Every answer is cross-verified by 3 leading AI models to ensure maximum accuracy. Get detailed per-option explanations and in-depth question analysis.
GPT Pro
Claude Opus
Gemini Pro
Per-option explanations
In-depth question analysis
3-model consensus accuracy
Practice Questions
1
Question 1
HOTSPOT -
To complete the sentence, select the appropriate option in the answer area.
Hot Area:
Part 1:
A banking system that predicts whether a loan will be repaid is an example of the ______ type of machine learning.
Correct answer: A (classification).
Predicting whether a loan will be repaid is a supervised learning problem where the target is a discrete label, such as “repaid” vs “not repaid” (default). That is the definition of classification: learning from historical labeled examples (past loans with known outcomes) to predict a category for a new case.
Why not regression (B): Regression is used when the output is a continuous numeric value, such as predicting the exact loan loss amount, interest rate, or customer lifetime value. While a credit model might output a probability (a number), the core task described—“whether a loan will be repaid”—is a categorical decision.
Why not clustering (C): Clustering is unsupervised and groups data points by similarity without known labels (e.g., segmenting customers into risk groups when you don’t have repayment outcomes). Here, the bank explicitly wants a known outcome prediction, which requires labeled training data and classification.
2
Question 2
(Select 2)
What are two tasks that can be performed by using the Computer Vision service? Each correct answer presents a complete solution.
NOTE: Each correct selection is worth one point.
Incorrect. Training a custom image classification model is typically done with Azure AI Custom Vision, which is designed for building and training custom classifiers and object detectors using labeled images. The Computer Vision service focuses on prebuilt analysis (tags, captions, OCR, etc.) rather than training your own classification model. In AI-900, “custom training” is a key clue pointing to Custom Vision.
Correct. Computer Vision can detect faces in an image by locating them and returning bounding boxes (and sometimes related attributes depending on the API/version and policy constraints). This is face detection, not identity recognition. For AI-900, treat “detect faces” as a Computer Vision capability, while person identification/verification is associated with the Face service and has additional governance requirements.
Correct. Recognizing handwritten text is supported through Computer Vision OCR (often referred to as the Read capability). It can extract both printed and handwritten text from images and documents, returning the text plus positional information. This is a common workload for digitizing notes, forms, and scanned paperwork, and it is a standard example of Computer Vision functionality in AI-900.
Incorrect. Translating text between languages is performed by Azure AI Translator (an NLP service). While Computer Vision can extract the text from an image using OCR, it does not translate it. A typical solution is a two-step pipeline: Computer Vision (OCR) to read the text, then Translator to translate the extracted text into the target language.
Question Analysis
Core concept:
This question tests what you can do with Azure AI Vision (commonly referred to in AI-900 as the Computer Vision service). Computer Vision provides prebuilt models for analyzing images and extracting information such as objects, tags, captions, OCR text, and face-related detection (depending on feature availability and responsible AI constraints).
Why the answer is correct:
B (Detect faces in an image) is a supported Computer Vision capability. The service can detect the presence and location of faces (for example, returning bounding boxes). In exam context, “detect faces” is distinct from “identify/verify a person,” which is handled by Azure Face (a separate capability/service area with stricter access requirements).
C (Recognize handwritten text) is also a core Computer Vision capability via Optical Character Recognition (OCR). Azure AI Vision Read/OCR can extract printed and handwritten text from images and documents, returning the recognized text and its layout/coordinates.
Key features and best practices:
- OCR/Read supports both printed and handwritten text and is commonly used for digitizing forms, receipts, notes, and scanned documents.
- Face detection returns geometric information (bounding boxes/landmarks) but does not inherently perform identity recognition.
- From an Azure Well-Architected perspective, use managed AI services to reduce operational burden (Operational Excellence) and ensure you follow Responsible AI and privacy requirements (Security).
Common misconceptions:
A seems plausible because “image classification” is a vision task, but training a custom image classification model is done with Azure AI Custom Vision (a separate service) rather than the prebuilt Computer Vision service.
D seems plausible because OCR extracts text, but translating text between languages is an NLP task performed by Azure AI Translator. A common pattern is: Computer Vision OCR to extract text, then Translator to translate it.
