Azure AI-900 Practice Test

Exam AI-900: Microsoft Azure AI Fundamentals

Azure AI-900 overview

  • 200 practice questions
  • 32 flashcards
  • 5 study guide topics
  • 40 key terms

What you'll study

  • Describe Artificial Intelligence workloads and considerations18%
  • Describe fundamental principles of machine learning on Azure18%
  • Describe features of computer vision workloads on Azure18%
  • Describe features of Natural Language Processing (NLP) workloads on Azure18%
  • Describe features of generative AI workloads on Azure23%

Exam AI-900: Microsoft Azure AI Fundamentals is Microsoft's entry-level AI credential exam, aimed at candidates at the beginning of a career in AI solution development. Microsoft describes the audience as people with conceptual knowledge of AI solutions in Azure plus foundational technical skills, Python coding familiarity, and comfort with Azure resources; data science and software engineering experience are not required, though basic cloud and client-server awareness helps. The exam measures five skill areas: AI workloads and considerations (15–20%), fundamental machine learning principles on Azure (15–20%), computer vision features (15–20%), natural language processing features (15–20%), and generative AI features (20–25%). Passing score is 700. Preparation is normally self-paced learning paths or instructor-led training, plus the free practice assessment available through AI Skills Navigator. Important: the certification page now lists AI-901 as the required exam, and a warning on the AI-900 page states that AI-900 was retired on June 30, 2026 and replaced by AI-901 - while the same page separately displays "Retirement date: none." Confirm which exam you are booking before you pay. This site is independent and not affiliated with Microsoft.

Common mistakes

Reads a scenario such as 'identify whether a photo of a manufactured part shows a defect' and selects the anomaly detection workload because the sentence contains the word 'detect', rather than selecting a computer vision workload.

Exam stems are written with action verbs (detect, extract, predict, classify) that appear across every workload family, so candidates pattern-match on the verb instead of on the input data type and the required output.

What to do instead

Before answering, write down two things: the input format (image, text, audio, numeric/tabular) and the output format (a label, a number, a location, a grouping). Map input to the workload family first, then use the output to pick the specific technique. Do not use the verb in the stem as evidence.

Treats the responsible AI principles as interchangeable, for example stating that 'inclusiveness' means the model produces equally good results for every demographic group, or that 'transparency' means the system is secure.

Several of the six principles are near-synonyms in everyday English (fairness vs inclusiveness, transparency vs accountability), so the wording feels familiar and the candidate answers from general intuition rather than from the formal definitions.

What to do instead

Memorise one distinguishing test word per principle: fairness = parity of outcomes across groups; inclusiveness = usable by people with the widest range of abilities and needs; reliability and safety = behaves correctly under unexpected or adverse conditions; privacy and security = data is protected and controlled; transparency = people can understand how and why a decision was made; accountability = humans remain responsible and there is governance. If the stem is about who is answerable, it is accountability; if it is about who is able to use it, it is inclusiveness.

Recommends Azure Machine Learning for every scenario, including ones that only need summarising text, translating speech, or detecting faces in an image.

Azure Machine Learning is the most prominent service name in the domain and feels like the 'safe' answer for anything involving AI, while the prebuilt Azure AI services are taught more briefly.

What to do instead

Apply a two-question rule. Does an existing prebuilt model already perform this task (language, vision, speech, document extraction)? If yes, the answer is an Azure AI service / prebuilt model. Does the organisation need to train a model on its own labelled data and manage the model lifecycle? If yes, the answer is Azure Machine Learning. Only pick Azure Machine Learning when the scenario mentions custom training, datasets, compute, or model evaluation.

Labels a scenario as clustering when it is classification, or as classification when it is regression, for example calling 'predict whether a customer will leave' clustering, or 'predict the price of a house' classification.

The everyday meaning of 'group', 'categorise' and 'predict' blurs together, and the candidate does not check whether labelled answer values exist in the historical data.

What to do instead

Ask whether the training data already contains the answer (a label). No labels, and the goal is to discover natural groupings: clustering (unsupervised). Labels present: look at the label's data type - a category, yes/no, or class name is classification; a numeric quantity is regression; a numeric value over time is forecasting. Anchor each option to the label type before choosing.

Chooses image classification when the stem requires locating several separate objects and returning coordinates, or chooses object detection when the stem asks for a per-pixel outline of each object.

All three techniques take images and produce object-related output, so they are remembered as one bucket; candidates rarely rehearse the output format that distinguishes them.

What to do instead

Decide by output shape, not by the object count. One label for the whole image: image classification. Many labels for the whole image with no location: multilabel classification. Bounding boxes with coordinates for each object: object detection. A mask that assigns every pixel to a class: semantic segmentation. If the stem mentions drawing a box, it is detection; if it mentions outlining the exact shape of each object, it is segmentation.

Answers with Azure AI Vision OCR for a scenario about pulling vendor name, invoice date and line-item totals out of scanned invoices, or answers with Face for a scenario about extracting printed text from a photograph.

Both services are described as 'reading things from images', and the general-purpose Vision resource does contain a Read capability, so the more specific document service is overlooked.

What to do instead

Separate 'where is the text' from 'what does the field mean'. If the task is to read text from a photo, sign, or scene: Vision OCR / Read. If the task is to extract named fields and values from structured or semi-structured forms, invoices, or receipts: Azure AI Document Intelligence. If the task is to detect or analyse human faces: Face. Check the stem for the words forms, fields, invoices, or receipts to trigger Document Intelligence.

Mismatches the NLP feature to the requested output, for example choosing key phrase extraction when the stem asks for the people, places and organisations named in the text, or choosing sentiment analysis when the stem asks what the user wants the application to do.

All of these features take text and return short text-based tags or scores, so the candidate remembers the family but not the specific return type.

What to do instead

Memorise the exact return shape of each feature: sentiment = positive/negative/neutral with confidence scores; key phrase extraction = the main concepts as a list; named entity recognition = tokens tagged by type such as person, location, organisation, datetime, quantity; language detection = a language name and code; intent recognition in conversational language understanding = the action the user wants plus extracted entities. Match the noun in the question to the return shape.

Treats conversational language understanding, question answering and translation as the same service, recommending question answering for a scenario where a user issues commands like 'book a flight to Paris tomorrow' that must be mapped to an action, or recommending conversational language understanding for an FAQ knowledge base.

All three consume natural language and appear in conversational solutions, and the marketing descriptions of each mention 'chatbots' and 'conversation'.

What to do instead

Ask what the system does with the utterance. Map it to an action or command with parameters: conversational language understanding (intents and entities). Retrieve the best answer from a curated knowledge base of question-and-answer pairs: question answering. Convert text between languages: translation. If the stem mentions utterances mapped to intents, it is conversational language understanding; if it mentions an FAQ or knowledge base, it is question answering.

Describes retrieval-augmented generation as fine-tuning, or says fine-tuning is the way to give a model access to a company's current internal documents.

Both are described as 'making the model know more about your data', and both appear near each other in generative AI guidance.

What to do instead

Use the weight-change test. If the model's parameters are unchanged and grounding documents are retrieved and inserted into the prompt at request time, that is retrieval-augmented generation - it is the correct choice for frequently changing or private factual content. If additional training on example data changes the model's behaviour, tone, or format, that is fine-tuning. Fine-tuning is not the tool for injecting current facts.

States that a well-written prompt makes a generative model's output reliable and factual, or that prompt engineering changes or retrains the underlying model.

Prompt engineering is presented as the primary skill for working with generative AI, so it is credited with guarantees it cannot provide; and because prompt quality visibly changes output, candidates infer that the model itself has been modified.

What to do instead

Treat generative output as unverified by default. Prompt engineering only shapes the input at inference time; it does not alter model weights and does not guarantee correctness. Reduce risk with grounding the prompt in retrieved, authoritative content, adding content filters and moderation, and keeping a human review step for consequential outputs. On the exam, when a stem asks how to make output trustworthy, prefer grounding and verification over 'write a better prompt'.

Exam requirements can change. Confirm current details with the exam provider.

Azure AI-900 practice

200 questions. Each one lists the choices, the correct answer, and an explanation. With JavaScript, this list opens one question at a time. The checker at the top of this page saves progress in this browser when it is available.

  1. Question 1Describe Artificial Intelligence workloads and considerations

    A company receives thousands of scanned purchase orders and wants key fields such as vendor name and total amount extracted automatically. Which AI workload should it use?

    1. Object detection in computer vision
    2. Document intelligence Correct answer.
    3. Anomaly detection
    4. Speech synthesis

    Explanation

    Document intelligence combines optical character recognition with trained models to extract structured fields, key-value pairs, and tables from forms and scanned documents. It is purpose-built for turning semi-structured paperwork into usable data.

  2. Question 2Describe Artificial Intelligence workloads and considerations

    Which responsible AI principle focuses on ensuring an AI system operates consistently and safely under both expected and unexpected conditions?

    1. Reliability and safety Correct answer.
    2. Accountability
    3. Transparency
    4. Inclusiveness

    Explanation

    The reliability and safety principle requires that AI systems perform as intended, handle unexpected inputs gracefully, and include safeguards against harmful outcomes. Transparency, inclusiveness, and accountability address different concerns such as understanding, accessibility, and ownership.

  3. Question 3Describe Artificial Intelligence workloads and considerations

    Which AI workload converts a recording of a customer phone call into a written transcript?

    1. Sentiment analysis
    2. Text-to-speech
    3. Speech recognition Correct answer.
    4. Machine translation

    Explanation

    Speech recognition, also called speech-to-text, transcribes spoken audio into written text. The reverse direction, generating audio from text, is text-to-speech.

  4. Question 4Describe Artificial Intelligence workloads and considerations

    Which responsible AI principle requires that an AI system treat people in similar circumstances in a similar way, regardless of demographic group?

    1. Accountability
    2. Inclusiveness
    3. Transparency
    4. Fairness Correct answer.

    Explanation

    Fairness means AI systems should not produce biased outcomes or discriminate against particular groups. It is measured by comparing model behavior and results across subgroups of people.

  5. Question 5Describe Artificial Intelligence workloads and considerations

    Which machine learning task predicts a numeric value, such as the expected selling price of a house?

    1. Anomaly detection
    2. Regression Correct answer.
    3. Clustering
    4. Classification

    Explanation

    Regression models learn to output a continuous numeric value from input features, such as a price, temperature, or demand quantity. Classification instead predicts a category label.

  6. Question 6Describe Artificial Intelligence workloads and considerations

    Which machine learning approach groups customers with similar purchasing behavior when no labels are available?

    1. Anomaly detection
    2. Regression
    3. Classification
    4. Clustering Correct answer.

    Explanation

    Clustering is an unsupervised technique that discovers natural groupings in data without predefined labels. The resulting groups are inferred from similarity between data points.

  7. Question 7Describe Artificial Intelligence workloads and considerations

    Which AI workload is used to flag unusual credit card transactions that differ from a customer's normal spending pattern?

    1. Anomaly detection Correct answer.
    2. Clustering
    3. Image classification
    4. Regression

    Explanation

    Anomaly detection identifies data points that deviate significantly from expected patterns, which makes it well suited to fraud and fault detection. It can work with few or no labeled examples of the unusual event.

  8. Question 8Describe Artificial Intelligence workloads and considerations

    Which AI workload lets a user ask questions and receive answers across several turns of dialogue with a software agent?

    1. Computer vision
    2. Conversational AI Correct answer.
    3. Knowledge mining
    4. Speech synthesis

    Explanation

    Conversational AI covers bots and agents that maintain a dialogue with a user, interpreting intent over multiple turns and responding appropriately. It often combines language understanding with a channel such as chat, voice, or messaging.

  9. Question 9Describe Artificial Intelligence workloads and considerations

    An organization wants employees to search a large archive of scanned contracts and internal reports using natural language and receive answers with links to the source documents. Which AI workload fits best?

    1. Speech recognition
    2. Knowledge mining Correct answer.
    3. Regression modeling
    4. Anomaly detection

    Explanation

    Knowledge mining uses OCR, enrichment, and search indexing to make large unstructured document stores queryable in natural language. It typically returns answers together with references back to the underlying documents.

  10. Question 10Describe Artificial Intelligence workloads and considerations

    Which description best matches the transparency principle of responsible AI?

    1. Every AI system must support every spoken language used by any of its potential users.
    2. Users can understand how a system works and its limitations. Correct answer.
    3. Every AI system must encrypt all of its data at rest and in transit at all times.
    4. Every AI system must be verified as completely error-free before it is released to production.

    Explanation

    Transparency means the behavior of an AI system, and the boundaries of what it can reliably do, should be understandable to the people who use and are affected by it. Documentation, explanations, and clear communication of limitations are typical practices.

  11. Question 11Describe Artificial Intelligence workloads and considerations

    A bank has years of historical loan records labeled as 'repaid' or 'defaulted' and wants a model that predicts the label for new applicants. Which approach should be used?

    1. Reinforcement learning that rewards an agent for good decisions
    2. Supervised learning trained on labeled historical examples Correct answer.
    3. Unsupervised learning that groups similar applications without using labels
    4. A static rule set written entirely by hand

    Explanation

    When historical data already contains the correct answer for each case, supervised learning can map input features to those known labels. This is a binary classification problem because there are exactly two possible labels.

  12. Question 12Describe Artificial Intelligence workloads and considerations

    A speech recognition service performs well for most users but frequently fails for people with strong regional accents. Which responsible AI principle is most directly unmet?

    1. Inclusiveness Correct answer.
    2. Transparency
    3. Accountability
    4. Reliability and safety

    Explanation

    Inclusiveness requires AI systems to work well for people with diverse abilities, backgrounds, and characteristics, including different speech patterns. A service that only serves a subset of users fails that goal.

  13. Question 13Describe Artificial Intelligence workloads and considerations

    Which responsible AI principle places responsibility on the people who build and operate an AI system for the outcomes it produces?

    1. Transparency
    2. Privacy and security
    3. Fairness
    4. Accountability Correct answer.

    Explanation

    Accountability means humans remain responsible for AI systems and their effects, with clear ownership, oversight, and governance. It is why organizations define review processes and sign-off roles for AI solutions.

  14. Question 14Describe Artificial Intelligence workloads and considerations

    A manufacturer wants a system that finds and locates cracks in photographs of metal parts. Which AI workload should be used?

    1. Computer vision Correct answer.
    2. Text analytics
    3. Speech synthesis
    4. Natural language processing

    Explanation

    Computer vision models interpret images, and object detection within computer vision can both locate and classify visual defects. The input is photographic, so the visual workload applies rather than a language workload.

  15. Question 15Describe Artificial Intelligence workloads and considerations

    A company wants to automatically identify the main topics discussed in thousands of free-text customer reviews. Which workload applies?

    1. Anomaly detection in numeric records
    2. Natural language processing Correct answer.
    3. Computer vision image processing
    4. Text-to-speech synthesis

    Explanation

    Natural language processing extracts meaning from written text, including key phrases, topics, sentiment, and entities. Topic discovery from free-text reviews is a classic NLP scenario.

  16. Question 16Describe Artificial Intelligence workloads and considerations

    A demand forecasting model trained on 2019 data performs poorly after a sudden change in buying habits. What is the most likely cause?

    1. Overfitting to the training set
    2. Data drift in the input data Correct answer.
    3. Using too few input features
    4. Applying regularization too strongly

    Explanation

    Data drift occurs when the statistical properties of real-world inputs change after training, so patterns the model learned no longer hold. Retraining on recent data is the usual remedy.

  17. Question 17Describe Artificial Intelligence workloads and considerations

    Which scenario is the best fit for reinforcement learning?

    1. Grouping news articles by topic without any labels
    2. Transcribing a recorded meeting into written text
    3. Predicting house prices from a large set of labeled historical sales records
    4. A robot learning maze navigation through trial and reward Correct answer.

    Explanation

    Reinforcement learning trains an agent to take actions in an environment so as to maximize a reward signal, which suits sequential decision problems such as navigation. Labeled-example problems instead call for supervised learning.

  18. Question 18Describe Artificial Intelligence workloads and considerations

    A team must limit how much personal data an AI system collects and how long it retains that data. Which responsible AI principle does this address?

    1. Fairness
    2. Transparency
    3. Inclusiveness
    4. Privacy and security Correct answer.

    Explanation

    Privacy and security covers collecting only necessary personal data, protecting it, and retaining it no longer than required. This includes controls such as encryption, access restrictions, and data minimization.

  19. Question 19Describe Artificial Intelligence workloads and considerations

    Which AI workload generates spoken announcements from written text in a self-service phone system?

    1. Key phrase extraction
    2. Speech-to-text
    3. Machine translation
    4. Text-to-speech Correct answer.

    Explanation

    Text-to-speech synthesis converts written text into audible speech. It is the counterpart of speech-to-text, which converts audio into written words.

  20. Question 20Describe Artificial Intelligence workloads and considerations

    Which statement correctly describes the relationship among AI, machine learning, and deep learning?

    1. The three terms describe unrelated disciplines with no overlap in practice.
    2. Machine learning is a superset of AI and also includes deep learning as a subset.
    3. AI contains machine learning, which in turn contains deep learning. Correct answer.
    4. Deep learning is the broadest of the three terms and includes every AI system.

    Explanation

    Artificial intelligence is the broad field of building systems that exhibit intelligent behavior. Machine learning is a subset of AI that learns from data, and deep learning is a further subset of machine learning based on multi-layer neural networks.

  21. Question 21Describe Artificial Intelligence workloads and considerations

    A pneumonia detection model reports 99% accuracy on a test set in which 99% of the images are healthy. Which conclusion is most justified?

