AI-901 Practice Test 4: 50 Questions and Answers for Azure AI Fundamentals

Published July 5, 2026

Practice AI-901 questions for Microsoft Azure AI Fundamentals. This test covers AI concepts, responsible AI, machine learning basics, natural language processing, computer vision, speech, generative AI, Microsoft Foundry, and AI agents. Select your answers, review the explanations, and track your progress in the Learning Dashboard.

Exam: AI-901Questions: 50Recommended score: 70%+Time: 60 minutes

Before you start

This independent practice test includes single-choice, multiple-response, and true/false questions. It is designed for Azure AI Fundamentals learners who want to prepare for AI-901 with clear explanations and Microsoft Learn references. When a question requires more than one answer, the question text tells you exactly how many answers to choose.

This content is an educational practice resource and is not a Microsoft exam dump. The questions are original and focus on the skills measured for AI-901, including AI capabilities and implementation concepts with Microsoft Foundry.

Question 1: Which statement best describes artificial intelligence in a business application?

A storage method for keeping files in cloud containers
A capability that allows software to perform tasks that normally require human intelligence
A network protocol for connecting virtual machines
A billing model for cloud resources

The correct answer is A capability that allows software to perform tasks that normally require human intelligence.

Artificial intelligence helps applications recognize patterns, understand language, make predictions, generate content, or support decisions. It is not limited to one Azure service or one type of workload. In business scenarios, AI is used to automate and improve tasks that previously required human reasoning or perception.

The correct answer is Classification.

Classification predicts a category or class, such as renew or not renew, approved or rejected, and high risk or low risk. It is different from regression, which predicts numeric values. Subscription renewal prediction is a typical classification scenario because the output is a discrete label.

The correct answer is False.

Regression models predict numeric values, such as sales revenue, delivery time, temperature, or expected demand. They are not limited to image or video data. Regression is a common machine learning approach for forecasting and estimating continuous values.

The correct answer is Supervised learning.

Supervised learning trains a model by using input examples that already include the expected answers. The model learns the relationship between inputs and labels, then applies that pattern to new data. Common supervised learning tasks include classification and regression.

The correct answer is Optical character recognition.

Optical character recognition, or OCR, detects text in images, scanned pages, receipts, and forms. After the text is extracted, other services can analyze or process it. OCR is a computer vision capability frequently used in document processing workflows.

The correct answer is True.

Sentiment analysis is an NLP capability that evaluates the emotional tone of text. It can be used with reviews, support tickets, survey comments, or social messages. The result helps organizations understand user feedback at scale.

The correct answer is Speech-to-text.

Speech-to-text converts spoken language from audio into written text. It is useful for transcription, captions, call analytics, and voice-enabled applications. After transcription, the text can also be processed by NLP services.

The correct answer is Named entity recognition.

Named entity recognition detects and categorizes important entities in text, such as names, locations, dates, and organizations. It is part of natural language processing. This helps applications extract structured information from unstructured text.

The correct answer is True.

Generative AI models produce new content from user instructions, examples, and context. They can draft text, summarize documents, generate code, create images, or answer questions. The quality of the result depends on the model, prompt, grounding data, and safety controls.

The correct answer is To connect model responses to relevant source data.

Grounding gives a generative AI model relevant context, such as retrieved documents, enterprise data, or approved knowledge sources. This helps make responses more specific and useful. Grounding is commonly used with retrieval augmented generation to reduce unsupported answers.

The correct answers are Fairness, Transparency, and Accountability.

Responsible AI focuses on building systems that are fair, transparent, reliable, secure, privacy-aware, and accountable. These principles help reduce harm and increase trust. Ignoring privacy or retaining unlimited data without purpose would conflict with responsible AI practices.

The correct answer is False.

Model choice depends on quality, latency, cost, context window, safety, region availability, and the task being solved. A smaller or specialized model may be better for simple, high-volume, or low-latency workloads. Production architecture should balance capability with operational requirements.

The correct answer is Tokenization.

Tokenization converts text into smaller units called tokens. Large language models process prompts and responses as tokens, and token usage affects context limits and cost. Understanding tokens helps with prompt design and capacity planning.

