Browse Certification Practice Tests by Exam Family

Free AI-900 Full-Length Practice Exam: 50 Questions

Try 50 free AI-900 questions across the exam domains, with explanations, then continue with full IT Mastery practice.

This free full-length AI-900 practice exam includes 50 original IT Mastery questions across the exam domains.

These questions are for self-assessment. They are not official exam questions and do not imply affiliation with the exam sponsor.

Count note: this page uses the full-length practice count maintained in the Mastery exam catalog. Some certification vendors publish total questions, scored questions, duration, or unscored/pretest-item rules differently; always confirm exam-day rules with the sponsor.

Need concept review first? Read the AI-900 Cheat Sheet on Tech Exam Lexicon, then return here for timed mocks and full IT Mastery practice.

Open the matching IT Mastery practice page for timed mocks, topic drills, progress tracking, explanations, and full practice.

Try AI-900 on Web View full AI-900 practice page

Exam snapshot

  • Exam route: AI-900
  • Practice-set question count: 50
  • Time limit: 45 minutes
  • Practice style: mixed-domain diagnostic run with answer explanations

Full-length exam mix

DomainWeight
Describe Artificial Intelligence Workloads and Considerations19%
Describe Fundamental Principles of Machine Learning on Azure19%
Describe Features of Computer Vision Workloads on Azure19%
Describe Features of Natural Language Processing (NLP) Workloads on Azure19%
Describe Features of Generative AI Workloads on Azure24%

Use this as one diagnostic run. IT Mastery gives you timed mocks, topic drills, analytics, code-reading practice where relevant, and full practice.

Practice questions

Questions 1-25

Question 1

Topic: Describe Features of Computer Vision Workloads on Azure

Which Azure service should you choose for a solution that must analyze images, generate tags, and extract printed text by using OCR?

Options:

  • A. Azure AI Language

  • B. Azure AI Speech

  • C. Azure OpenAI Service

  • D. Azure AI Vision

Best answer: D

Explanation: Azure AI Vision is the correct choice for this computer vision workload. It supports image analysis features such as tagging, captioning, and OCR, which directly match a requirement to inspect images and read printed text.

Azure AI Vision is the Azure AI service family used for common visual analysis tasks at the AI-900 level. When the input is an image and the goal is to identify visual content, generate tags, describe what is shown, or extract printed text with OCR, Azure AI Vision is the best fit. The deciding clue in this question is that the solution must work directly with images and return visual insights plus text read from those images.

Services that focus on text, speech, or generative responses are different categories. If the requirement starts with photos, scanned images, or visual content, Azure AI Vision is the service candidates should recognize first.

  • Text-only focus Azure AI Language analyzes written text, such as sentiment or entities, rather than image pixels.
  • Audio focus Azure AI Speech is for spoken language scenarios like speech-to-text and text-to-speech.
  • Generative focus Azure OpenAI Service is mainly used for generative AI tasks, not as the primary service for standard image analysis and OCR.

Question 2

Topic: Describe Features of Computer Vision Workloads on Azure

Which Azure service is the best choice when a solution must specifically detect and analyze human faces in images?

Options:

  • A. Azure AI Vision

  • B. Azure AI Speech

  • C. Azure AI Language

  • D. Azure AI Face detection service

Best answer: D

Explanation: Use Azure AI Face detection service when the requirement is explicitly about finding and analyzing faces in images. Azure AI Vision handles broader image tasks, but the face-specific service is the clearest match here.

The core idea is to choose the Azure AI service that most directly matches the primary task. When the stated need is to detect or analyze human faces in images, Azure AI Face detection service is the best fit because it is dedicated to face-related computer vision scenarios. A broader service like Azure AI Vision is used for more general image analysis tasks such as tagging, captioning, OCR, or object detection. Azure AI Language is for text workloads, and Azure AI Speech is for audio workloads. For AI-900 questions, match the service to the input type and the specific capability being requested. If the requirement names faces directly, the face-specific service is the strongest answer.

  • Broader image tasks Azure AI Vision is a general computer vision service, so it is less specific when face detection is the explicit goal.
  • Text analysis Azure AI Language is used for processing written text such as sentiment or entity extraction.
  • Speech processing Azure AI Speech is for recognizing, synthesizing, or translating spoken audio, not analyzing faces in images.

Question 3

Topic: Describe Fundamental Principles of Machine Learning on Azure

A retail company builds a custom machine learning model to forecast daily product demand from historical sales data. The team needs to manage model versions, deploy the model for predictions, and monitor its performance over time. Which Azure service should they use?

Options:

  • A. Azure AI Language

  • B. Azure AI Vision

  • C. Azure Machine Learning

  • D. Azure OpenAI Service

Best answer: C

Explanation: Azure Machine Learning is the Azure service for the lifecycle of custom ML models. It supports managing model versions, deploying models for inference, and monitoring them after deployment, which matches the retail forecasting scenario.

Azure Machine Learning is designed for end-to-end management of custom machine learning solutions on Azure. In this scenario, the company already has a forecasting model and needs high-level lifecycle capabilities: model management, deployment for predictions, and ongoing monitoring. Those are core Azure Machine Learning functions.

Prebuilt Azure AI services are better when you want ready-made AI capabilities such as image analysis, language understanding, or generative text without managing a custom predictive model lifecycle. When the requirement is to work with your own trained model and keep track of it after release, Azure Machine Learning is the appropriate choice.

  • Vision mismatch Azure AI Vision is for image-based tasks such as OCR, image analysis, and object detection.
  • Language mismatch Azure AI Language focuses on NLP tasks like sentiment analysis and entity extraction, not custom model lifecycle management.
  • Generative mismatch Azure OpenAI Service is for generative AI scenarios such as chat, summarization, and content generation.

Question 4

Topic: Describe Features of Generative AI Workloads on Azure

A support app must take a short prompt such as “Write an empathetic response to a delayed shipment” and generate a new customer reply. The company does not need sentiment analysis, key phrase extraction, or entity detection on existing text. Which Azure service best fits this requirement?

Options:

  • A. Azure AI Speech

  • B. Azure Machine Learning

  • C. Azure OpenAI Service

  • D. Azure AI Language

Best answer: C

Explanation: This scenario is about generative AI: creating a new reply from a prompt. Azure OpenAI Service is designed for prompt-based text generation with large language models, while classic text analytics focuses on analyzing text that already exists.

The core distinction here is generating text versus analyzing text. Azure AI Language is mainly used for NLP tasks on existing content, such as sentiment analysis, key phrase extraction, entity recognition, classification, and summarization. Azure OpenAI Service is the Azure service for generative AI scenarios in which a model creates new content from a prompt, such as drafting replies, rewriting text, or producing chatbot responses.

In this scenario, the app must produce an original customer message from a short instruction. That makes the requirement generative text creation, not classic text analytics. Azure Machine Learning is a broader platform for building custom ML solutions, but it is not the most direct fundamentals-level choice for prompt-based text generation. That is why Azure OpenAI Service fits better than Azure AI Language here.

  • Text analytics confusion The option for Azure AI Language is tempting because it handles NLP, but its main role is analyzing existing text rather than generating new prompt-based replies.
  • Wrong modality The option for Azure AI Speech focuses on speech recognition, speech synthesis, and translation, not primary text generation.
  • Too broad The option for Azure Machine Learning can support custom AI solutions, but it is not the simplest best-fit service for this stated requirement.

Question 5

Topic: Describe Features of Generative AI Workloads on Azure

Which situation most clearly indicates that a generative AI solution should use grounding and human review before users rely on its output?

Options:

  • A. When the system must convert a recorded call into text

  • B. When the system must group customers by similar buying patterns

  • C. When the system must identify products in warehouse images

  • D. When replies must reflect current company policy and may affect customer decisions

Best answer: D

Explanation: Grounding is most important when a generative AI system must answer from trusted, up-to-date information rather than only from its pretrained patterns. Human review is also needed when incorrect output could influence customer actions or business decisions.

Generative AI can produce fluent answers that sound correct even when they are incomplete, outdated, or inaccurate. Grounding reduces that risk by anchoring responses to approved source content, such as current company policy documents. Human oversight is especially important when the output has real-world consequences, like informing customers about rules, eligibility, or next steps.

The other situations describe different AI workloads:

  • Converting audio to text is speech recognition.
  • Identifying products in images is computer vision.
  • Grouping similar customers is clustering.

The key takeaway is that grounding and review matter most when generative output must be trustworthy, current, and safe to act on.

  • Converting a recorded call into text is an Azure AI Speech scenario, not a grounding decision for generative output.
  • Identifying products in warehouse images is a computer vision task, typically handled by Azure AI Vision.
  • Grouping customers by similar buying patterns describes clustering, which is an unsupervised machine learning task.

Question 6

Topic: Describe Features of Computer Vision Workloads on Azure

A retailer wants to analyze shelf images and receive a rectangle around each product it finds, along with the product type. Which computer vision approach best fits this requirement?