    1. The data is imbalanced, so accuracy can hide poor detection of the minority class. Correct answer.
    2. The model is ready for clinical use because its accuracy is above the 95% threshold commonly required.
    3. The model has almost certainly overfitted the training data and must be retrained from scratch.
    4. Accuracy is a trustworthy metric in this case because the test set contains a large number of images.

    Explanation

    With severe class imbalance, a model that always predicts the majority class can reach very high accuracy while never identifying a positive case. Metrics such as recall and precision for the minority class reveal the real performance.

  22. Question 22Describe Artificial Intelligence workloads and considerations

    An audit finds that a resume-screening model recommends fewer interviews for one demographic group. Which action best follows the fairness principle?

    1. Measure model outcomes across groups and remediate the disparity before deployment. Correct answer.
    2. Remove candidate names from the input data and deploy the model without further analysis.
    3. Lower the decision threshold so that a greater number of candidates pass the screening stage.
    4. Collect a much larger number of resumes so the training set becomes larger overall.

    Explanation

    Fairness requires measuring outcomes across subgroups and actively correcting disparities, for example through data remediation, feature changes, or threshold adjustments validated per group. Removing one sensitive field or simply enlarging the dataset does not confirm that the disparity is gone.

  23. Question 23Describe Artificial Intelligence workloads and considerations

    A team improves a hiring model's overall accuracy, but its outcomes become less equal across groups. What does this best illustrate?

    1. The model should be retrained repeatedly until both accuracy and fairness reach perfect scores.
    2. Accuracy and fairness can conflict, so teams must evaluate trade-offs. Correct answer.
    3. Accuracy is the only metric that matters when judging whether an AI system is responsible.
    4. Fairness always improves whenever a model's overall accuracy increases, so no further work is needed.

    Explanation

    Responsible AI involves balancing competing objectives, and optimizing one metric can degrade another such as fairness across groups. Teams therefore track multiple metrics and make deliberate, documented trade-off decisions.

  24. Question 24Describe Artificial Intelligence workloads and considerations

    Which statement distinguishes classification from clustering?

    1. Classification requires no training data at all, while clustering always needs a fully labeled dataset.
    2. Classification uses predefined labels; clustering discovers groups in unlabeled data. Correct answer.
    3. The two terms name the same machine learning task, differing only in the file format of the data.
    4. Clustering is used to predict numeric values, while classification is used to predict discrete categories.

    Explanation

    Classification is supervised and learns from examples that already carry the correct label, so the categories are known in advance. Clustering is unsupervised and infers groupings from similarity without any labels.

  25. Question 25Describe Artificial Intelligence workloads and considerations

    Why is a purely rule-based system often unsuitable for tasks such as recognizing objects in photographs?

    1. Rule-based systems cannot be deployed on a web server or in a cloud environment without modification.
    2. Rules cannot generalize from examples and must be rewritten as conditions change. Correct answer.
    3. Rule-based systems require large labeled training datasets before any rule can be evaluated.
    4. Rule-based systems always output numeric values, so they cannot be used for categorization tasks.

    Explanation

    Hand-written rules must anticipate every case explicitly and do not learn patterns from data, which makes them brittle for complex perceptual tasks. Machine learning models instead generalize from examples and adapt when retrained on new data.

  26. Question 26Describe Artificial Intelligence workloads and considerations

    A company wants software to read the printed text on millions of scanned receipts and store that text in a searchable database. Which AI workload provides this capability?

    1. Computer vision Correct answer.
    2. Natural language processing
    3. Speech recognition
    4. Anomaly detection

    Explanation

    Reading characters from a scanned image is optical character recognition, which is a computer vision capability that interprets pixels in an image. Natural language processing begins only after the text already exists in a machine-readable form.

  27. Question 27Describe Artificial Intelligence workloads and considerations

    A team building a self-driving shuttle must ensure that the system behaves predictably and does not cause harm when a sensor returns missing or unusual readings. Which responsible AI principle does this requirement describe?

    1. Reliability and safety Correct answer.
    2. Transparency of decisions
    3. Accountability for outcomes
    4. Privacy and security of data

    Explanation

    Reliability and safety requires that an AI system perform consistently and fail safely when it encounters conditions outside its normal operating range. Continuing to operate without harm when sensor input is imperfect is the defining concern of this principle.

  28. Question 28Describe Artificial Intelligence workloads and considerations

    A model is trained to predict the selling price of a house in dollars from its size, age, and location. Which type of machine learning task is this?

    1. Binary classification
    2. Clustering
    3. Reinforcement learning
    4. Regression Correct answer.

    Explanation

    Regression models predict a continuous numeric quantity, and a price in dollars is continuous. Classification would instead predict a discrete category, and clustering would group similar records without a target value.

  29. Question 29Describe Artificial Intelligence workloads and considerations

    A media company wants to generate searchable written transcripts from hundreds of hours of recorded interviews. Which AI workload should it use?

    1. Speech recognition Correct answer.
    2. Anomaly detection
    3. Text-to-speech synthesis
    4. Computer vision

    Explanation

    Speech recognition transcribes spoken audio into written text, which is exactly what producing transcripts from recordings requires. Text-to-speech works in the opposite direction, turning text into audio.

  30. Question 30Describe Artificial Intelligence workloads and considerations

    An accounts payable team receives thousands of scanned invoices with varying layouts and needs the vendor name, invoice date, and total amount captured automatically. Which AI workload is designed for this?

    1. Knowledge mining
    2. Generative AI
    3. Anomaly detection
    4. Document intelligence Correct answer.

    Explanation

    Document intelligence extracts structured fields such as names, dates, and totals from forms and invoices, including documents with differing layouts. It returns field-level values rather than a searchable index or newly generated content.

  31. Question 31Describe Artificial Intelligence workloads and considerations

    Which AI workload responds to a natural-language prompt by producing original content such as a paragraph of text or an illustration?

    1. Anomaly detection
    2. Computer vision
    3. Speech synthesis
    4. Generative AI Correct answer.

    Explanation

    Generative AI models create new content, including text and images, in response to a prompt. The other workloads analyze, classify, or convert existing data rather than create original output.

  32. Question 32Describe Artificial Intelligence workloads and considerations

    In supervised machine learning, what is the term for the measurable input attributes, such as age and income, that a model uses to make a prediction?

    1. Hyperparameters
    2. Features Correct answer.
    3. Labels
    4. Evaluation metrics

    Explanation

    Features are the input variables that describe each observation and drive the prediction, while the label is the known target value the model is trained to output. Hyperparameters and metrics relate to how a model is configured and scored, not to its input attributes.

  33. Question 33Describe Artificial Intelligence workloads and considerations

    A public self-service kiosk is designed so that people with visual impairments can complete every task using audio prompts and screen-reader support. Which responsible AI principle does this best demonstrate?

    1. Inclusiveness Correct answer.
    2. Transparency
    3. Reliability and safety
    4. Accountability

    Explanation

    Inclusiveness requires that AI systems empower and engage everyone, including people with disabilities, by offering usable alternative modes of interaction. Providing audio and screen-reader paths so all users can finish tasks is a direct application of that principle.

  34. Question 34Describe Artificial Intelligence workloads and considerations

    Which natural language processing capability classifies a piece of text as positive, negative, or neutral?

    1. Key phrase extraction
    2. Speech synthesis
    3. Sentiment analysis Correct answer.
    4. Object detection

    Explanation

    Sentiment analysis scores text for the attitude it expresses, commonly positive, negative, or neutral. Key phrase extraction instead returns the main terms in the text without judging opinion.

  35. Question 35Describe Artificial Intelligence workloads and considerations

    A subscription service trains a model that outputs a probability between 0 and 1 indicating whether each customer will cancel in the next month. Which type of machine learning task is this?

    1. Reinforcement learning
    2. Binary classification Correct answer.
    3. Regression
    4. Clustering

    Explanation

    The model chooses between two discrete outcomes, will cancel or will not cancel, which makes it binary classification; the probability is simply a confidence score for one of those classes. Regression would apply only if the target were a continuous quantity such as a revenue amount.

  36. Question 36Describe Artificial Intelligence workloads and considerations

    A drone must assign a class label to every individual pixel in each camera frame so that roads, trees, and buildings are each outlined. Which computer vision capability is required?

    1. Object detection
    2. Semantic segmentation Correct answer.
    3. Image classification
    4. Optical character recognition

    Explanation

    Semantic segmentation classifies each pixel, producing a mask that outlines the region belonging to every class. Image classification returns a single label for the whole picture, and object detection returns bounding boxes rather than pixel-level masks.

  37. Question 37Describe Artificial Intelligence workloads and considerations

    Which statement best describes what distinguishes deep learning from classical machine learning?

    1. It requires labeled training data, whereas classical methods do not
    2. It always produces models that are simpler to explain than classical ones
    3. It learns features automatically with deep neural networks Correct answer.
    4. It can be applied only to image data, while classical methods handle text

    Explanation

    Deep learning uses neural networks with many layers that learn increasingly abstract feature representations directly from raw data, reducing the need for hand-crafted features. Classical machine learning typically depends on features that a person defines in advance.

  38. Question 38Describe Artificial Intelligence workloads and considerations

    A live event requires spoken English to appear as written French subtitles within a few seconds. Which combination of AI capabilities is required?

    1. Speech synthesis and image classification
    2. Sentiment analysis and text translation
    3. Speech synthesis and optical character recognition
    4. Speech recognition and text translation Correct answer.

    Explanation

    The audio must first be transcribed into text by speech recognition, and that text is then translated into French by a text translation capability. Each stage transforms the content into the input format that the next stage requires.

  39. Question 39Describe fundamental principles of machine learning on Azure

    A data scientist trains a model by providing historical sales records that include the actual sales amount for each month. Which type of machine learning does this describe?

    1. Reinforcement learning
    2. Unsupervised learning
    3. Supervised learning Correct answer.
    4. Clustering analysis

    Explanation

    Supervised learning trains a model on examples that include both input features and the known outcome, called the label. Because historical sales amounts are supplied, the model can learn the mapping from inputs to the numeric outcome. Unsupervised methods work without labels.

  40. Question 40Describe fundamental principles of machine learning on Azure

    In a dataset used to train a model that predicts house prices, what is the term for the actual recorded price of each house?

    1. Feature
    2. Hyperparameter
    3. Label Correct answer.
    4. Algorithm

    Explanation

    The label is the target value the model is trained to predict, such as the recorded price. Features are the input columns used to make the prediction. Hyperparameters and algorithms are configuration choices, not data values.

  41. Question 41Describe fundamental principles of machine learning on Azure

    Which machine learning approach groups customers into segments without using any predefined categories?

    1. Classification
    2. Forecasting
    3. Regression
    4. Clustering Correct answer.

    Explanation

    Clustering is an unsupervised technique that groups records by similarity without predefined categories. Classification requires known classes, and regression predicts numeric values.

  42. Question 42Describe fundamental principles of machine learning on Azure

    A model is required to predict tomorrow's temperature in degrees Celsius. Which type of machine learning task is this?

    1. Regression Correct answer.
    2. Clustering
    3. Classification
    4. Anomaly detection

    Explanation

    Regression tasks predict a continuous numeric quantity such as a temperature. Classification predicts a category, and clustering discovers groups without labels.

  43. Question 43Describe fundamental principles of machine learning on Azure

    What is the primary purpose of the designer in Azure Machine Learning?

    1. To host trained models as web services for real-time scoring
    2. To monitor the cost of compute resources used by experiments
    3. To build pipelines visually by connecting modules Correct answer.
    4. To generate labeled training data automatically from raw images

    Explanation

    The designer is a visual authoring surface where modules representing data and model steps are connected into a pipeline. Hosting models and monitoring costs are separate capabilities of the workspace.

  44. Question 44Describe fundamental principles of machine learning on Azure

    What does automated machine learning (AutoML) in Azure Machine Learning do?

    1. It converts every trained model into a REST endpoint automatically
    2. It labels images and text using pretrained foundation models
    3. It rewrites notebooks so they run on graphics processing units
    4. It tries many algorithms and settings to find a good model Correct answer.

    Explanation

    Automated machine learning runs many combinations of algorithms, preprocessing steps, and hyperparameter settings, then ranks the resulting models by a chosen metric. It does not label data or convert models into endpoints.

  45. Question 45Describe fundamental principles of machine learning on Azure

    What is the purpose of training data in the machine learning process?

    1. To teach the model the relationship between features and labels Correct answer.
    2. To tune the number of layers in a deep neural network
    3. To evaluate the model after it has been deployed to production
    4. To store the version history of the model in a registry

    Explanation

    Training data provides the examples, including features and labels, from which the model learns patterns. Evaluation and versioning happen around training rather than through the training data itself.

  46. Question 46Describe fundamental principles of machine learning on Azure

    Why is a separate test dataset used when evaluating a model?

    1. It estimates how the model performs on data it has never seen Correct answer.
    2. It automatically balances the classes in the training data
    3. It provides additional labeled examples for the training step
    4. It reduces the amount of compute needed to train the model

    Explanation

    A test set is held out from training so the evaluation measures generalization to data the model has not seen. Reusing training data for evaluation would overstate performance.

  47. Question 47Describe fundamental principles of machine learning on Azure

    A classification model scores 99 percent on training data but only 62 percent on new data. What does this indicate?

    1. The model has too few features to learn
    2. The model was trained on too much data
    3. The model is overfitting the training data Correct answer.
    4. The model is underfitting the training data

    Explanation

    A large gap between training and holdout performance is the signature of overfitting, where the model memorizes training-specific noise. Underfitting would instead show poor results on both sets.

  48. Question 48Describe fundamental principles of machine learning on Azure

    A binary classification model has an accuracy of 0.85. What does this mean?

    1. 85 percent of predicted positives are actually positive
    2. 85 percent of the features are useful for prediction
    3. 85 percent of actual positives were correctly identified
    4. 85 percent of all predictions match the true labels Correct answer.

    Explanation

    Accuracy is the ratio of correct predictions to all predictions. Recall and precision describe performance on the positive class only.

  49. Question 49Describe fundamental principles of machine learning on Azure

    Which metric reports the proportion of actual positive cases that the model correctly identified?

    1. Accuracy
    2. Recall Correct answer.
    3. Specificity
    4. Precision

    Explanation

    Recall, also called the true positive rate, is true positives divided by all actual positives. Precision instead uses the number of positive predictions as its denominator.

  50. Question 50Describe fundamental principles of machine learning on Azure

    Which metric reports the proportion of positive predictions that were actually correct?

    1. Precision Correct answer.
    2. Recall
    3. Accuracy
    4. AUC

    Explanation

    Precision is true positives divided by all predicted positives, so it measures how trustworthy a positive prediction is. Recall measures coverage of actual positives.

  51. Question 51Describe fundamental principles of machine learning on Azure

    What does root mean squared error (RMSE) express about a regression model?

    1. The percentage of predictions that fall within one unit
    2. The average magnitude of error in the units of the label Correct answer.
    3. The ratio of correct predictions to total predictions
    4. The proportion of rows the model classified incorrectly

    Explanation

    RMSE is the square root of the mean squared difference between predicted and actual values, so it is expressed in the same units as the label. It summarizes the typical size of the error rather than a percentage of correct classifications.

  52. Question 52Describe fundamental principles of machine learning on Azure

    A data scientist converts a customer's birth date into an age column. What is this an example of?

    1. Model deployment
    2. Data labeling
    3. Hyperparameter tuning
    4. Feature engineering Correct answer.

    Explanation

    Feature engineering creates new input columns, such as age derived from a birth date, that make patterns easier for the model to learn. Labeling, deployment, and tuning are different stages of the workflow.

  53. Question 53Describe fundamental principles of machine learning on Azure

    Which Azure Machine Learning resource is used to run a training experiment that requires scalable compute?

    1. A datastore
    2. An online endpoint
    3. A model registry
    4. A compute cluster Correct answer.

    Explanation

    A compute cluster provides scalable, multi-node compute for training runs and can scale down when idle. Registries, datastores, and endpoints serve different purposes.

  54. Question 54Describe fundamental principles of machine learning on Azure

    Which Azure Machine Learning resource provides a REST endpoint for real-time predictions from a trained model?

    1. A compute instance
    2. A pipeline
    3. A managed online endpoint Correct answer.
    4. A dataset

    Explanation

    A managed online endpoint hosts a model and exposes a REST interface for low-latency, real-time scoring. Compute instances are for development, and pipelines orchestrate workflow steps.

  55. Question 55Describe fundamental principles of machine learning on Azure

    Why should the test dataset be kept separate from the training dataset?

    1. So the labels can be generated automatically
    2. So the model can store feature values in memory
    3. So training finishes faster on smaller datasets
    4. So the evaluation reflects performance on unseen data Correct answer.

    Explanation

    Holding out data prevents the model from being evaluated on examples it memorized during training. The split is about the validity of the performance estimate, not speed or memory.

  56. Question 56Describe fundamental principles of machine learning on Azure

    What is the main purpose of the model registry in Azure Machine Learning?

    1. To track and version trained models across workspaces Correct answer.
    2. To provide a code editor for authoring pipelines
    3. To store datasets that are used in training experiments
    4. To schedule notebooks to run at a specific time

    Explanation

    The registry stores model versions with their metadata so they can be tracked, compared, and deployed. Datasets, schedules, and authoring tools are handled by other workspace features.

  57. Question 57Describe fundamental principles of machine learning on Azure

    In a binary classification confusion matrix, what is a false positive?

    1. A negative case that the model labeled positive Correct answer.
    2. A positive case that the model labeled positive
    3. A positive case that the model labeled negative
    4. A negative case that the model labeled negative

    Explanation

    A false positive is a negative case that the model incorrectly labels as positive. A positive case labeled negative is a false negative.

  58. Question 58Describe fundamental principles of machine learning on Azure

    Which principle of responsible AI is supported by tools that explain why a model made a particular prediction?

