Common AI Workloads Explained: Generative, Agentic, Text, Speech, Vision and Extraction (AI-901)

A big share of AI-901 questions describe a business need and ask which kind of AI fits. These are called workloads. Learn the six below, the techniques inside each, and the keywords that give them away.

1. Generative AI

Creates new content from a prompt: text, code, summaries, images and more.

Typical scenarios:

• Drafting emails, product descriptions or reports
• Answering questions over company documents (with grounding)
• Writing and explaining code
• Chat assistants and copilots

Keywords: generate, draft, write, create, chat, summarise in your own words.

2. Agentic AI

An agent combines a generative model with instructions and tools, so it can decide what to do and take actions, often over several steps.

Typical scenarios:

• A support agent that looks up an order, checks the return policy and creates a refund
• An IT helpdesk agent that resets passwords through an API
• A research agent that searches the web, reads results and writes a report
• Several agents working together, each with its own role (multi-agent systems)

Keywords: take actions, call APIs, complete tasks, multi-step, autonomous, tools.

The difference from a chatbot: a chatbot answers; an agent does.

3. Text analysis (natural language processing)

Extracts meaning and structure from text. Core techniques:

• Keyword (key phrase) extraction: the main talking points. "The hotel room was clean but the Wi-Fi was slow" → hotel room, Wi-Fi.
• Entity detection (named entity recognition): people, places, organisations, dates, amounts. Entity linking connects them to a knowledge base, such as a Wikipedia page.
• Sentiment analysis: positive, negative, neutral or mixed, with confidence scores. Opinion mining ties sentiment to specific aspects ("room: positive, Wi-Fi: negative").
• Summarization: extractive picks the most important sentences; abstractive writes a new summary.
• Also common: language detection, PII detection and redaction, translation, text classification and question answering.

Classic techniques such as counting word frequency, removing stop words ("the", "and") and TF-IDF (scoring words that are frequent in one document but rare across others) are the ancestors of these features, and still appear in fundamentals questions.

Keywords: reviews, opinions, mentions, extract names, key points, language, redact.

4. Speech

Speech recognition (speech to text): turns spoken audio into text.

• Real-time transcription (live captions) and batch transcription of recordings
• Diarization: labelling who spoke when
• Custom models for industry vocabulary or noisy environments

Speech synthesis (text to speech): turns text into natural-sounding audio.

• Neural voices in many languages and styles
• SSML to control pronunciation, pauses, speed and emphasis
• Custom voices (with strict responsible AI controls)

Also: speech translation (speech in one language → text or speech in another), keyword recognition ("Hey…"), and speaker recognition.

Keywords: transcribe, captions, call recordings, read aloud, voice assistant, accents.

5. Computer vision and image generation

Understanding images and video:

• Image classification: what is this image of? (one label for the whole image)
• Object detection: what objects are there, and where? (bounding boxes)
• Semantic segmentation: which pixels belong to which object?
• Optical character recognition (OCR): read printed or handwritten text
• Image analysis: captions, tags, people detection
• Face detection and analysis: finding faces and attributes. Identification and verification features are restricted under Microsoft's Limited Access policy.

Multimodal models can now answer free-form questions about images: "What's wrong with this machine part?"

Image generation creates new images from text prompts, and can edit existing ones. Many generators use diffusion.

Keywords: photo, camera, detect, count objects, read text from images, generate an image, logo, illustration.

6. Information extraction

Pulls structured data out of unstructured content: documents, forms, images, audio and video.

• Invoice number, totals and line items from invoices
• Fields from receipts, IDs and contracts
• A summary, topics and action items from a meeting recording
• Chapters and key frames from a video

In Foundry this is the job of Azure Content Understanding (covered in its own lesson), which combines OCR, speech transcription and generative models.

Keywords: forms, invoices, receipts, fields, structured data, automate data entry.

Practice: match the scenario

Combining workloads

Real solutions often chain several:

• Call centre analytics: speech to text → sentiment analysis → summarization
• Accessible shopping app: image analysis describes products → text to speech reads the description
• Expense agent: Content Understanding extracts receipt fields → an agent checks policy and submits the claim

If a question describes several steps, look for the answer that lists the right workloads in the right order.

Key takeaways

• Six workloads: generative, agentic, text analysis, speech, computer vision / image generation, information extraction.
• Know the techniques inside each and their keywords.
• Classification labels a whole image; detection locates objects. Recognition is speech to text; synthesis is text to speech.

Next lesson: writing system and user prompts.