Options:

  • A. Object detection

  • B. Image classification

  • C. Image captioning

  • D. Optical character recognition (OCR)

Best answer: A

Explanation: Object detection is the right choice when you need both what an item is and where it appears in the image. The requirement for rectangles around each product means the solution must return bounding boxes, not just a single label or a text description.

This scenario is about matching the required output to the correct computer vision approach. When a solution must find multiple items in an image and show each item’s position, the key clue is the need for rectangles around objects, which are bounding boxes.

  • Image classification returns a category for the whole image.
  • Object detection returns object labels plus bounding boxes.
  • OCR extracts printed or handwritten text.
  • Image captioning generates a natural-language description.

Because the retailer needs both product identification and object locations, object detection is the best fit; OCR and captioning produce different kinds of output, and image classification does not locate individual items.

  • Whole-image label image classification can label an image but does not return coordinates for each detected product.
  • Text extraction OCR is for reading text, such as signs or labels, rather than locating general objects as products.
  • Natural-language summary image captioning describes what is in an image but does not provide structured bounding boxes.

Question 7

Topic: Describe Features of Generative AI Workloads on Azure

An HR department wants to improve its employee help portal. The new AI solution must:

  • answer policy questions in natural language
  • summarize long handbook sections
  • draft first-pass email responses for HR staff

Which Azure solution is the best fit?

Options:

  • A. Use Azure AI Language custom question answering for the portal.

  • B. Use Azure AI Speech to transcribe employee phone calls.

  • C. Use Azure OpenAI Service to build an HR assistant.

  • D. Use Azure Machine Learning to train a request classification model.

Best answer: C

Explanation: This is a generative AI scenario because the system must create new text, not just return a label or score. Azure OpenAI Service is designed for chat, summarization, and drafting content, so it best matches all of the stated needs.

The key clue is that the portal must produce original language output, such as summaries and draft replies. That makes this a generative AI workload rather than a traditional predictive machine learning workload. Predictive ML usually predicts a category or value, such as classifying a request or forecasting demand. In contrast, generative AI creates new content from prompts and provided context.

Azure OpenAI Service is the best fit because it supports chatbot experiences, summarization, and text generation in one solution family. A more limited text-analysis or question-answering tool might help with one requirement, but it would not be the best overall match for all three. The main takeaway is that generating natural-language content points to generative AI, not predictive ML.

  • Classification only A request classification model predicts labels, but it does not generate summaries or draft email text.
  • Too narrow Custom question answering can help answer known questions, but it is not the best fit for broad text generation and drafting.
  • Wrong modality Speech transcription focuses on converting audio to text, which does not address the main portal requirements.

Question 8

Topic: Describe Features of Natural Language Processing (NLP) Workloads on Azure

A retailer stores customer support tickets as plain text from a web form. It wants to automatically extract entities such as product names and cities from each ticket and determine whether the customer sentiment is positive, neutral, or negative. Which Azure service should the retailer choose?

Options:

  • A. Azure AI Vision

  • B. Azure OpenAI Service

  • C. Azure AI Language

  • D. Azure AI Speech

Best answer: C

Explanation: This scenario is a standard NLP task on text data. Azure AI Language is the Azure service designed for text analysis features such as entity extraction and sentiment analysis, so it best matches the stated requirements.

The core concept is choosing the Azure AI service that matches the workload. Support tickets in this scenario are already plain text, and the company wants two common NLP capabilities: extracting named entities and detecting sentiment. Azure AI Language includes prebuilt text analysis features for exactly these tasks, making it the best fundamentals-level choice.

Azure AI Vision is for image-based analysis, and Azure AI Speech is for spoken audio tasks such as speech-to-text or text-to-speech. Azure OpenAI Service is used for generative AI scenarios like chat, summarization, and content generation, but it is not the most direct fit when the requirement is standard prebuilt entity and sentiment analysis.

When the need is to analyze written text for meaning, Azure AI Language is usually the right starting point.

  • The computer vision option fits images and video, not written support tickets.
  • The speech option fits audio input and spoken output, not text entity extraction.
  • The generative AI option is useful for chat or summarization, but this scenario asks for classic prebuilt NLP analysis.

Question 9

Topic: Describe Fundamental Principles of Machine Learning on Azure

A retailer plans to use Azure Machine Learning. Which business problem is a likely example of regression?

Options:

  • A. Classify customer reviews as positive or negative

  • B. Forecast monthly sales revenue for each store

  • C. Group shoppers into segments by purchase behavior

  • D. Generate new product descriptions from keywords

Best answer: B

Explanation: Regression is used to predict a numeric value on a continuous scale. Forecasting monthly sales revenue fits regression because the output is an amount, not a category, group, or generated text.

The core idea of regression is predicting a continuous number. In this case, monthly sales revenue is a numeric amount, so forecasting it is a classic regression task. Common business examples include predicting sales, house prices, demand, or temperature. By contrast, assigning reviews to positive or negative classes is classification, grouping shoppers by similar behavior is clustering, and writing product descriptions is a generative AI task. A simple way to recognize regression is to ask whether the expected result is a number that can vary across a range. If it is, regression is usually the best fit.

  • Sentiment categories uses classification because the output is a label such as positive or negative.
  • Customer segments uses clustering because it groups similar records without predefined target labels.
  • Generated descriptions is a generative AI task because it creates new text rather than predicts a numeric value.

Question 10

Topic: Describe Features of Natural Language Processing (NLP) Workloads on Azure

A travel company is building a kiosk that must listen to customers speaking German and return the request in English in near real time. The team wants a prebuilt Azure AI service and does not want to train a custom model. Which Azure service category should they choose?

Options:

  • A. Azure Machine Learning

  • B. Azure AI Speech

  • C. Azure AI Language

  • D. Azure AI Vision

Best answer: B

Explanation: This scenario requires translating spoken audio from one language to another, which is a speech workload. Azure AI Speech is the prebuilt Azure service family designed for speech recognition, synthesis, and speech translation.

Azure AI Speech is the best fit when the input is audio and the system must recognize, synthesize, or translate speech. Here, the kiosk must listen to spoken German and provide the result in English, which matches speech translation at a fundamentals level. Because the company wants a prebuilt Azure AI service, it should use a service that already supports speech capabilities instead of building and training its own model.

Services centered on text analysis or image analysis do not match an audio-based translation requirement. The key clue is that the source content is spoken language, not text or images.

  • Text analytics mismatch The language service is mainly for analyzing and understanding text, not translating live spoken audio.
  • Vision mismatch The vision service works with images and video, such as OCR or object detection, not speech.
  • Custom ML not needed The machine learning platform can build custom models, but it is not the best choice when a prebuilt speech service already fits the requirement.

Question 11

Topic: Describe Artificial Intelligence Workloads and Considerations

A bank uses Azure OpenAI Service to build a generative AI assistant that summarizes loan products and answers customer questions. During pilot testing, employees notice that the assistant gives less complete guidance when applicants mention being self-employed. The bank wants a high-level responsible AI mitigation step before broader rollout, without redesigning the system in detail. Which action is the best choice?

Options:

  • A. Test diverse applicant scenarios and review outputs for bias

  • B. Use Azure AI Language to measure question sentiment

  • C. Use Azure AI Vision to classify uploaded documents

  • D. Cluster chat transcripts into common topic groups

Best answer: A

Explanation: This is a responsible AI fairness issue: one customer group is receiving lower-quality answers. The best first step is to evaluate the assistant with diverse, representative scenarios and review whether outputs differ unfairly before wider release.

When an AI system gives weaker results for a particular group, the main responsible AI concern is fairness. At a fundamentals level, the best mitigation is to test the system with representative prompts and scenarios that reflect different user types, then compare the quality of the outputs. That helps identify whether the model is treating groups differently and supports follow-up actions such as refinement, review, or human oversight.

Changing to a different Azure AI service does not address the fairness problem if the workload is still a generative assistant. Similarly, organizing transcripts by topic may be useful for analysis, but it does not directly check whether one group is receiving worse answers. The key idea is to evaluate for bias where the issue appears.

  • Wrong workload the document-classification option uses a vision service for images, not a mitigation for unfair text responses from a generative assistant.
  • Wrong signal measuring sentiment can detect opinion or emotion in text, but it does not determine whether answers differ unfairly by applicant type.
  • Wrong ML goal clustering transcripts can group similar chats, but it does not directly test or reduce biased responses.

Question 12

Topic: Describe Features of Natural Language Processing (NLP) Workloads on Azure

A retailer records live product briefings in Spanish. Employees want a searchable written transcript of each briefing in Spanish. The company does not need the content translated into another language or converted back into audio. Which Azure AI Speech capability should they use?

Options:

  • A. Speech recognition

  • B. Speech translation

  • C. Speech synthesis

  • D. Key phrase extraction

Best answer: A

Explanation: Speech recognition is used when spoken words must be converted into text. In this scenario, the goal is a transcript in the original language, not translated output or generated speech.

The key decision is the required output. Speech recognition takes audio input and produces text in the same language, which is exactly what a transcript requires. Speech translation is for converting spoken content into another language, so it adds a capability the retailer does not need. Speech synthesis does the reverse direction by generating spoken audio from text. Key phrase extraction can analyze text that already exists, but it does not first turn speech into text. When a scenario asks for spoken language to become searchable written text without changing languages, speech recognition is the best fit.

