AI-900: Microsoft Azure AI Fundamentals

AI-900 exam objectives and study guide

The skills measured on the AI-900 exam, by objective domain. Percentages are the share of the exam each domain carries.

AI-900 practice questions and answers

10 free sample questions from the AI-900 bank, with the correct answer and a full explanation for each. These are original questions written to the Microsoft objective domains — not real exam content.

1. Question 1Describe Artificial Intelligence workloads and considerations

   Customer service agents report delays when they manually enter data from payer-issued coverage cards. The automation must capture plan, member, and coverage details from scanned card images without creating a labeled training set. Which model should be used?

   • AAzure AI Document Intelligence prebuilt Health insurance card model
   • BAzure AI Document Intelligence prebuilt Identity model
   • CAzure AI Document Intelligence prebuilt Business card model
   • DAzure AI Document Intelligence prebuilt Receipt model

   Show answer and explanation

   Correct answer

   • Azure AI Document Intelligence prebuilt Health insurance card model

   Explanation

   The correct answer is the prebuilt Health insurance card model. Azure AI Document Intelligence includes a prebuilt model for extracting insurance coverage details from health insurance cards, so a labeled custom training set is not required.

   • The prebuilt Identity model extracts identification and verification details, not health coverage fields.
   • The prebuilt Business card model extracts business contact details, not member coverage information.
   • The prebuilt Receipt model extracts sales transaction details, not insurance plan or member fields.

   References:

   • What Is Azure Document Intelligence in Foundry Tools?
   • Document Processing Models
2. Question 2Describe features of computer vision workloads on Azure

   A developer asks which Azure AI service is designed to detect, recognize, and analyze human faces in images. Which service should you recommend?

   • AAzure AI Face
   • BAzure AI Language
   • CAzure AI Speech
   • DAzure AI Document Intelligence

   Show answer and explanation

   Correct answer

   • Azure AI Face

   Explanation

   Azure AI Face provides prebuilt algorithms for detecting, recognizing, and analyzing human faces in images.

   • Azure AI Language is for text analytics and natural language processing.
   • Azure AI Speech is for speech-to-text, text-to-speech, and speech translation scenarios.
   • Azure AI Document Intelligence extracts structured information from documents and forms.

   References:

   • Azure Face documentation
   • Face detection, attributes, and input data
3. Question 3Describe features of computer vision workloads on Azure

   Litware plans to crop profile pictures around visible faces before users save them. Which statement describes the core capability of Azure AI Face detection?

   • AIt locates human faces and can return landmarks or supported face attributes.
   • BIt translates spoken audio into another language in real time.
   • CIt extracts printed and handwritten text from scanned documents.
   • DIt trains custom models to recognize product labels in photos.

   Show answer and explanation

   Correct answer

   • It locates human faces and can return landmarks or supported face attributes.

   Explanation

   Azure AI Face detection locates human faces in images and can return additional analysis such as landmarks and supported face attributes.

   • Translating spoken audio is a Speech service capability.
   • Extracting printed or handwritten text is an OCR capability.
   • Training product-label detection models is a Custom Vision object detection scenario, not the prebuilt Face detector.

   References:

   • Azure Face documentation
   • Face detection, attributes, and input data
4. Question 4Describe features of computer vision workloads on Azure

   Photos submitted for a facilities badge must be rejected when more than one person appears or the visible face is turned too far to one side. Dara, an app designer, also needs coordinates to display an on-screen guide, but the app must not compare the photo with anyone. Which approach should be used?

   • AUse face detection to count faces and return a face rectangle with head-pose attributes.
   • BUse face verification to compare each submitted photo with a stored employee portrait.
   • CUse face identification against a trained PersonGroup to determine whether the person is an employee.
   • DUse Custom Vision object detection trained on sample portraits to locate people in each photo.

   Show answer and explanation

   Correct answer

   • Use face detection to count faces and return a face rectangle with head-pose attributes.

   Explanation

   The correct answer is to use face detection and request the face rectangle and head-pose attributes. Detection can count detected faces, return the face location, and provide analysis attributes without performing identity matching.

