AI-900 Microsoft Azure AI Fundamentals Free Practice Exam Questions (2026 Updated)

The correct mapping is based on how each Azure Cognitive Service functions within the Microsoft AI ecosystem, as detailed in the Microsoft Azure AI Fundamentals (AI-900) study guide and Microsoft Learn Cognitive Services documentation.

Convert spoken requests into text → Azure AI SpeechThe Azure AI Speech service provides speech-to-text (STT) capabilities, which enable an application to recognize spoken language and convert it into written text. This functionality is foundational in voice-enabled applications like digital assistants or transcription services. When a user speaks, this service captures the audio signal and produces an accurate textual representation that can then be processed by other AI services.

Identify the intent of a user’s requests → Azure AI LanguageThe Azure AI Language service (which includes Conversational Language Understanding, formerly LUIS) is designed to extract meaning from text. It identifies intents—the goals or actions a user wants to perform—and entities, which are key details within that request. For example, in the command “Book a flight to Paris,” the intent is “book a flight,” and the entity is “Paris.”

Apply intent to entities and utterances → Azure AI LanguageAgain, the Language service performs this deeper contextual analysis. It not only identifies what the user wants (intent) but also applies it to utterances (specific user expressions) and entities (data elements extracted from text). This helps conversational AI systems take meaningful actions, such as fulfilling user requests.

In summary, Azure AI Speech handles audio-to-text conversion, while Azure AI Language performs natural language understanding, mapping intents and entities—a workflow essential in intelligent conversational applications.

According to the Microsoft Azure AI Fundamentals (AI-900) Official Study Guide and the Microsoft Learn module “Explore fundamental principles of machine learning,” anomaly detection is a specialized machine learning technique used to identify data points, patterns, or events that deviate significantly from normal behavior.

Anomaly detection is widely used for monitoring time-series data and detecting unexpected or rare occurrences that may indicate problems, opportunities, or fraud. For example:

Detecting fraudulent transactions in banking systems.

Identifying equipment malfunctions in industrial IoT applications.

Monitoring network intrusions in cybersecurity.

Detecting unexpected spikes or drops in web traffic or sales.

In Azure, this workload is supported by the Azure AI Anomaly Detector service, which uses statistical and machine learning algorithms to learn from historical data and establish a baseline of normal behavior. When the system detects data points that fall outside expected patterns, it flags them as anomalies.

Let’s evaluate the incorrect options:

A. A machine learning technique that understands written and spoken language → This describes Natural Language Processing (NLP), not anomaly detection.

B. A machine learning technique that classifies objects based on user-supplied images → This refers to image classification, typically using computer vision.

D. A machine learning technique that classifies images based on their contents → Also describes computer vision, not anomaly detection.

Therefore, the correct answer is C, since anomaly detection specifically refers to analyzing data over time and identifying unusual or abnormal patterns that differ from the expected trend.

The correct answer is “An embedding.”

In the context of large language models (LLMs) such as GPT-3, GPT-3.5, or GPT-4, an embedding refers to a multi-dimensional numeric vector representation assigned to each word, token, or phrase. According to the Microsoft Azure AI Fundamentals (AI-900) study guide and Microsoft Learn documentation for Azure OpenAI embeddings, embeddings are used to represent textual or semantic meaning in a numerical form that a machine learning model can process mathematically.

Each embedding captures the semantic relationships between words. Words or tokens with similar meanings (for example, “car” and “automobile”) are represented by vectors that are close together in the multi-dimensional space, while unrelated words (like “tree” and “laptop”) are farther apart. This vector representation enables the model to understand context, similarity, and relationships between different pieces of text.

Embeddings are fundamental in tasks such as:

Semantic search: Finding documents or sentences with similar meaning.

Clustering: Grouping related concepts together.

Recommendation systems: Suggesting similar content based on text meaning.

Contextual understanding: Helping generative models produce coherent and context-aware text.

Option review:

Attention: A mechanism used within transformers to focus on relevant parts of input sequences but not a representation of words.

A completion: Refers to the generated text output from a model, not the internal representation.

A transformer: The architecture that powers models like GPT, not the vector representation of tokens.

Therefore, the correct term for a multi-dimensional vector assigned to each word or token in a large language model (LLM) is An embedding, which represents how meaning is numerically encoded and compared within language models.

According to the Microsoft Azure AI Fundamentals (AI-900) official study guide and Microsoft Learn documentation on Azure AI Vision services, different Azure AI Vision capabilities are suited for different use cases such as object detection, brand recognition, facial recognition, and spatial analysis.

