Study Azure AI-900 Workload Fit: key concepts, common traps, and exam decision cues.
AI-900 starts by asking a simple question that many candidates skip: what kind of problem is this really? If you classify the workload correctly, many service-choice questions become much easier. If you classify it badly, even the correct Azure product name can still be wrong because it solves a different problem.
Primary workload: The main AI task being tested, even if the broader business process includes several smaller steps.
| If the scenario wants… | Primary workload | Strong first thought |
|---|---|---|
| categories, objects, image content, OCR, or facial analysis | computer vision | what should the system see or extract from visuals? |
| sentiment, entities, translation, transcript, or text meaning | natural language processing or speech | is the input text-first or audio-first? |
| text and layout from forms, invoices, or receipts | document processing | the system is reading existing structure, not generating content |
| a predicted number or label from historical data | machine learning | the system is making a prediction from examples |
| a draft, summary, answer, rewrite, or conversation | generative AI | the system is creating new content |
The exam often hides the answer in the output shape. If the desired result is a category, a number, extracted fields, or generated text, that output tells you the workload faster than the Azure product names do.
| Confusion pair | Keep this distinction clear |
|---|---|
| document processing vs generic computer vision | form and layout reading vs broad image analysis |
| document processing vs generative AI | extracting existing information vs creating new language |
| NLP vs speech | text-first workload vs audio-first workload |
| machine learning vs prebuilt AI service | custom predictive modeling vs ready-made capability |
| generative AI vs classification | content creation vs label prediction |
| Scenario clue | Likely lane |
|---|---|
| “summarize these case notes” | generative AI |
| “extract invoice totals and vendor names” | document processing |
| “detect whether a review is positive or negative” | NLP |
| “transcribe the meeting audio” | speech |
| “predict whether the customer will churn” | machine learning classification |
| “tag the objects in an image” | computer vision |
AI-900 sometimes gives you a workflow that includes more than one valid AI step. For example, a support application might transcribe speech, analyze sentiment, and then summarize the interaction. The exam still expects you to identify the primary workload named in the question stem.
Not every business automation problem is an AI problem. If the scenario can be solved by fixed if-then rules with no learning, perception, or language interpretation, the exam may be testing whether you avoid overclassifying the requirement as AI.