AI-901
Deep Dive
Azure offers multiple computer vision services, each optimized for a specific use case. The AI-901 tests which service to choose given a scenario — know the key differences.
Contents
Azure AI Vision
Image Analysis + Read API (OCR)
General image analysis and document OCR
Azure AI Vision Custom Model
Custom classification + custom object detection
When you need business-specific categories that general models don't recognize
Face API
Facial detection, identification, verification
Biometric access systems, identity verification, facial attribute analysis
Document Intelligence
Structured OCR + form extraction
Invoices, receipts, contracts, ID cards — when you need specific fields from the document
Video Indexer
Deep video analysis
Video transcription, speaker identification, scene detection, OCR in video
General image analysis service based on Microsoft's pre-trained models. Used via REST API — no training of your own required.
Automatic description
Generates descriptive text of an image's content in natural language.
Object detection
Locates common objects (people, cars, animals, furniture) with bounding boxes.
Tagging
A list of descriptive tags for the image with confidence level.
Categorization
Classifies the image into a hierarchy of 86 predefined categories.
Adult content detection
Identifies inappropriate, violent, or adult content for moderation.
Dominant color
Extracts the main colors of the image, including whether it is color or grayscale.
Brand detection
Identifies well-known brand logos in images.
Video with Content Understanding
Segments and describes video recordings with generative models: transcription, scenes and custom fields. Spatial Analysis, which counted people in real-time video, was retired in 2025.
The Read API (part of Azure AI Vision) is Azure's most advanced OCR service. It handles dense documents, multiple languages, text in various orientations, and multi-page PDFs.
Key features
Handwritten and typed text
Documents with complex layouts (tables, columns)
Multi-page PDFs
170+ languages supported
Text in different orientations and rotations
Returns bounding box for each word
Ideal use cases
Bulk digitization of archived documents
Automatic processing of invoices and receipts
Extracting data from scanned forms
Making historical documents searchable
Capturing text from photos of signs or whiteboards
Convert a photo of a receipt into structured text in seconds, with per-line confidence scores.
+1 more...
To classify images or detect objects with your own categories you used Custom Vision, which trained models on your labeled images with no code. Microsoft announced its retirement (supported until September 25, 2028) and recommends Content Understanding (custom classification), a multimodal model from the Foundry catalog, or Azure Machine Learning AutoML. What AI-901 tests is interpreting images with multimodal models and building apps with vision capabilities in Foundry.
Custom image classification
Trains a model that recognizes YOUR specific categories.
Example: a pharmaceutical company wants to classify its 50 types of pills by visual appearance. General models don't know those specific pills.
Recommended minimum: 30 images per class
Custom object detection
Detects and locates specific objects that aren't in pre-trained models.
Example: detecting and locating defective screws on a manufacturing line — the general model doesn't know that specific defect.
Requires manually labeling bounding boxes
Azure AI Vision vs Custom Model — when to use each
Azure AI Vision (pre-trained model)
When the categories are general and the pre-trained model recognizes them. No data or training of your own required.
Azure AI Vision Custom Model
When you need domain-specific categories that general models don't know. You have your own images to train on.
Create your own Custom Vision project, upload tagged images, and train a model to detect specific defects.
+1 more...
No guided steps: read a scenario with several stacked requirements and pick, out of 4 vision services, the only one that meets them all.
+1 more...
A model barely reaches 61% accuracy. Find the cause (it's not the algorithm) and fix it — no guided steps.
+1 more...
Detect
No person identificationDetects faces in an image. Returns bounding boxes and optional attributes (estimated age, expression, accessories).
Identify
Requires a database of peopleCompares a face against a group of registered people (PersonGroup). Returns who it is with a confidence level.
Verify
1 vs 1Compares two facial images and says whether they are the same person. For 1:1 verification (are you who you say you are?).
Find Similar
No registered identitiesGiven a face, finds the most similar ones in a list. Does not identify by name — only by visual similarity.
Group
UnsupervisedGroups unidentified faces by visual similarity. Useful for organizing photo collections.
Locate faces in an image and understand why Microsoft restricts person identification for ethical reasons.
+1 more...
Document Intelligence goes beyond simple OCR: it doesn't just extract text, it understands the document's structure and extracts specific fields. It has pre-built models for common document types and supports custom models.
🧾
Invoices
Extracts fields: vendor, amount, date, line items, subtotals, taxes, due date.
🏪
Receipts
Extracts: merchant name, purchased items, total, taxes, date, payment method.
🪪
ID cards
Extracts: name, date of birth, document number, expiration date, issuing country.
📇
Business cards
Extracts: name, company, title, email, phone, address.
📋
W-2 / tax forms
Extracts specific fields from U.S. tax forms.
⚙️
Custom model
You train it with your own documents to extract specific fields from proprietary forms.
Use the prebuilt-invoice model to extract named fields (vendor, total, date) instead of plain text.
+1 more...
Video Indexer combines multiple AI models to extract complete insights from video content — it doesn't just analyze a single frame, but the whole video with its temporal context.
Automatic audio-to-text transcription (Speech-to-Text)
Speaker identification (who spoke when)
Automatic translation of transcriptions
OCR in video (text that appears on screen)
Scene detection and camera cuts
Person identification (with a facial database)
Extraction of key topics from the content
Detection of the speaker's emotions and sentiment
Recognition of mentioned or visible brands
Moderation of inappropriate content in video
| Scenario | Recommended service |
|---|---|
| Automatically tag photos from a family album | Azure AI Vision — Image Analysis |
| Summarize meeting recordings by scene for an internal search tool | Azure Content Understanding (video) |
| Read text from multi-page scanned invoices | Azure AI Vision — Read API |
| Classify photos of your company's products into 20 of your own categories | Azure AI Vision Custom Model — custom classification |
| Detect and locate defective screws on a manufacturing line | Azure AI Vision Custom Model — object detection |
| Verify that a photo on a bank ID is the same person registering | Face API — Verify |
| Automatically extract amount, vendor, and date from invoices | Azure AI Document Intelligence |
| Transcribe and make searchable an archive of 500 training videos | Video Indexer |
Did you get this topic?
Test what you just learned
An insurance company wants to automate claims processing. Customers submit photos of their scanned medical invoices. The system must automatically extract the medical provider's name, the date of service, and the total amount. Which Azure service is most appropriate?
Put what you learned to the test with original AI-901 exam questions.
Start AI-901 mock examUnlimited mock exams with the Pro plan
Log in to track your progress.