Microsoft Azure AI Fundamentals Training Course

Most computer vision items on AI-900 are service-selection questions. They give you a business scenario and four Azure services that all sound reasonable, and the mark goes to whoever can eliminate three of them quickly. This lesson gives you a procedure and then runs it against the patterns the exam actually uses.

The decision procedure

Work through these in order. The first question that applies determines the family; the second narrows to the service.

Step 1 — What is the required output?

• A label for the whole image → classification
• Labels plus locations, or a count → object detection
• Exact pixel boundaries or area → segmentation
• The text that appears in the image → OCR
• Named business fields from a document → document intelligence
• Identity of a person from their face → face verification or identification
• A description, tags, moderation flags, or a smart thumbnail → general image analysis
• A searchable index across a large mixed collection → knowledge mining

Step 2 — Are the categories generic or organization-specific?

• Generic — people, vehicles, animals, furniture, printed text, standard invoices → prebuilt: Azure AI Vision, or a Document Intelligence prebuilt model.
• Organization-specific — your products, your defect types, your own form → custom: Azure AI Custom Vision, or a Document Intelligence custom model.

Step 3 — Are there constraints that override the obvious choice?

• Must run offline or on a device → a Custom Vision compact domain with model export.
• Faces with identity involved → Azure AI Face, and check whether Limited Access approval is required.
• Pixel-accurate output needed → beyond Custom Vision; this belongs in Azure Machine Learning.
• Labeling effort is the binding constraint → prefer the lowest rung on the precision ladder that meets the requirement.

The service map in one pass

• Azure AI Vision — prebuilt analysis of general images: captions, tags, objects, people, smart crops, moderation flags, OCR, image retrieval, spatial analysis. No training data.
• Azure AI Custom Vision — classification and object detection on your own categories, from your own labeled images. Exportable for edge deployment.
• Azure AI Face — face detection with attributes, verification, identification, grouping, liveness. Access-restricted for identity features.
• Azure AI Document Intelligence — named fields, key-value pairs, tables, and selection marks from documents; prebuilt, layout, and custom models.
• Azure Machine Learning — anything the above cannot do, including segmentation and any bespoke architecture.

Worked patterns

"An insurer wants to process vehicle damage photographs, identifying which panels are damaged and how severely." Output is what and where on a vehicle, with categories specific to the insurer's damage taxonomy. Detection, custom categories → Azure AI Custom Vision object detection. If the requirement said "measure the affected area in square centimeters," it would move to segmentation and out to Azure Machine Learning.

"A retailer wants to know how many customers enter each hour and how long they stay in each department." Counting and positions over video, generic subject — people. Prebuilt → Azure AI Vision people detection and spatial analysis. Custom Vision is wrong here because people are not an organization-specific category.

"A hospital wants to digitize decades of handwritten patient notes so clinicians can search them." Output is text; requirement is search, not named fields. → OCR, with the extracted text going into a search index. Document Intelligence would be over-specified. Note the handwriting caveat and the PII obligation on the output.

"A finance team wants supplier invoices posted to the ledger automatically." Named fields from a standard commercial document. → Document Intelligence prebuilt invoice model, with confidence thresholds routing uncertain documents to a person.

"A bank wants customers to open accounts by photographing their ID and taking a selfie." Two capabilities: read the ID document, and confirm the selfie matches its photograph. → Document Intelligence prebuilt identity document model plus Azure AI Face verification, with liveness detection to defeat photograph-of-a-photograph attacks. This is the classic two-service answer.

"A social platform must stop unsuitable images appearing in public feeds." Generic moderation. → prebuilt content moderation, with human review for borderline severity.

"A manufacturer wants to spot cracked welds on the line using cameras, and the plant network is unreliable." Custom defect categories, plus an offline constraint. → Custom Vision with a compact domain, exported and run at the edge.

The traps

• Choosing custom when prebuilt suffices. The most common error. If the categories are ones any general model knows, prebuilt wins on speed, cost, and maintenance.
• Choosing detection when classification suffices. Detection costs far more labeling effort. Only pay it if you need location or a count.
• Confusing OCR with Document Intelligence. Words versus fields. Look for whether the requirement names specific fields.
• Using Azure AI Vision for identity. It detects that a face is present; it does not tell you whose.
• Forgetting the second service. Many scenarios need a pipeline — OCR then translation, detection then OCR, extraction then a business system.
• Ignoring Responsible AI cues. When a stem mentions biometrics, consent, monitoring, or demographic groups, the correct answer usually includes a governance step, not just a service name.

How to use this under time pressure

Read the stem once and write two words in your head: the output shape, and generic or specific. Those two words eliminate most options before you have read them carefully. Then check for an overriding constraint — offline, biometric, pixel-accurate, labeling-limited — and confirm your choice survives it. That sequence handles the large majority of domain 3, and it is fast enough to leave time for the generative AI questions, which is where the marks are heaviest.