The correct answer is Clustering.

Clustering is an unsupervised learning technique that groups similar items based on patterns in the data. It does not require predefined labels. Customer segmentation is a common clustering use case.

The correct answer is Anomaly detection.

Anomaly detection identifies data points or patterns that deviate from expected behavior. It is useful for fraud detection, monitoring, predictive maintenance, and security analytics. The goal is to surface unusual activity that may require investigation.

The correct answers are Detecting objects in photos, Extracting text from scanned receipts, and Classifying images by product type.

Computer vision workloads analyze images and visual content. Object detection, image classification, and OCR are common examples. Speech translation and customer clustering are AI workloads, but they are not computer vision scenarios.

The correct answer is Text-to-speech.

Text-to-speech converts written text into spoken audio. It is commonly used in accessibility tools, virtual assistants, announcements, and multilingual voice experiences. It is the opposite of speech-to-text.

The correct answer is True.

Retrieval augmented generation combines information retrieval with generative AI. The system searches relevant data sources and passes the retrieved context to the model. This approach improves answer relevance and can make responses easier to trace to source content.

The correct answer is A numeric value.

Revenue prediction is a regression scenario because the expected output is numeric. Regression is used when the result is a continuous value. Examples include price, demand, revenue, time, or temperature predictions.

The correct answers are They represent content as vectors that capture semantic meaning, and they can support semantic search and similarity comparison.

Embeddings transform text or other content into vector representations. Similar meanings usually appear closer together in vector space, which makes embeddings useful for semantic search, recommendations, clustering, and retrieval augmented generation. They do not replace security or governance controls.

The correct answer is False.

Responsible AI should be considered throughout the full lifecycle, including design, data preparation, model selection, testing, deployment, monitoring, and retirement. Waiting until production increases risk. Early planning helps teams build safer and more reliable systems.

The correct answer is Building, deploying, managing, and monitoring AI apps and agents.

Microsoft Foundry is a platform for working with AI models, tools, projects, agents, evaluations, and monitoring. It helps teams build and operate AI applications at scale. It is not a DNS, networking, or subscription replacement service.

The correct answer is To make a selected model available for an application to call.

A model deployment exposes a selected model configuration for use by applications, agents, or APIs. Deployments are important because they define the model endpoint and operational settings. Applications should call the deployment rather than an abstract model name.

The correct answer is True.

Foundry Agent Service is designed for agent-based AI solutions. Agents can use instructions, models, tools, and knowledge sources to complete tasks. The service helps developers build and operate agents with managed platform capabilities.

The correct answer is Agent instructions or system message.

Agent instructions, often implemented as system-level guidance, define how the agent should behave. They can specify role, tone, allowed actions, limitations, and response style. Clear instructions help keep agent behavior consistent across conversations.

The correct answers are Approved documents, Search indexes or knowledge bases, and Permission-aware organizational data.

Grounding should use trusted and governed data sources. Approved documents, search indexes, and permission-aware enterprise data can help agents provide relevant answers while respecting access controls. Random ungoverned content and secrets in prompts create security and quality risks.

The correct answer is Retrieval augmented generation with approved sources.

Retrieval augmented generation retrieves relevant information from approved sources before the model generates an answer. This is useful when answers must reflect internal policy or current documentation. It helps reduce generic or unsupported responses.

The correct answer is Azure AI Content Safety.

Azure AI Content Safety helps detect potentially harmful content such as hate, sexual, violent, or self-harm related material. It can be used to moderate user input and model output. Content safety is an important control for responsible AI applications.

The correct answer is Evaluation.

Evaluation measures how well an AI solution performs against defined criteria such as relevance, groundedness, correctness, fluency, or safety. Teams can use evaluations to compare prompts, models, and configurations. This supports safer release decisions.

The correct answer is False.

Production AI systems should be monitored for quality, latency, cost, errors, safety issues, and user feedback. Application behavior can change because prompts, user inputs, data sources, or model versions change. Monitoring is part of operating AI responsibly.

The correct answer is Model catalog.