App1: Uses a set of images and photos to extract brand names → Image AnalysisThe Azure AI Vision – Image Analysis service (formerly part of Computer Vision) can detect and extract brands, objects, text, and other visual features from images. It uses advanced image classification and object detection models to recognize logos and identify brand names (for example, “Microsoft” or “Coca-Cola”) in photos. The Image Analysis API can also return descriptive tags, scene descriptions, and confidence scores. Therefore, since App1 analyzes static images to extract brand names, it specifically relies on the Image Analysis feature of Azure AI Vision.

App2: Enables touchless access control for buildings → FaceThe Azure AI Face service is designed for facial detection, verification, and identification. It can recognize and match faces in real time, making it ideal for access control, identity verification, and attendance tracking systems. A “touchless access control” system uses a camera to detect a person’s face and verify identity against a stored profile, allowing or denying entry without physical interaction.

The other options are not suitable:

Optical Character Recognition (OCR) extracts text, not brand logos.

Spatial Analysis is for detecting movement or presence in video feeds.

Video Analysis is for analyzing dynamic video content rather than still images.

The question describes a smart device that can understand and respond to a spoken or written question such as, “What is the stock price of Contoso, Ltd.?” This scenario directly maps to the Natural Language Processing (NLP) workload in Microsoft Azure AI.

According to the Microsoft AI Fundamentals (AI-900) study guide and the Microsoft Learn module “Describe features of common AI workloads,” NLP enables systems to understand, interpret, and generate human language. Azure AI Language and Azure Speech services are examples of NLP-based solutions.

In this case, the smart device performs several NLP tasks:

Speech recognition – converts spoken input into text.

Language understanding – interprets the user’s intent, i.e., retrieving the stock price of a specific company.

Response generation – formulates a meaningful answer that can be presented back as text or speech.

This process shows a full pipeline of natural language understanding (NLU) and conversational AI. It does not involve visual data (computer vision), data pattern analysis (anomaly detection), or document search (knowledge mining).

Hence, the correct AI workload is D. Natural Language Processing.

According to the Microsoft Azure AI Fundamentals (AI-900) official study materials and the Microsoft Learn module “Describe features of Natural Language Processing (NLP) workloads on Azure”, this scenario combines two major capabilities of AI: speech recognition and natural language understanding.

Speech-to-Text (E) – This is the first step in processing verbal commands. The Azure Speech service converts the spoken words of a user into textual data that can be understood and processed by downstream components. This workload is commonly referred to as speech recognition, and it falls under the speech capabilities of Azure Cognitive Services. Without this transcription process, the system could not interpret the user’s voice input.

Language Modeling (C) – After the speech input is converted into text, the next step is to interpret the meaning of the text so the system can take appropriate action. Language modeling, also known as language understanding, is responsible for identifying the user’s intent (for example, “turn on the lights” or “set the thermostat to 72 degrees”) and extracting entities (such as device name or temperature value). In Azure, this function is handled by Language Understanding (LUIS) or Conversational Language Understanding (CLU). These models allow smart systems to process commands and map them to defined actions.

Other options are not correct:

A. Text-to-speech converts text output into spoken language, which is not mentioned as a requirement.

B. Translation converts text from one language to another, irrelevant to this scenario.

D. Key phrase extraction identifies important terms in text but doesn’t interpret or execute commands.

Therefore, the solution uses speech-to-text to transcribe verbal commands and language modeling to understand and act upon them — the two key NLP workloads enabling voice-controlled smart devices.

Box 3: Natural language processing

Natural language processing (NLP) is used for tasks such as sentiment analysis, topic detection, language detection, key phrase extraction, and document categorization.

Confidence.

According to the Microsoft Azure AI Fundamentals (AI-900) Official Study Guide and the Microsoft Learn module “Explore computer vision in Microsoft Azure,” the confidence score represents the calculated probability that a model’s prediction is correct. In image classification, when an AI model analyzes an image and assigns it to a specific category, it also produces a confidence value—a numerical probability (usually between 0 and 1) indicating how certain the model is about its prediction.

For example, if an image classification model identifies an image as a “cat” with a confidence of 0.92, it means the model is 92% certain that the image depicts a cat. The confidence value helps developers and users understand the model’s certainty level about its classification output.