    1. Accountability
    2. Privacy and security
    3. Reliability and safety
    4. Transparency Correct answer.

    Explanation

    Transparency means people can understand how a model behaves and why it produced a result, which is what explanation tools support. Fairness, privacy, and reliability address other concerns.

  59. Question 59Describe fundamental principles of machine learning on Azure

    A model has an AUC of 0.9. Which interpretation is correct?

    1. The model's predictions are within 0.9 units of the true value
    2. A random positive ranks above a random negative 90 percent of the time Correct answer.
    3. The model identifies 90 percent of all positive cases at every possible threshold
    4. The model classifies exactly 90 percent of all cases in the dataset correctly

    Explanation

    AUC is the probability that a randomly chosen positive case receives a higher score than a randomly chosen negative case. It is threshold independent and is not an accuracy percentage.

  60. Question 60Describe fundamental principles of machine learning on Azure

    A linear regression model has high error on both the training set and the validation set. What is the most likely cause?

    1. The model is too simple for the data Correct answer.
    2. The validation set was shuffled before use
    3. The training set contains too many rows
    4. The model is too complex for the data

    Explanation

    High error on both training and validation data indicates the model is too simple or lacks useful features, which is underfitting. Overfitting would instead show low training error and high validation error.

  61. Question 61Describe fundamental principles of machine learning on Azure

    Why is k-fold cross-validation used when evaluating a model?

    1. It removes the need to label the training data
    2. It increases the number of features available for training
    3. It guarantees that the model will never overfit the data
    4. It averages results across several train-test splits Correct answer.

    Explanation

    Cross-validation trains and evaluates the model several times on different folds, then averages the scores to reduce the influence of one particular split. It does not add data or guarantee any outcome.

  62. Question 62Describe fundamental principles of machine learning on Azure

    What is a hyperparameter in machine learning?

    1. A value that the model learns from the training data
    2. A column in the dataset that is used as an input
    3. The known output value recorded for each training row
    4. A setting chosen before training to control learning Correct answer.

    Explanation

    Hyperparameters, such as learning rate or tree depth, are set before training and govern how learning proceeds. Model parameters and labels are learned from or supplied by the data.

  63. Question 63Describe fundamental principles of machine learning on Azure

    A fraud dataset contains 1 percent fraudulent cases. A model that predicts 'not fraud' for every row reports 99 percent accuracy. What should the data scientist conclude?

    1. The dataset should be split into more training folds
    2. The model is ready for deployment because accuracy is high
    3. Accuracy is misleading here; review recall or precision Correct answer.
    4. The model needs more features to improve its accuracy

    Explanation

    With severe class imbalance, a trivial model can score high accuracy while missing every positive case. Metrics focused on the positive class, such as recall and precision, reveal the failure.

  64. Question 64Describe fundamental principles of machine learning on Azure

    Which machine learning approach trains a model on data that includes both input features and known target values?

    1. Supervised learning Correct answer.
    2. Transfer learning
    3. Reinforcement learning
    4. Unsupervised learning

    Explanation

    Supervised learning uses labeled examples that pair features with known targets, allowing the model to learn a mapping from inputs to outputs. Unsupervised learning works without labels, and reinforcement learning learns from reward signals gathered through interaction. Transfer learning reuses a pretrained model but is not defined by the presence of labels.

  65. Question 65Describe fundamental principles of machine learning on Azure

    A marketing team wants to group customers into segments based on similarity of purchase behavior, without predefined segment labels. Which type of machine learning should they use?

    1. Named entity recognition
    2. Binary classification
    3. Clustering Correct answer.
    4. Regression

    Explanation

    Clustering is an unsupervised technique that groups similar observations together when no predefined categories exist. Regression and classification are supervised tasks that require known target values, while named entity recognition is a text extraction task.

  66. Question 66Describe fundamental principles of machine learning on Azure

    Which machine learning task predicts a continuous numeric value, such as the expected number of units sold next month?

    1. Multi-class classification
    2. Binary classification
    3. Clustering
    4. Regression Correct answer.

    Explanation

    Regression models output a continuous numeric quantity, such as a count, price, or demand forecast. Classification tasks assign items to discrete categories, and clustering groups unlabeled observations by similarity.

  67. Question 67Describe fundamental principles of machine learning on Azure

    In a dataset used to train a model that predicts house prices, the column containing the actual sale price is known as the:

    1. feature
    2. label Correct answer.
    3. hyperparameter
    4. weight

    Explanation

    The label, also called the target, is the value the model is trained to predict. Features are the input columns, while hyperparameters and weights are model configuration or learned values rather than dataset columns.

  68. Question 68Describe fundamental principles of machine learning on Azure

    What is the primary purpose of a validation dataset during model development?

    1. To increase the volume of data used for model training
    2. To hold raw records before feature engineering is applied
    3. To tune hyperparameters and compare candidate models Correct answer.
    4. To provide the final unbiased estimate of model performance

    Explanation

    The validation set guides model selection and hyperparameter tuning so that the test set stays untouched for a final unbiased estimate. Training on validation data would undermine its purpose of checking generalization during development.

  69. Question 69Describe fundamental principles of machine learning on Azure

    Why are numeric features often scaled to a similar range before training certain models?

    1. It automatically converts every categorical column into a numeric value for the model
    2. It removes the need to split the available data into training and test sets
    3. It guarantees that the trained model will not overfit the training dataset
    4. It stops large-magnitude features from dominating distance calculations Correct answer.

    Explanation

    Scaling puts features on comparable ranges so that distance metrics and gradient-based optimization are not controlled by columns with large numeric values. It does not replace data splitting, guarantee generalization, or encode categorical data.

  70. Question 70Describe fundamental principles of machine learning on Azure

    A dataset contains a 'City' column with values such as Seattle, Paris, and Osaka. Which technique best represents this feature for a model that cannot use raw text?

    1. Mean imputation of missing city values
    2. Min-max normalization of the city codes
    3. Binning the cities into equal-width ranges
    4. One-hot encoding Correct answer.

    Explanation

    One-hot encoding creates a separate binary column for each category, letting the model use nominal data without implying an order. Scaling, binning, and imputation address numeric range, numeric grouping, and missing values respectively.

  71. Question 71Describe fundamental principles of machine learning on Azure

    Why is a stratified split recommended when creating training and test sets from an imbalanced classification dataset?

    1. It preserves the original class proportions in both resulting sets Correct answer.
    2. It reduces the number of features that the model must process
    3. It guarantees that the model will achieve a higher accuracy score
    4. It removes duplicate rows from the dataset before training

    Explanation

    Stratification keeps each class's share of the data consistent across splits, so rare classes remain represented in both training and test data. It does not change feature count, remove duplicates, or ensure a higher accuracy score.

  72. Question 72Describe fundamental principles of machine learning on Azure

    Which Azure Machine Learning capability lets you build training workflows visually by connecting data and prebuilt module nodes?

    1. Compute instance
    2. Designer Correct answer.
    3. Model registry
    4. Data store

    Explanation

    Designer provides a drag-and-drop canvas for composing pipelines from modules. Compute instances are development virtual machines, the model registry tracks model versions, and data stores hold connection information for storage services.

  73. Question 73Describe fundamental principles of machine learning on Azure

    A team builds a model to predict whether a loan will default and includes a 'collection agency referral date' feature that is recorded only after default. What problem does this feature introduce?

    1. Data leakage that inflates measured performance Correct answer.
    2. Class imbalance caused by the new feature
    3. Underfitting caused by too few features
    4. Feature scaling drift between training and scoring

    Explanation

    A feature recorded only after the outcome is known leaks the target into training, so the model appears accurate but cannot be used at prediction time. Such features must be removed or replaced with information available before the event being predicted.

  74. Question 74Describe fundamental principles of machine learning on Azure

    A company must score several million records every night and does not need immediate responses. Which deployment approach is most appropriate?

    1. A local container started manually by an operator each evening
    2. An interactive notebook that a user runs on demand
    3. A batch endpoint run on a nightly schedule Correct answer.
    4. A real-time endpoint with autoscaling enabled for peak load

    Explanation

    Batch endpoints score large datasets on a schedule without keeping a live service waiting for requests. Real-time endpoints are designed for low-latency individual requests and cost more when idle.

  75. Question 75Describe fundamental principles of machine learning on Azure

    A deployed model's input data gradually shifts so that it no longer resembles the training data, and prediction quality falls. What is this phenomenon called?

    1. Overfitting
    2. Regularization
    3. Hyperparameter tuning
    4. Data drift Correct answer.

    Explanation

    Data drift is a change in the distribution of model inputs relative to the training data, which can degrade predictions over time. Monitoring input distributions helps detect drift and trigger retraining.

  76. Question 76Describe fundamental principles of machine learning on Azure

    What is the role of an Azure Machine Learning workspace?

    1. It provides a central place to manage experiments, models, and compute Correct answer.
    2. It stores only the raw data files that training scripts read
    3. It is a REST endpoint that serves model predictions to client applications
    4. It is a virtual machine used exclusively for interactive notebook development

    Explanation

    A workspace is the top-level Azure Machine Learning resource that organizes experiments, models, compute, and related assets. Data storage, development virtual machines, and scoring endpoints are separate resources managed from it.

  77. Question 77Describe features of computer vision workloads on Azure

    Which image analysis capability produces a short natural-language sentence describing what an image contains?

    1. Image captioning Correct answer.
    2. Face verification
    3. Optical character recognition
    4. Object detection

    Explanation

    Captioning generates a readable sentence that summarizes the whole image. OCR extracts characters from the picture, object detection returns boxes and labels for localized items, and face verification compares two faces for identity. Only captioning yields a descriptive sentence.

  78. Question 78Describe features of computer vision workloads on Azure

    Which computer vision capability returns a class label together with bounding box coordinates for each item found in an image?

    1. Image tagging
    2. Image classification
    3. OCR
    4. Object detection Correct answer.

    Explanation

    Object detection both identifies what is present and localizes it, so every result includes a label and a rectangle. Classification assigns a label to the image as a whole without coordinates, tagging returns descriptive words without positions, and OCR extracts text.

  79. Question 79Describe features of computer vision workloads on Azure

    Which capability should you use to extract printed and handwritten text from scanned documents and photographs?

    1. Custom Vision
    2. Read OCR Correct answer.
    3. Face detection
    4. Spatial analysis

    Explanation

    The Read capability of Azure AI Vision extracts printed and handwritten text and returns it organized into pages, lines, and words. Custom Vision trains classifiers or detectors on your own images, spatial analysis tracks people in video, and face detection locates faces.

  80. Question 80Describe features of computer vision workloads on Azure

    What does the Azure AI Face service detect and locate in photographs?

    1. Human faces in photographs Correct answer.
    2. Text lines in scanned pages
    3. Barcodes on product packaging
    4. Emotions in written customer reviews

    Explanation

    The Face service finds faces and returns a rectangle for each one, optionally along with attributes such as head pose, blur, or glasses. It does not read text, analyze written sentiment, or decode barcodes. Those tasks belong to other vision capabilities.

  81. Question 81Describe features of computer vision workloads on Azure

    Which Azure service is designed to extract structured fields such as vendor name, invoice date, and total amount from business documents?

    1. Azure AI Vision Image Analysis
    2. Azure Video Indexer
    3. Azure AI Face
    4. Document Intelligence Correct answer.

    Explanation

    Document Intelligence provides pretrained and custom models that return named fields and values from forms, invoices, and receipts. Image Analysis describes picture content, Video Indexer works on media files, and the Face service analyzes faces.

  82. Question 82Describe features of computer vision workloads on Azure

    Which computer vision workload assigns one or more labels to a whole image without indicating where objects are located?

    1. Classification Correct answer.
    2. Spatial analysis
    3. Optical character recognition
    4. Object detection

    Explanation

    Classification answers the question of what the image shows by returning labels for the entire picture. Detection adds coordinates for each item, OCR converts text, and spatial analysis tracks people across video frames.

  83. Question 83Describe features of computer vision workloads on Azure

    Which Azure service analyzes video files to produce transcripts, speaker insights, and detected topics?

    1. Azure AI Vision
    2. Azure AI Document Intelligence
    3. Azure AI Face
    4. Azure Video Indexer Correct answer.

    Explanation

    Video Indexer extracts insights from media, combining speech transcription, speaker identification, topic extraction, and visual analysis. The other services target still images, documents, or faces rather than full media files.

  84. Question 84Describe features of computer vision workloads on Azure

    What is Azure Custom Vision used for?

    1. Converting speech to text in real time
    2. Summarizing long documents
    3. Training models on your own images Correct answer.
    4. Detecting sentiment in customer reviews

    Explanation

    Custom Vision lets you upload and label your own images, then train classification or object detection models tuned to your specific categories. It does not perform speech conversion, sentiment analysis, or document summarization.

  85. Question 85Describe features of computer vision workloads on Azure

    A retailer wants to count how many customers enter a store and how long they wait in a checkout queue using existing ceiling cameras. Which Azure AI Vision feature is the best fit?

    1. Custom Vision classification
    2. Image captioning
    3. Read OCR
    4. Spatial analysis Correct answer.

    Explanation

    Spatial analysis processes streams from fixed cameras and emits events such as person counts, dwell time, and queue length. Captioning and OCR do not track movement over time, and a classification model would label frames without counting people.

  86. Question 86Describe features of computer vision workloads on Azure

    In Custom Vision, which project type should you choose when every image belongs to exactly one of several mutually exclusive categories?

    1. Document Intelligence
    2. Object detection
    3. Multilabel classification
    4. Multiclass classification Correct answer.

    Explanation

    Multiclass classification assigns a single label from a set of mutually exclusive classes. Multilabel classification allows several tags at once, object detection adds regions, and Document Intelligence handles forms rather than image categories.

  87. Question 87Describe features of computer vision workloads on Azure

    Azure AI Vision returns tags with an associated confidence value between 0 and 1. How should that value be used?

    1. It indicates the image resolution
    2. It ranks images by file size
    3. It counts how many objects exist
    4. It sets the acceptance threshold Correct answer.

    Explanation

    The confidence score expresses how certain the model is about a prediction, so an application can accept only results above a chosen threshold. It carries no information about file size, pixel dimensions, or the number of objects.

  88. Question 88Describe features of computer vision workloads on Azure

    A news website needs square thumbnail images generated automatically from large photos while keeping the main subject visible. Which capability should be used?

    1. Face verification
    2. Smart cropping Correct answer.
    3. OCR
    4. Object detection

    Explanation

    Smart cropping, available through the thumbnail generation feature of Image Analysis, analyzes the image and chooses a crop region that keeps the most important content. Object detection, OCR, and face verification produce labels, text, or identity comparisons instead of cropped images.

  89. Question 89Describe features of computer vision workloads on Azure

    A factory needs to know whether workers are wearing safety helmets and where each helmet appears in each frame. Which approach is most appropriate?

    1. Use Document Intelligence on the video
    2. Use classification with one 'factory' label
    3. Train an object detection model on labeled images Correct answer.
    4. Use OCR on every frame

    Explanation

    Because the solution must locate each helmet, a Custom Vision object detection model trained on labeled helmet images is required. Classification only labels the frame, OCR reads characters, and Document Intelligence handles documents.

  90. Question 90Describe features of computer vision workloads on Azure

    Which capability reports quality attributes such as blur, exposure, and occlusion for a detected face?

    1. Face quality attributes Correct answer.
    2. OCR
    3. Object detection
    4. Image tagging

    Explanation

    Face detection can return attributes that describe image quality for each face, which helps filter out photos unsuitable for recognition. OCR and tagging describe text or scene content, and generic object detection reports no facial quality measures.

  91. Question 91Describe features of computer vision workloads on Azure

    What is the main purpose of the prebuilt receipt model in Document Intelligence?

    1. Extract fields like date and total Correct answer.
    2. Translate receipts into other languages
    3. Verify the signature on a receipt
    4. Detect the merchant logo as an object

    Explanation

    The prebuilt receipt model recognizes common receipt structures and returns named fields such as merchant name, transaction date, line items, and total. It is not a translation tool, a logo detector, or a signature validator.

  92. Question 92Describe features of computer vision workloads on Azure

    How does Custom Vision differ from the pretrained Image Analysis models in Azure AI Vision?

    1. Custom Vision lets you train on your own labeled images Correct answer.
    2. Custom Vision cannot be used for images
    3. Custom Vision requires no training data at all
    4. Custom Vision only works on invoices

    Explanation

    Pretrained Image Analysis models are ready to use and cover broad, general concepts, whereas Custom Vision trains a model from images you label so it recognizes your specific classes. Custom Vision therefore depends on training data and is not limited to documents.

  93. Question 93Describe features of computer vision workloads on Azure

    Which Azure AI Vision feature helps a social media app flag potentially unsuitable images before they are published?

    1. Adult and racy content filtering Correct answer.
    2. Face verification
    3. Custom Vision classification
    4. Spatial analysis

    Explanation

    Image Analysis can return adult and racy content flags so an application can block or review risky images. Spatial analysis tracks people in video, face verification compares identities, and a custom classifier would need its own labeled training data.

  94. Question 94Describe features of computer vision workloads on Azure

    How does an application typically invoke Azure AI Vision's Image Analysis capability?

    1. Only through the Azure portal
    2. Only through Power BI reports
    3. Only by exporting models to Excel
    4. REST API calls to the resource endpoint Correct answer.

    Explanation

    The service is exposed through REST endpoints secured by a key or Microsoft Entra authentication, and client SDKs wrap those calls. Portal testing is convenient but not the integration path, and Excel or Power BI are not required to consume the API.

  95. Question 95Describe features of computer vision workloads on Azure

    An online catalogue needs to place products on clean white backgrounds. Which Image Analysis feature supports this?