Exam tips:
- Remember the split: prebuilt image analysis/OCR/face detection = Computer Vision; custom training for classification/detection = Custom Vision.
- Translation is not a vision feature; it’s handled by Translator.
- Watch for wording: “detect faces” (yes) vs “recognize/identify people” (not Computer Vision in AI-900 framing).
Want to practice all questions on the go?
Download Cloud Pass for free — includes practice tests, progress tracking & more.
Download Cloud Pass and access all Microsoft AI-900 practice questions for free.
Want to practice all questions on the go?
Get the free app
Download Cloud Pass for free — includes practice tests, progress tracking & more.
3
Question 3
Your company wants to build a recycling machine for bottles. The recycling machine must automatically identify bottles of the correct shape and reject all other items.
Which type of AI workload should the company use?
Anomaly detection focuses on identifying rare or unusual patterns compared to a baseline, commonly in numeric data such as sensor telemetry, logs, or transactions. It can be used for defect detection in manufacturing when you mostly know what “normal” looks like. However, the question’s core requirement is visual identification of bottle shape from images, which is better addressed by computer vision classification/detection.
Conversational AI is designed for interactive experiences using natural language, such as chatbots and voice bots. It involves intent recognition, dialog management, and generating responses (for example, with Azure Bot Service). A recycling machine identifying bottle shapes is not a dialog or conversation problem, so conversational AI does not fit the workload described.
Computer vision is the correct workload because the machine must visually recognize whether an item matches the correct bottle shape and reject others. This maps to image classification and/or object detection. In Azure, you could use Azure AI Vision for prebuilt capabilities or Custom Vision to train a model on your specific bottle shapes under different lighting and orientations, enabling accurate automated acceptance/rejection.
Natural language processing (NLP) deals with understanding and generating human language in text or speech-to-text scenarios—tasks like sentiment analysis, entity extraction, translation, and summarization. Since the recycling machine’s input is the physical appearance (shape) of items rather than text or spoken language, NLP is not the appropriate AI workload.
Question Analysis
Core Concept:
This scenario is a computer vision workload: using images/video to detect, classify, or verify physical objects based on visual characteristics (shape, size, contours). In Azure AI terms, this aligns with services like Azure AI Vision (image analysis) and Custom Vision (train a model to classify items or detect objects).
Why the Answer is Correct:
A recycling machine that must “identify bottles of the correct shape and reject all other items” needs to visually inspect items. The key requirement is recognizing an object’s shape and deciding whether it matches an accepted class (correct bottle) or not. That is classic image classification (Is this an accepted bottle type?) and/or object detection (Where is the bottle and what type is it?)—both are computer vision tasks.
Key Features / How it’s commonly implemented:
- Image classification: Train a model with labeled images of acceptable bottles vs. non-bottles/other items. The model outputs probabilities for each class.
- Object detection: If items appear in different positions/orientations, detection can locate the object and classify it.
- Custom Vision: Often used when you need domain-specific recognition (your bottle shapes) rather than generic labels.
- Operational considerations: Use adequate training data across lighting, angles, occlusions, and bottle variations; validate with a test set; monitor drift (new bottle designs). For edge/real-time needs, deploy to edge devices (e.g., Azure IoT Edge) to reduce latency.
Common Misconceptions:
Anomaly detection can sound plausible because “reject all other items” resembles “detect outliers.” However, anomaly detection is typically used for unusual patterns in numeric/time-series data (sensor readings, transactions) or for “defect detection” when you mostly have examples of normal and few of abnormal. Here, the primary signal is visual shape recognition, which is computer vision.
Exam Tips:
When the input is images/video (cameras) and the goal is to recognize objects, shapes, text in images, or scenes, choose computer vision. Reserve NLP for text/language, conversational AI for chatbots/voice assistants, and anomaly detection for outlier detection in metrics or sensor/telemetry streams.
(Well-Architected tie-in: For reliability and performance, consider edge inference for low latency and resilience; for security, protect camera feeds and model endpoints; for cost optimization, right-size compute and retrain only as needed.)
4
Question 4
HOTSPOT -
For each of the following statements, select Yes if the statement is true. Otherwise, select No.
NOTE: Each correct selection is worth one point.
Hot Area:
Part 1:
Labelling is the process of tagging training data with known values.
