Choosing the Right AI Workload
Level Up Team
This post summarizes Choosing the Right AI Workload, module 2.1 of Level Up’s Azure AI Fundamentals series. The lesson is on YouTube as Azure AI Fundamentals AI-901 Exam: 2.1: Choosing the Right AI Workload. Module 1.2 is The Anatomy of an Azure AI Solution.
The workload is the kind of problem being solved
Exam scenarios often include distracting details: an industry, an Azure resource, a file type, or a model name. Isolate the required outcome first. Then choose the workload whose defining capability matches that outcome.
Microsoft’s current AI-901 objectives name six workload families. A single solution can use several. A question usually emphasizes one primary capability. Ask what transformation the application performs.
| Workload | Defining capability |
|---|---|
| Generative AI | Creates new content |
| Agentic AI | Pursues goals through decisions and actions |
| Text analysis | Discovers meaning in language |
| Speech | Processes or produces spoken language |
| Computer vision | Interprets or creates visual content |
| Information extraction | Converts unstructured content into structured data |
Generative AI creates something new
Generative AI produces new content from instructions and context: answers, summaries, emails, reports, images, code, translations, or rewritten text.
Scenario language such as “generate,” “compose,” “draft,” “rewrite,” or “create” points here. A support assistant that drafts a personalized answer from the customer’s question and company documentation is generative AI.
Summarization can appear under both generative AI and text analysis. A fluent new response is generative AI. Identifying or analyzing meaning in existing text is text analysis.
Agentic AI pursues a goal and takes action
Agentic AI uses a model to decide what to do next and to interact with tools or systems. Possible actions include searching a knowledge source, calling an API, updating a record, creating a support ticket, asking for missing information, or handing work to a person.
The typical loop is goal, reason, act, observe, and continue. An assistant that checks an order, requests a replacement, updates the CRM, and tells the customer what happened is agentic.
If the system only produces an answer, classify it as generative AI. If it chooses and performs actions toward a goal, classify it as agentic AI.
Text analysis discovers meaning in existing language
Text analysis finds meaning, patterns, or classifications in text. Core techniques include keyword extraction, entity detection, sentiment analysis, language detection, classification, and summarization.
The desired result describes or labels existing text. Analyzing a customer message for negative sentiment, a product name, an account number, and cancellation intent is text analysis. If the application must determine what text means, text analysis is usually the match.
Speech processes or produces spoken language
Speech capabilities include speech recognition (speech to text), speech synthesis (text to speech), speech translation, and spoken interaction with multimodal models.
If the audio represents human language, think speech. A caller who describes a problem aloud, and an assistant that answers with a synthesized voice, is a speech scenario. Analyzing nonverbal audio — alarms, machinery, or environmental sounds — may be information extraction.
Computer vision interprets or creates visual content
Computer vision classifies images, detects objects, describes scenes, answers visual questions, reads text in images, or generates new images.
Choose computer vision when the answer depends on what is visually present. A customer photograph that the assistant uses to identify product damage is computer vision. Pulling named fields out of that image is information extraction.
Information extraction returns structured fields
Information extraction converts unstructured or semi-structured content into structured fields. Sources include documents, forms, images, audio, video, and text. The current objectives explicitly include extraction from text, images, audio, and video, not only traditional documents.
An invoice that becomes Customer: Maria Chen, Invoice: 10482, Date: 2026-09-12, and Total: $847.50 is information extraction. Explaining what a document says is text analysis or generative AI. Capturing specific fields from it is information extraction.
Workloads compose
A customer-support assistant can use all six in one case:
- Transcribe the caller — speech
- Detect frustration — text analysis
- Inspect a damage photo — computer vision
- Capture invoice fields — information extraction
- Draft the response — generative AI
- Issue a refund through an API — agentic AI
An application does not have to be “a vision application” or “a generative-AI application.” On the exam, identify the capability associated with the step being described.
Classify the transformation
Find the verb, identify the input, and inspect the expected output. Classify what the AI must do.
| Scenario language | Likely workload |
|---|---|
| Draft an answer | Generative AI |
| Complete the task | Agentic AI |
| Detect sentiment | Text analysis |
| Transcribe a call | Speech |
| Identify an object | Computer vision |
| Populate invoice fields | Information extraction |
A support system that reads emailed invoices and stores the invoice number, supplier, date, and total is information extraction. The decisive requirement is converting document content into predefined structured fields. The presence of text does not make the scenario text analysis.
Bring this to your team
Level Up can run a personalized Azure AI Fundamentals workshop for your team. Tell us a little about your group and we’ll reply with pricing.