AI-901 lesson 1: the AI workload map
Artificial intelligence is the broad idea of software doing tasks that normally need human judgement, and every AI-901 scenario can be filed under one of five workload families: generative and agentic AI, text analysis, speech, computer vision, or information extraction. Each family answers a different question, such as “write something”, “what does this text say”, “what was spoken”, “what is in this picture” or “which fields are in this document”. Picking the family is the first step of almost every concepts question, and the whole course is organised around it.
From AI to agents: four nested ideas
The terms are nested rather than competing.
- Artificial intelligence is the umbrella: any system that perceives, predicts, decides or generates in a way that feels intelligent.
- Machine learning is the dominant way of building AI today. Instead of a developer writing rules, a model learns patterns from example data and then applies them to new data. Predicting which customers will cancel is a classic case.
- Generative AI is machine learning that produces new content, such as text, images, code or audio, rather than a label or a number. A large language model drafting a reply is generative.
- Agentic AI adds action. An agent uses a model to decide what to do next, calls tools (for example a search or a database lookup), and works through steps toward a goal. The model alone only answers; the agent can act. The topic page on AI agents goes further into this.
A useful test: if the output is a prediction or classification, think machine learning; if it is newly composed content, generative; if the system chooses and performs steps, agentic.
The five workload families
The figure above gives the overview. The table adds the question each family answers and the shape of its output.
| Family | Question it answers | Typical output |
|---|---|---|
| Generative and agentic AI | Can the system compose or do something? | New text, code, images, or completed steps |
| Text analysis | What does this written text contain? | Key phrases, entities, a sentiment label, a summary |
| Speech | What was said, or how do we say this aloud? | A transcript, or synthesised audio |
| Computer vision | What is in this image or video frame? | Descriptions, tags, detected objects, or a generated image |
| Information extraction | Which specific fields sit in this content? | Structured fields from documents, images, audio or video |
In Microsoft’s current naming, the ready-made services behind several of these families are grouped as Foundry Tools (Azure Language, Azure Speech, Azure Vision and Azure Content Understanding), while models for generating content are deployed in Microsoft Foundry. The name “Azure AI services” is still common in older articles and refers to the same group. Later lessons cover each family in detail.
Worked example: a 200-person retailer lists its needs
A 200-person online retailer lists five projects and files each one.
- A chat assistant that drafts friendly answers to customers: generative AI. If the assistant must also look up an order and issue a refund by itself, it becomes agentic.
- A report showing how customers feel about the new returns policy based on thousands of written reviews: text analysis, specifically sentiment analysis.
- A feature that reads product descriptions aloud for visually impaired shoppers: speech, specifically speech synthesis.
- A tool that checks warehouse photos for crushed boxes: computer vision.
- A process that pulls the order number, date and total from supplier PDFs into the finance system: information extraction.
Two of these show where thinking is needed. Project 2 starts with text, but the output is a label, so it is analysis, not generation. Project 5 could be attempted with a chat model, yet the requirement is named fields in a predictable shape, which points to extraction.
Easy to confuse
| Often mixed up | The difference |
|---|---|
| Text analysis vs generative AI | Analysis extracts or labels what is already in the text; generation composes new text |
| Summarisation vs information extraction | A summary is shorter prose; extraction returns named fields |
| Speech recognition vs speech synthesis | Recognition turns audio into text; synthesis turns text into audio |
| Vision analysis vs image generation | Analysis reads an existing image; generation creates a new one |
| A model vs an agent | A model produces a response; an agent also chooses steps and uses tools |
When two answers both seem possible, read the required output. A field, a label, a transcript and a paragraph each point to a different family.
Key points
- AI is the umbrella; machine learning learns patterns from data; generative AI creates new content; agentic AI chooses and performs steps with tools.
- The five workload families are generative and agentic AI, text analysis, speech, computer vision and information extraction.
- Choose a family by the required output: new content, a label, a transcript or audio, an image result, or structured fields.
- Foundry Tools is the current name for the group once called Azure AI services.
- Analysing existing content and generating new content are different workloads even when both involve text.
Check yourself
Question 1. A hotel chain wants a system that reads guest comments and marks each one as positive, neutral or negative. Which workload family fits best?
- A. Generative AI
- B. Computer vision
- C. Text analysis
- D. Speech
Show answer
Answer: C
Labelling the feeling in written comments is sentiment analysis, which belongs to text analysis. Generative AI would compose new text rather than label existing text, computer vision works on images, and speech workloads deal with audio.
Want more questions like this? Full AI-901 practice tests →
Question 2. A support assistant must not only answer a question about a late parcel but also look up the shipment in a tracking system and open a replacement order without a person stepping in. What best describes this solution?
- A. An agentic AI solution
- B. A text analysis solution
- C. A speech synthesis solution
- D. An information extraction solution
Show answer
Answer: A
Choosing steps and calling tools such as a tracking lookup and an order system is what makes a solution agentic. A plain text analysis feature returns labels, a speech service converts audio, and information extraction pulls fields out of content; none of these act on other systems.
Want more questions like this? Full AI-901 practice tests →
Question 3. A company receives scanned delivery notes and needs the delivery date and recipient name from each as separate values for a database. Which workload family should be chosen?
- A. Speech
- B. Image generation within generative AI
- C. Summarisation within text analysis
- D. Information extraction
Show answer
Answer: D
The requirement is named fields from scanned content, which is information extraction. Speech synthesis produces audio, image generation creates pictures, and summarisation would give shorter prose rather than separate database values.
Want more questions like this? Full AI-901 practice tests →
This lesson opens the first exam domain, Identify AI concepts and capabilities, which carries 40 to 45% of AI-901, and the recap page AI concepts explained condenses the same material for revision.