  • Speech translation is for changing spoken content into another language, which the scenario explicitly does not need.
  • Speech synthesis creates audio from text, so it does not produce a transcript from recorded speech.
  • Key phrase extraction can analyze text after transcription, but it cannot convert the original audio into text.

Question 13

Topic: Describe Features of Natural Language Processing (NLP) Workloads on Azure

A company wants a prebuilt Azure service to analyze written customer reviews for sentiment, extract key phrases, and identify entities such as product names and cities. Which service should the company use?

Options:

  • A. Azure AI Vision

  • B. Azure Machine Learning

  • C. Azure AI Speech

  • D. Azure AI Language

Best answer: D

Explanation: Azure AI Language is the Azure service designed for prebuilt NLP analysis of text. It can evaluate sentiment, extract key phrases, and recognize named entities from written content without requiring a custom model.

The key concept is choosing the Azure service that matches the workload. When the input is written text and the goal is to detect sentiment, pull out key phrases, or identify named entities, Azure AI Language is the best fit because those are built-in text analysis capabilities. This is the standard AI-900 service choice for common NLP analytics tasks on text.

Azure AI Speech is meant for spoken language scenarios such as speech-to-text or text-to-speech. Azure AI Vision focuses on images and video. Azure Machine Learning is used to build and manage custom machine learning solutions when a prebuilt service does not already meet the requirement. The deciding clue in the question is the need for prebuilt text analytics on written reviews.

  • The speech option fits audio workloads, such as recognizing or generating spoken language, not analyzing written reviews.
  • The vision option is for image and video analysis, so it does not match a text-only requirement.
  • The machine learning option is for custom model development, which is unnecessary when a prebuilt text analytics service already fits.

Question 14

Topic: Describe Artificial Intelligence Workloads and Considerations

A finance department receives thousands of scanned invoices each month. They need an AI solution that can read printed and handwritten text, identify fields such as invoice number, total, and due date, and send the results to an accounting system as structured data. Which AI workload category is the best fit?

Options:

  • A. Natural language processing

  • B. Generative AI

  • C. Document processing

  • D. Computer vision

Best answer: C

Explanation: The goal is to turn scanned invoices into structured data by reading text and locating specific fields. That matches document processing, which combines OCR and document understanding for forms, invoices, and similar business records.

Document processing is the best fit when the input is a business document and the desired output is structured data. In this scenario, the invoices are scanned files, but the company does not just want to recognize that an invoice exists. It wants to read printed and handwritten text, find fields such as invoice number, total, and due date, and return those values in a usable format. That is a document processing workload, commonly associated with services like Azure AI Document Intelligence.

Computer vision is broader image analysis, NLP focuses on understanding language in text that is already available, and generative AI is for creating new content such as summaries or chat responses. The key clue is extracting named fields from forms.

  • Computer vision is a broader image-analysis category, but the scenario needs form field extraction rather than general image labeling or detection.
  • Natural language processing works on language meaning after text is available; it is not the best match for reading scanned invoices and capturing fields.
  • Generative AI is useful for chat and content creation, but it is not the primary workload for reliable invoice data extraction.

Question 15

Topic: Describe Features of Generative AI Workloads on Azure

A company wants an Azure-based solution for its employee portal. Which requirement is the best fit for a generative AI assistant rather than simple search, retrieval, or a rules-only workflow?

Options:

  • A. Return ranked documents that match the user’s keywords.

  • B. Create a natural-language reply using multiple policy documents and chat context.

  • C. Show a predefined article for each known error code.

  • D. Apply fixed if-then rules to approve requests.

Best answer: B

Explanation: A generative AI assistant creates new natural-language output from prompts and supplied context. Combining multiple policy documents and chat history into a tailored reply is generation, while the other choices only find stored content or execute predefined logic.

The key feature of a generative AI assistant is that it produces a fresh response, usually in natural language, based on a user prompt and relevant context. Here, the system must synthesize information from several policy documents and the ongoing conversation to create an answer that fits the user’s specific question. That is different from search, which returns matching sources; retrieval, which surfaces existing stored content; or rules-based workflows, which follow explicit if-then conditions. In AI-900, prompts to compose, summarize, rephrase, or continue a conversation usually indicate a generative AI workload. A ranked result list may support an assistant, but by itself it is still search rather than generation.

  • The ranked-document idea is search because it returns matching sources instead of composing an answer.
  • The predefined-article idea is retrieval or scripted lookup because the response already exists.
  • The fixed approval logic is deterministic automation based on explicit rules, not generative AI.

Question 16

Topic: Describe Fundamental Principles of Machine Learning on Azure

Which Azure service supports data and compute resources for machine learning experimentation, training, and model development?

Options:

  • A. Azure AI Language

  • B. Azure Machine Learning

  • C. Azure AI Vision

  • D. Azure AI Speech

Best answer: B

Explanation: Azure Machine Learning is the Azure platform for creating and training machine learning models. It supports the data and compute resources needed for experiments and model development rather than a single prebuilt AI task.

Azure Machine Learning is designed for machine learning workflows in which you develop models by using your own data and Azure-managed resources. At the AI-900 level, the key idea is that it supports experimentation, training, and model development by giving you access to data assets and compute resources for training runs. This is different from prebuilt Azure AI services, which focus on ready-made capabilities such as analyzing images, processing natural language, or recognizing speech. When the requirement is to build or train a model instead of using a single prebuilt AI feature, Azure Machine Learning is the best fit. The key distinction is custom ML development versus prebuilt AI capabilities.

  • The speech option is for audio tasks such as speech recognition and speech synthesis, not model experimentation.
  • The vision option is for computer vision tasks such as image analysis and OCR.
  • The language option is for NLP tasks such as sentiment analysis and entity extraction.

Question 17

Topic: Describe Features of Generative AI Workloads on Azure

A company is choosing one Azure AI service for an internal help desk portal. Which requirement is the best fit for Azure OpenAI Service?

Options:

  • A. A service that detects sentiment and extracts entities from emails

  • B. A tool that translates spoken requests into another language

  • C. A chat assistant that summarizes past tickets and drafts reply suggestions

  • D. A feature that reads printed text from scanned forms

Best answer: C

Explanation: Azure OpenAI Service is the best match when a solution must generate natural-language content during a conversation. Summarizing ticket history and drafting reply suggestions for agents are classic generative AI tasks supported by Azure-hosted large language models.

The deciding factor is whether the requirement is to generate new content rather than only analyze or recognize existing content. Azure OpenAI Service supports generative AI scenarios such as chat, summarization, drafting, and code generation by using large language models in Azure. In this scenario, the portal must both summarize earlier tickets and create suggested replies, which makes it a generative assistant use case.

The other requirements describe specialized AI tasks: sentiment and entity extraction are text analytics, reading printed text is OCR, and spoken-language translation is a speech workload. Those are useful Azure AI capabilities, but they are not the primary choice for a chat-based system that produces new natural-language responses.

  • Text analytics matches Azure AI Language because sentiment detection and entity extraction analyze text instead of generating reply content.
  • OCR task matches Azure AI Vision because the goal is to read printed text from images or scanned files.
  • Speech translation matches Azure AI Speech because the input is spoken language that must be recognized and translated.

Question 18

Topic: Describe Features of Computer Vision Workloads on Azure

A retailer wants to automate several image tasks in Azure. Which requirement should use the OCR capability of Azure AI Vision rather than general image classification?

Options:

  • A. Classify store photos as indoor or outdoor

  • B. Classify shelf photos as empty or stocked

  • C. Read totals from scanned receipt images

  • D. Classify product photos as shirts or shoes

Best answer: C

Explanation: OCR is used when the business needs to read characters or numbers that appear inside an image, such as text on a receipt. Image classification is used when the goal is to assign a label to the overall image, such as product type or scene category.

OCR and image classification solve different computer vision problems. OCR reads text that appears within an image or document, so it fits scenarios involving receipts, invoices, forms, and labels where the business needs actual words or numbers. In this scenario, reading the total from a scanned receipt requires extracting text from the image.

General image classification does not return the text inside an image. Instead, it predicts a category for the image as a whole, such as a clothing type, shelf status, or scene type. A simple AI-900 rule is: if the requirement is to read text, use OCR; if the requirement is to label the picture, use image classification.

  • Classifying product photos as shirts or shoes is a whole-image labeling task, so it fits image classification.
  • Classifying shelf photos as empty or stocked predicts a visual category from the scene, not text from the image.
  • Classifying store photos as indoor or outdoor also assigns an overall image label rather than extracting characters.

Question 19

Topic: Describe Features of Computer Vision Workloads on Azure

A presenter is introducing Azure AI services in an AI-900 workshop. Which statement about Azure AI Vision is appropriate because it focuses on service capability rather than implementation detail?

Options:

  • A. It can extract text from images and analyze visual content.

  • B. It trains tabular prediction models from labeled rows.

  • C. It converts spoken audio into text transcripts.