Yes. Labeling is the process of associating each training example with the correct known value (the “ground truth”). In supervised learning, these labels are what the model learns to predict—for example, tagging images as “cat” or “dog,” marking emails as “spam” or “not spam,” or assigning a numeric target value for regression such as “house price.” Labeling can be done manually by humans, programmatically (when rules or existing systems provide ground truth), or via assisted labeling tools. Without labels, you typically move into unsupervised learning (clustering, dimensionality reduction) or self-supervised approaches, which are different problem types. Therefore, the statement correctly describes labeling as tagging training data with known values.
Part 2:
You should evaluate a model by using the same data used to train the model.
No. You should not evaluate a model using the same data used to train it because it does not measure generalization. Training performance often looks better than real-world performance, especially if the model overfits (learns noise or memorizes examples). Proper evaluation uses a separate test dataset (or validation dataset) that the model has not seen during training. Common approaches include train/test splits, train/validation/test splits, and cross-validation. This helps estimate how the model will perform on new data and supports better decisions about model selection and hyperparameter tuning. Using training data for evaluation can lead to deploying a model that appears accurate in development but fails in production.
Part 3:
Accuracy is always the primary metric used to measure a model’s performance.
No. Accuracy is not always the primary metric; the best metric depends on the task and the consequences of different error types. In imbalanced classification (e.g., fraud detection), a model can achieve high accuracy by always predicting the majority class while missing the minority class entirely. In such cases, recall (sensitivity), precision, F1-score, and PR-AUC are often more informative. If false positives are costly (flagging legitimate transactions as fraud), precision may matter more; if false negatives are costly (missing fraud), recall may be prioritized. For regression problems, accuracy is not the standard metric at all—metrics like MAE or RMSE are used. Therefore, accuracy is not universally the primary measure of model performance.
5
Question 5
Which metric can you use to evaluate a classification model?
True positive rate (TPR) is a classification metric derived from the confusion matrix. It measures the proportion of actual positives correctly predicted as positive: TP/(TP+FN). TPR is also known as recall or sensitivity and is especially useful when the cost of missing positive cases (false negatives) is high, such as medical screening or fraud detection.
Mean absolute error (MAE) is primarily a regression metric. It measures the average absolute difference between predicted numeric values and actual numeric values. Because classification outputs are discrete labels (or class probabilities that are typically thresholded into labels), MAE is not the standard way to evaluate classification performance in AI-900 contexts.
Coefficient of determination (R2) is a regression metric that indicates how well a model explains the variance of a continuous target variable. Values closer to 1 generally indicate a better fit for regression problems. It is not used to evaluate discrete class predictions and does not come from a confusion matrix, so it is not appropriate for classification evaluation.
Root mean squared error (RMSE) is a regression metric that measures the square root of the average squared differences between predicted and actual numeric values. RMSE penalizes larger errors more heavily than MAE, making it useful for continuous prediction tasks. It is not a standard metric for classification models, which are evaluated with confusion-matrix metrics like precision and recall.
Question Analysis
Core Concept:
This question tests your ability to choose appropriate evaluation metrics for different machine learning problem types. In AI-900, you must distinguish classification (predicting discrete classes/labels) from regression (predicting continuous numeric values). Classification models are commonly evaluated using confusion-matrix-derived metrics such as accuracy, precision, recall, F1-score, and true/false positive/negative rates.
Why the Answer is Correct:
True positive rate (TPR), also called recall or sensitivity, measures how well a classification model identifies actual positive cases: TPR = TP / (TP + FN). Because it is derived from the confusion matrix and directly reflects classification performance (especially for the positive class), it is a valid metric for evaluating a classification model. TPR is particularly important when missing positives is costly (for example, fraud detection, disease screening, safety alerts).
Key Features / Best Practices:
In real-world classification, you often tune the decision threshold (e.g., probability cutoff) to trade off TPR (recall) against false positive rate (FPR). This is visualized with ROC curves and summarized with AUC. For imbalanced datasets, accuracy can be misleading, so metrics like TPR/recall, precision, F1-score, and PR-AUC are often more informative. In Azure ML or Azure AI services, classification evaluation commonly reports these confusion-matrix metrics.
