AI-901 Practice Test.
65 free practice questions with answers and explanations.
No signup required. Choose a topic and review each answer.
Start practicing →About these practice questions
These are original study questions written from published exam objectives—not recalled, copied, or confidential live-exam items. Always confirm current coverage with the official sources linked on this page.
Exam format and study resources
The AI-901 is administered by Microsoft, with a 700/1000 result.
This free AI-901 practice test has 65 original questions written to Microsoft's official content outline, last checked against it on July 18, 2026. Every question shows a worked explanation, and nothing here requires a signup.
Browse all questions & answers
1. A retailer wants a system that can look at photos of shelves and count how many bottles of a specific product are present. Which Azure AI workload category best describes this requirement?
- A. Computer vision
- B. Natural language processing
- C. Knowledge mining
- D. Generative AI
Show answer & explanation
Answer: A
Counting objects in images is an object detection task, which falls under computer vision. Natural language processing deals with text and speech, not images; knowledge mining is about extracting structured information from large document repositories; generative AI creates new content rather than analyzing existing images for object counts.2. A hospital chatbot needs to summarize a patient's symptoms described in free-text messages and route the message to the correct department. Which core AI capability is primarily being used?
- A. Natural language processing
- B. Computer vision
- C. Anomaly detection
- D. Reinforcement learning
Show answer & explanation
Answer: A
Understanding and summarizing free-text human language, and classifying its intent for routing, are classic natural language processing tasks. Computer vision applies to images, anomaly detection identifies unusual data patterns, and reinforcement learning trains agents through trial-and-error rewards rather than text understanding.3. Which of the following best describes the 'fairness' principle of responsible AI as commonly defined by Microsoft?
- A. AI systems should always produce the same output for the same input regardless of context
- B. AI systems should treat all people fairly and not reinforce or amplify societal biases against particular groups
- C. AI systems should maximize accuracy scores even if it means excluding certain demographic groups from training data
- D. AI systems should be open source so any developer can inspect the code
Show answer & explanation
Answer: B
Fairness means an AI system's decisions or predictions should not disadvantage or discriminate against particular groups based on characteristics such as gender, ethnicity, or age. Deterministic output is a reproducibility property, not fairness; being open source relates to transparency-adjacent practices, not fairness itself; and deliberately excluding demographic groups is the opposite of fairness.4. A loan-approval model built with machine learning denies a disproportionate number of applications from a particular zip code, even after removing zip code as a direct feature. Which responsible AI principle is most directly at risk here?
- A. Fairness
- B. Inclusiveness
- C. Reliability and safety
- D. Privacy and security
Show answer & explanation
Answer: A
A model producing systematically different outcomes correlated with a protected or proxy attribute (like zip code standing in for demographics) is a fairness problem, even if the sensitive feature was removed, because proxy variables can still encode bias. Inclusiveness concerns designing for people with diverse abilities and needs; reliability and safety concerns consistent, safe operation under expected conditions; privacy and security concerns data protection, not biased outcomes.5. In the context of Azure AI, what is the primary purpose of the 'transparency' principle of responsible AI?
- A. Ensuring the AI system runs on transparent, glass-walled data centers
- B. Ensuring the AI system never uses encrypted data
- C. Ensuring the AI model's source code is published publicly
- D. Ensuring people understand how and why an AI system makes the decisions or predictions it does
Show answer & explanation
Answer: D
Transparency in responsible AI is about making a system's behavior understandable and communicating its capabilities, limitations, and decision rationale to stakeholders. It does not require publishing source code, has nothing to do with physical data center design, and is unrelated to encryption practices, which fall under security.6. A company trains a machine learning model to predict the price of a used car based on features like mileage, age, and brand. What type of machine learning task is this?
- A. Regression
- B. Classification
- C. Clustering
- D. Reinforcement learning
Show answer & explanation
Answer: A
Predicting a continuous numeric value, like price, is a regression task. Classification predicts a discrete category or label rather than a continuous number; clustering groups similar unlabeled data points without predicting a target value; reinforcement learning trains an agent through rewards and penalties over time, not by fitting a single continuous output.7. A retailer has unlabeled customer purchase histories and wants to discover natural groupings of customers with similar buying behavior, without predefining the groups. Which machine learning approach is appropriate?