  • D. Its key distinction is the SDK package used to call image-analysis methods.

Best answer: A

Explanation: AI-900 emphasizes matching business needs to Azure AI service capabilities. For Azure AI Vision, that means recognizing image-related tasks such as OCR and image analysis, not focusing on SDK, API, or deployment specifics.

The core AI-900 skill is identifying the right Azure AI service by workload and capability. Azure AI Vision is used for computer vision tasks such as analyzing images and extracting text from them. That is the correct level of description for a fundamentals exam.

Details like SDK packages, method names, endpoints, or deployment choices are implementation concerns. They matter when building a solution, but they are not the main focus when you are learning how to recognize what service fits a requirement. In AI-900, you should usually ask, “What kind of task is needed?” and then map that task to the right service family.

A good rule is: describe the service by what it can do, not by how a developer calls it.

  • SDK focus: The option about an SDK package shifts to implementation detail instead of describing the service capability.
  • Wrong workload: The option about converting spoken audio to text describes Azure AI Speech, not a vision service.
  • Wrong service type: The option about training tabular prediction models describes machine learning, typically associated with Azure Machine Learning rather than Azure AI Vision.

Question 20

Topic: Describe Artificial Intelligence Workloads and Considerations

A logistics company wants a mobile app that takes photos of shipping labels and automatically captures the printed tracking number and destination address for indexing. The company does not need to identify package types or generate a summary. Which Azure AI capability is the best fit?

Options:

  • A. Sentiment analysis in Azure AI Language

  • B. OCR in Azure AI Vision

  • C. Text summarization in Azure OpenAI Service

  • D. Image classification in Azure AI Vision

Best answer: B

Explanation: This scenario is about extracting text from images. OCR in Azure AI Vision reads printed characters from a photo and returns the text so the company can store tracking numbers and addresses. The requirement is not to categorize the image or generate new content.

The core computer vision workload here is OCR. OCR is used when the input is an image or scanned document and the goal is to read text, such as label numbers, addresses, or invoice fields. Azure AI Vision includes OCR capabilities for extracting printed text from photos.

Image classification would label an image as a package, label, or form, but it would not return the actual tracking number or address. Sentiment analysis is an NLP workload for opinions in text, and text summarization with Azure OpenAI Service is a generative AI task that shortens existing text rather than reading text from an image. When the need is “take a picture and pull out the words,” think OCR.

  • Image labeling mismatch image classification can categorize a label image, but it does not extract the printed characters.
  • Wrong AI domain sentiment analysis works on text meaning and opinion, not on reading words from an image.
  • Generative AI confusion text summarization can shorten text after it is available, but it does not perform image-to-text extraction.

Question 21

Topic: Describe Features of Natural Language Processing (NLP) Workloads on Azure

A company stores customer support emails as free text. It wants to automatically capture customer names, order numbers, product names, and dates from each email into separate database fields. Which NLP task is most appropriate?

Options:

  • A. Sentiment analysis

  • B. Language detection

  • C. Key phrase extraction

  • D. Named entity recognition

Best answer: D

Explanation: This scenario is about turning unstructured text into structured data fields. Named entity recognition is the NLP task designed to find and label specific facts such as people, dates, products, and identifiers in text.

The core concept is extracting structured facts from free-form text. Named entity recognition (NER) is used when you need to identify specific items in text and classify them into useful categories, such as person names, organizations, dates, locations, or other domain-relevant values like order numbers and product names. That makes it a good fit when the goal is to populate database columns from emails, notes, or documents.

By contrast, some NLP tasks summarize meaning without producing labeled fields. Key phrase extraction finds important terms, and sentiment analysis detects opinion or emotion. Language detection only identifies which language the text uses. When the requirement is to pull out specific facts and store them in structured form, NER is the best match.

  • Sentiment analysis focuses on whether the message is positive, negative, or neutral, not on pulling out fields like dates or order numbers.
  • Language detection only determines the language of the email, which does not extract business facts.
  • Key phrase extraction can surface important terms, but it does not reliably label them into structured categories for separate fields.

Question 22

Topic: Describe Artificial Intelligence Workloads and Considerations

A retailer is comparing possible Azure AI projects. Which requirement represents a content-generation use case instead of a predictive AI use case?

Options:

  • A. Generate first-draft product descriptions from item attributes

  • B. Predict whether a shipment will arrive late

  • C. Estimate the probability that a customer will cancel

  • D. Forecast next month’s sales for each store

Best answer: A

Explanation: Generating first-draft product descriptions is a generative AI workload because the system creates new natural-language content. The other scenarios use data to estimate future values or outcomes, which makes them predictive AI use cases.

The core difference is whether the AI system is predicting something or creating something. Predictive AI uses existing data to estimate a future value, class, or probability, such as sales forecasts, late shipments, or customer churn. Content-generation use cases ask the AI to produce new content such as text, images, or code.

In this scenario, drafting product descriptions from item attributes means the system must generate original text. That fits a generative AI workload. By contrast, forecasting sales, predicting delays, and estimating cancellations all produce predictions about future business outcomes.

A simple test is: if the output is a forecast, score, or label, think predictive AI; if the output is new content, think generative AI.

  • Forecasting sales fails because it predicts a future numeric value rather than creating new content.
  • Predicting late shipments fails because it estimates an operational outcome from existing data.
  • Estimating customer cancellation fails because churn probability is a prediction, not generated content.

Question 23

Topic: Describe Fundamental Principles of Machine Learning on Azure

Which Azure service is Microsoft’s cloud platform for data science, model training, deployment, and lifecycle management?

Options:

  • A. Azure OpenAI Service

  • B. Azure AI Vision

  • C. Azure AI Language

  • D. Azure Machine Learning

Best answer: D

Explanation: Azure Machine Learning is the end-to-end machine learning platform in Azure. It is designed for data science work such as training custom models, deploying them, and managing the ML lifecycle.

The key distinction is between a general machine learning platform and task-specific AI services. Azure Machine Learning is Microsoft’s Azure service for end-to-end ML work: data scientists can use it to build experiments, train and validate models, deploy models to endpoints, and manage lifecycle activities such as versioning and monitoring. That makes it the right choice when an organization needs a platform for creating and operating custom machine learning solutions.

By contrast, services focused on vision, language, or generative AI provide prebuilt capabilities for particular workloads rather than the main Azure platform for custom ML lifecycle management.

  • Vision mismatch supports image analysis workloads, not full custom model training and lifecycle management.
  • Language mismatch provides NLP capabilities such as sentiment analysis and entity extraction, not a general ML platform.
  • Generative AI mismatch gives access to foundation models for chat and content generation, not the core Azure platform for custom ML operations.

Question 24

Topic: Describe Features of Generative AI Workloads on Azure

A retail company wants an internal assistant that summarizes product manuals and drafts customer-service replies. The team plans to use Azure OpenAI Service, but a manager says responsible AI review is unnecessary because the assistant generates new text instead of only analyzing existing data. Which approach is best?

Options:

  • A. Use Azure AI Language because summarization and reply drafting are non-generative NLP tasks.

  • B. Use Azure OpenAI Service and include safety testing, transparency, and human review.

  • C. Use Azure AI Vision because responsible AI concerns mainly apply to image workloads.

  • D. Use Azure OpenAI Service and skip responsible AI review because the outputs are original.

Best answer: B

Explanation: Responsible AI principles apply to generative AI as well as analytical AI. An assistant that summarizes content and drafts replies can still create unsafe, misleading, or biased output, so it should be reviewed for safety, transparency, and appropriate human oversight.

Responsible AI is not limited to systems that classify or predict from existing data. A generative AI solution that creates new text can still introduce risk, such as inaccurate answers, biased wording, or unsafe content. In this scenario, Azure OpenAI Service fits the requirement to summarize manuals and draft reply suggestions, but the team should also apply responsible AI practices before deployment.

  • Test for harmful or incorrect outputs.
  • Be transparent that the assistant is AI-generated.
  • Keep humans involved in reviewing important responses.

The key point is that generating original content does not remove the need for responsible AI controls.

  • Skipping review fails because generative output can still be harmful, biased, or incorrect.
  • Using Azure AI Language misses the drafting requirement because that service focuses on language analysis and prebuilt NLP features rather than general text generation.
  • Using Azure AI Vision is a workload mismatch because the scenario is about generating and summarizing text, not analyzing images.

Question 25

Topic: Describe Features of Generative AI Workloads on Azure

A company is comparing several AI features for a support portal. Which requirement is the best fit for a generative AI solution such as Azure OpenAI Service?

Options:

  • A. Classify each review as positive or negative

  • B. Detect damaged products in return photos

  • C. Extract printed text from scanned warranty cards

  • D. Draft a reply to a customer question from product documentation

Best answer: D

Explanation: Drafting a reply from a customer question and supporting documents is a classic generative AI scenario. Generative models create new natural-language content, such as answers, summaries, and assistant responses, while the other options classify data or extract existing information.

Generative AI is used when a system must create new content from a prompt and context. In this scenario, producing a reply from a customer question and product documentation fits Azure OpenAI Service because the model generates natural-language text for an agent to review or use.