   • Verification compares one face with another face or an enrolled person, which the requirement explicitly does not need.
   • Identification searches a trained person group to determine who appears, which violates the no-comparison requirement.
   • Custom Vision object detection can be trained for custom objects, but it does not provide the prebuilt facial head-pose analysis expected from Azure AI Face.

   References:

   • Face detection, attributes, and input data
   • What is the Azure Face service?
5. Question 5Describe features of computer vision workloads on Azure

   A mobile app needs a prebuilt computer vision capability that finds human faces in images and can return face-related details such as landmarks or head pose. Which capability should be used?

   • AAzure AI Face detection
   • BAzure AI Face identification
   • CAzure AI Document Intelligence Read
   • DAzure AI Custom Vision object detection

   Show answer and explanation

   Correct answer

   • Azure AI Face detection

   Explanation

   The correct answer is Azure AI Face detection. It is designed to locate human faces in images and can return face rectangles, landmarks, and selected facial attributes.

   • Face identification is used to determine who a person is from an enrolled group.
   • Document Intelligence Read extracts text from documents and images.
   • Custom Vision object detection requires training a custom model and is not the prebuilt facial analysis capability described.

   References:

   • Face detection, attributes, and input data
   • What is the Azure Face service?
6. Question 6Describe features of computer vision workloads on Azure

   Lamna Healthcare must design a patient check-in flow. The solution has consent and approval for face recognition. It must first ensure that the camera sees exactly one unoccluded face with acceptable capture quality, and then compare the selfie only with the claimed patient record. Which two features should be used?

   • AUse face detection and request attributes such as `occlusion` and `qualityForRecognition` before accepting the capture.
   • BUse one-to-one face verification between the selfie and the claimed patient enrollment.
   • CUse one-to-many face identification against all patient records to discover the identity.
   • DUse Azure AI Vision Image Analysis `/imageanalysis` for both OCR and biometric matching.
   • EUse Custom Vision object detection with patient names as tags.

   Show answer and explanation

   Correct answers

   • Use face detection and request attributes such as `occlusion` and `qualityForRecognition` before accepting the capture.
   • Use one-to-one face verification between the selfie and the claimed patient enrollment.

   Explanation

   The correct answers are face detection with analysis attributes and one-to-one face verification. Detection attributes such as occlusion and qualityForRecognition can support the capture-quality gate, while verification compares the selfie with the claimed patient record without searching all patients.

   • One-to-many identification would search an enrolled population, which conflicts with the claimed-record-only requirement.
   • Image Analysis /imageanalysis is not the recommended service for face recognition and is subject to retirement.
   • Custom Vision object detection is not a biometric verification service and does not provide the Azure AI Face recognition workflow.

   References:

   • Face detection, attributes, and input data
   • Face recognition
   • Migrate from Azure Vision in Foundry Tools
7. Question 7Describe features of Natural Language Processing (NLP) workloads on Azure

   Imani, a data engineer, is testing the Azure AI Language REST API with one English sentence. The request must return standard categories such as people, locations, organizations, dates, and quantities. Complete the request body.

   {
     "kind": "1",
     "parameters": {
       "modelVersion": "latest"
     },
     "analysisInput": {
       "documents": [
         { "id": "1", "language": "en", "text": "Marta plans to visit Berlin on 12 September with 3 colleagues." }
       ]
     }
   }

   Blank 1

   • AEntityRecognition
   • BPiiEntityRecognition
   • CEntityLinking
   • DKeyPhraseExtraction

   Show answer and explanation

   Correct answer

   • Blank 1: EntityRecognition

   Explanation

   The correct value is `EntityRecognition`. In the Azure AI Language analyze-text request body, this kind selects the prebuilt capability that extracts named entities and returns their categories.

   • `PiiEntityRecognition` detects personally identifiable information and supports redaction scenarios, which is narrower than the requested standard entity categories.
   • `EntityLinking` disambiguates recognized entities and can return external links, rather than simply returning standard entity categories.
   • `KeyPhraseExtraction` extracts important phrases, but it does not return named entity categories such as person, location, organization, date, or quantity.

   References:

   • What is the named entity recognition (NER) feature in Azure Language?
   • What is Azure Language in Foundry Tools
   • What is the Personally Identifiable Information (PII) detection feature in Azure Language?
8. Question 8Describe features of Natural Language Processing (NLP) workloads on Azure

   You are asked to process short equipment inspection notes that contain proprietary component names and failure codes. The built-in Azure AI Language categories do not return the labels that inspectors need, and the team can provide labeled examples for training. Which capability should you choose?