- A. Clustering
- B. Classification
- C. Regression
- D. Supervised learning with labeled outcomes
Show answer & explanation
Answer: A
Clustering is an unsupervised technique that groups unlabeled data points based on similarity, which fits discovering unknown customer segments. Classification and supervised learning with labeled outcomes both require predefined labels, which don't exist here; regression predicts a continuous numeric value rather than discovering groupings.8. Which statement best distinguishes a large language model (LLM) used in generative AI from a traditional supervised classification model?
- A. An LLM requires no training data, while classification models always require millions of labeled examples
- B. An LLM generates new, coherent content such as text based on learned patterns, while a classification model assigns input data to one of a fixed set of predefined categories
- C. An LLM and a classification model are functionally identical and differ only in naming convention
- D. An LLM can only process numeric data, while classification models can only process text
Show answer & explanation
Answer: B
Generative models like LLMs produce new content (text, code, etc.) by predicting likely continuations based on patterns learned from vast training data, whereas classification models output a label from a fixed, predefined set of categories. LLMs are trained on massive text corpora, not 'no data'; they typically process text (and increasingly other modalities), not just numbers; and the two model types serve fundamentally different purposes, not just different names.9. What best describes a 'prompt' in the context of working with a generative AI language model?
- A. The input text or instruction a user provides to guide the model's generated output
- B. A fixed configuration file that must be compiled before the model can run
- C. The numeric weight assigned to each neuron in the model
- D. A hardware accelerator used to train the model faster
Show answer & explanation
Answer: A
A prompt is the natural-language input or instruction given to a generative model to elicit a desired response; prompt design and phrasing directly influence output quality. It is not a compiled configuration file, not a neural network weight, and not a piece of training hardware.10. What is the key difference between an 'agentic AI' system and a standard single-turn generative AI chatbot?
- A. An agentic AI system cannot access any external data source
- B. An agentic AI system is always slower and less accurate than a chatbot
- C. An agentic AI system only accepts voice input, never text
- D. An agentic AI system can autonomously plan and execute multi-step actions, potentially invoking tools, to accomplish a goal
Show answer & explanation
Answer: D
Agentic AI extends generative models with the ability to reason over a goal, break it into steps, and autonomously call tools or APIs to complete tasks, rather than just responding to a single prompt. It is not limited to voice input, it can access external data sources (often through connected tools), and speed or accuracy depends on implementation, not an inherent property of being agentic.11. A team is deciding whether their new customer support automation project qualifies as an 'AI workload.' Which of these scenarios is the clearest example of an AI workload?
- A. A system that analyzes incoming support tickets to automatically detect sentiment and route urgent complaints
- B. A script that copies files from one folder to another on a fixed nightly schedule
- C. A web page that displays static company contact information
- D. A spreadsheet formula that sums a column of expense totals
Show answer & explanation
Answer: A
Automatically detecting sentiment and routing based on inferred meaning requires a trained model interpreting unstructured text, which is a natural language processing AI workload. File copying on a schedule, a static webpage, and a spreadsheet sum are all deterministic, rule-based operations with no learning or inference involved.12. Within Microsoft Foundry, what is the primary purpose of the model catalog?
- A. To let developers browse, compare, and select from a curated collection of foundation models to use in their applications
- B. To store billing invoices for Azure subscriptions
- C. To provide a directory of certified human customer support agents
- D. To list available virtual machine sizes for general compute workloads
Show answer & explanation
Answer: A
The model catalog in Microsoft Foundry gives developers a searchable collection of foundation models (from Microsoft and partners) with details on capabilities and benchmarks, so they can select the right model for their scenario. It is not a billing tool, a staffing directory, or a general-purpose VM size list, which are unrelated Azure concerns.13. A developer wants to build a solution using Microsoft Foundry that combines a language model with the ability to call external APIs and retain conversational context across steps. Which Foundry capability supports this scenario?