    1. Object detection
    2. Face verification
    3. OCR
    4. Background removal Correct answer.

    Explanation

    Background removal produces a foreground matte of the main subject so it can be composited onto a new background. Object detection, OCR, and face verification return labels, text, or identity results rather than modified images.

  96. Question 96Describe features of computer vision workloads on Azure

    What does the dense captions feature of Image Analysis return?

    1. A list of faces with names
    2. Captions for regions, each with a bounding box Correct answer.
    3. A translation of text found in the image
    4. A single label for the whole image

    Explanation

    Dense captioning describes multiple areas of an image rather than summarizing it once, returning a caption and coordinates for each region. It is not a whole-image label, a translator, or a face identification feature.

  97. Question 97Describe features of computer vision workloads on Azure

    Which statement about face identification and verification in Azure AI Face is accurate?

    1. It only works on video streams
    2. It requires Limited Access approval Correct answer.
    3. It is available to any subscriber without approval
    4. It is only supported for handwritten text

    Explanation

    Face identification and verification are Limited Access capabilities: an organization must register a use case and be approved before using them. Face detection is broadly available, but identity matching is gated to support responsible AI commitments.

  98. Question 98Describe features of computer vision workloads on Azure

    An image classification model returns 'cat' with a confidence of 0.45 for a photo that actually shows a dog. What is the most reasonable interpretation?

    1. OCR failed on the image
    2. The image file is corrupt
    3. The model detected a cat and is certain
    4. The prediction is below a likely threshold Correct answer.

    Explanation

    Confidence values near 0.5 indicate weak certainty, so a well-chosen threshold would reject or flag this result. A corrupt file would typically cause an error rather than a confident label, and OCR plays no role in scene classification.

  99. Question 99Describe features of computer vision workloads on Azure

    A set of contracts has tables, checkboxes, and handwritten notes in a layout unique to one company. Which Document Intelligence approach is most suitable?

    1. Prebuilt receipt model
    2. A custom Document Intelligence model Correct answer.
    3. Spatial analysis
    4. Azure AI Vision image tagging

    Explanation

    Custom Document Intelligence models are trained on labeled samples of your own layout so they can extract the fields you define. Prebuilt models cover standard document types, while tagging and spatial analysis do not extract structured fields.

  100. Question 100Describe features of computer vision workloads on Azure

    Which statement best explains why a team would choose the Read capability of Azure AI Vision instead of a basic optical character recognition call?

    1. OCR only works on PDF files
    2. Read cannot return bounding boxes
    3. Read supports only English text
    4. Read handles print and handwriting Correct answer.

    Explanation

    The Read capability is designed for both printed and handwritten content and returns structured output such as pages, lines, and words. Basic OCR is more limited in scope, especially for handwriting, and is not restricted to PDF files.

  101. Question 101Describe features of computer vision workloads on Azure

    A company must make recorded meetings searchable by both spoken words and text shown on slides during the meeting. Which solution fits best?

    1. Video Indexer with transcription and OCR of frames Correct answer.
    2. Custom Vision classification alone
    3. Spatial analysis alone
    4. Document Intelligence alone

    Explanation

    Video Indexer combines speech transcription with analysis of video frames, including reading on-screen text, and exposes the extracted insights for search. The other services cover only documents, image labels, or live camera tracking and cannot capture both modalities.

  102. Question 102Describe features of computer vision workloads on Azure

    A development team is building a Custom Vision model for photographs of hotel rooms. Each photo may show several amenities at once, such as a bathtub, a balcony, and a coffee machine, and every amenity that appears must be reported. Which Custom Vision project type should they create?

    1. A general image analysis project using prebuilt tags
    2. A multilabel classification project Correct answer.
    3. An object detection project
    4. A multiclass classification project

    Explanation

    A multilabel classification project allows one image to carry several tags simultaneously, so a room photo can be tagged with every amenity it contains. Multiclass projects force one exclusive label per image, and object detection would require locations that this scenario does not need.

  103. Question 103Describe features of computer vision workloads on Azure

    What information does the face detection operation in Azure AI Face return for each face it finds in an image?

    1. The full name and email address of the person in the photo
    2. A ranked list of close matches from a person directory that was built earlier
    3. A natural-language caption describing the person's appearance
    4. A bounding rectangle for each face, plus optional attributes Correct answer.

    Explanation

    Detection locates faces and returns a rectangle for each one, and can optionally include attributes such as head pose, blur, or accessories when requested. It does not establish who the person is; resolving identity is a separate recognition step that uses a person directory.

  104. Question 104Describe features of computer vision workloads on Azure

    Which Image Analysis feature produces a single human-readable sentence that describes the overall content of a photo and is often used to create alt text?

    1. Captioning Correct answer.
    2. Tagging
    3. Optical character recognition
    4. Smart cropping

    Explanation

    Captioning generates one sentence summarizing the whole image, which reads naturally as alternative text. Tagging returns a set of content keywords and OCR returns characters found in the image, neither of which forms a descriptive sentence.

  105. Question 105Describe features of computer vision workloads on Azure

    To train a Custom Vision object detection model, what must each training image include for every object the model should learn to find?

    1. A confidence score supplied by a human reviewer
    2. A drawn region around each object plus a tag Correct answer.
    3. A text transcript of any writing visible in the image
    4. A single tag applied to the whole image

    Explanation

    Object detection learns from labeled regions, so each object instance must be enclosed in a box and assigned its class tag. A whole-image tag alone trains a classification model and conveys no location information.

  106. Question 106Describe features of computer vision workloads on Azure

    A team has trained a Custom Vision classification model and needs it to run on a mobile device that often has no internet connection. What should they do?

    1. Call the prediction endpoint with the device's cellular connection whenever it is available
    2. Export the trained iteration to a mobile-friendly format and embed it in the app Correct answer.
    3. Publish the iteration and give the app the training key
    4. Move the project to a general-purpose prebuilt model in Azure AI Vision

    Explanation

    Custom Vision can export a trained iteration in formats such as TensorFlow, CoreML, ONNX, or as a Docker image, which lets inference run locally without a service call. The exported model keeps the team's own classes, whereas prebuilt models do not learn custom classes.

  107. Question 107Describe features of computer vision workloads on Azure

    Which credential and endpoint does an application use to obtain predictions from a published Custom Vision iteration?

    1. The subscription ID and the resource group name
    2. The prediction key and the prediction endpoint URL Correct answer.
    3. The training key and the training endpoint
    4. The project ID and the iteration ID only

    Explanation

    Scoring requests are served by the prediction resource, so the app authenticates with the prediction key and sends images to the prediction endpoint URL. The training resource and its key are used to build the model, not to classify new images.

  108. Question 108Describe features of computer vision workloads on Azure

    In Custom Vision object detection, what does mean average precision (mAP) summarize?

    1. The number of images the model can process per second during a prediction call
    2. How closely predicted boxes match labeled boxes across classes Correct answer.
    3. The percentage of images that were rejected during the training process
    4. The cost of each prediction call at the selected pricing tier

    Explanation

    Mean average precision compares predicted boxes with the labeled boxes over a range of confidence levels and then averages the result across classes. It gives a broader view of detection quality than precision or recall measured at a single threshold.

  109. Question 109Describe features of computer vision workloads on Azure

    What is the purpose of liveness detection in Azure AI Face?

    1. To compare two stored faces and return a similarity score
    2. To remove blur and red-eye from a photograph before it is stored
    3. To estimate the age of a person so the app can apply content restrictions
    4. To confirm a face comes from a real person, not a photo or mask Correct answer.

    Explanation

    Liveness detection verifies that the face presented to a camera belongs to a live person, which blocks spoofing attempts using printed photos, replayed video, or masks. Comparing two faces or scoring appearance attributes are different operations.

  110. Question 110Describe features of computer vision workloads on Azure

    A solution must extract text from scanned multi-page PDF files that are several megabytes each. Which approach fits the Read capability of Azure AI Vision?

    1. Train an image classification model on the pages that contain text
    2. Submit the document to the asynchronous Read operation and poll for results Correct answer.
    3. Submit each page as a separate synchronous request and discard the results of failed pages
    4. Use the smart cropping feature to isolate the text areas

    Explanation

    The Read capability exposes an asynchronous operation that accepts a document, returns an operation location, and stores results for retrieval by polling. That pattern is the supported way to process large, multi-page documents.

  111. Question 111Describe features of computer vision workloads on Azure

    Which insight in Azure Video Indexer makes text that appears on screen, such as slides or shop signs, searchable in a video?

    1. Speaker diarization
    2. Optical character recognition of on-screen text Correct answer.
    3. Keyword extraction from the audio transcript
    4. Sentiment analysis of spoken sentences

    Explanation

    Video Indexer applies optical character recognition to frames so that text displayed visually becomes part of the searchable index. Audio-based insights describe what was said, not what was shown on screen.

  112. Question 112Describe features of computer vision workloads on Azure

    A Custom Vision model recognizes four product tags. One tag was trained on six images while the other three used more than three hundred each, and the model rarely recognizes the small tag. What is the most appropriate next step?

    1. Delete images from the larger tags until all tags have six images
    2. Add more varied images for the small tag and retrain the model Correct answer.
    3. Switch the project to the Compact domain so the model runs faster
    4. Publish a new iteration without retraining and rely on a higher probability threshold

    Explanation

    Recognition quality depends on enough representative labeled examples, so the weak tag needs more varied images followed by a new trained iteration that can be compared with the current one. Threshold changes and removing data do not teach the model missing visual patterns.

  113. Question 113Describe features of computer vision workloads on Azure

    A warehouse solution must find each pallet in a photo and then read the serial number printed on its label. Which combination of capabilities meets both requirements?

    1. Face detection to locate the pallets and tagging to read the labels
    2. Smart cropping to locate the pallets and moderation to read the labels
    3. Object detection to locate the pallets and OCR to read the label text Correct answer.
    4. Image classification to locate the pallets and captioning to read the labels

    Explanation

    Two distinct capabilities are required: locating each pallet is a detection task, and converting printed label characters into text is an OCR task. Classification, captioning, tagging, and moderation return neither object locations together with transcribed text.

  114. Question 114Describe features of computer vision workloads on Azure

    A defect detector built with Custom Vision returns a probability for each image. Missing a defect costs far more than sending a good part for manual review. How should the team tune the decision threshold?

    1. Lower the threshold so borderline images are flagged Correct answer.
    2. Raise the threshold above 0.9 so that only highly confident detections are accepted
    3. Set the threshold to 0.5 because that value is always optimal
    4. Remove the threshold entirely and treat every prediction as a defect

    Explanation

    Lowering the threshold makes the model act on weaker evidence, which increases the proportion of real defects that are caught while raising the number of false alarms. Since a missed defect is more costly than an extra manual review, that trade-off suits this scenario.

  115. Question 115Describe features of Natural Language Processing (NLP) workloads on Azure

    What does the sentiment analysis feature of Azure AI Language report about a piece of text?

    1. The main talking points that the text discusses in detail
    2. The named people, places, and organizations in the text
    3. Which natural language the text is written in
    4. Whether the text is positive, negative, or neutral Correct answer.

    Explanation

    Sentiment analysis evaluates opinion tone and returns a label of positive, negative, or neutral, usually with confidence scores. It does not identify language, topics, or named entities.

  116. Question 116Describe features of Natural Language Processing (NLP) workloads on Azure

    Which Azure service should you use to convert a recorded meeting into a text transcript?

    1. Azure AI Document Intelligence
    2. Azure AI Speech Correct answer.
    3. Azure AI Language
    4. Azure AI Translator

    Explanation

    Speech-to-text capability in Azure AI Speech transcribes spoken audio into written text. The other services operate on text or documents rather than audio.

  117. Question 117Describe features of Natural Language Processing (NLP) workloads on Azure

    What is the purpose of key phrase extraction in Azure AI Language?

    1. To translate text into a target language
    2. To score each sentence for opinion tone
    3. To detect and mask personal information
    4. To identify the main concepts mentioned in a document Correct answer.

    Explanation

    Key phrase extraction returns the salient terms and concepts in unstructured text, which helps summarize what a document is about. It does not translate, score sentiment, or redact data.

  118. Question 118Describe features of Natural Language Processing (NLP) workloads on Azure

    You have a document and need to know whether it is written in French or Portuguese. Which Azure AI Language feature should you use?

    1. Sentiment analysis
    2. Named entity recognition
    3. Key phrase extraction
    4. Language detection Correct answer.

    Explanation

    Language detection examines the text and returns the detected language along with a confidence score. The other features assume the language is already known.

  119. Question 119Describe features of Natural Language Processing (NLP) workloads on Azure

    What is the primary function of Azure AI Translator?

    1. To extract entities from unstructured text
    2. To translate text between languages Correct answer.
    3. To summarize long documents automatically
    4. To convert spoken audio into written text

    Explanation

    Azure AI Translator is the dedicated translation service and converts text from a source language into one or more target languages. Transcription, summarization, and entity extraction belong to other services or features.

  120. Question 120Describe features of Natural Language Processing (NLP) workloads on Azure

    Which capability produces audible speech from a written script?

    1. Text to speech Correct answer.
    2. Language detection
    3. Speech to text
    4. Speech translation

    Explanation

    Text-to-speech synthesis converts written text into spoken audio using a chosen voice. It is the reverse of speech-to-text and is unrelated to translation or language identification.

  121. Question 121Describe features of Natural Language Processing (NLP) workloads on Azure

    A news organization wants to automatically list the people, organizations, and locations mentioned in its articles. Which Azure AI Language feature should it use?

    1. Text summarization
    2. Named entity recognition Correct answer.
    3. Key phrase extraction
    4. Sentiment analysis

    Explanation

    Named entity recognition classifies words in text into categories such as person, organization, location, and date. The other features do not produce typed entity lists.

  122. Question 122Describe features of Natural Language Processing (NLP) workloads on Azure

    Which Azure AI Language feature lets you build a knowledge base from existing FAQ pages and answer user questions in a chat?

    1. Question answering Correct answer.
    2. Custom named entity recognition
    3. Conversational language understanding
    4. Sentiment analysis

    Explanation

    Question answering ingests sources such as FAQ documents or web pages and returns the best matching answer to a user's natural language question. Intent classification is the role of conversational language understanding.

  123. Question 123Describe features of Natural Language Processing (NLP) workloads on Azure

    Which Azure AI Language feature identifies credit card numbers and phone numbers in a document so that they can be masked?

    1. Sentiment analysis
    2. PII detection Correct answer.
    3. Key phrase extraction
    4. Entity linking

    Explanation

    PII detection finds sensitive personal data such as phone numbers, card numbers, and addresses and can return masked versions of them. The other features do not target sensitive data.

  124. Question 124Describe features of Natural Language Processing (NLP) workloads on Azure

    A law firm needs to extract clause names that are unique to its own contract templates. Which approach fits best?

    1. Use sentiment analysis to score the contract text
    2. Use the prebuilt named entity recognition model on each contract
    3. Use key phrase extraction on each contract to find clauses
    4. Train a custom named entity recognition model Correct answer.

    Explanation

    Custom named entity recognition lets you label examples of your own entity types and train a model to find them in new documents. Prebuilt models cover general categories and will not recognize firm-specific clause names.

  125. Question 125Describe features of Natural Language Processing (NLP) workloads on Azure

    What does a conversational language understanding model predict from a user's utterance?

    1. The translation of the utterance
    2. The sentiment score of the user
    3. The intent and any entities contained in it Correct answer.
    4. The language in which the user typed

    Explanation

    Conversational language understanding classifies an utterance into an intent and extracts the entities it contains, which lets a bot decide what the user wants. Tone, language, and translation are handled by other features.

  126. Question 126Describe features of Natural Language Processing (NLP) workloads on Azure

    A call center wants customers to speak in their own language and have their words returned as text in another language. Which Azure AI Speech capability provides this?

    1. Speech translation Correct answer.
    2. Speech to text
    3. Speaker recognition
    4. Text to speech

    Explanation

    Speech translation combines speech recognition with translation so that spoken input is returned as text in a target language. Plain speech-to-text keeps the original language.

  127. Question 127Describe features of Natural Language Processing (NLP) workloads on Azure

    When sentiment analysis returns a document-level result, what additional information is provided for the document?

    1. A readability grade level computed for the whole document
    2. A full translation of the document into English
    3. Confidence scores for positive, neutral, and negative Correct answer.
    4. A list of the key phrases found on each page of the file

    Explanation

    Sentiment results include the predicted label and confidence scores for each possible label, at both document and sentence level. Translation, key phrase listing, and readability grading are not part of the sentiment response.

  128. Question 128Describe features of Natural Language Processing (NLP) workloads on Azure

    Which statement describes abstractive summarization?

    1. It joins the first sentence of every paragraph
    2. It extracts key phrases and lists them as bullet points
    3. It returns the highest-scoring sentences copied from the source
    4. It generates new sentences that paraphrase the source Correct answer.

    Explanation

    Abstractive summarization produces new wording that captures the meaning of the source rather than reusing its sentences. Approaches that reuse sentences from the source are extractive.

  129. Question 129Describe features of Natural Language Processing (NLP) workloads on Azure

    Which feature resolves the word "Paris" in a text to a specific known identity, such as the French city rather than a person's name?

    1. Named entity recognition
    2. Key phrase extraction
    3. Entity linking Correct answer.
    4. Sentiment analysis

    Explanation

    Entity linking disambiguates a mention and associates it with a specific known entity, typically with a reference entry. Named entity recognition only tags the mention as a type such as location.

  130. Question 130Describe features of Natural Language Processing (NLP) workloads on Azure

    Your company's domain terminology is consistently mistranslated by the standard translation model. What should you use?