Common Misconceptions:
MAE and RMSE are popular metrics, so learners sometimes apply them universally. However, they measure average numeric error magnitude and are designed for regression. R2 (coefficient of determination) also applies to regression, describing how much variance in a continuous target is explained by the model. These do not directly measure correctness of discrete class predictions.
Exam Tips:
First identify the workload: classification vs regression. If the options include MAE/RMSE/R2, those are regression. If you see precision/recall/TPR/F1/ROC-AUC, those are classification. Also remember that “true positive rate” is essentially “recall” for the positive class, a key metric when false negatives matter.
6
Question 6
You build a machine learning model by using the automated machine learning user interface (UI).
You need to ensure that the model meets the Microsoft transparency principle for responsible AI.
What should you do?
Setting Validation type to Auto controls how AutoML splits data for training/validation (or uses cross-validation) to estimate generalization performance. This improves evaluation quality and can reduce overfitting risk, but it does not provide interpretability or explanations of model decisions. Therefore, it does not directly address the Responsible AI transparency principle.
Enabling “Explain best model” generates interpretability outputs for the best model found by AutoML, such as feature importance and explanation artifacts. This helps users understand which inputs drive predictions and supports communicating model behavior to stakeholders. These explanations are a direct mechanism to meet the Responsible AI transparency principle in Azure ML AutoML.
Setting Primary metric to accuracy only changes the optimization target used to select the best model during AutoML (e.g., accuracy vs. AUC vs. F1). While choosing an appropriate metric is important for model quality, it does not make the model more understandable or provide insight into how predictions are made. Transparency requires explanations, not just performance metrics.
Max concurrent iterations controls how many AutoML training iterations run in parallel, impacting training speed and potentially cost/resource usage. A value of 0 is not a transparency feature; it relates to compute scheduling and throughput. Concurrency settings do not produce interpretability artifacts or improve stakeholders’ understanding of model behavior.
Question Analysis
Core concept:
This question tests the Responsible AI principle of transparency and how Azure Machine Learning Automated ML (AutoML) supports it. Transparency means stakeholders can understand how and why a model makes predictions, including which features most influence outcomes and what patterns the model learned.
Why the answer is correct:
In the Automated ML UI, enabling “Explain best model” generates model explanations (interpretability artifacts) for the selected/best-performing model. This produces feature importance and related interpretability outputs that help users understand the model’s behavior. Providing explanations is a direct implementation of transparency: it makes the model less of a “black box,” supports documentation, and helps communicate decision logic to business users, auditors, and affected individuals.
Key features and best practices:
When “Explain best model” is enabled, Azure ML can compute global feature importance (overall drivers of predictions) and, depending on configuration/model type, local explanations for individual predictions. These artifacts can be reviewed in the studio UI and used in reports. From an Azure Well-Architected Framework perspective, transparency supports Operational Excellence (better monitoring and understanding of system behavior) and Reliability/Security (easier to detect unexpected or risky model behavior).
Common misconceptions:
It’s easy to confuse “better evaluation” with “transparency.” Validation type and primary metric affect how you measure performance, not how you explain decisions. Similarly, concurrency settings affect speed/cost, not interpretability. A highly accurate model can still be non-transparent.
Exam tips:
For AI-900, map Responsible AI principles to concrete Azure features:
- Transparency/Interpretability: model explanations (e.g., “Explain best model” in AutoML).
- Fairness: fairness assessment/metrics.
- Privacy & security: data protection controls.
- Reliability & safety: robust evaluation/monitoring.
When you see “transparency,” look for “explain,” “interpretability,” or “feature importance.”
7
Question 7
HOTSPOT -
To complete the sentence, select the appropriate option in the answer area.
Hot Area:
Part 1:
Predicting how many hours of overtime a delivery person will work based on the number of order received is an example of ______
Correct answer: C (regression).
Predicting “how many hours of overtime” produces a numeric quantity (for example, 0, 2.5, 6). When the target/output variable is a continuous number, the ML task is regression. The input feature here is the number of orders received, and the model learns a relationship between orders and overtime hours using labeled historical data (supervised learning).
Why not A (classification)? Classification predicts discrete labels or categories, such as “will work overtime: yes/no” or “overtime band: low/medium/high.” If the question had asked whether the person will work overtime, that would be classification. But it asks for the number of hours.
