasks normally requiring human intelligence—learning, prediction, recognition, language understanding, and reasoning. AI-901 is the entry-level credential: it asks for no deep math. Instead it expects a map of which technique solves which problem, plus the basics of how to implement it in Microsoft Foundry. The key skill is choosing the right workload from a scenario.
1.1.1AI, machine learning, and deep learning
First, the nesting. AI is the broadest concept; inside it sits machine learning (ML); inside ML sits deep learning. AI is "software that behaves intelligently," ML is "techniques that learn rules from data," and deep learning is "ML using multi-layer neural networks." Most of today’s generative AI, speech, and image recognition rest on deep learning. For AI-901, remember that ML is the foundation running inside many workloads.
1.1.2Common workloads emphasized in AI-901
- Generative AI: generate new content—text, images, code—via large language models. The central theme of AI-901.
- Agentic AI: given a goal, an AI that autonomously plans steps and calls tools to act—extending generative AI into "reason then act."
- Text analysis (NLP): extract the meaning, sentiment, entities, and intent of text.
- Speech: speech-to-text and text-to-speech, plus translation.
- Computer vision: analyze images/video to recognize objects, faces, and text (OCR).
- Information extraction: pull structured fields/insights from invoices, forms, images, audio, and video (Azure Content Understanding).
Organize these by input data and goal: create new content → generative AI; act autonomously toward a goal → agentic AI; language (text) → text analysis; audio → speech; images/video → computer vision; pull key fields from documents/files → information extraction. The exam asks you to judge the workload from both the input and the goal.
| Scenario (example) | Workload | How it’s done in Foundry |
|---|---|---|
| Draft an email | Generative AI | Chat completion on a Foundry model (e.g., GPT) |
| Handle a booking change autonomously | Agentic AI | A tool-using agent in Foundry Agent Service |
| Detect sentiment in a message | Text analysis | Azure AI Language |
| Transcribe meeting audio | Speech | Azure AI Speech |
| Recognize objects/text in photos | Computer vision | Azure AI Vision |
| Extract amounts from invoices | Information extraction | Azure Content Understanding |
1.1.3Microsoft Foundry as the unified platform
AI-900 centered on a dichotomy: prebuilt (Azure AI services) vs custom (Azure Machine Learning). AI-901 evolves this: the center of gravity is Microsoft Foundry (formerly Azure AI Foundry, earlier Azure AI Studio)—one platform to select, deploy, build, evaluate, and monitor models, apps, and agents end to end. Within Foundry you pick pretrained models from a catalog (Foundry Models) provided by Microsoft, OpenAI, and others, and embed them into generative AI apps or agents. Task-focused Azure AI services (Language, Speech, Vision, Content Understanding) remain available, and Foundry is the unified front door to them.
Naming note: the development platform was renamed "Azure AI Studio → Azure AI Foundry → Microsoft Foundry" (Ignite 2025). This book uses the latest name Microsoft Foundry with older names in parentheses. Note Azure OpenAI models are not retired—they remain available as part of the Foundry model catalog (Foundry Models).
Scenario: automating support. Classify incoming email urgency with text analysis → check defects in attached images with computer vision → extract amounts from invoice PDFs with information extraction → draft a reply with generative AI → and have an agentic AI autonomously confirm, reply, and change a booking. Multiple workloads chain together on Foundry to automate one process.
Watch the mix-ups: (1) generative AI (creates content) vs agentic AI (autonomously plans and acts toward a goal)—the latter adds tool use and multi-step action. (2) computer vision (general image recognition) vs information extraction / Content Understanding (structured extraction from documents/files). (3) OCR is a computer-vision capability. (4) Do not confuse a workload (what you do) with a service/platform (product name).
Questions often map a scenario → the right workload. E.g., "draft an email" = generative AI; "given a goal, autonomously execute steps" = agentic AI; "detect sentiment" = text analysis; "transcribe audio" = speech; "identify products in a photo" = computer vision; "extract amounts from invoices" = information extraction. Remember implementation happens on Microsoft Foundry.
1.1.4Section summary
- AI = software performing human-like intelligent tasks; AI ⊃ machine learning ⊃ deep learning
- Key AI-901 workloads: generative AI / agentic AI / text analysis / speech / computer vision / information extraction
- Implementation centers on Microsoft Foundry (select model → deploy → build app/agent → evaluate → monitor); Azure OpenAI models continue as Foundry Models
- The exam centers on scenario → workload; workloads combine
Sign in to track progress — Log in.
Quick check
(just a quick review)Q1. Which kind of AI, given a goal, autonomously plans steps and calls tools to execute them?
Q2. Which workload creates new content such as text or images?
Q3. Which Azure capability extracts structured fields/insights from invoices, forms, images, audio, and video?
Q4. In AI-901, which platform unifies selecting/deploying models and building, evaluating, and monitoring apps/agents?
Q5. Which correctly describes the nesting of AI, machine learning, and deep learning?
Q6. Which workload best fits transcribing meeting audio to text?