The model catalog helps developers discover and compare available models for different use cases. Teams can choose a model based on task fit, performance, cost, availability, and responsible AI requirements. The selected model can then be deployed for use.

The correct answer is Quality, latency, cost, region availability, and context requirements.

Model selection is an engineering decision. A low-latency chatbot needs a model that provides acceptable answer quality while meeting response-time, cost, region, and context-size needs. Larger models are not automatically the best option.

The correct answer is Azure AI Document Intelligence.

Azure AI Document Intelligence is designed to extract structured information from documents, forms, receipts, invoices, and identity documents. It can identify fields, tables, and key-value pairs. This makes it suitable for document automation scenarios.

The correct answers are Content filtering, Grounded answers with source data, and Human review for sensitive workflows.

Safety and reliability require technical and operational controls. Content filtering can reduce harmful content, grounding can improve answer traceability, and human review can reduce risk in sensitive decisions. Skipping testing and ignoring evaluations increase risk.

The correct answer is Text summarization.

Text summarization creates a shorter version of longer text while preserving important information. It can help users process support cases, meeting notes, reports, and articles. It is part of natural language processing capabilities.

The correct answer is Image analysis.

Image analysis can identify visual features, objects, tags, captions, and other image information. It is used to understand image content programmatically. This differs from speech, translation, or text-only NLP services.

The correct answer is Azure AI Language.

Azure AI Language includes natural language processing capabilities such as language detection, sentiment analysis, key phrase extraction, and entity recognition. Language detection helps route messages or select the correct downstream translation and processing steps.

The correct answer is Speech translation.

Speech translation combines speech recognition and translation to convert spoken content into another language. It can be used for captions, meetings, call centers, and multilingual experiences. It is different from text translation alone because the source is audio.

The correct answers are Tools can allow agents to retrieve information or call external systems, and tool use should be governed by permissions, validation, and monitoring.

Tools allow agents to interact with APIs, workflows, search systems, or other external capabilities. Because tool actions can affect data and business processes, they should be controlled with authorization, validation, logging, and monitoring. Tool use does not remove the need for grounding or testing.

The correct answer is Use permission-aware data sources and enforce access control.

AI systems should respect the same access boundaries as other enterprise applications. Permission-aware retrieval and proper authorization help prevent unauthorized disclosure. The model should not be trusted to infer or enforce permissions without application-level controls.

The correct answer is To test prompts and model behavior against representative examples.

Evaluation datasets provide representative inputs and expected quality criteria. They help teams compare prompts, models, grounding strategies, and safety controls before release. This makes testing repeatable and reduces reliance on subjective manual checks only.

The correct answer is A vector index built from document embeddings.

Semantic search commonly uses embeddings stored in a vector index. User queries are also converted to vectors, and the system retrieves content with similar meaning. This is a common foundation for retrieval augmented generation over enterprise documents.

The correct answer is False.

Prompt engineering is useful but not sufficient by itself. Production AI systems also need grounding, evaluations, security controls, content safety, monitoring, and responsible AI practices. A well-written prompt cannot replace architecture and governance.

The correct answers are Create or select a project, Choose and deploy a model, and Evaluate and monitor the solution.

A typical Foundry workflow includes organizing work in a project, selecting a model, deploying it, building the application or agent, evaluating behavior, and monitoring production usage. Disabling safety controls or placing secrets in prompts are unsafe practices.

The correct answer is Azure AI Vision.

Azure AI Vision provides capabilities for analyzing images and extracting visual information. Depending on the feature, it can identify objects, tags, captions, text, or other visual attributes. Language and Speech services focus on text and audio workloads.

The correct answer is An endpoint and authorized credentials.

Applications call deployed models through an endpoint and must authenticate using an allowed method, such as keys or managed identity depending on configuration. Access should be controlled and monitored. Anonymous administrative access is not appropriate for production AI workloads.

The correct answer is True.

Human oversight helps reduce risk when AI affects sensitive, regulated, or high-impact decisions. AI systems can make mistakes, reflect bias, or produce unsupported output. Review workflows, escalation paths, and accountability are important responsible AI controls.