Why not B (clustering)? Clustering is unsupervised and groups similar records without a known target label (for example, grouping delivery people into behavioral segments). This scenario explicitly predicts a known numeric target, so clustering is not appropriate.
8
Question 8
HOTSPOT -
For each of the following statements, select Yes if the statement is true. Otherwise, select No.
NOTE: Each correct selection is worth one point.
Hot Area:
Part 1:
Azure Machine Learning designer provides a drag-and-drop visual canvas to build, test, and deploy machine learning models.
Yes. Azure Machine Learning designer is specifically a drag-and-drop visual canvas for building machine learning workflows. You can add and connect modules to perform tasks such as importing data, cleaning/transforming it, training models (e.g., classification/regression modules), scoring, and evaluating results. After validating the workflow, you can operationalize it by publishing it as a pipeline and deploying it (commonly as a real-time endpoint or batch pipeline, depending on the workflow).
This aligns with AI-900’s emphasis on identifying no-code/low-code ML tooling in Azure. The “No” option would be incorrect because the designer’s primary value proposition is exactly this visual approach to build, test, and deploy ML models without requiring you to write training code from scratch (though code-based approaches also exist in Azure ML).
Part 2:
Azure Machine Learning designer enables you to save your progress as a pipeline draft.
Yes. Azure ML designer supports iterative development, including saving your work-in-progress. In the designer experience, you can save changes as you build out the pipeline graph, and you can work on a pipeline in a draft state before publishing/submitting it as a runnable pipeline. This is important for experimentation and collaboration: you can incrementally adjust modules, parameters, and connections, then run the pipeline to compare results.
The “No” option would be incorrect because saving progress is a standard capability of the designer authoring experience. From an exam perspective, remember that designer is meant to support rapid prototyping and repeatable workflows (pipelines), which naturally includes the ability to save drafts/edits prior to publishing and executing at scale.
Part 3:
Azure Machine Learning designer enables you to include custom JavaScript functions.
No. Azure Machine Learning designer does not provide a mechanism to include custom JavaScript functions as part of the ML workflow. Designer is focused on ML pipeline composition using built-in modules and ML-specific extensibility. When you need custom logic, the typical supported approach is to use Python-based customization (for example, modules like Execute Python Script) or create custom components that run code in the Azure ML execution environment.
Choosing “Yes” would confuse Azure ML designer with web development or client-side extensibility patterns. JavaScript is not the standard way to extend training/scoring logic inside Azure ML pipelines. For AI-900, the key takeaway is: designer is low-code with ML modules, and custom code integration is generally done via Python (or packaged components), not JavaScript functions.
9
Question 9
You need to determine the location of cars in an image so that you can estimate the distance between the cars.
Which type of computer vision should you use?
OCR extracts text from images (printed or handwritten), returning the recognized characters and often their positions in the text layout. It’s used for scenarios like reading signs, invoices, forms, or license plate text. It does not detect or localize non-text objects like cars, so it won’t help determine car positions for distance estimation.
Object detection is the correct choice because it identifies each car in the image and returns its location, usually as a bounding box. That location data is necessary if you want to estimate the distance between cars, even if the initial estimate is only in image or pixel space. Unlike classification, object detection can handle multiple cars in the same image and distinguish where each one appears. In AI-900 terms, this is the standard workload for determining both the type of object and its position.
Image classification assigns a label (or set of tags) to an entire image, such as “contains cars,” but it does not provide the coordinates of each car. Since estimating distance between cars requires knowing where each car is located in the image, classification alone is insufficient even if it correctly recognizes the presence of cars.
Face detection is a specialized computer vision task focused on locating human faces and sometimes analyzing facial attributes. While it does return locations (bounding boxes), it is trained and optimized for faces, not vehicles. For locating cars and measuring distance between them, you should use general object detection trained to detect cars.
Question Analysis
Core concept:
This question is about choosing the correct computer vision workload for identifying both what an object is and where it appears in an image. In AI-900, the correct workload for locating cars is object detection because it returns positions for each detected object, typically as bounding boxes.
Why correct:
To estimate the distance between cars, you first need to know the location of each car in the image. Object detection provides coordinates for each detected car, which makes it possible to compare their relative positions. Although estimating true physical distance may require additional calibration or depth information, the vision task that supplies object locations is still object detection.