Microsoft Learn emphasizes that in Azure Cognitive Services—such as the Custom Vision Service—each prediction result includes both the predicted label (class) and a confidence score. These confidence scores are essential for evaluating model performance and determining thresholds for automated decisions (e.g., accepting predictions only above a 0.8 probability).

Let’s evaluate the other options:

Accuracy: This is an overall performance metric measuring the percentage of correct predictions across the dataset, not a probability for a single prediction.

Root Mean Square Error (RMSE): This is a metric for regression models, not classification tasks. It measures average error magnitude between predicted and actual values.

Sentiment: This is a type of prediction (positive, negative, neutral) in text analysis, not a probability metric.

Therefore, based on Microsoft’s AI-900 training materials and Azure Cognitive Services documentation, the calculated probability of a correct image classification is called Confidence, which expresses how sure the model is about its prediction for a specific input.

According to the Microsoft Azure AI Fundamentals (AI-900) official study guide and the Microsoft Learn module “Identify features of common machine learning types”, the classification technique is a type of supervised machine learning used to predict which category or class a new observation belongs to, based on patterns learned from labeled training data.

In this scenario, a banking system that predicts whether a loan will be repaid is dealing with a binary outcome—either the loan will be repaid or will not be repaid. These two possible results represent distinct classes, making this problem a classic example of binary classification. During training, the model learns from historical data containing features such as customer income, credit score, loan amount, and repayment history, along with labeled outcomes (repaid or defaulted). After training, it can classify new applications into one of these two categories.

The AI-900 curriculum distinguishes between three key supervised and unsupervised learning approaches:

Classification: Predicts discrete categories (e.g., spam/not spam, fraud/not fraud, will repay/won’t repay).

Regression: Predicts continuous numerical values (e.g., house prices, sales forecast, temperature).

Clustering: Groups data based on similarity without predefined labels (e.g., customer segmentation).

Since the banking problem focuses on predicting a categorical outcome rather than a continuous numeric value, it fits squarely into the classification domain. In Azure Machine Learning, such tasks can be performed using algorithms like Logistic Regression, Decision Trees, or Support Vector Machines (SVMs), all configured for categorical prediction.

Therefore, per Microsoft’s official AI-900 learning objectives, a banking system predicting whether a loan will be repaid represents a classification type of machine learning problem.

According to the Microsoft Azure AI Fundamentals (AI-900) study guide, the three main types of supervised and unsupervised machine learning models—classification, clustering, and regression—are used for distinct problem types depending on the structure of the data and the prediction goal.

Clustering is an unsupervised learning technique used when the goal is to group items with similar characteristics without predefined labels. In this scenario, “Assign categories to passengers based on demographic data” implies automatically grouping passengers based on patterns such as age, income, or travel frequency, without any prior labeling. This directly maps to clustering, which discovers hidden groupings (for example, segmenting customers into categories like business travelers or vacationers).

Regression is a supervised learning method used to predict continuous numerical values. The scenario “Predict the amount of consumed fuel based on flight distance” is a classic regression problem because the output (fuel consumption) is a continuous variable dependent on another continuous variable (distance). Regression models, such as linear regression, are trained to estimate numeric outputs.

Classification is also a supervised learning approach, but it predicts discrete categories or outcomes. The scenario “Predict whether a passenger will miss their flight based on demographic data” involves a binary decision (missed or not missed), which is typical of classification tasks. These models learn from labeled examples to assign new instances to specific categories.

In summary, Clustering groups similar passengers, Regression predicts continuous numerical outcomes, and Classification determines categorical outcomes. This alignment precisely matches the definitions in Microsoft’s AI-900 learning objectives under “Describe common machine learning types and scenarios.”

Azure AI Form Recognizer extracts information from structured and semi-structured documents. A custom model in Form Recognizer allows an organization to train the system on its specific document layouts and data fields.

As per the AI-900 study guide, a key advantage of a custom model is its flexibility. It can be trained with a set of labeled examples (e.g., invoices, purchase orders, receipts) that match the company’s format. Once trained, the model learns where to locate and extract fields such as invoice numbers, dates, or totals—regardless of layout differences between form types.

Option B is correct because a custom model can be trained to recognize a variety of form types, making it adaptable for diverse business processes.

Options A, C, and D are incorrect:

A: Both prebuilt and custom models are cloud-based; on-premises deployment is not an exclusive feature.

C: Custom models are not cheaper; they may involve additional training costs.