The other requirements are different AI workloads. Sentiment classification assigns a label such as positive or negative. OCR reads existing text from scanned documents or images. Damage detection in photos is a computer vision task that analyzes image content. The key clue is that only one option asks the AI to author a new response rather than identify, classify, or extract information.

  • Sentiment label classifying reviews as positive or negative is an NLP analysis task, not text generation.
  • OCR task extracting printed text from scanned cards reads existing content instead of drafting new content.
  • Vision analysis detecting damage in photos is a computer vision use case focused on image interpretation.

Questions 26-50

Question 26

Topic: Describe Features of Computer Vision Workloads on Azure

Contoso is reviewing several image-related requirements for Azure. Which requirement is primarily a facial detection use case?

Options:

  • A. Match a visitor to a stored employee identity record

  • B. Detect faces in event photos so they can be blurred

  • C. Determine which employee entered a restricted room

  • D. Enforce consent rules for uploaded profile photos

Best answer: B

Explanation: Facial detection is about finding whether faces are present and where they are in an image. The photo-blurring requirement depends on locating faces, while the other scenarios focus on identity verification, security investigation, or compliance policy.

Face detection is a computer vision workload that answers questions like: Is there a face in this image, and where is it located? That makes it a good fit for tasks such as blurring, cropping, or counting faces in photos. In this scenario, automatically blurring faces requires the system to detect each face first.

The other requirements are different kinds of problems. Matching a person to an existing record is identity verification or recognition. Determining which employee entered a secure room is mainly a security and identity investigation. Enforcing consent rules for profile photos is a compliance and governance process. The key distinction is whether the primary need is simply to detect faces, rather than decide who the person is or whether a policy has been followed.

  • Identity matching goes beyond detection because it tries to link a face to a known person.
  • Security investigation focuses on proving who entered a restricted area, not just locating faces.
  • Compliance policy is about consent and governance, even though photos are involved.

Question 27

Topic: Describe Features of Generative AI Workloads on Azure

A company wants to add a virtual assistant to its employee portal. The assistant must chat with users, generate draft replies from prompts, and summarize long policy documents. Which Azure service is the most appropriate choice?

Options:

  • A. Azure AI Speech

  • B. Azure OpenAI Service

  • C. Azure AI Language

  • D. Azure Machine Learning

Best answer: B

Explanation: Azure OpenAI Service is designed for generative AI workloads that create new text, support chat experiences, and summarize content from prompts. The scenario describes a copilot-style assistant, which aligns directly with large language models.

This scenario is about a generative AI workload. When the requirement is to chat with users, draft natural-language responses, and summarize long text, the best Azure fit is Azure OpenAI Service because it provides access to large language models for prompt-based generation. These models are commonly used for conversational assistants, text summarization, and drafting helpful responses. By contrast, Azure AI Language is primarily associated with prebuilt NLP analysis capabilities such as sentiment analysis or entity extraction, even though it can also perform some text tasks. Azure AI Speech focuses on spoken audio, and Azure Machine Learning is the broader platform for building and managing custom ML solutions. The key clue is the need for generated conversational output, not just text analysis.

  • Language analysis fits many NLP tasks, but the stem centers on prompt-based chat and generated replies.
  • Speech processing is for recognizing or synthesizing audio, not for a text-first generative assistant.
  • Custom ML platform is broader than needed when a dedicated generative AI service already matches the requirement.

Question 28

Topic: Describe Fundamental Principles of Machine Learning on Azure

An organization uses Azure OpenAI Service to summarize support tickets and draft replies. Which model architecture underpins many modern generative AI and language models for these tasks?

Options:

  • A. Linear regression model

  • B. K-means clustering

  • C. Transformer architecture

  • D. Convolutional neural network

Best answer: C

Explanation: Modern generative AI for text commonly relies on transformer architecture. Transformers underpin many large language models used for tasks such as summarization, drafting, and question answering. In Azure contexts, this is the core model family behind many language-generation capabilities.

Transformer architecture is the foundation of many modern generative AI and language models. It is designed to work with sequences of tokens and learn relationships across text, which makes it effective for tasks such as text generation, summarization, translation, and question answering. In AI-900, the important idea is recognizing transformers as the model family behind many large language models, including models accessed through Azure OpenAI Service. By contrast, k-means groups similar data points, linear regression predicts numeric values, and convolutional neural networks are most closely associated with image analysis. For modern text-centric generative AI, think transformers rather than classic prediction or clustering methods.

  • Vision focus The convolutional neural network option is mainly associated with image analysis rather than most modern text-generation models.
  • Unsupervised grouping The k-means option clusters similar items but does not generate natural-language content.
  • Numeric prediction The linear regression option predicts continuous values, not summaries or drafted replies.

Question 29

Topic: Describe Features of Computer Vision Workloads on Azure

A retailer wants to review each returned product photo and label it as damaged or not damaged. The solution does not need to show where the damage appears in the image. Which computer vision task is most appropriate?

Options:

  • A. Face detection

  • B. Optical character recognition (OCR)

  • C. Image classification

  • D. Object detection

Best answer: C

Explanation: This scenario is about assigning one label to an entire product photo. Because the retailer only needs to decide whether the photo shows damage and does not need the damage location, image classification is the best fit.

Image classification is the computer vision task used when you want to predict a category for a whole image. In this case, each product photo should be labeled as damaged or not damaged, so the model only needs to choose the correct class for the image.

Object detection would be more appropriate if the retailer needed boxes around the damaged area or needed to locate multiple defects within the photo. OCR is for reading printed or handwritten text from images, and face detection is for finding human faces. The key distinction is whether you need a single label for the entire image or the location of something inside it.

  • Object location mismatch The option about detecting objects is unnecessary because the stem says the solution does not need to show where the damage is.
  • Text extraction mismatch The option about OCR fits reading labels, receipts, or packaging text, not judging visible product damage.
  • Wrong image target The option about face detection is specialized for identifying faces, which is unrelated to inspecting product condition.

Question 30

Topic: Describe Fundamental Principles of Machine Learning on Azure

A delivery company wants to use package weight, route distance, and traffic conditions to predict how many minutes a new delivery will take. Which machine learning technique best fits this goal?

Options:

  • A. Clustering

  • B. Regression

  • C. Classification

  • D. Reinforcement learning

Best answer: B

Explanation: This scenario requires predicting a numeric value. Because the target is delivery time measured in minutes, the best-fit machine learning technique is regression.

Regression is the correct choice when a model must predict a continuous number, such as time, cost, temperature, or sales. In this scenario, package weight, route distance, and traffic conditions are the features, and the output is delivery time in minutes. That output is numeric and can take many possible values, so this is not a category-labeling problem.

Classification would fit if the goal were to predict a label such as late or on time. Clustering would fit if the company wanted to group deliveries by similarity without predefined labels. Reinforcement learning is used when a system learns actions through rewards over time. When the target is a number, regression is the best fit.

  • Classification mismatch fits predicting labels or categories, not an exact number of minutes.
  • Clustering mismatch groups similar records without using a known target value.
  • Reinforcement learning mismatch is for learning decision policies from rewards, not direct numeric prediction from labeled examples.

Question 31

Topic: Describe Fundamental Principles of Machine Learning on Azure

A retailer plans to use Azure OpenAI Service to build an assistant that can chat with customers, answer free-form product questions, and summarize long return-policy documents. Which model architecture underpins many modern generative AI and language models used for this kind of solution?

Options:

  • A. Convolutional neural network (CNN)

  • B. Transformer architecture

  • C. K-means clustering

  • D. Linear regression

Best answer: B

Explanation: The scenario requires generative language capabilities such as conversation, question answering, and summarization. Many modern language models that perform these tasks are based on transformer architecture, which makes it the best match.

Transformer architecture is the foundation for many large language models used in generative AI. In a scenario like this one, the assistant must understand natural language, keep track of context, and generate useful text responses and summaries. Transformers are well suited to those language tasks, so they commonly underpin modern generative AI solutions, including models used for chat-style experiences.

  • Chatting with users is a generative language task.
  • Answering free-form questions requires language understanding and text generation.
  • Summarizing documents is also a common transformer-based capability.

The key distinction is that the scenario needs generative language modeling, not image analysis, data grouping, or numeric prediction.

  • Image focus The convolutional neural network option is mainly associated with computer vision tasks such as image classification.
  • Unsupervised grouping The k-means clustering option groups similar data points, but it does not generate conversational text.
  • Numeric prediction The linear regression option predicts continuous numbers, not natural-language responses or summaries.

Question 32

Topic: Describe Artificial Intelligence Workloads and Considerations

A retail company receives thousands of supplier invoices as PDFs. It needs to extract the vendor name, invoice number, and total amount from each file and store them as structured fields in an accounting system. The company does not want a chatbot or generated summaries; it only wants the values already shown in each document. Which Azure service is the best fit?

Options:

  • A. Azure AI Document Intelligence

  • B. Azure AI Vision

  • C. Azure OpenAI Service

  • D. Azure AI Language

Best answer: A

Explanation: This is a document processing requirement, not a generative AI requirement. Azure AI Document Intelligence is built to extract existing text and fields from business documents and return them as structured data.