Key features:
Object detection can find multiple instances of the same object class in a single image and provide a bounding box for each one. This makes it suitable for scenarios such as counting cars, locating pedestrians, or tracking items in a scene. For exam purposes, remember object detection as the workload that answers both “what is it?” and “where is it?”
Common misconceptions:
Image classification only tells you what the image contains, not where objects are located. OCR is only for reading text, and face detection is limited to detecting human faces rather than general objects like vehicles. A common trap is choosing classification because cars are present, but classification does not provide coordinates.
Exam tips:
If the question mentions location, coordinates, bounding boxes, counting separate objects, or measuring distance between items in an image, object detection is usually the right answer. If it asks only for the overall subject of the image, think classification. If it asks to read words, think OCR.
10
Question 10
You have a frequently asked questions (FAQ) PDF file.
You need to create a conversational support system based on the FAQ.
Which service should you use?
QnA Maker is built to create a question-and-answer knowledge base from FAQ-style sources (including PDFs) and publish it as an endpoint for conversational experiences. It supports ranking answers, confidence scores, multi-turn follow-ups, and active learning. It commonly integrates with Azure Bot Service to deliver a chatbot that responds using the FAQ content with minimal custom model training.
Text Analytics (part of Azure AI Language) extracts insights from text such as sentiment, key phrases, named entities, and language detection. It does not natively build a knowledge base that maps user questions to best-matching FAQ answers. It can complement a support system (e.g., analyze customer messages), but it is not the primary service for FAQ-based conversational Q&A.
Computer Vision analyzes images and video (OCR, image classification, object detection, etc.). A conversational support system based on an FAQ PDF is primarily a natural language question-answering problem, not an image understanding workload. While OCR could extract text from scanned documents, it still wouldn’t provide the Q&A matching and knowledge base capabilities required for a chatbot.
Language Understanding (LUIS) is used to identify user intent and extract entities to drive actions in an application (e.g., “book a flight,” “reset password”). It is not optimized for retrieving answers from an FAQ knowledge base. LUIS might be used alongside QnA Maker to route between “answer a question” vs. “perform an action,” but alone it doesn’t satisfy FAQ-based conversational support.
Question Analysis
Core concept:
This question tests selecting the right Azure Natural Language Processing service for building a conversational support experience from an existing FAQ document. The key requirement is “conversational support system based on the FAQ PDF,” which maps to knowledge base + question answering.
Why the answer is correct:
QnA Maker (and its successor capability, Question Answering in Azure AI Language) is designed to create a knowledge base from semi-structured content such as FAQ pages, documents, and PDFs, then expose it through a chat-style interface. You can ingest the FAQ PDF, extract question/answer pairs (or learn them from headings and sections), and publish an endpoint that returns the best answer with a confidence score. This directly satisfies the requirement without needing to build and train a custom NLP model.
Key features, configurations, and best practices:
QnA Maker supports importing from URLs, files (including PDF), and manual Q&A entry. It provides multi-turn prompts (follow-up questions), active learning to improve answer quality based on user feedback, and integration patterns with Azure Bot Service for a full conversational experience. For production readiness (Azure Well-Architected Framework), you’d consider reliability (deploy in a supported region, monitor with Application Insights), security (use managed identity/keys, restrict access), and operational excellence (iterate the knowledge base, use active learning, versioning, and testing). Also note the exam-era naming: AI-900 content often references “QnA Maker,” while newer implementations use “Azure AI Language – Question Answering.” The concept remains the same: FAQ-to-chat knowledge base.
Common misconceptions:
Text Analytics sounds relevant because it’s NLP, but it focuses on extracting insights (sentiment, key phrases, entities) rather than answering user questions. LUIS is for intent classification and entity extraction to route actions (e.g., “reset password”), not for retrieving answers from an FAQ knowledge base. Computer Vision is unrelated because the input is text content, not images.
Exam tips:
When you see “FAQ,” “knowledge base,” “chatbot answers from documents,” or “question answering,” choose QnA Maker/Question Answering. When you see “detect intent” or “extract entities to trigger actions,” choose LUIS. When you see “sentiment/key phrases/entities,” choose Text Analytics. When you see “analyze images,” choose Computer Vision.