    1. A custom question answering project
    2. Custom Translator Correct answer.
    3. A custom named entity recognition model
    4. Speech translation with a custom voice

    Explanation

    Custom Translator lets you train a translation model on your own parallel documents so that domain vocabulary is translated correctly. The other options address different workloads entirely.

  131. Question 131Describe features of Natural Language Processing (NLP) workloads on Azure

    A hospital wants to extract medical conditions, medications, and dosages from clinical notes. Which Azure AI Language feature is designed for this?

    1. Language detection
    2. Sentiment analysis
    3. Text Analytics for health Correct answer.
    4. Key phrase extraction

    Explanation

    Text Analytics for health is a specialized Azure AI Language feature that extracts clinical entities such as conditions, medications, and dosages from medical text. General-purpose features are not tuned for clinical terminology.

  132. Question 132Describe features of Natural Language Processing (NLP) workloads on Azure

    What distinguishes question answering from conversational language understanding?

    1. Question answering returns an answer from a knowledge base instead of an intent Correct answer.
    2. Question answering returns the intent that best matches the request
    3. Question answering translates the question before answering it
    4. Question answering trains only on labeled utterances supplied by users

    Explanation

    Question answering retrieves the best matching answer from curated sources such as FAQs or documents. Conversational language understanding instead classifies utterances into intents and extracts entities.

  133. Question 133Describe features of Natural Language Processing (NLP) workloads on Azure

    You have 1,000 hours of archived call recordings that must be transcribed overnight. Which approach is most appropriate?

    1. Real-time speech to text through a microphone
    2. Text to speech using a neural voice
    3. Pronunciation assessment on each call
    4. Batch transcription with the Azure AI Speech API Correct answer.

    Explanation

    Batch transcription is designed for large volumes of stored audio that are processed asynchronously rather than in real time. Microphone-based recognition and pronunciation scoring are intended for interactive scenarios.

  134. Question 134Describe features of Natural Language Processing (NLP) workloads on Azure

    A language-learning app must score how closely a learner's spoken sentence matches a reference text. Which Azure AI Speech feature should it use?

    1. Text to speech
    2. Pronunciation assessment Correct answer.
    3. Speech translation
    4. Speaker recognition

    Explanation

    Pronunciation assessment compares spoken audio with reference text and returns accuracy, fluency, and completeness scores. The other speech features do not evaluate pronunciation quality.

  135. Question 135Describe features of Natural Language Processing (NLP) workloads on Azure

    An organization wants to transcribe customer calls, detect the sentiment of the transcripts, and translate the transcripts to English. Which combination of Azure services meets the requirement?

    1. Azure AI Speech, Azure AI Document Intelligence, Azure AI Vision
    2. Azure AI Language, Azure AI Translator, Azure AI Vision
    3. Azure AI Speech, Azure AI Language, Azure AI Translator Correct answer.
    4. Azure AI Translator, Azure AI Vision, Azure AI Document Intelligence

    Explanation

    Azure AI Speech provides transcription, Azure AI Language provides sentiment analysis over the resulting text, and Azure AI Translator performs the translation. Vision and Document Intelligence are not needed for audio transcription or text sentiment.

  136. Question 136Describe features of Natural Language Processing (NLP) workloads on Azure

    A compliance team requires every summary to quote sentences exactly as they appear in the source document. Which summarization approach should you choose?

    1. Sentiment analysis, because it ranks the sentences
    2. Abstractive summarization, because it rewrites source text in new words
    3. Extractive summarization, because it selects existing sentences Correct answer.
    4. Key phrase extraction, because it keeps the original words

    Explanation

    Extractive summarization builds the summary from sentences that already exist in the source, so quotations remain verbatim. Abstractive methods generate new phrasing that would not match the original wording.

  137. Question 137Describe features of Natural Language Processing (NLP) workloads on Azure

    You plan to use both conversational language understanding and question answering in one application. What is the simplest way to provision resources?

    1. Use one Azure AI Language resource for both features Correct answer.
    2. Create a separate Azure AI Vision resource for each feature
    3. Create an Azure Machine Learning workspace and register both models
    4. Create an Azure AI Speech resource and enable both features

    Explanation

    A single Azure AI Language resource exposes multiple language features, including conversational language understanding and question answering, so one resource covers both. Vision and Speech resources do not provide these text capabilities.

  138. Question 138Describe features of Natural Language Processing (NLP) workloads on Azure

    You are training a conversational language understanding model, and users often say things unrelated to any of your intents. What should you do to improve the model?

    1. Label those utterances with the None intent Correct answer.
    2. Delete the utterances from the training set entirely
    3. Add those utterances to the intent that is closest in meaning
    4. Convert them to entities in a custom named entity recognition project

    Explanation

    The None intent captures utterances that do not belong to any defined intent, which teaches the model to recognize out-of-scope requests and avoid false matches. Assigning them to a real intent would blur the boundary between intents.

  139. Question 139Describe features of Natural Language Processing (NLP) workloads on Azure

    You need to control how a text-to-speech voice pronounces a product name and where it pauses. What should you use?

    1. The key phrase extraction feature
    2. A custom named entity recognition model
    3. Speech Synthesis Markup Language (SSML) Correct answer.
    4. Batch transcription with a custom speech model

    Explanation

    SSML is the markup used with text-to-speech to adjust pronunciation, pauses, emphasis, and voice selection. The other options address text analysis or transcription and cannot influence synthesized audio.

  140. Question 140Describe features of Natural Language Processing (NLP) workloads on Azure

    What is the purpose of tokenization in a natural language processing pipeline?

    1. Converting recorded speech into written text
    2. Breaking text into smaller units such as words Correct answer.
    3. Scoring how positive or negative a sentence is
    4. Reducing every word to its dictionary base form

    Explanation

    Tokenization divides a piece of text into tokens, typically words, punctuation marks, or subwords, so that later processing steps can operate on discrete units. It is normally the first step in a text analysis pipeline. The other listed tasks belong to different NLP capabilities, such as speech recognition, normalization, or sentiment analysis.

  141. Question 141Describe features of Natural Language Processing (NLP) workloads on Azure

    A team wants a quick list of the main talking points in customer feedback, without assigning any categories. Which Azure AI Language capability should they use?

    1. Language detection
    2. Key phrase extraction Correct answer.
    3. Named entity recognition
    4. Sentiment analysis

    Explanation

    Key phrase extraction identifies the salient terms and concepts in a document, which is exactly a list of main talking points. It requires no category labels or custom training. Entity recognition returns people, places, and similar items rather than general topics.

  142. Question 142Describe features of Natural Language Processing (NLP) workloads on Azure

    Which information does Azure AI Language language detection return for an input document?

    1. A list of people, places, and organizations found in the text
    2. The text translated into every supported target language, with scores
    3. The sentiment label together with confidence scores for each sentence
    4. The language name, its ISO code, and a confidence score Correct answer.

    Explanation

    Language detection returns the name of the detected language, its ISO 639-1 code, and a confidence score for the prediction. It does not translate or interpret the content. The result is typically used to route text to a correctly configured analyzer or translator.

  143. Question 143Describe features of Natural Language Processing (NLP) workloads on Azure

    Why might an NLP pipeline remove stop words such as "the" and "and" before analysis?

    1. Because these words are the most likely to be named entities
    2. Because they are the only words that carry sentiment polarity
    3. Because they must be translated before any other word is processed
    4. Because they are very common and add little analytical value Correct answer.

    Explanation

    Stop words occur so frequently that they carry little information for tasks such as frequency analysis or key phrase extraction, so removing them reduces noise and data volume. They are ordinary function words rather than content words. The decision is a design choice and depends on the task.

  144. Question 144Describe features of Natural Language Processing (NLP) workloads on Azure

    How does lemmatization differ from stemming?

    1. It removes stop words before analysis, which stemming performs afterwards
    2. It translates each word into English, whereas stemming keeps the original language
    3. It returns a valid dictionary word; stemming only trims word endings Correct answer.
    4. It assigns parts of speech to entities, while stemming labels their types

    Explanation

    Lemmatization considers vocabulary and grammatical context to produce a meaningful base form, so the result is always a real word. Stemming applies crude suffix-trimming rules and can produce non-words. Both are normalization techniques applied to tokens.

  145. Question 145Describe features of Natural Language Processing (NLP) workloads on Azure

    A term appears often in one document but rarely anywhere else in a collection. What does TF-IDF indicate about that term?

    1. It is probably a misspelled word that appears across the collection
    2. It is highly indicative of that document's content Correct answer.
    3. It has a low weight because very few documents mention it
    4. It is likely a common stop word that should be removed first

    Explanation

    TF-IDF multiplies term frequency within a document by inverse document frequency across the collection. A term frequent in one document but rare elsewhere gets a high weight, marking it as characteristic of that document. A term appearing in many documents is down-weighted instead.

  146. Question 146Describe features of Natural Language Processing (NLP) workloads on Azure

    A hotel chain wants sentiment for each specific aspect mentioned in reviews, such as room cleanliness and staff friendliness. Which Azure AI Language capability fits?

    1. Language detection combined with entity linking results
    2. Abstractive summarization of every submitted review
    3. Opinion mining, or aspect-based sentiment analysis Correct answer.
    4. Key phrase extraction followed by a custom text classification model

    Explanation

    Opinion mining, also called aspect-based sentiment analysis, associates opinions with the targets or aspects they refer to, so each aspect receives its own sentiment. Plain sentiment analysis returns only an overall document or sentence score. This makes opinion mining the right choice for per-aspect feedback.

  147. Question 147Describe features of Natural Language Processing (NLP) workloads on Azure

    Which Azure AI Translator feature converts the letters of a word into a different script without translating its meaning?

    1. Transliteration Correct answer.
    2. Dictionary lookup of alternative translations
    3. Document translation
    4. Custom translator with a phrase dictionary

    Explanation

    Transliteration maps the characters of a source word into the script of a target language, preserving pronunciation rather than meaning. It is useful for names and terms that should sound the same in another alphabet. Translation, by contrast, changes the meaning into the target language.

  148. Question 148Describe features of Natural Language Processing (NLP) workloads on Azure

    What data do you provide to train a conversational language understanding model to recognize intents and entities?

    1. Audio recordings paired with human-written transcripts and timestamps
    2. Example utterances labeled with intents and entities Correct answer.
    3. A glossary of domain terms used to customize the translation model
    4. Knowledge base articles paired with alternate questions and answers

    Explanation

    A conversational language understanding project is trained on example utterances; each utterance is assigned an intent and has the relevant words labeled as entities. The model learns from these labeled examples to classify new utterances. Knowledge bases, audio transcripts, and glossaries belong to other services.

  149. Question 149Describe features of Natural Language Processing (NLP) workloads on Azure

    A speech-to-text model repeatedly misrecognizes rare product names, treating them as unrelated everyday words. Which customization addresses this?

    1. Train a custom neural voice from recordings of each product name
    2. Apply speech synthesis markup language to control pronunciation
    3. Enable pronunciation assessment scoring for the audio stream
    4. Use custom speech with a phrase list of the product names Correct answer.

    Explanation

    A phrase list supplied to a custom speech model biases the recognizer toward specific terms, so rare product names are transcribed as written rather than as similar common words. This is a language-model adaptation of speech-to-text. The other options concern speech output or evaluation rather than recognition accuracy.

  150. Question 150Describe features of Natural Language Processing (NLP) workloads on Azure

    A conversational language understanding model returns the top intent with a confidence score of 0.35, with the next intent at 0.31. How should the application treat this result?

    1. The two candidate intents should be merged into one intent
    2. A score of 0.35 exceeds the default acceptance threshold of 0.3
    3. Low confidence; fall back or ask the user to rephrase Correct answer.
    4. The request failed and should be retried with the same utterance

    Explanation

    A low top score with a close runner-up means the model cannot distinguish the intents reliably, so the application should not act on the guess. Typical designs set a threshold and route weak predictions to a fallback or a clarification prompt. Merging intents would not resolve the underlying ambiguity in the training data.

  151. Question 151Describe features of Natural Language Processing (NLP) workloads on Azure

    When you submit text to Azure AI Translator without specifying a source language, what happens?

    1. The service automatically detects the source language first Correct answer.
    2. The service assumes the text is English and translates from English
    3. The request fails with an error until a source language code is supplied
    4. The service returns the original text unchanged with a warning

    Explanation

    When the source language is omitted, Translator performs language detection on the input and then translates from the detected language. Specifying the source explicitly is optional and can improve reliability for short or ambiguous text. Translation still returns a target-language result either way.

  152. Question 152Describe features of Natural Language Processing (NLP) workloads on Azure

    A bank wants every incoming email labeled as one of its own five internal categories. No prebuilt model matches those categories. Which Azure AI Language capability should be used?

    1. The prebuilt key phrase extraction capability
    2. Prebuilt named entity recognition over the message body
    3. A conversational language understanding project with one intent
    4. A custom text classification model Correct answer.

    Explanation

    Custom text classification lets you train a model on your own labeled examples so documents are assigned to your own categories, including multiple labels per document if desired. Prebuilt models only cover fixed, general-purpose tasks. The bank supplies labeled emails for each category during training.

  153. Question 153Describe features of generative AI workloads on Azure

    Which Azure service provides REST API access to models such as GPT-4, DALL-E, and Whisper?

    1. Azure AI Vision
    2. Azure AI Speech
    3. Azure OpenAI Service Correct answer.
    4. Azure Machine Learning

    Explanation

    Azure OpenAI Service exposes OpenAI generative models through REST APIs and SDKs, covering text, image, and speech scenarios. It is the entry point for consuming GPT, DALL-E, and Whisper models in Azure. The other services address narrower vision, speech, or model-lifecycle needs.

  154. Question 154Describe features of generative AI workloads on Azure

    Which model available in Azure OpenAI generates images from a natural-language description?

    1. GPT-4
    2. Whisper
    3. A text embedding model
    4. DALL-E Correct answer.

    Explanation

    DALL-E is the image-generation model in Azure OpenAI and produces images from text prompts. GPT-4 works with text and image understanding, Whisper handles speech, and embedding models produce vectors. Image creation from a description is specifically a DALL-E capability.

  155. Question 155Describe features of generative AI workloads on Azure

    Which Azure OpenAI model is designed to transcribe spoken audio into text?

    1. DALL-E
    2. Ada embeddings
    3. GPT-3.5 Turbo
    4. Whisper Correct answer.

    Explanation

    Whisper is the speech model offered through Azure OpenAI and supports transcription and translation of audio. GPT-3.5 Turbo handles text completions, DALL-E creates images, and Ada embeddings produce vectors. Speech-to-text conversion is the specific role of Whisper.

  156. Question 156Describe features of generative AI workloads on Azure

    In a generative AI workload, what is a token?

    1. A piece of text, such as a word or part of a word, processed by a model Correct answer.
    2. A unit of compute allocated to a model deployment
    3. A security key used to authenticate API calls
    4. A billing region for an Azure OpenAI resource

    Explanation

    Models break input and output text into tokens, which may be whole words, word fragments, or punctuation. Usage and cost are measured in tokens, and context limits are expressed in tokens. Tokens are therefore a unit of text processing rather than a credential, compute allocation, or region.

  157. Question 157Describe features of generative AI workloads on Azure

    What is a completion in an Azure OpenAI request?

    1. The response that a model returns for a given prompt Correct answer.
    2. The verdict produced by the content filter
    3. A status reported by a fine-tuning job
    4. The process of deploying a model to an endpoint

    Explanation

    A completion is the model-generated output that follows the prompt you submit. The term comes from the model completing the text based on the input. Deployment, fine-tuning status, and content filtering are separate operations.

  158. Question 158Describe features of generative AI workloads on Azure

    Azure OpenAI content filters evaluate prompts and completions for which set of harmful content categories?

    1. Grammar, style, and tone
    2. Hate, sexual, violence, and self-harm Correct answer.
    3. Sentiment, language, and key phrases
    4. Currency, date, and location

    Explanation

    The default Azure OpenAI content filters assess four harm categories: hate, sexual, violence, and self-harm, each with a severity level. These filters run on both input prompts and output completions. The other listed sets belong to text analytics or writing assistance rather than safety filtering.

  159. Question 159Describe features of generative AI workloads on Azure

    Before an application can call a model through the Azure OpenAI REST API, what must exist in the Azure OpenAI resource?

    1. A model deployment Correct answer.
    2. A virtual network gateway
    3. A storage account
    4. A logic app

    Explanation

    Calls are made against a deployment, which gives a deployed model a name and endpoint used in API requests. Creating the resource alone is not sufficient; the model must be deployed. Networking, storage, and workflow resources are optional and unrelated to model invocation.

  160. Question 160Describe features of generative AI workloads on Azure

    Which responsible AI principle emphasizes that people should understand an AI system's capabilities and limitations?

    1. Transparency Correct answer.
    2. Privacy and security
    3. Fairness
    4. Reliability and safety

    Explanation

    Transparency means informing users about how an AI system works, what it can and cannot do, and its limitations. Fairness concerns equitable treatment, reliability concerns consistent safe operation, and privacy concerns protection of data. Understanding capabilities and limits is the defining focus of transparency.

  161. Question 161Describe features of generative AI workloads on Azure

    A team needs highly consistent, repeatable answers for a text classification task. Which parameter setting best supports that goal?

    1. Temperature set to 1.5
    2. Temperature set close to 0 Correct answer.
    3. Top_p set to 1.0
    4. Max_tokens set to 4000

    Explanation

    Temperature controls randomness in token selection; values near 0 make the model favor the most likely tokens, producing stable outputs. Higher temperatures widen sampling and increase variation. Top_p and max_tokens do not by themselves make responses more deterministic.

  162. Question 162Describe features of generative AI workloads on Azure

    Which message role is used to define a model's persona and behavioral instructions across a conversation?