The key decision is whether the system must extract existing information or create new content. In this scenario, the company wants specific values already printed on invoices—such as vendor name, invoice number, and total amount—and wants those values stored in structured fields. That is a document processing workload.

Azure AI Document Intelligence is the best fit because it is designed to read forms and business documents and pull out relevant fields. Generative AI services, such as Azure OpenAI Service, are more appropriate when the goal is to draft summaries, answer questions in natural language, or produce new text. When the requirement is “read the document and return the data already on it,” choose document processing rather than generative AI.

  • Generative output The option using Azure OpenAI Service is meant for creating summaries, answers, or other new text rather than extracting fixed invoice fields.
  • Language analysis The option using Azure AI Language focuses on text understanding tasks such as sentiment and entity extraction, not document field capture from PDFs.
  • General vision The option using Azure AI Vision can analyze images and read text, but it is not the best choice for structured invoice data extraction.

Question 33

Topic: Describe Artificial Intelligence Workloads and Considerations

A company wants an AI solution that can summarize meeting notes, draft follow-up emails, and answer employee questions in a chat experience by generating new text. Which AI workload type best fits this need?

Options:

  • A. Computer vision workload

  • B. Predictive machine learning workload

  • C. Generative AI workload

  • D. Natural language processing workload

Best answer: C

Explanation: The key requirement is generating new text for summaries, drafts, and chat responses. That makes generative AI the best fit because it creates content from prompts and context instead of only analyzing existing text.

Generative AI workloads are used when a solution must produce original content, such as summaries, drafted emails, or conversational answers. In this scenario, the system is not just detecting sentiment or extracting facts from text; it must generate human-like responses based on provided information and user questions.

Traditional NLP workloads usually focus on analyzing language, such as identifying key phrases, entities, sentiment, or intent. Computer vision workloads analyze images and video, and predictive machine learning workloads forecast values or assign labels from data. When the main goal is to create new text content for users, generative AI is the clearest match.

  • NLP analysis is tempting because it works with text, but classic NLP usually analyzes or extracts information instead of drafting new content.
  • Computer vision applies to images and video, so it does not fit a text summarization and chat scenario.
  • Predictive ML is used to predict categories or numbers, not to compose summaries or conversational replies.

Question 34

Topic: Describe Features of Natural Language Processing (NLP) Workloads on Azure

A company is building an airport kiosk. The kiosk must convert travelers’ spoken questions to text, speak responses aloud, and translate spoken conversations between English and French in real time. Which Azure service category should the company choose?

Options:

  • A. Azure AI Vision

  • B. Azure OpenAI Service

  • C. Azure AI Language

  • D. Azure AI Speech

Best answer: D

Explanation: The scenario requires three speech capabilities: speech-to-text, text-to-speech, and speech translation. Azure AI Speech is the Azure service category built for spoken-language input, output, and translation, so it best matches the kiosk requirements.

Azure AI Speech is the Azure service category for applications that work with spoken language. In this scenario, the kiosk must recognize audio from travelers, generate spoken replies, and translate speech between languages. Those needs map directly to the core Azure AI Speech capabilities.

  • Spoken audio to text uses speech recognition.
  • Text to natural-sounding audio uses speech synthesis.
  • Real-time spoken language conversion uses speech translation.

Services centered on text analysis, images, or generative text do not directly provide this full speech workflow. When the main requirement is understanding and producing human speech, Azure AI Speech is the best fit.

  • Text analytics The option based on Azure AI Language fits tasks like sentiment analysis or entity extraction on text, not core audio processing.
  • Image analysis The option based on Azure AI Vision is for images and video rather than recognizing or generating speech.
  • Generative AI The option based on Azure OpenAI Service can generate text, but it is not the primary Azure service category for speech recognition and speech synthesis.

Question 35

Topic: Describe Fundamental Principles of Machine Learning on Azure

A team uses Azure Machine Learning to build a simple image classification model. The training files are already stored and registered in the workspace. The team now needs a resource that supplies CPU or GPU capacity to run training jobs. Which Azure Machine Learning resource is a compute service?

Options:

  • A. Label

  • B. Compute cluster

  • C. Data asset

  • D. Datastore

Best answer: B

Explanation: In Azure Machine Learning, a compute cluster is used to run training workloads by providing managed compute capacity. Datastores and data assets help you connect to and organize data, while labels are target values used in supervised learning.

The key distinction is between resources that store or reference data and resources that execute work. In Azure Machine Learning, a compute cluster is a compute service: it provides managed CPU or GPU resources for training and other jobs. By contrast, a datastore is a connection to a storage location, and a data asset is a registered reference to data used in the workspace. A label is not a service at all; it is the value a model is trained to predict in supervised learning.

If the need is to run training jobs, choose a compute resource rather than a data resource. That is why the scalable training option is the best fit here.

  • Datastore confusion: a datastore connects Azure Machine Learning to storage, but it does not run jobs.
  • Data asset confusion: a data asset registers and organizes data for reuse, but it is not compute.
  • ML term mix-up: a label is part of training data, not an Azure Machine Learning service resource.

Question 36

Topic: Describe Features of Generative AI Workloads on Azure

A company wants to build a generative HR assistant in Azure. The assistant must answer employee questions in a chat window, summarize long policy updates, and draft tailored email replies from company documents. The team first needs to make the service-selection decision, not a model-behavior or responsible-AI decision. Which recommendation is the BEST fit?

Options:

  • A. Use Azure OpenAI Service for the assistant.

  • B. Use Azure AI Language for sentiment analysis.

  • C. Tune the model to reduce hallucinations first.

  • D. Prioritize transparency notices for AI-generated answers.

Best answer: A

Explanation: This scenario asks for the Azure service family that fits a generative AI assistant. Azure OpenAI Service matches chat, summarization, and text drafting, while transparency and hallucination control are important considerations but not the service-selection answer.

The core concept is separating a service choice from a model-behavior or responsible AI choice. Here, the workload is clearly generative AI because the solution must chat with users, summarize long text, and create new email drafts. In Azure, Azure OpenAI Service is the best fit for that kind of assistant.

Reducing hallucinations is about reliability and safety, and telling users that content is AI-generated supports transparency. Those are both valid design considerations, but they do not answer the question of which Azure service family to select. Azure AI Language is better aligned to traditional NLP tasks such as sentiment analysis, entity extraction, and similar text-analysis capabilities rather than a full generative assistant that produces new conversational and drafted content.

When the requirement is generative chat plus summaries and drafted text, start with Azure OpenAI Service.

  • Text analysis mismatch The Azure AI Language option focuses on traditional NLP tasks like sentiment analysis, not a full generative assistant.
  • Responsible AI, not service The transparency option is good governance practice, but it does not choose the Azure service family.
  • Behavior, not platform The hallucination option addresses model quality and safety after or alongside service selection, not instead of it.

Question 37

Topic: Describe Artificial Intelligence Workloads and Considerations

A retail company receives thousands of supplier invoices as scanned PDFs. It needs to extract the text and capture fields such as invoice number, invoice date, and total amount from these structured or semi-structured documents so the data can be sent to an accounting system. Which Azure AI service is the best fit?

Options:

  • A. Azure AI Vision

  • B. Azure AI Language

  • C. Azure OpenAI Service

  • D. Azure AI Document Intelligence

Best answer: D

Explanation: Azure AI Document Intelligence is designed for document processing scenarios such as invoices, receipts, and forms. It can extract both raw text and specific fields from structured or semi-structured files, which matches this accounting workflow better than general vision, language analysis, or generative AI.

The core concept is the difference between document processing and general computer vision. When the goal is to pull text and named fields from business documents like invoices, receipts, or forms, the best fit is Azure AI Document Intelligence. It is designed to understand document layout and extract items such as dates, totals, and identifiers from structured or semi-structured files.

General computer vision focuses more on analyzing image content, such as tags, objects, or basic OCR scenarios. Language services work after text is already available, and generative AI is for creating or transforming content rather than serving as the primary tool for reliable field extraction from forms.

If the need were image tagging or object detection instead of invoice field extraction, a general vision service would be a better match.

  • General vision is useful for image analysis and some OCR tasks, but it is not the best choice for extracting named fields from forms and invoices.
  • Language analysis works on text that has already been extracted, such as sentiment or entity detection, so it does not specialize in document layouts.
  • Generative AI can summarize or draft content, but it is not the primary service for dependable extraction of invoice fields.

Question 38

Topic: Describe Features of Computer Vision Workloads on Azure

A company is building an employee badge photo-check app. The team wants a prebuilt Azure AI service that can confirm exactly one face is present in each uploaded photo, return the face location, and flag photos where the face is turned too far away from the camera. The app does not need to identify the person, read text, or generate captions. Which Azure AI solution is the BEST fit?