D: Custom models do not always guarantee higher accuracy—accuracy depends on the training data quality.

facial analysis.

According to the Microsoft Azure AI Fundamentals (AI-900) official study guide and the Microsoft Learn module “Describe features of computer vision workloads on Azure,” facial analysis is a computer vision capability that detects faces and extracts attributes such as facial expressions, emotions, pose, occlusion, and image quality factors like exposure and noise. It does not identify or verify individual identities; rather, it interprets facial features and image characteristics to analyze conditions in an image.

In this question, the AI solution helps photographers take better portrait photos by providing feedback on exposure, noise, and occlusion — tasks directly linked to facial analysis. The model analyzes the detected face to determine if the image is well-lit, clear, and unobstructed, thereby improving photo quality. These capabilities are part of the Azure Face service in Azure Cognitive Services, which includes both facial detection and facial analysis functionalities.

Here’s how the other options differ:

Facial detection only identifies that a face exists in an image and provides its location using bounding boxes, without further interpretation.

Facial recognition goes a step further — it attempts to identify or verify a person’s identity by comparing the detected face with stored images. This is not what the scenario describes.

Thus, when an AI solution evaluates image quality aspects like exposure, noise, and occlusion, it’s performing facial analysis, which focuses on understanding image and facial characteristics rather than identification.

In summary, based on Microsoft’s AI-900 study material, this scenario demonstrates facial analysis, a subcategory of computer vision tasks within Azure Cognitive Services.

Box 1: Entity recognition

the Named Entity Recognition module in Machine Learning Studio (classic), to identify the names of things, such as people, companies, or locations in a column of text.

Named entity recognition is an important area of research in machine learning and natural language processing (NLP), because it can be used to answer many real-world questions, such as:

Which companies were mentioned in a news article?

Does a tweet contain the name of a person? Does the tweet also provide his current location?

Were specified products mentioned in complaints or reviews?

Box 2: Sentiment Analysis

The Text Analytics API ' s Sentiment Analysis feature provides two ways for detecting positive and negative sentiment. If you send a Sentiment Analysis request, the API will return sentiment labels (such as " negative " , " neutral " and " positive " ) and confidence scores at the sentence and document-level.

Accelerate your business processes by automating information extraction. Form Recognizer applies advanced machine learning to accurately extract text, key/value pairs, and tables from documents. With just a few samples, Form Recognizer tailors its understanding to your documents, both on-premises and in the cloud. Turn forms into usable data at a fraction of the time and cost, so you can focus more time acting on the information rather than compiling it.

In Natural Language Processing (NLP), entities are pieces of information extracted from text, such as names, locations, or phone numbers. According to the Microsoft Learn module “Explore natural language processing in Azure,” Azure’s Language Understanding (LUIS) supports several entity types:

Machine-learned entities – Automatically learned based on context in training data.

List entities – Used for predefined, limited sets of values (e.g., colors or product names).

Pattern.any entities – Capture flexible, unstructured phrases in user input.

Regular expression entities – Use regex patterns to match specific data formats such as phone numbers, postal codes, or dates.

A regular expression is ideal for recognizing phone numbers because phone numbers follow specific numeric or symbol-based patterns (e.g., (555)-123-4567 or +1 212 555 0199). By defining a regex pattern, the AI model can accurately extract phone numbers regardless of text context.

To improve the quality and relevance of responses generated by a generative AI solution using GPT-3.5, the following three actions are emphasized in the Microsoft Learn: Azure OpenAI Service best practices and AI-900/AI-102 training materials:

(A) Add grounding data to prompts: Grounding ensures that the model’s output is based on factual, domain-specific information. By adding context or external knowledge sources, responses become more accurate and aligned with the organization’s data rather than relying on the model’s general training corpus.

(B) Provide additional examples to prompts: Also known as few-shot prompting, this method demonstrates desired response patterns by including examples in the prompt. This significantly improves output quality, consistency, and adherence to desired formats.

(E) Modify system messages: In Azure OpenAI chat completions, system messages define the model’s behavior, style, and tone. Adjusting system messages allows fine-tuning of the model’s response quality, ensuring it follows context or persona guidelines.

The remaining options are incorrect:

(C) Modify tokenization is a low-level text-processing technique not used to improve model response quality directly.

(D) Add training data to prompts is not possible at runtime since GPT-3.5 models are pre-trained; only prompt engineering can influence output behavior.

Therefore, based on Azure OpenAI and AI-900 guidance, the three best ways to enhance generative