    1. tool
    2. system Correct answer.
    3. user
    4. assistant

    Explanation

    The system message sets overarching instructions such as tone, constraints, and role for the model. User messages carry the person's input, and assistant messages hold prior model replies. Setting persistent behavior is the purpose of the system role.

  163. Question 163Describe features of generative AI workloads on Azure

    A company wants a chat solution to answer questions using its own internal documents, which change frequently. Which approach should be used?

    1. Generate illustrated answers with DALL-E
    2. Increase the temperature of the model
    3. Fine-tune the model on public web data
    4. Retrieve relevant passages from an Azure AI Search index and include them in the prompt Correct answer.

    Explanation

    Retrieval-augmented generation grounds responses in current documents by adding retrieved passages to the prompt. This lets answers reflect content that changes without retraining the model. Temperature and public-data fine-tuning do not supply the internal knowledge, and image generation does not provide factual grounding.

  164. Question 164Describe features of generative AI workloads on Azure

    What is the primary purpose of an embeddings model in Azure OpenAI?

    1. To summarize long documents into bullet points
    2. To translate speech into another language
    3. To convert text into numeric vectors that capture semantic meaning Correct answer.
    4. To generate images from text prompts

    Explanation

    Embeddings map text to vectors positioned so that semantically similar content is close together. They enable similarity search, clustering, and retrieval for retrieval-augmented generation. They do not themselves translate, draw, or summarize.

  165. Question 165Describe features of generative AI workloads on Azure

    A developer includes three example input-and-output pairs in a prompt to show the model the desired pattern. What is this technique called?

    1. Few-shot prompting Correct answer.
    2. Fine-tuning
    3. Retrieval augmentation
    4. Zero-shot prompting

    Explanation

    Few-shot prompting supplies a small number of examples within the prompt so the model can infer the desired format or behavior. Zero-shot prompting gives instructions without examples. Fine-tuning changes model weights through training, and retrieval augmentation adds external reference content.

  166. Question 166Describe features of generative AI workloads on Azure

    Which request parameter caps the length of the text that the model generates in its response?

    1. max_tokens Correct answer.
    2. temperature
    3. frequency_penalty
    4. top_p

    Explanation

    Max_tokens sets an upper bound on how many tokens the model may produce in the completion. Temperature and top_p influence randomness, and frequency_penalty discourages repetition. Only max_tokens directly limits response length.

  167. Question 167Describe features of generative AI workloads on Azure

    For which scenario is fine-tuning a base model the most appropriate choice?

    1. Reducing the network latency of API calls
    2. Teaching a consistent response style or format using labeled examples Correct answer.
    3. Creating illustrations for a product catalog
    4. Adding newly published facts that change every day

    Explanation

    Fine-tuning adapts a model's behavior using labeled examples so it follows a desired style, tone, or output structure consistently. It is not a good way to keep up with fast-changing facts, which is better handled with retrieval. Latency tuning and image creation are unrelated to fine-tuning.

  168. Question 168Describe features of generative AI workloads on Azure

    Which Azure AI Content Safety capability detects attempts to manipulate a generative model with jailbreak or prompt injection attacks?

    1. Prompt shields Correct answer.
    2. Personally identifiable information detection
    3. Text summarization
    4. Groundedness detection

    Explanation

    Prompt shields identify adversarial input designed to bypass a model's safeguards, such as jailbreak or prompt injection attempts. Groundedness detection checks whether responses are supported by source material, and PII detection finds personal data. Blocking manipulative prompts is the specific role of prompt shields.

  169. Question 169Describe features of generative AI workloads on Azure

    What does the model catalog in Azure AI Foundry provide?

    1. A registry of virtual machine images for training clusters
    2. A gallery of prebuilt Power BI reports
    3. A list of compliance certifications for Azure regions
    4. A collection of models from multiple providers that can be deployed for AI applications Correct answer.

    Explanation

    The model catalog offers a range of models, including OpenAI and third-party options, that can be explored, compared, and deployed. It is a starting point for selecting a model for a generative AI application. It is not a compute-image registry, certification list, or reporting gallery.

  170. Question 170Describe features of generative AI workloads on Azure

    Which Azure OpenAI deployment option reserves dedicated model-processing capacity measured in provisioned throughput units?

    1. Global standard
    2. Provisioned Correct answer.
    3. Standard
    4. Batch

    Explanation

    Provisioned deployments allocate dedicated capacity measured in provisioned throughput units, giving predictable performance for steady workloads. Standard and global standard deployments use pay-as-you-go token billing, and batch is intended for asynchronous bulk processing. Reserved capacity is the defining trait of provisioned deployments.

  171. Question 171Describe features of generative AI workloads on Azure

    What is a hallucination in a generative AI workload?

    1. Content that is fluent but factually incorrect or unsupported Correct answer.
    2. A failure caused by exceeding the token quota
    3. A delay in model response during peak demand
    4. An error raised when the content filter blocks a prompt

    Explanation

    Hallucination describes output that reads plausibly but contains fabricated or unsupported claims. It occurs because models predict likely text rather than verifying facts. Quota errors, latency, and filter blocks are operational conditions, not hallucinations.

  172. Question 172Describe features of generative AI workloads on Azure

    What is the ability of a model to accept both text and images in the same prompt called?

    1. Tokenization
    2. Multimodality Correct answer.
    3. Fine-tuning
    4. Orchestration

    Explanation

    A multimodal model can process more than one type of input, such as combined text and images. Tokenization is how text is split for processing, fine-tuning adapts a model with training data, and orchestration coordinates workflow steps. Handling multiple input modalities is multimodality.

  173. Question 173Describe features of generative AI workloads on Azure

    A legal team needs answers that reflect internal policy documents updated weekly, and the answers must cite the source passages. Which solution best meets these requirements?

    1. Retrain the base model from scratch each week
    2. Raise the temperature so the model explores more options
    3. Use an image-generation model to render the policies
    4. Ground the model with a search index that is refreshed from the document repository Correct answer.

    Explanation

    Refreshing a search index and retrieving passages at request time keeps answers current and allows citations to the retrieved sources. This avoids the cost and delay of retraining and gives traceability. Temperature changes and image generation do not deliver current, citable policy text.

  174. Question 174Describe features of generative AI workloads on Azure

    A developer observes that a submitted prompt is rejected before any model output is produced. Which component most likely caused the rejection?

    1. The max_tokens setting
    2. The deployment quota
    3. The temperature value
    4. The content filter applied to the prompt Correct answer.

    Explanation

    Content filters evaluate the input prompt before generation, so a prompt classified as harmful can be blocked before the model runs. Max_tokens and temperature affect output behavior rather than pre-screening. Quota limits produce capacity errors, not content rejection.

  175. Question 175Describe features of generative AI workloads on Azure

    Which evaluation metric measures how well a model's response is supported by the source documents supplied in the prompt?

    1. Fluency
    2. Coherence
    3. Relevance
    4. Groundedness Correct answer.

    Explanation

    Groundedness assesses whether the generated response is backed by the provided source material rather than invented. Fluency measures language quality, coherence measures logical flow, and relevance measures how well the response addresses the question. Support by sources is specifically groundedness.

  176. Question 176Describe features of generative AI workloads on Azure

    A solution sends entire long documents with every request, raising token usage and cost. Which change reduces token consumption while preserving answer quality?

    1. Send each document twice for emphasis
    2. Increase max_tokens for each request
    3. Set temperature to the maximum value
    4. Retrieve only the most relevant document chunks and include those in the prompt Correct answer.

    Explanation

    Chunking documents and retrieving only the passages relevant to the question shrinks the prompt while keeping the needed context. This is the core efficiency benefit of retrieval-augmented generation. Raising output limits, maximizing temperature, or duplicating text increases cost or reduces focus.

  177. Question 177Describe features of generative AI workloads on Azure

    Which Azure AI Foundry capability lets developers build, test, and deploy multi-step LLM workflows as a connected graph?

    1. Content safety
    2. Model catalog
    3. Provisioned throughput
    4. Prompt flow Correct answer.

    Explanation

    Prompt flow provides a visual and code-based way to author flows that chain prompts, models, and tools into an executable sequence. It supports testing and deployment of those multi-step workflows. The model catalog lists models, content safety filters content, and provisioned throughput controls capacity.

  178. Question 178Describe features of generative AI workloads on Azure

    A marketing team wants to create original illustrations from short written descriptions by using Azure OpenAI. Which model type should the team use?

    1. An image-generation model such as DALL-E Correct answer.
    2. A speech-to-text model such as Whisper
    3. A chat model such as GPT-4o
    4. An embeddings model such as text-embedding-3-large

    Explanation

    Image-generation models in Azure OpenAI, such as DALL-E, accept a text description and produce a new image. Chat models produce text, embeddings models produce numeric vectors, and Whisper works with audio.

  179. Question 179Describe features of generative AI workloads on Azure

    In a large language model, what is a token?

    1. A credential that authorizes an API request
    2. A chunk of text, such as part of a word, processed by the model Correct answer.
    3. A record of billable usage that the service creates for each deployment
    4. A single message exchanged in a chat session

    Explanation

    Models split text into tokens, which are words, word fragments, or characters, and both processing limits and billing are measured in these units. Tokens are not credentials, messages, or billing records.

  180. Question 180Describe features of generative AI workloads on Azure

    What is a prompt in a generative AI workload?

    1. A cached response that the service reuses
    2. The name of the deployment that routes a request
    3. A safety filter applied to model output
    4. The instruction or question sent to a model Correct answer.

    Explanation

    The prompt is the input the user supplies, which can include instructions, context, examples, or a question, and the model responds with a completion. Deployment names and filters are configuration elements, not the input itself.

  181. Question 181Describe features of generative AI workloads on Azure

    Which of the following is a Microsoft generative AI assistant embedded in productivity apps such as Word, Excel, and Outlook?

    1. Azure Machine Learning
    2. Azure AI Content Safety
    3. Microsoft 365 Copilot Correct answer.
    4. Azure AI Search

    Explanation

    Microsoft 365 Copilot brings generative AI assistance into Microsoft 365 applications to draft, summarize, and analyze content. The other services provide safety filtering, search, or model training infrastructure rather than an in-app assistant.

  182. Question 182Describe features of generative AI workloads on Azure

    When an Azure OpenAI content filter flags text, it also assigns a severity level. What does the severity level describe?

    1. The model version that generated the text
    2. The number of tokens the request consumed
    3. How strongly the text matches a harmful category Correct answer.
    4. The language in which the text was written

    Explanation

    Azure OpenAI content filters score each harmful category on a severity scale, commonly safe, low, medium, and high. Applications can use that score to decide whether content should be blocked.

  183. Question 183Describe features of generative AI workloads on Azure

    Which responsible AI principle is most directly concerned with avoiding biased outcomes that disadvantage particular groups of people?

    1. Transparency
    2. Fairness Correct answer.
    3. Reliability and safety
    4. Accountability

    Explanation

    Fairness focuses on treating people equitably so that an AI system does not produce systematically worse outcomes for some groups. Accountability, transparency, and reliability address responsibility, understanding, and consistent safe operation instead.

  184. Question 184Describe features of generative AI workloads on Azure

    A chat application must begin displaying a model's reply before the full answer has been generated so that users see progress quickly. Which feature should the developer enable?

    1. Set the n parameter to request several completions at once
    2. Enable content filtering annotations
    3. Submit the requests through a batch job
    4. Enable streaming so tokens are returned as they are generated Correct answer.

    Explanation

    Streaming returns output incrementally as each token is produced, allowing the interface to render text progressively. Batch jobs and multiple completions change throughput or variety, not the time to first visible text.

  185. Question 185Describe features of generative AI workloads on Azure

    A prompt and the requested completion together exceed the maximum context length of an Azure OpenAI model. What is the typical outcome?

    1. The model automatically compresses the prompt into a shorter summary
    2. The request is rejected with an error before output is produced Correct answer.
    3. The service silently drops the oldest messages and continues processing
    4. The service reroutes the request to a model with a larger context window

    Explanation

    Every model has a context window that must hold both the input and the generated output. When the combined size exceeds that limit, the call fails and the application must shorten the prompt or reduce the output allowance.

  186. Question 186Describe features of generative AI workloads on Azure

    How does Azure OpenAI treat the prompts and completions that a customer submits?

    1. They are published in the model catalog for other customers
    2. They are not used to train the models unless the customer opts in Correct answer.
    3. They are shared with the public OpenAI service to improve its models
    4. They are deleted before the model processes the request

    Explanation

    Customer prompts and completions are not used to train or improve the underlying foundation models. Content may be logged for abuse monitoring, but it stays within the customer's Azure environment rather than being shared publicly.

  187. Question 187Describe features of generative AI workloads on Azure

    What is prompt engineering?

    1. Adding throughput units to a deployment to reduce latency
    2. Writing and refining the instructions given to a model Correct answer.
    3. Encrypting prompts while they are in transit
    4. Compiling a model into a container image for deployment

    Explanation

    Prompt engineering is the practice of designing, structuring, and iterating the instructions, context, and examples in a prompt to obtain the desired response. It is independent of deployment packaging, security, and capacity settings.

  188. Question 188Describe features of generative AI workloads on Azure

    A team needs a cost-efficient model for high-volume, straightforward text summarization. Which model choice best fits this requirement?

    1. An image-generation model
    2. An embeddings model
    3. The largest multimodal model available in the catalog
    4. A small language model such as GPT-4o mini Correct answer.

    Explanation

    Smaller language models are designed for higher-volume, less complex tasks at lower cost and lower latency. Large multimodal models cost more per token, while image and embedding models serve different purposes.

  189. Question 189Describe features of generative AI workloads on Azure

    A prompt includes retrieved document text followed by the user's question. Which technique helps the model distinguish the supplied content from the instructions?

    1. Increasing the max_tokens value
    2. Wrapping each section in markers such as triple quotes or XML tags Correct answer.
    3. Sending each section to a separate deployment and merging the replies
    4. Setting the temperature to zero

    Explanation

    Clear delimiters separate instructions from referenced material, so the model is less likely to treat document text as a command. Output length and sampling settings do not clarify the structure of the prompt.

  190. Question 190Describe features of generative AI workloads on Azure

    A developer wants a model to solve multi-step arithmetic word problems more reliably by writing out intermediate reasoning steps. Which prompting technique describes this?

    1. Content filtering
    2. Retrieval augmentation
    3. Zero-shot prompting
    4. Chain-of-thought prompting Correct answer.

    Explanation

    Chain-of-thought prompting asks the model to articulate intermediate steps, which improves results on problems that require several reasoning stages. Zero-shot prompting gives no examples or reasoning guidance.

  191. Question 191Describe features of generative AI workloads on Azure

    A developer wants to compare how two models respond to the same prompt and adjust parameters interactively before writing code. Which Azure AI Foundry feature supports this?

    1. Content filters in Azure OpenAI
    2. The chat playground in Azure AI Foundry Correct answer.
    3. The model catalog in Azure AI Foundry
    4. Azure Monitor

    Explanation

    The chat playground provides an interactive environment for testing prompts and settings against deployed models so behavior can be compared before implementation. The catalog lists models, filters screen content, and monitoring reports telemetry.

  192. Question 192Describe features of generative AI workloads on Azure

    Which factor most directly determines the cost of a single Azure OpenAI text-generation request?

    1. The number of deployments in the resource
    2. The tokens processed in the prompt and the reply Correct answer.
    3. The region where the resource is created
    4. The number of chat messages retained in the session history

    Explanation

    Text-generation charges are based on tokens consumed, counting both the input sent to the model and the output it returns. Deployment count and stored conversation history do not drive per-request token charges.

  193. Question 193Describe features of generative AI workloads on Azure

    Before an AI-generated clinical recommendation is shown to a patient, a clinician reviews and approves it. Which responsible AI practice does this describe?

    1. Human oversight of AI-generated output Correct answer.
    2. Fine-tuning with labeled examples
    3. Provisioned throughput capacity
    4. Prompt shielding against jailbreak attempts

    Explanation

    Human oversight, often called human-in-the-loop, keeps a person responsible for reviewing and approving AI output in consequential decisions. Prompt shields, fine-tuning, and capacity settings serve different purposes.

  194. Question 194Describe features of generative AI workloads on Azure

    A team must generate one million product descriptions overnight with no requirement for immediate results. Which approach is most cost-effective?

    1. Fine-tuning a separate model for each request
    2. Provisioned throughput with reserved capacity
    3. Asynchronous batch processing Correct answer.
    4. Streaming the responses to the client

    Explanation

    Batch deployments accept large volumes of requests and process them asynchronously at a lower cost when latency is not important. Reserved throughput and streaming target interactive, low-latency scenarios.

  195. Question 195Describe features of generative AI workloads on Azure

    Which parameter discourages a model from repeating the same tokens by penalizing them in proportion to how often they have already appeared?

    1. The frequency penalty parameter Correct answer.
    2. The top_p parameter
    3. The max_tokens parameter
    4. The stop sequence parameter

    Explanation

    The frequency penalty reduces the likelihood of tokens based on their existing occurrence count in the generated text, which curbs verbatim repetition. max_tokens sets response length and stop sequences end generation at a marker.

  196. Question 196Describe features of generative AI workloads on Azure

    Which capability allows an assistant built on Azure OpenAI to run Python code in a sandbox to analyze a file that a user uploaded?

    1. The content filter
    2. The system message
    3. The model catalog
    4. The Code Interpreter tool Correct answer.

    Explanation

    The Code Interpreter tool executes Python in a sandboxed environment so an assistant can perform calculations, transform data, and analyze uploaded files. The catalog, filters, and system message do not execute code.