Options:

  • A. Use Azure AI Vision OCR

  • B. Use Azure AI Vision image classification

  • C. Use Azure OpenAI Service to generate captions

  • D. Use the Azure AI Face detection service

Best answer: D

Explanation: The requirement is face detection, not general image understanding. The Azure AI Face detection service can find faces, return their location, and provide face-related attributes such as pose, which matches the badge photo-check scenario.

Face detection is the most appropriate computer vision task when a solution must determine whether faces are present, count how many appear, and return information about each detected face. In this scenario, the app must confirm that exactly one face exists, get the face location, and check a face-related attribute such as pose. That maps directly to the Azure AI Face detection service.

  • Detects one or more faces in an image
  • Returns a face rectangle for each detection
  • Supports face-related analysis needed for basic photo checks

A general image classification capability labels an entire image, but it does not directly solve the need to locate a face and evaluate face-specific details.

  • OCR mismatch extracts printed or handwritten text, which does not help find or assess a face.
  • Whole-image labels image classification can categorize an image, but it does not directly return face location or pose.
  • Generative mismatch caption generation creates natural-language descriptions, but it is not the best tool for precise face detection.

Question 39

Topic: Describe Fundamental Principles of Machine Learning on Azure

A company wants a model to predict whether an incoming customer email should be labeled as sales, billing, or technical support. Which machine learning approach should it use?

Options:

  • A. Regression

  • B. Classification

  • C. Clustering

  • D. Anomaly detection

Best answer: B

Explanation: Classification is used when the prediction target is a named category rather than a numeric value. Here, the model must choose among discrete labels for each email, so classification is the best fit.

Classification is a supervised machine learning technique used to predict which category an item belongs to. In this scenario, each email must be assigned one label from a fixed set: sales, billing, or technical support. Because the output is a discrete class, not a number, the task is classification.

A simple way to distinguish common techniques is:

  • Classification predicts labels such as yes/no or named categories.
  • Regression predicts continuous numeric values such as revenue or temperature.
  • Clustering groups similar items when no labels are provided.
  • Anomaly detection finds unusual patterns or outliers.

When the goal is to predict a category or label, think classification first.

  • Regression confusion fails because regression predicts a number, not a support category.
  • Clustering confusion fails because clustering discovers groups without using known target labels.
  • Outlier focus fails because anomaly detection looks for unusual emails, not normal category assignment.

Question 40

Topic: Describe Features of Generative AI Workloads on Azure

A company is designing an assistant that drafts email replies for support agents. Which decision is primarily a service-selection decision rather than a model-behavior or responsible-AI decision?

Options:

  • A. Requiring human approval before sending replies

  • B. Choosing Azure OpenAI Service for reply drafting

  • C. Informing agents that replies are AI-generated

  • D. Adjusting prompts for a more formal tone

Best answer: B

Explanation: A service-selection decision is about choosing the Azure product that fits the workload. In this scenario, selecting Azure OpenAI Service chooses the generative AI service, while the other choices deal with output behavior or responsible AI practices.

The core idea is to separate which service you use from how the model behaves and how you govern its use. A generative AI assistant that drafts email replies needs a generative AI service, so choosing Azure OpenAI Service is the service-selection decision. Adjusting prompts changes the style or behavior of the model’s output. Informing users that content is AI-generated is a transparency practice. Requiring a person to approve messages before they are sent is an accountability safeguard.

When AI-900 asks about service selection, focus on the Azure service family that matches the workload; when it asks about behavior or responsible AI, focus on output control or governance instead.

  • Adjusting prompts changes how the model responds, so it is a model-behavior decision.
  • Informing agents that text is AI-generated addresses transparency, not Azure service selection.
  • Requiring human approval is an accountability control, not a choice of generative AI service.

Question 41

Topic: Describe Features of Natural Language Processing (NLP) Workloads on Azure

An online retailer stores customer review comments as text in Azure. The support team needs a prebuilt AI capability that labels each comment as positive, neutral, or negative so unhappy customers can be contacted quickly. Which solution is the best fit?

Options:

  • A. Use OCR in Azure AI Vision.

  • B. Use Azure OpenAI Service to generate reply text.

  • C. Use sentiment analysis in Azure AI Language.

  • D. Use key phrase extraction in Azure AI Language.

Best answer: C

Explanation: The requirement is to determine the opinion expressed in written customer reviews and sort it into sentiment categories. That is exactly what sentiment analysis does, and Azure provides it as a prebuilt NLP capability in Azure AI Language.

Sentiment analysis is used when you need to understand the emotional tone or opinion in text, such as customer reviews, survey comments, or social posts. In this scenario, the company wants each written review labeled as positive, neutral, or negative so it can spot unhappy customers quickly. That makes sentiment analysis the best fit, and in Azure this capability is part of Azure AI Language.

  • The input is text, not images or audio.
  • The goal is to classify opinion, not extract topics or generate new content.
  • A prebuilt NLP service is sufficient; no custom model is required.

Key phrase extraction can show what customers talk about, but it does not tell you whether the feedback is favorable or unfavorable.

  • Topic vs. tone key phrase extraction finds important terms, but it does not label overall opinion.
  • Image mismatch OCR reads text from images, while the reviews are already plain text.
  • Generative mismatch generating reply text creates new content, but the need here is to classify existing feedback.

Question 42

Topic: Describe Artificial Intelligence Workloads and Considerations

An Azure-based AI system helps pre-screen loan applications. The team notices that response times increase during busy periods, and applicants with similar income and debt levels receive different recommendations when their postal codes differ. Which concern is most directly a responsible AI principle rather than a general software-quality concern?

Options:

  • A. Latency during peak usage

  • B. Fairness of automated decisions

  • C. Maintainability of the solution

  • D. Scalability of the service

Best answer: B

Explanation: Fairness is the responsible AI principle concerned with whether similar people receive similar treatment and whether model outputs are biased. Here, different recommendations for comparable applicants point to fairness, while slow peak-time responses are standard software performance concerns.

Fairness is the responsible AI principle that focuses on avoiding unjustified differences in outcomes for similar individuals or groups. In this scenario, applicants with comparable financial data receive different recommendations when postal code changes, which suggests the model may be using a location-based proxy in a biased way. That is a responsible AI issue because it affects how people are treated by the system.

By contrast, slower responses during busy periods describe engineering quality concerns about speed and capacity, not ethical treatment. Maintainability is about how easily the application can be updated and supported over time.

When a scenario includes both ethical and technical problems, identify the option tied to the impact of the AI decision on people.

  • Latency fits the slow-response symptom, but it is about speed, not unequal decision outcomes.
  • Scalability concerns handling increased demand, which explains peak-load issues rather than biased recommendations.
  • Maintainability is about how easy the system is to update or fix, not whether similar applicants are treated differently.

Question 43

Topic: Describe Features of Natural Language Processing (NLP) Workloads on Azure

An insurance company is building a phone self-service system. Customers will ask questions by speaking, and the system must convert the audio to text and read approved answers back aloud. The company does not need image analysis or generative content creation. Which Azure AI service family is the best fit?

Options:

  • A. Azure AI Language

  • B. Azure AI Speech

  • C. Azure OpenAI Service

  • D. Azure AI Vision

Best answer: B

Explanation: Azure AI Speech is the best choice when the main interaction is audio. This scenario requires both converting callers’ speech into text and turning approved text responses back into spoken audio, which is different from text-only language analysis.

The key decision is the modality of the input and output. Azure AI Speech is designed for audio-based experiences, such as transcribing spoken words into text and generating spoken audio from text. In this scenario, customers speak their requests and expect spoken replies, so speech processing is the primary requirement.

Azure AI Language is used when the content is already text and you want to analyze or understand that text, such as sentiment analysis, entity extraction, or classification. Azure OpenAI Service is intended for generative AI tasks like drafting or summarizing content, and Azure AI Vision focuses on images and video. When the scenario centers on spoken interaction rather than text-only analysis, Azure AI Speech is the right service family.

  • The text-analysis option fits tasks like sentiment analysis or entity extraction after text already exists, not voice capture and spoken playback.
  • The generative AI option is for creating or transforming content, but the scenario says approved answers already exist.
  • The vision option works with images and video, not spoken customer calls.

Question 44

Topic: Describe Fundamental Principles of Machine Learning on Azure

What is the main purpose of a validation dataset in machine learning?

Options:

  • A. To check how well the model generalizes to data not used for learning

  • B. To supply live production data for generating predictions

  • C. To provide the final unbiased score after all tuning is complete

  • D. To adjust the model’s weights by learning from labeled examples

Best answer: A

Explanation: A validation dataset is used to evaluate a model on data it did not train on. This helps estimate whether the model is generalizing well instead of only memorizing the training data.

The core idea is to separate the data used for learning from the data used for checking performance. A model learns patterns from the training dataset, but the validation dataset is kept aside during that learning step. You then use the validation results to compare versions of a model, tune settings, and watch for overfitting.

If performance is strong on training data but weak on validation data, the model may not generalize well. The validation dataset is therefore a development-time checkpoint for unseen data, while a test dataset is usually reserved for the final evaluation after tuning is finished.

