Practice for the AI-900: Microsoft Azure AI Fundamentals exam with 15 exam-style practice questions, instant answer reveals, and concise explanations of every correct answer. Topics include: A global enterprise is developing an accessibility application that must convert users' spoken commands into text in rea. Follow @CertPunch and visit certpunch.com for more certification practice exams and study content.
Prefer hands-on? Take this round as an interactive practice test — answer every question, get instant feedback, and see your score: Start the AI-900: Microsoft Azure AI Fundamentals practice test →
What you will practice
- A global enterprise is developing an accessibility application that must convert users' spoken commands into…
- Which Azure AI Vision service should a retail company use to count specific products on store shelves from ca…
- Which scenario requires reinforcement learning?
- A model initially predicts house prices (regression) but is retrained to classify houses into "High," "Medium…
- A developer is building a system to match employee badge photos with live camera feeds for secure facility ac…
- A retail company has gathered thousands of customer reviews—both structured and unstructured—and wants to gai…
Answers and explanations
Tap a question to expand the answer and the exam reasoning. Try to commit to your own pick first.
Q1. A global enterprise is developing an accessibility application that must convert users' spoken commands into text in real time and then transform text-based responses into natural, lifelike audio. Which Azure AI Speech services should be c…
Answer: D. Speech to Text and Text to Speech
Speech to Text handles transcribing the spoken commands into written text, while Text to Speech generates the lifelike audio responses needed for the application. Speech translation would incorrectly convert the input into a different target language rather than creating audio output.
Q2. Which Azure AI Vision service should a retail company use to count specific products on store shelves from camera feeds?
Answer: A. Object Detection
Object detection is the correct choice because it locates and identifies multiple distinct items within an image using bounding boxes, allowing the system to tally specific retail items. Optical character recognition strictly extracts text and cannot identify physical objects.
Q3. Which scenario requires reinforcement learning?
Answer: B. Training a robot to navigate uneven terrain
Training a robot to navigate terrain requires reinforcement learning because the agent learns by receiving rewards or penalties through trial and error in a dynamic environment. Classifying emails or grouping articles uses supervised and unsupervised learning instead.
Q4. A model initially predicts house prices (regression) but is retrained to classify houses into "High," "Medium," or "Low" price categories. Which evaluation metric becomes irrelevant after this change?
Answer: A. Mean Absolute Error (MAE)
Mean Absolute Error measures continuous value differences, making it irrelevant once the task changes to predicting discrete categories. For classification tasks, rely on metrics derived from a confusion matrix, such as accuracy or F1 score.
Q5. A developer is building a system to match employee badge photos with live camera feeds for secure facility access. Which Azure Face service should they use?
Answer: C. Face Verification
Face Verification directly compares two images to confirm if they belong to the same person, which is ideal for access control scenarios. Do not confuse this with Identification, which searches a database to find a specific person, or Liveness, which only checks for a physical presence.
Q6. A retail company has gathered thousands of customer reviews—both structured and unstructured—and wants to gain insights into customer sentiment while also extracting domain-specific entities (like unique product codes and service types) th…
Answer: D. Sentiment Analysis and Custom Named Entity Recognition
Sentiment analysis evaluates text tone, while custom named entity recognition extracts specialized domain terms. Rely on prebuilt entities only for standard categories like dates or locations, and switch to custom NER when the business needs to identify unique proprietary items.
Q7. A logistics company scans damaged shipping labels with skewed text and faint ink. They want to automate data entry from these labels. Which Azure AI capability best addresses this challenge?
Answer: C. Optical Character Recognition (OCR)
Optical character recognition extracts printed and handwritten text from images, handling skewed orientation and varying clarity. Object detection identifies items within an image but fails to transcribe text content, making it the wrong tool for data entry.
Q8. A company is developing a chatbot to categorize customer feedback into predefined sentiment categories (e.g., "positive," "neutral," "negative"). The model struggles to group phrases like "extremely satisfied" and "delighted" into the same…
Answer: D. Use vector embeddings to represent the text.
Vector embeddings map text into multidimensional spaces, placing phrases with similar meanings mathematically close together. Lowercasing standardizes characters but misses deeper semantic relationships, which makes it an inferior choice for improving comprehension.
Q9. A developer is preprocessing images for a fruit classification model. They convert all images to grayscale to reduce computational complexity. What is a potential drawback of this approach?
Answer: C. The model may lose the ability to distinguish fruits with similar shapes but different colors (e.g., lemons vs. bananas).
Converting images to grayscale discards critical color features needed to visually distinguish items like yellow lemons and bananas. Grayscale actually decreases file size, and convolutional neural networks remain fully compatible with standard RGB channels.
Q10. A model achieves 95% training accuracy but 60% validation accuracy. What is the MOST likely issue?
Answer: A. Overfitting
Overfitting happens when a model memorizes training noise, preventing generalization to new data and causing large accuracy gaps. High bias indicates underfitting, where the model fails to learn the training data effectively during initial training.
Q11. A retail company's AI system for inventory prediction shows these characteristics: misses major seasonal trends, performs similarly on both training and test data, generates overly simplified predictions, and ignores important feature rela…
Answer: C. High bias requiring more complex modeling and additional features
The model is underfitting the data, which indicates high bias because it fails to capture the underlying trends. You should increase model complexity or add relevant features. Overfitting would mean performing well on training data but poorly on test data.
Q12. A developer needs to build an application that processes customer reviews by detecting sentiment, extracting key phrases, and translating text. The goal is to use Azure's prebuilt AI services to minimize development complexity. Which Azure…
Answer: D. A multi‐service language resource that provides integrated NLP capabilities.
A multi-service language resource is designed to handle multiple natural language processing tasks through a single endpoint, reducing development overhead. Using separate single-service resources would increase management complexity and fail to minimize integration effort.
Q13. An e-commerce platform uses Azure AI Vision to analyze product images. They need precise outlines of each product's shape for augmented reality previews. Which capability should they use?
Answer: A. Image Segmentation
Image segmentation identifies exact pixel boundaries for objects, making it crucial for accurate augmented reality overlays. Object detection only provides bounding boxes, which are rectangular areas that do not follow the precise contours of the product shape.
Q14. Document Intelligence combines _______ technology with AI to recognize both printed and handwritten text while extracting key-value pairs from documents.
Answer: C. Optical Character Recognition
Optical character recognition serves as the foundational technology that converts pictures of printed or handwritten text into machine-readable data. Document Intelligence layers machine learning on top of this to understand document layouts and extract structured key-value pairs.
Q15. A bank wants an AI loan approval system that auditors can easily understand. Which trade-off is MOST relevant?
Answer: D. Interpretability vs. Performance
Prioritizing interpretability ensures that auditors can clearly trace and understand how the model makes its decisions. While simpler models offer this transparency, they often sacrifice some predictive performance compared to highly complex, opaque deep learning models.
More AI-900: Microsoft Azure AI Fundamentals drills and other practice exams are on @CertPunch. New rounds drop every few days at certpunch.com.