  197. Question 197Describe features of generative AI workloads on Azure

    A team wants to compare candidate models objectively against a labeled set of sample inputs before choosing one for production. Which capability should they use?

    1. Content filters with severity thresholds
    2. Provisioned throughput units reserved for the deployment
    3. The model catalog of available models
    4. The evaluation feature with a test dataset Correct answer.

    Explanation

    Evaluation runs a model or prompt flow against a dataset and reports quality metrics, so alternatives can be compared on the same inputs. Catalog browsing, filters, and capacity settings do not produce comparative quality scores.

  198. Question 198Describe features of generative AI workloads on Azure

    Which practice involves deliberately probing a generative AI system with adversarial prompts to uncover unsafe behavior before it is released?

    1. Red teaming Correct answer.
    2. Prompt caching
    3. Fine-tuning
    4. Vector indexing

    Explanation

    Red teaming is structured adversarial testing in which people attempt to make the system produce harmful, biased, or insecure output so weaknesses can be fixed before launch. The other options are optimization or adaptation techniques.

  199. Question 199Describe features of generative AI workloads on Azure

    A developer needs a model's response to always conform to a supplied JSON schema so downstream code can parse it reliably. Which approach should be used?

    1. Increasing the temperature
    2. Using structured outputs with a supplied JSON schema Correct answer.
    3. Sending the same prompt twice and comparing the two answers
    4. Enabling streaming of the response

    Explanation

    Structured outputs constrain generation to a supplied JSON schema, guaranteeing that the response matches the expected shape. Sampling changes and repeated calls improve variety or confidence but cannot guarantee schema conformance.

  200. Question 200Describe features of generative AI workloads on Azure

    A marketing team generates slogan ideas from the same prompt every morning and finds that each run returns nearly identical wording. The team wants more varied, creative phrasing across runs. Which configuration change should they make?

    1. Raise the max tokens value so that the model produces longer completions.
    2. Add a stop sequence so that the completion ends at the first line break.
    3. Lower the top-p value so that only the most probable tokens are considered.
    4. Raise the temperature value used for the request. Correct answer.

    Explanation

    Temperature is the sampling parameter that controls how much randomness is applied when the model selects the next token; a higher value flattens the probability distribution so less likely words become more reachable, producing more varied output. Lower temperatures (near zero) make responses deterministic and repeatable, which is why the same prompt keeps yielding nearly identical slogans. Parameters such as max tokens and stop sequences only bound the length of the output, not its diversity.

Azure AI-900 flashcards

32 cards. Read the term and the definition here. The deck at the top of this page is optional.

Fairness
A Responsible AI principle requiring that AI systems treat all groups of people equitably and avoid bias that favors or disadvantages particular populations.
Reliability and safety
A Responsible AI principle stating that AI systems must operate consistently, handle unexpected conditions, and avoid causing harm under normal and adverse circumstances.
Transparency
A Responsible AI principle requiring that people understand how an AI system works, what data it uses, and how it reaches its decisions.
Accountability
A Responsible AI principle meaning humans remain responsible for AI systems and their outcomes, with governance and oversight built into the design.
Inclusiveness
A Responsible AI principle requiring AI systems to empower and benefit everyone, including people with disabilities and diverse backgrounds, rather than exclude them.
Privacy and security
A Responsible AI principle requiring that data used by AI systems is protected, handled lawfully, and secured against unauthorized access throughout the system lifecycle.
Knowledge mining
An AI workload that uses cognitive skills such as OCR, entity recognition, and key phrase extraction to extract insights from large volumes of unstructured documents.
Supervised learning
A machine learning approach that trains a model on labeled data, where each input example is paired with the known correct output called a label.
Unsupervised learning
A machine learning approach that trains on unlabeled data to discover patterns or groupings, such as clusters, without any predefined correct answers.
Regression
A supervised learning task that predicts a continuous numeric value, such as a price or a temperature, based on the input features of the data.
Classification
A supervised learning task that predicts which category or class an item belongs to, such as spam versus not spam in an email dataset.
Clustering
An unsupervised learning technique that groups similar data points together based on feature similarity, without using any predefined labels.
Overfitting
A modeling problem where a model fits the training data too closely, including its noise, and therefore performs poorly on new, unseen data.
Automated machine learning
An Azure Machine Learning feature that automatically tries many algorithms and preprocessing options to find the best model for a dataset with minimal code.
Image classification
A computer vision task that assigns one or more class labels to an entire image, such as tagging a photograph as containing a dog.
Object detection
A computer vision task that identifies the location of objects in an image using bounding boxes and assigns a class label to each detected object.
Semantic segmentation
A computer vision task that classifies every pixel in an image into a category, producing a mask that shows the exact shape of each region.
Optical character recognition
A computer vision technique that reads printed or handwritten text in images and documents and converts it into machine-readable characters.
Azure AI Vision
An Azure cognitive service providing prebuilt computer vision capabilities such as image tagging, captioning, object detection, and OCR through simple API calls.
Custom Vision
An Azure service that lets you train custom image classification and object detection models using your own set of labeled images.
Sentiment analysis
A natural language processing task that evaluates text and returns a sentiment label such as positive, neutral, or negative, usually with confidence scores.
Key phrase extraction
A natural language processing task that identifies the main concepts or topics in a body of text, returning a list of significant terms and phrases.
Named entity recognition
A natural language processing task that finds and classifies entities in text, such as people, places, organizations, dates, and quantities.
Azure AI Language
An Azure service that bundles NLP capabilities including sentiment analysis, key phrase extraction, entity recognition, summarization, and question answering behind one resource.
Speech synthesis
The text-to-speech capability of Azure AI Speech that converts written text into spoken audio using configurable synthetic voices.
Question answering
An Azure AI Language capability that builds a knowledge base of question-and-answer pairs and returns the best answer to a user's natural language question.
Large language model
A generative AI model trained on vast amounts of text to predict the next token, enabling it to write, summarize, translate, and answer questions.
Prompt
The input text or instruction given to a generative AI model that describes the task and guides the response the model produces.
Retrieval augmented generation
A technique that grounds a generative model's response by retrieving relevant data from an external source and including that data in the prompt.
Hallucination
A generative AI failure where the model produces plausible-sounding but factually incorrect or fabricated information that is not supported by source data.
Azure OpenAI Service
An Azure service that provides access to OpenAI generative models such as GPT, with Azure security, compliance, networking, and content filtering built in.
Content filtering
A safety feature in Azure OpenAI that detects and blocks harmful content categories such as hate, violence, sexual content, and self-harm in prompts and responses.

Azure AI-900 study guide

Describe Artificial Intelligence Workloads and Considerations

The workload families

AI-900 expects you to match a scenario to a workload family: machine learning (predicting a value or category from data), computer vision (interpreting images and video), natural language processing (understanding, translating and producing text and speech), document intelligence (pulling structured fields out of forms, invoices and receipts), knowledge mining (building a searchable index over unstructured content), and generative AI (creating new text, images or code).

The six responsible AI principles

  • Fairness - similar outcomes for people in similar situations, regardless of group.
  • Reliability and safety - the system behaves predictably and fails safely.
  • Privacy and security - personal data is protected and used with consent.
  • Inclusiveness - everyone can use the system, including people with disabilities.
  • Transparency - people understand how the system works and what data it uses.
  • Accountability - humans remain responsible for the system's decisions and can be asked to answer for them.

Traps that show up repeatedly

Transparency versus accountability is the classic swap: explaining a model is transparency; signing off on its output is accountability. Fairness versus inclusiveness is the second: fairness is about comparable performance across groups, inclusiveness is about accessibility. Also remember that Microsoft's guidance applies to the application you build, not only to the Azure service.

Worked example

A bank's loan model rejects a higher proportion of applicants in one age band: that is a fairness issue. Leadership asks how the model scores applicants: that is transparency. Requiring a human officer to review every rejection is the accountability control.

Rule of thumb: if the question is about who is answerable, answer accountability; if it is about who can see how it works, answer transparency.

Describe Fundamental Principles of Machine Learning on Azure

Learning types

Supervised learning trains on labelled data - each row has features (inputs) and a label (the answer). Two flavours: regression predicts a numeric value (price, units, temperature); classification predicts a category (yes/no, or one of several classes). Unsupervised learning has no label; clustering groups similar rows. Reinforcement learning learns by reward and penalty through trial and error. A quick grep of the scenario is usually enough: continuous answer = regression, discrete answer = classification, no answer column = clustering.

The Azure Machine Learning toolset

An Azure Machine Learning workspace organises experiments, data, compute and models. Automated ML sweeps algorithms and settings for you and exposes a leaderboard of models with their metrics. The designer gives a drag-and-drop pipeline canvas. Pipelines automate repeatable steps; compute clusters scale training; a model registry versions artefacts; an endpoint serves a deployed model.

Evaluation and traps

Split data into training, validation and test sets - never evaluate on data the model trained on, which is overfitting plus data leakage. Regression metrics: MAE, RMSE, R² (higher R² is better). Classification metrics: accuracy, precision, recall, F1, AUC. Classic traps: choosing classification for a numeric target, treating clustering as supervised, and assuming accuracy is a good metric on imbalanced data (prefer precision/recall/AUC).

Worked example

Predicting "how many bikes will be rented tomorrow" needs a numeric label, so it is regression - even though the count is a whole number. Predicting "will this customer churn" is binary classification. Grouping customers with no churn label is clustering.

Describe Features of Computer Vision Workloads on Azure

The four analysis tasks

  • Image classification - one or more labels describing the whole image ("dog").
  • Object detection - a class label plus a bounding box for each instance, and more than one object can be returned.
  • Semantic segmentation - a pixel-level mask per class, so the exact outline is known; instance segmentation separates individual objects of the same class.
  • OCR - reading printed or handwritten text, with positions, from images and PDFs.

Azure services to know

Azure AI Vision provides prebuilt capabilities: image tagging, captioning, OCR, spatial analysis and face detection. Custom Vision lets you train your own image classification or object detection models from labelled images. Azure AI Face provides detection (where a face is) and, under a restricted-access review, verification and identification - but face detection alone does not disclose identity, emotion or age with certainty. Azure AI Document Intelligence extracts fields from invoices, receipts, IDs and custom forms. Azure AI Video Indexer analyses video and audio.

Traps that appear in questions

"How many objects and where" always means object detection, not classification. Custom Vision offers classification and object detection only - not segmentation. Captioning gives a sentence; tagging gives keywords. Predicting a single attribute of a person from an image is a limited-access or restricted scenario, so the compliant answer is usually detection, not inference.

Worked example

A warehouse wants a camera to flag each empty pallet slot on a shelf and draw a rectangle around it. One image contains many slots, so the workload is object detection trained with Custom Vision; classification would return a single label for the whole photo and could not count or locate the slots.

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

Core Azure AI Language capabilities

  • Language detection - which language is this text?
  • Sentiment analysis and opinion mining - positive/neutral/negative, and at aspect level ("the food was great but the service was slow").
  • Key phrase extraction - the main concepts in a document.
  • Named entity recognition (NER) - people, places, organisations, dates, quantities.
  • Entity linking - resolving mentions to a known knowledge-base entry.
  • PII detection - finding and redacting personal data.
  • Summarisation, question answering, and conversational language understanding (CLU) - CLU maps an utterance to an intent and extracts entities.

Speech and translation

Azure AI Speech covers speech-to-text, text-to-speech and speech translation; custom neural voice is limited access. Azure AI Translator handles text translation and can be combined with speech for real-time interpretation.

Traps that appear in questions

"What is this document about" is key phrase extraction, not entity recognition. Extracting a person's name or a date is NER. Routing a chatbot request such as "book a flight" is intent classification in CLU, and "to Paris" is an entity. Matching a user's free-text question to an existing FAQ is question answering. Detecting customer frustration in a support ticket is sentiment analysis. Also, sentiment applies to text, not to images, and PII detection is about redaction, not sentiment.

Worked example

A hotel wants to know whether guests praise the room but complain about breakfast. Whole-document sentiment would average this away, so the correct feature is opinion mining within sentiment analysis, which reports sentiment per aspect. If the goal were simply "list the topics mentioned", the answer would be key phrase extraction instead.

Describe Features of Generative AI Workloads on Azure

How generative AI works

Large language models (LLMs) are trained on huge text corpora to predict the next token. At inference you send a prompt and receive a completion. Prompts have roles: a system message sets behaviour and guardrails, user messages are the request, assistant messages are prior model turns. Parameters matter: temperature controls randomness - higher is more creative and varied, lower is more deterministic.

Azure offerings

Azure OpenAI hosts Microsoft-managed models (GPT text and code models, image generation, embeddings) with Azure security, networking and content filters. Azure AI Foundry / model catalog adds third-party and open models. Embeddings convert text to vectors that power semantic search. A copilot is an application pattern that embeds a model into a workflow, not a model type.

Grounding: RAG versus fine-tuning

Retrieval-augmented generation (RAG) retrieves relevant documents at query time and puts them in the prompt, so answers are grounded in your current data - the model's weights never change. Fine-tuning trains a base model further on your examples to shape style, tone or format. Fine-tuning is a poor way to inject facts that change often, and it cannot guarantee freshness.

Traps and limitations

Hallucinations (confident but wrong output), bias, data cut-off, latency and cost are inherent. Prompt injection - malicious instructions hidden in retrieved content - is a generative AI-specific risk that content filters alone do not fully solve. Classic exam trap: choose RAG for "our product documentation changed yesterday"; choose fine-tuning for "always answer in our brand voice and output JSON".

Rule of thumb

Facts and freshness → RAG. Behaviour and format → fine-tuning.

Azure AI-900 cheat sheet

Quick reference

Describe Artificial Intelligence workloads and considerations

  • AI workload families in scope: machine learning, computer vision, natural language processing, document intelligence, knowledge mining, and generative AI.

  • The six Responsible AI principles: fairness, reliability and safety, privacy and security, inclusiveness, transparency, and accountability.

  • Fairness means equitable treatment across groups; biased or unrepresentative training data is the usual root cause of unfair model outcomes.

  • Reliability and safety means the system operates consistently and does not cause harm, which requires testing, monitoring, and fail-safe design.

  • Transparency means stakeholders can understand the system's purpose, data, and decision logic; explainability is the mechanism that supports it.

  • Accountability means people and organizations remain answerable for AI outcomes, implemented through governance boards, approvals, and audit logs.

  • Inclusiveness means the solution works for all users, including people with disabilities; speech, vision, and language features support accessibility.

  • Privacy and security means personal data is collected lawfully, protected in transit and at rest, and used only for its stated purpose.

  • Scalability and cost-optimization are engineering qualities, not one of the six Responsible AI principles.

Quick reference

Describe fundamental principles of machine learning on Azure

  • Supervised learning uses labeled data; its two core tasks are regression (numeric output) and classification (categorical output).

  • Unsupervised learning uses unlabeled data; the main task in scope is clustering, which groups similar records by feature similarity.

  • Regression is evaluated with MAE, MSE, RMSE, and R-squared; classification is evaluated with accuracy, precision, recall, and F1 score.

  • Precision = true positives / (true positives + false positives); recall = true positives / (true positives + false negatives).

  • Overfitting shows as strong training performance with weak validation or test performance; remedies include more data, a simpler model, or regularization.

  • Data is normally split into training, validation, and test sets; the test set estimates performance on unseen data.

  • An Azure Machine Learning workspace is the top-level resource that holds compute targets, datastores, datasets, models, and pipelines.

  • Automated ML selects algorithms and hyperparameters automatically, while the Azure Machine Learning designer provides a drag-and-drop pipeline canvas.

  • Feature engineering creates or transforms input columns (for example normalizing values) so the model can learn the relationship more effectively.

Quick reference

Describe features of computer vision workloads on Azure

  • Image classification assigns one or more labels to an entire image; object detection returns labels plus bounding boxes locating each object.

  • Semantic segmentation classifies every pixel into a category, producing a mask that shows the exact shape of each region.

  • OCR extracts printed or handwritten text from images and PDF files and converts it into machine-readable characters.

  • Prebuilt Azure AI Vision capabilities include image tagging, captioning, object detection, face detection, OCR/Read, and spatial analysis.

  • Custom Vision trains image classification or object detection models on your own labeled images and is a separate service from Azure AI Vision.

  • Azure AI Document Intelligence (formerly Form Recognizer) extracts fields, key-value pairs, and tables from forms, receipts, and invoices.

  • Face detection returns face locations, verification compares two faces, and identification matches a face to a known person; these are limited-access features.

  • Object detection accuracy is summarized with mean average precision (mAP), computed across classes and intersection-over-union thresholds.

  • A bounding box is defined by its top-left x and y coordinates plus a width and a height.

Quick reference

Describe features of Natural Language Processing (NLP) workloads on Azure

  • Azure AI Language capabilities include language detection, sentiment analysis, key phrase extraction, named entity recognition, summarization, PII detection, and question answering.

  • Sentiment analysis returns positive, neutral, or negative labels with confidence scores between 0 and 1 at both document and sentence level.

  • Named entity recognition classifies entities such as person, location, organization, datetime, and quantity; PII detection flags personal identifiers for redaction.

  • Common text preprocessing steps are tokenization (splitting text into tokens), stop-word removal, and stemming or lemmatization.

  • TF-IDF scores a term by its frequency in a document, weighted down by the number of documents that contain it.

  • Azure AI Translator translates text between languages; Azure AI Speech provides speech-to-text, text-to-speech, and speech translation.

  • Conversational Language Understanding (CLU) extracts intents and entities from user utterances so a bot can decide what to do next.

  • Question answering in Azure AI Language returns the best answer from a curated knowledge base instead of generating free-form text.

Quick reference

Describe features of generative AI workloads on Azure

  • Generative models create new content such as text, images, or code by learning patterns in training data; text models work by predicting the next token.