The correct answers are Apply data minimization, Use access control for retrieval sources, and Redact or avoid unnecessary personal data.

Sensitive data should be handled with the same care in AI systems as in other applications. Data minimization, access control, and redaction reduce privacy and security risk. Secrets and unnecessary personal data should not be placed into prompts or exposed in logs.

The correct answers are Some agents can be created and tested in the Foundry portal, and Portal-based tools can help prototype agent behavior.

Foundry provides portal experiences that can help teams create, test, and refine agents. No-code or low-code development can speed up prototyping, but agents still need governance, testing, access control, and monitoring. The development method does not automatically guarantee safety.

The correct answers are The agent purpose and limitations, The data sources the agent can use, and Escalation or human review path.

Users and operators should understand what the agent is designed to do, what data it can access, where its limits are, and when humans should review or intervene. Clear documentation supports transparency and accountability. Credentials and permission bypasses should never be documented as part of user-facing behavior.

AI-901 Practice Test 4 – 50 Questions and Answers (Azure AI Fundamentals)

AI-901 practice test questions

The correct answer is A capability that allows software to perform tasks that normally require human intelligence.

Related Microsoft Learn topic

The correct answer is Classification.

Related Microsoft Learn topic

The correct answer is False.

Related Microsoft Learn topic

The correct answer is Supervised learning.

Related Microsoft Learn topic

The correct answer is Optical character recognition.

Related Microsoft Learn topic

The correct answer is True.

Related Microsoft Learn topic

The correct answer is Speech-to-text.

Related Microsoft Learn topic

The correct answer is Named entity recognition.

Related Microsoft Learn topic

The correct answer is True.

Related Microsoft Learn topic

The correct answer is To connect model responses to relevant source data.

Related Microsoft Learn topic

The correct answers are Fairness, Transparency, and Accountability.

Related Microsoft Learn topic

The correct answer is False.

Related Microsoft Learn topic

The correct answer is Tokenization.

Related Microsoft Learn topic

The correct answer is Clustering.

Related Microsoft Learn topic

The correct answer is Anomaly detection.

Related Microsoft Learn topic

The correct answers are Detecting objects in photos, Extracting text from scanned receipts, and Classifying images by product type.

Related Microsoft Learn topic

The correct answer is Text-to-speech.

Related Microsoft Learn topic

The correct answer is True.

Related Microsoft Learn topic

The correct answer is A numeric value.

Related Microsoft Learn topic

The correct answers are They represent content as vectors that capture semantic meaning, and they can support semantic search and similarity comparison.

Related Microsoft Learn topic

The correct answer is False.

Related Microsoft Learn topic

The correct answer is Building, deploying, managing, and monitoring AI apps and agents.

Related Microsoft Learn topic

The correct answer is To make a selected model available for an application to call.

Related Microsoft Learn topic

The correct answer is True.

Related Microsoft Learn topic

The correct answer is Agent instructions or system message.

Related Microsoft Learn topic

The correct answers are Approved documents, Search indexes or knowledge bases, and Permission-aware organizational data.

Related Microsoft Learn topic

The correct answer is Retrieval augmented generation with approved sources.

Related Microsoft Learn topic

The correct answer is Azure AI Content Safety.

Related Microsoft Learn topic

The correct answer is Evaluation.

Related Microsoft Learn topic

The correct answer is False.

Related Microsoft Learn topic

The correct answer is Model catalog.

Related Microsoft Learn topic

The correct answer is Quality, latency, cost, region availability, and context requirements.

Related Microsoft Learn topic

The correct answer is Azure AI Document Intelligence.

Related Microsoft Learn topic

The correct answers are Content filtering, Grounded answers with source data, and Human review for sensitive workflows.

Related Microsoft Learn topic

The correct answer is Text summarization.

Related Microsoft Learn topic

The correct answer is Image analysis.

Related Microsoft Learn topic

The correct answer is Azure AI Language.

Related Microsoft Learn topic

The correct answer is Speech translation.

Related Microsoft Learn topic

The correct answers are Tools can allow agents to retrieve information or call external systems, and tool use should be governed by permissions, validation, and monitoring.

Related Microsoft Learn topic