   • AUse custom named entity recognition in Azure AI Language.
   • BUse prebuilt named entity recognition in Azure AI Language.
   • CUse entity linking in Azure AI Language.
   • DUse key phrase extraction in Azure AI Language.
   • EUse language detection in Azure AI Language.

   Show answer and explanation

   Correct answer

   • Use custom named entity recognition in Azure AI Language.

   Explanation

   The correct answer is custom named entity recognition. It is designed for extracting domain-specific entities from unstructured text by training a model with labeled examples.

   • Prebuilt named entity recognition returns standard categories, such as person, organization, location, date, and quantity, but it is not trained with your proprietary labels.
   • Entity linking disambiguates entities and returns links to known entries; it does not learn inspection-specific labels from training examples.
   • Key phrase extraction can surface important phrases, but it does not create a supervised custom entity schema.
   • Language detection identifies the language of text and does not extract component names or failure codes.

   References:

   • What is the custom named entity recognition (CNER) feature in Azure Language?
   • What is the named entity recognition (NER) feature in Azure Language?
   • What is language detection in Azure Language?
9. Question 9Describe fundamental principles of machine learning on Azure

   Jamal, a data analyst, must choose which business questions should be solved by regression instead of classification or clustering. Which two questions should he choose?

   • AHow many kilowatt-hours will a building use next month based on weather and occupancy?
   • BWhat payout amount should be estimated for an insurance claim based on claim details?
   • CWill a subscriber renew the service contract at the end of the month?
   • DShould a support ticket be assigned low, medium, or high priority?
   • EWhich natural groups exist in customers who have no predefined segment labels?

   Show answer and explanation

   Correct answers

   • How many kilowatt-hours will a building use next month based on weather and occupancy?
   • What payout amount should be estimated for an insurance claim based on claim details?

   Explanation

   A regression model predicts a numeric quantity. The electricity-use and claim-payout scenarios both require numeric outputs.

   • Monthly electricity use in kWh is a numeric value, so it is a regression target.
   • Insurance claim payout is a numeric amount, so it is a regression target.
   • Renewal yes/no is binary classification.
   • Support priority is multiclass classification.
   • Discovering customer groups without a target label is clustering.

   References:

   • Introduction to Machine Learning Concepts
   • What is automated ML? AutoML
   • Get Started with AI Architecture Design
10. Question 10Describe fundamental principles of machine learning on Azure

    Marta, an AI trainer, is explaining why modern generative language models can focus on relevant tokens in a long prompt when predicting the next token. Which Transformer component assigns higher or lower importance to tokens in the context?

    • AAttention mechanism
    • BActivation function
    • CLoss function
    • DConfusion matrix

    Show answer and explanation

    Correct answer

    • Attention mechanism

    Explanation

    The correct answer is the attention mechanism. In a Transformer, attention assigns weights to tokens so the model can focus on the parts of the context that are most relevant when building token representations or predicting output.

    • An activation function introduces nonlinearity in a neural network, but it is not the component that weighs contextual tokens.
    • A loss function measures training error; it does not choose which context tokens are important.
    • A confusion matrix evaluates classification results; it is not part of Transformer attention.

    References:

    • Large language models (LLMs)
    • Prompt engineering techniques

Membership includes 193 questions and explanations aligned to the AI-900 curriculum.

Other Microsoft certifications

Every one of these is included with the same membership as AI-900.

AB-900Microsoft

Microsoft 365 Copilot and Agent Administration Fundamentals

This exam is intended for candidates who are familiar with Microsoft 365, including core services, security, identity and access, data…

161 practice questionsView

AZ-900Microsoft

Microsoft Azure Fundamentals

This exam is intended for technology professionals who want to demonstrate foundational knowledge of cloud concepts and Microsoft Azure.…

170 practice questionsView

DP-900Microsoft

Microsoft Azure Data Fundamentals

This exam measures knowledge of core data concepts and related Microsoft Azure data services. It is intended for candidates who are…

198 practice questionsView

AI-900 exam FAQ