  • A prompt typically contains a system message (behavior rules) and a user message (the task); the model responds with a completion.

  • Tokens are chunks of text, roughly four characters of English each; max tokens and model choice drive cost and latency.

  • Temperature and top_p control randomness: lower temperature gives more deterministic output, higher temperature gives more varied output.

  • Retrieval augmented generation (RAG) retrieves grounding data from a search index and injects it into the prompt, reducing hallucination without retraining.

  • Fine-tuning retrains a base model on your own examples to change style or format, while RAG supplies current facts at query time.

  • Azure OpenAI Service hosts OpenAI models (GPT, DALL-E, embeddings) inside Azure with private networking, compliance coverage, and content filters.

  • Content filtering in Azure OpenAI scores prompts and completions against hate, sexual, violence, and self-harm categories and can block them.

  • Hallucination is fluent but unsupported or fabricated output; mitigations include grounding with RAG, clear system instructions, and human review.

Azure AI-900 key terms

Artificial intelligence (AI)

The field of building computer systems that perform tasks normally requiring human intelligence, such as understanding language, recognizing images, making recommendations, or supporting decisions.

Related: Machine learning, Responsible AI
Machine learning

A branch of AI where a model learns patterns from data instead of following hand-written rules. The trained model then makes predictions or decisions on new, unseen data.

Related: Supervised learning, Deep learning
Deep learning

A machine learning approach that uses artificial neural networks with many layers to learn complex patterns from large amounts of data, powering tasks such as image recognition and language models.

Related: Machine learning, Large language model (LLM)
Anomaly detection

An AI workload that finds unusual items or events differing from the normal pattern in data, such as fraudulent transactions, failing equipment, or unusual sign-in activity.

Related: Unsupervised learning, Clustering
Responsible AI

A set of principles and practices for designing and deploying AI systems so they are fair, safe, private, inclusive, transparent, and accountable to the people affected by them.

Related: Fairness, Transparency
Fairness

A responsible AI principle requiring that an AI system treats all groups of people equitably and does not produce biased outcomes for people with similar relevant characteristics.

Related: Responsible AI, Inclusiveness
Inclusiveness

A responsible AI principle about designing AI systems so that people with different abilities, languages, and backgrounds can all use them and benefit from them.

Related: Fairness, Responsible AI
Reliability and safety

A responsible AI principle requiring that AI systems operate consistently under normal and unexpected conditions, and that they are tested and controlled so they do not cause harm.

Related: Responsible AI, Accountability
Privacy and security

A responsible AI principle about protecting personal and organizational data used by AI systems, keeping it secure, and giving people control over how their information is used.

Related: Responsible AI, Transparency
Transparency

A responsible AI principle about making it understandable how an AI system works and why it produced a given result, so people can trust and challenge its output.

Related: Accountability, Responsible AI
Accountability

A responsible AI principle stating that the people and organizations who design and operate an AI system remain answerable for how it behaves and for the outcomes it causes.

Related: Transparency, Responsible AI
Azure AI services

A family of prebuilt cloud APIs, formerly called Cognitive Services, that add ready-made vision, speech, language, and decision capabilities to applications without training models from scratch.

Related: Computer vision, Sentiment analysis
Supervised learning

A machine learning method where the training data includes both inputs and the correct answers, so the model learns to predict those answers for new inputs it has not seen.

Related: Classification, Label
Unsupervised learning

A machine learning method that trains on data with no labels, letting the model discover structure on its own, such as natural groupings or unusual values in the data.

Related: Clustering, Anomaly detection
Reinforcement learning

A machine learning method where an agent learns by acting in an environment and receiving rewards or penalties, gradually discovering the behavior that earns the highest total reward.

Related: Supervised learning
Regression

A supervised learning task that predicts a continuous numeric value, such as a price, temperature, or duration, rather than placing the item into a category.

Related: Classification, Supervised learning
Classification

A supervised learning task that predicts which category or class an item belongs to, such as spam versus not spam, or one of several product types.

Related: Regression, Label
Clustering

An unsupervised learning task that groups similar items together based on their features, without being told in advance what the groups are or how many exist.

Related: Unsupervised learning, Anomaly detection
Feature

An individual measurable property of the data used as input to a machine learning model, such as age, price, or the pixel values of an image.

Related: Label, Clustering
Label

The known answer or target value attached to a training example in supervised learning, telling the model what the correct prediction should be.

Related: Supervised learning, Feature
Overfitting

A modeling problem where a model memorizes the training data too closely, including its noise, and therefore performs poorly on new, unseen data.

Related: Supervised learning, Automated machine learning (AutoML)
Automated machine learning (AutoML)

A feature of Azure Machine Learning that automatically tries many algorithms and settings, selects the best-performing model for your data, and can deploy it for you.

Related: Machine learning, Overfitting
Computer vision

An AI workload in which systems interpret and analyze images and video to classify scenes, find objects, read text, or detect faces and movement.

Related: Image classification, Optical character recognition (OCR)
Image classification

A computer vision task that assigns one or more labels to a whole image, answering the question of what the image shows overall, for example cat or vehicle.

Related: Object detection, Computer vision
Object detection

A computer vision task that identifies where objects are in an image by returning a bounding box and a class label for each detected object.

Related: Image classification, Semantic segmentation
Semantic segmentation

A computer vision task that classifies every pixel in an image into a category, producing a mask that shows the exact shape of each region, such as road or sky.

Related: Object detection, Computer vision
Optical character recognition (OCR)

A computer vision capability that extracts printed or handwritten text from images and documents and converts it into machine-readable, searchable characters.

Related: Computer vision, Azure AI services
Natural language processing (NLP)

An AI workload that enables computers to understand, interpret, and generate human language in text or speech, covering tasks such as sentiment analysis, translation, and question answering.

Related: Sentiment analysis, Named entity recognition
Sentiment analysis

A natural language processing task that evaluates text and reports whether the opinion expressed in a piece of text is positive, negative, neutral, or mixed.

Related: Natural language processing (NLP), Azure AI services
Named entity recognition

A natural language processing task that finds and categorizes named things in text, such as people, organizations, locations, dates, and quantities.

Related: Natural language processing (NLP), Sentiment analysis
Speech recognition

A natural language processing capability, also called speech-to-text, that transcribes spoken audio into written text so it can be searched, analyzed, or acted on.

Related: Speech synthesis, Natural language processing (NLP)
Speech synthesis

A natural language processing capability, also called text-to-speech, that converts written text into natural-sounding spoken audio in a chosen voice and language.

Related: Speech recognition, Natural language processing (NLP)
Translation

A natural language processing capability that converts text or speech from one language into another while preserving meaning, used to localize content and enable cross-language communication.

Related: Natural language processing (NLP)
Conversational language understanding

A natural language processing service that interprets a user's utterance to detect its intent and extract key entities, so an application or bot can respond appropriately.

Related: Named entity recognition, Speech recognition
Generative AI

A category of AI that creates new content, such as text, images, code, or audio, by learning patterns from large amounts of existing data rather than only classifying or predicting.

Related: Large language model (LLM), Prompt
Large language model (LLM)

A very large neural network trained on huge amounts of text so it can predict and generate language, supporting tasks such as summarization, question answering, and conversation.

Related: Generative AI, Prompt
Prompt

The input text or instruction a person gives to a generative AI model, which guides what the model produces; clearer, more specific prompts generally yield better results.

Related: Large language model (LLM), Hallucination
Retrieval-augmented generation (RAG)

A technique that first retrieves relevant documents or data from a trusted source and then gives them to a generative model, so answers are grounded in real, current information.

Related: Prompt, Azure OpenAI Service
Azure OpenAI Service

A Microsoft Azure service that provides access to OpenAI's generative models, with Azure security, compliance, private networking, and content filtering built in so organizations can use them safely.

Related: Generative AI, Large language model (LLM)
Hallucination

A failure of a generative AI model where it produces content that sounds plausible but is factually wrong or invented, such as citing a source that does not exist.

Related: Generative AI, Retrieval-augmented generation (RAG)

Azure AI-900 study plan

8 weeks · About 6 hours per week

Week 1

Describe features of generative AI workloads on Azure

  • Define generative AI and distinguish it from classical ML and discriminative workloads in your own words

  • Identify the core capabilities of large language models on Azure (text generation, summarization, code, image generation) and match three business scenarios to the right model family

  • Explain what a copilot is and list three ways Microsoft copilots differ from a plain chatbot

  • Label the parts of a prompt (system message, user message, few-shot examples, output constraints) and rewrite three vague prompts into well-structured ones

Your checkpoint

A one-page generative AI glossary of 10 terms plus 5 rewritten prompts, self-checked against a 10-question quiz at 80% or better.

Week 2

Describe features of generative AI workloads on Azure & Describe Artificial Intelligence workloads and considerations

  • Sequence the steps of a retrieval-augmented generation flow (ingest, chunk, embed, index, retrieve, ground, generate) from memory

  • Compare prompt engineering, retrieval-augmented generation, and fine-tuning, then choose the best approach for four stated scenarios

  • Describe how to evaluate a generative AI response for groundedness, relevance, coherence, and harmful content

  • Map each of the six Microsoft responsible AI principles to a concrete generative AI failure it prevents

Your checkpoint

A one-page architecture sketch with written justification for a document question-answering solution (RAG vs. fine-tuning) plus a 15-question quiz scored at 80% or better.

Week 3

Describe Artificial Intelligence workloads and considerations

  • Classify twelve business scenarios into the correct AI workload category (computer vision, NLP, document processing, generative AI, anomaly detection)

  • State the six Microsoft responsible AI principles and name one process or tool that supports each

  • Distinguish a prebuilt AI service, a custom trained model, and a no-code studio experience such as Azure AI Foundry

  • Explain the roles of an Azure AI resource, endpoint, and key, and describe when a responsible AI impact assessment is required

Your checkpoint

A scenario-classification table of 12 items with zero misses, plus a one-page plan for governing a single high-risk use case.

Week 4

Describe fundamental principles of machine learning on Azure

  • Define machine learning and distinguish supervised, unsupervised, and reinforcement learning with one example of each

  • Compare regression, classification, and clustering, then select the correct technique for six stated problems

  • Explain the roles of features, labels, training data, validation data, and test data in a training pipeline

  • Describe what automated machine learning and the drag-and-drop designer in Azure Machine Learning do, and state when to use each

Your checkpoint

A 40-question mixed practice quiz covering domains 1, 2, and 5 scored at 80% or better, plus a decision flowchart mapping problem types to ML techniques.

Week 5

Describe fundamental principles of machine learning on Azure

  • Calculate accuracy, precision, recall, F1 score, RMSE, and R-squared from a small confusion matrix and error table

  • Explain overfitting and underfitting and list three techniques that mitigate each

  • Describe the Azure Machine Learning workflow: workspace, compute, data asset, experiment run, model registration, and endpoint deployment

  • Place fairness assessment, error analysis, and interpretability at the correct stage of a responsible ML workflow

Your checkpoint

Hand-calculated metrics for a 10-row confusion matrix with written interpretation of each metric, plus 20 practice questions scored at 80% or better.

Week 6

Describe features of computer vision workloads on Azure

  • List the core computer vision capabilities in Azure (image classification, object detection, optical character recognition, facial detection) and match each to a scenario

  • Distinguish image classification from object detection and select the right one for four cases

  • Describe how optical character recognition and document intelligence extract printed text, key-value pairs, and tables

  • Explain the responsible AI constraints Microsoft places on face-related workloads and why they exist

Your checkpoint

A 12-scenario matching sheet plus a comparison table of computer vision capabilities, with a 15-question quiz scored at 80% or better.

Week 7

Describe features of Natural Language Processing (NLP) workloads on Azure

  • List the Azure AI Language features (sentiment analysis, key phrase extraction, named entity recognition, language detection, summarization, conversational language understanding, question answering)

  • Distinguish conversational language understanding from question answering and choose the right one for three stated scenarios

  • Describe speech-to-text, text-to-speech, speech translation, and text translation scenarios and name the service that handles each

  • Explain how NLP features combine to support a bot or a document-processing pipeline

Your checkpoint

A feature-to-scenario map covering 12 NLP scenarios plus a 15-question quiz scored at 80% or better.

Week 8

Describe features of generative AI workloads on Azure & Describe Artificial Intelligence workloads and considerations

  • Re-derive the six responsible AI principles and the main generative AI design patterns from memory without notes

  • Complete two timed full-length practice exams under exam conditions

  • Review every missed item and rewrite the correct explanation in your own words

  • Confirm a per-domain accuracy of 75% or better in each of the five exam domains

Your checkpoint

Two timed practice exams scored at 85% or better with a written error log explaining every missed question.

Milestones

Week 2: Score 80% or better on a 30-question quiz covering generative AI and responsible AI topics without using notes.

Week 4: Score 80% or better on a 40-question mixed quiz spanning domains 1, 2, and 5, and explain two missed answers aloud.

Week 6: Score 80% or better on a 40-question mixed quiz spanning domains 2 and 3, with no domain scoring below 70%.

Week 8: Score 85% or better on two full-length timed practice exams, with no individual domain below 75%.

Azure AI-900 exam day

Your checklist

Government-issued photo ID

Bring a valid, unexpired photo ID whose name matches your exam registration exactly. If your registered name differs even slightly, fix it before test day.

Second form of ID (online proctored or test center)

Some delivery providers ask for a second ID or a signature-bearing card. Check the ID requirements listed in your appointment confirmation and confirm with the provider if unclear.

Appointment confirmation / registration number

Have your confirmation email or registration ID accessible (printed or on your phone before you sit down). Test centers may ask to see it at check-in.

Test center location and travel plan

If testing in person, confirm the exact address, parking/transit options, and the arrival time stated in your confirmation. Plan to arrive with buffer time for check-in.

Computer with working webcam, microphone, and speakers

For online proctored delivery only. Use a desktop or laptop (not a tablet or phone), and make sure the camera can show your face and your room.

Completed system/pre-check test

Run the provider's system test on the same machine and network you will use on exam day, ideally 72 hours ahead so you have time to fix problems.

Quiet, private, well-lit room

For online proctored delivery: no other people, no interruptions, clear desk, door closed, and lighting that keeps your face visible.

Stable wired-or-strong internet connection

Have a backup plan (mobile hotspot or phone number for provider support) in case your connection drops at start time.

Knowledge of allowed scratch materials

On-screen noteboard/whiteboard availability and physical scratch paper rules vary by delivery method. Confirm what is permitted with the provider before test day.

Accommodation approval (if applicable)

If you requested extra time or other accommodations, confirm it is approved and attached to your appointment before test day.

Water and a light snack for before/after

Eat and hydrate before you start. Personal items generally cannot be at your desk during the exam, so plan for before and after.

Notifications and background apps silenced

Turn off phone alerts, messaging apps, and automatic updates so nothing interrupts the session or triggers a proctor flag.

Plan the day

72 hours before

Run the provider's system test on your exam machine and network; resolve any camera, microphone, browser, or bandwidth issues now. Confirm the appointment date, time, and time zone in your confirmation email, and check the provider's reschedule/cancellation window if your plans are at risk.

72 hours before

Do a final targeted review instead of a full re-read: AI workloads (machine learning, computer vision, NLP, conversational AI), responsible AI principles, and Azure AI service names and what each is for. Stop studying new topics you cannot reinforce.

24 hours before

Lay out your ID, confirm travel or room setup, check the exact start time in local time, and read the provider's rules on what is allowed at your desk. Get a normal night's sleep rather than cramming.

24 hours before

Light review only: skim your own summary notes and any missed practice questions. Avoid memorizing exact question counts or score thresholds - those can vary.

Morning of

Eat, hydrate, and avoid starting a big new topic. For online delivery, close all other applications, plug in power, and start the check-in process early as instructed in your confirmation. For a test center, arrive at the time your confirmation states, allowing buffer for check-in.

During the test

Read each question fully and watch for words like 'not' and 'best'. Answer confidently-known items first, flag uncertain ones, and use any review screen to revisit them. Pace yourself against the timer - roughly a minute per question is comfortable for a 45-minute exam - and leave time to review before submitting.

During the test

If you hit a technical problem, stop and contact the proctor or provider support rather than troubleshooting alone; they can document the issue and advise on next steps.

Rules to check

  • Your registered name must match your photo ID exactly; mismatches can prevent you from testing.

  • For online proctored exams, you must be alone in a closed, well-lit room with a clear desk; no notes, phones, or second screens unless the provider explicitly allows them.

  • Do not talk aloud, read questions aloud, or leave the camera view during an online proctored session - this can be flagged as a policy violation.

  • Breaks are generally not offered during a fundamentals-level exam of this length; if you need an accommodation, arrange and confirm it in advance with the provider.

  • Exam content is confidential: do not screenshot, copy, record, or share questions in any form.

  • Scoring and any retake policy are set by the provider; confirm the current passing score, score report timing, and retake rules with the provider rather than relying on secondhand reports.

  • Follow all proctor instructions immediately; unresolved or repeated violations can result in the session being ended and the exam voided.

  • If you need to reschedule, do it within the provider's stated window (commonly a set number of hours before the appointment) - confirm the exact deadline with the provider.

Azure AI-900 FAQ

How many Azure AI-900 practice questions are available?

The practice set has 200 Azure AI-900 questions. Each one includes the choices, the correct answer, and an explanation, and the set is free to read with no signup.

How many Azure AI-900 flashcards are available?

The flashcard set has 32 Azure AI-900 cards. Each card shows the term and its definition.

Which topics does the Azure AI-900 study material cover?

The Azure AI-900 study material has 5 sections, including Describe Artificial Intelligence Workloads and Considerations, Describe Fundamental Principles of Machine Learning on Azure, and Describe Features of Computer Vision Workloads on Azure.