  • Learning data confusion: the option about adjusting weights describes the training dataset, not the validation dataset.
  • Final score confusion: the option about the final unbiased score matches the role of a test dataset.
  • Production data confusion: the option about live prediction input refers to inference data used after deployment.

Question 45

Topic: Describe Features of Natural Language Processing (NLP) Workloads on Azure

A retail company wants to process customer reviews as they are submitted on its website. The solution must identify whether each review is positive, neutral, or negative and extract key phrases from the text, but it does not need to draft replies or generate new content. Which Azure service family is the best fit?

Options:

  • A. Azure AI Vision

  • B. Azure OpenAI Service

  • C. Azure AI Language

  • D. Azure AI Speech

Best answer: C

Explanation: Azure AI Language is the best fit because the company needs to analyze existing text rather than generate new text. Sentiment analysis and key phrase extraction are core NLP capabilities, while generative AI services are better for drafting or conversational output.

When a scenario focuses on understanding text that already exists, think NLP. In this case, the company needs two classic NLP tasks: assign sentiment labels to reviews and extract important phrases. Azure AI Language is designed for those text analysis workloads with prebuilt capabilities.

Generative AI is a better fit when the main goal is to create new content, such as drafting replies, answering open-ended questions, or acting as an assistant. Azure OpenAI Service can work with text, but its core value is generation rather than standard text analytics. Azure AI Speech is for spoken language, and Azure AI Vision is for images and visual content.

A simple rule is: analyze existing text with NLP services; create new text with generative AI services.

  • The Azure OpenAI choice is tempting because it also works with text, but it is mainly used to generate or transform content rather than provide built-in sentiment and key phrase analysis.
  • The Azure AI Speech choice fits audio scenarios such as speech recognition or spoken translation, not text analytics on written reviews.
  • The Azure AI Vision choice is for images and visual content, including OCR, not sentiment analysis of review text.

Question 46

Topic: Describe Artificial Intelligence Workloads and Considerations

A company is evaluating Azure AI solutions for several tasks. Which task represents a generative AI workload rather than only analyzing existing content?

Options:

  • A. Assigning each product review a positive or negative label

  • B. Identifying whether a photo contains a bicycle

  • C. Drafting a support response from a customer’s issue description

  • D. Extracting printed text from scanned invoices

Best answer: C

Explanation: Generative AI is used when a system creates new content such as text, code, images, or summaries. Drafting a support response produces new text, while the other tasks only classify or extract information from existing content.

Generative AI workloads create original output based on a prompt or source material. That output can be a draft email, a support response, a summary, code, or an image. In this scenario, drafting a support response requires the system to compose new wording from the customer’s issue description, so it is a generative task.

The other options are analysis workloads. Sentiment labeling assigns a category, OCR extracts text that already exists in a document, and image identification classifies visual content. A good test is to ask whether the system must produce new content or only analyze what is already there. If it is creating new text or other content, it is generative AI.

  • Assigning positive or negative labels is classification because the output is a predefined category.
  • Extracting printed text from invoices is OCR, which reads existing content instead of creating new content.
  • Identifying a bicycle in a photo is computer vision classification, not content generation.

Question 47

Topic: Describe Features of Generative AI Workloads on Azure

A company wants to build a customer-support copilot in Azure. Before development starts, the team wants to review available foundation models, compare candidates, and choose the best one for the solution. Which Azure capability should they use?

Options:

  • A. Azure OpenAI Service

  • B. Azure AI Language

  • C. Azure AI Foundry model catalog

  • D. Azure Machine Learning

Best answer: C

Explanation: Azure AI Foundry model catalog is the Azure capability most directly focused on browsing available generative AI models and helping teams evaluate which one fits a scenario. The stem is about model discovery and selection, not about consuming a specific prebuilt AI service or building a full custom ML workflow.

For generative AI in Azure, the model-selection task is most directly supported by the Azure AI Foundry model catalog. It provides a central place to explore foundation models and compare candidates before choosing one for an app such as a copilot, chatbot, or summarization solution.

This matches the stem because the team wants to:

  • review available models
  • evaluate candidate models
  • select a model before development

Azure OpenAI Service gives access to generative models, but the question asks for the capability centered on comparing and selecting models. Azure AI Language targets prebuilt NLP workloads, and Azure Machine Learning is a broader ML platform rather than the most direct catalog capability for this task.

  • The option naming Azure OpenAI Service is tempting because it provides generative models, but the stem focuses on model discovery and comparison.
  • The option naming Azure AI Language fits text analytics workloads, not foundation model selection for a copilot.
  • The option naming Azure Machine Learning is broader and can support ML projects, but it is not the most direct Azure catalog for evaluating generative AI models.

Question 48

Topic: Describe Features of Computer Vision Workloads on Azure

An online retailer stores one photo for each returned product. The retailer wants AI to label each photo as damaged or not damaged so employees can prioritize inspections. They do not need to read text from packaging or generate repair notes. Which computer vision task is the best fit?

Options:

  • A. Object detection

  • B. Optical character recognition (OCR)

  • C. Image classification

  • D. Generative text summarization

Best answer: C

Explanation: This scenario asks for one label for each whole product photo: damaged or not damaged. That makes image classification the best fit because the goal is to categorize the image, not read text or generate new content.

Image classification is the right choice when you want to assign a category to an entire image. In this scenario, the retailer wants each return photo sorted into a simple business label so staff can review damaged items first. The requirement is about recognizing the overall condition shown in the photo, not extracting text, locating specific regions, or creating written output.

  • Use image classification for whole-image labels.
  • Use object detection when you must find and locate items within an image.
  • Use OCR when the important information is printed or handwritten text.
  • Use generative AI when the system must create new text such as summaries or notes.

The deciding clue is that the system needs a category for each photo, not coordinates, text extraction, or generated content.

  • Object detection is for locating objects or regions in an image, which is unnecessary when a single label for the whole photo is enough.
  • OCR would help read packaging text, but the scenario explicitly says text extraction is not needed.
  • Generative summarization creates new text output, which does not match a visual labeling task.

Question 49

Topic: Describe Artificial Intelligence Workloads and Considerations

A company receives scanned invoices from many vendors. It wants an Azure-based AI solution that can read each invoice and extract the vendor name, invoice date, total amount, and line items into structured records. Which AI workload is the best fit for this goal?

Options:

  • A. Document processing

  • B. Object detection

  • C. Image classification

  • D. Optical character recognition (OCR)

Best answer: A

Explanation: Document processing is the best fit because the goal is to extract structured invoice data, including fields and line items. OCR alone reads text, but document processing also uses document layout and structure to produce usable records.

Document processing is the right workload when the input is a business document and the goal is to extract structured content from it. An invoice is more than a picture of text: it contains labeled fields, totals, and tabular line items. A document processing solution is designed to combine text reading with layout understanding so the result can be stored as structured data.

OCR is only part of that job. OCR can recognize printed or handwritten text, but by itself it does not fully handle document-specific structure such as key-value pairs and tables. Image classification labels an entire image, and object detection finds objects or regions in an image, so neither is the best match for extracting invoice details.

  • Image labeling mismatch image classification can categorize a whole image, but it does not return invoice fields or table data.
  • Region finding only object detection can locate items in an image, but it is not aimed at extracting structured business document content.
  • Text only the OCR option is tempting because invoices contain text, but the scenario requires structured extraction beyond plain text recognition.

Question 50

Topic: Describe Features of Generative AI Workloads on Azure

A bank plans to use Azure OpenAI Service for several tasks. In which scenario is human-in-the-loop review the most important safeguard before the generated content is sent to a customer?

Options:

  • A. Suggesting new website marketing headlines

  • B. Preparing loan denial explanations for applicants

  • C. Drafting internal meeting summaries for employees

  • D. Generating product descriptions for online catalogs

Best answer: B

Explanation: Human-in-the-loop review matters most when generative AI output supports a high-impact decision about a person. A loan denial explanation can affect fairness, compliance, and customer trust, so a human should verify the content before it is used.

In generative AI, human-in-the-loop review is especially important when generated content could influence consequential outcomes for people, such as finance, healthcare, employment, or legal decisions. A draft explanation for a denied loan is not just general content creation; it communicates a decision that may require accuracy, consistent policy use, and careful wording. A human reviewer can check that the explanation is correct, appropriate, and aligned to policy before it reaches the applicant. Lower-risk tasks like internal summaries, marketing headlines, or catalog descriptions may still benefit from review, but they usually do not carry the same direct impact on a person’s access to services.

  • Internal meeting summaries are useful drafting tasks, but they usually do not determine a person’s eligibility or rights.
  • Marketing headlines may need brand review, but they are not typically high-impact decisions about an individual.
  • Product descriptions should be checked for accuracy, yet they are generally lower risk than explaining a denied financial application.

Continue with full practice

Use the AI-900 Practice Test page for the full IT Mastery route, mixed-topic practice, timed mock exams, explanations, and web/mobile app access.

Try AI-900 on Web View AI-900 Practice Test

Focused topic pages

Free review resource

Read the AI-900 Cheat Sheet on Tech Exam Lexicon for concept review before another timed run.

Revised on Thursday, May 14, 2026