- A. Agent and orchestration capabilities for building agentic applications
- B. Static image thumbnail generation only
- C. Manual SQL query execution with no model involvement
- D. A fixed spreadsheet template with no code
Show answer & explanation
Answer: A
Microsoft Foundry provides agent-building and orchestration features that let a language model call tools/APIs and maintain state across a multi-step interaction, enabling agentic applications. Thumbnail generation is unrelated to conversational orchestration, manual SQL execution bypasses model involvement entirely, and a static spreadsheet template has no dynamic model or tool-calling capability.14. A developer wants to add a feature to an app that converts spoken customer calls into written transcripts in near real time. Which category of Foundry capability should they use?
- A. Speech-to-text (speech recognition) services
- B. Text-to-image generation services
- C. Optical character recognition for scanned documents only
- D. Anomaly detection for time-series sensor data
Show answer & explanation
Answer: A
Converting spoken audio into written text is the function of speech-to-text (speech recognition) services available through Microsoft Foundry's speech capabilities. Text-to-image generation creates pictures from text prompts, not transcripts; OCR extracts text from images of documents, not live audio; and anomaly detection identifies unusual patterns in numeric time-series data, unrelated to transcription.15. A developer needs their application to read text aloud to visually impaired users in a natural-sounding voice. Which Foundry capability addresses this need?
- A. Text-to-speech synthesis
- B. Speech-to-text transcription
- C. Named entity recognition
- D. Image captioning
Show answer & explanation
Answer: A
Text-to-speech synthesis converts written text into spoken audio output, which is exactly what's needed to read content aloud. Speech-to-text does the reverse (audio to text); named entity recognition extracts entities like names and dates from text, it doesn't produce audio; and image captioning describes images in text, unrelated to reading text aloud.16. A developer is building a solution to extract vendor names, invoice totals, and dates from scanned PDF invoices and load them into a database. Which Foundry capability is best suited to this task?
- A. Document intelligence / information extraction for structured data extraction from documents
- B. Text-to-speech synthesis
- C. Image generation from text prompts
- D. Sentiment analysis of social media posts
Show answer & explanation
Answer: A
Extracting structured fields like vendor name, totals, and dates from scanned documents is the core purpose of document intelligence and information extraction capabilities, which combine OCR with field-level understanding. Text-to-speech produces audio, not structured data; image generation creates new pictures rather than reading existing ones; sentiment analysis judges emotional tone in text, not extracting invoice fields.17. A marketing team wants to generate several original product photos from text descriptions for a new ad campaign, without hiring a photographer. Which Foundry capability fits this need?
- A. Image generation from text prompts
- B. Optical character recognition
- C. Speech translation
- D. Key phrase extraction
Show answer & explanation
Answer: A
Generating new images from text descriptions is precisely what text-to-image generation models provide, letting teams create original visuals without traditional photography. OCR reads text out of existing images rather than creating new ones, speech translation converts spoken language between languages, and key phrase extraction pulls important terms out of text, none of which produce images.18. A developer building a computer vision feature needs the model to draw bounding boxes around every car, pedestrian, and traffic sign it detects in a street photo. Which computer vision capability is required?
- A. Object detection
- B. Image classification
- C. Optical character recognition
- D. Image captioning
Show answer & explanation
Answer: A
Object detection identifies multiple objects within an image and returns their locations as bounding boxes, matching the need to locate cars, pedestrians, and signs individually. Image classification assigns a single label to the whole image without locating objects; OCR reads text characters, not general objects; and image captioning produces a single descriptive sentence rather than bounding boxes for each object.19. A city planning department wants a model that labels every individual pixel in aerial photos as 'road,' 'building,' 'vegetation,' or 'water.' Which computer vision technique should be used?
- A. Semantic segmentation
- B. Image classification
- C. Optical character recognition
- D. Facial recognition
Show answer & explanation
Answer: A
Semantic segmentation classifies every pixel in an image into a category, producing a pixel-level map, which fits labeling roads, buildings, vegetation, and water across an aerial photo. Image classification only assigns one label to the whole image, not per pixel; OCR is for reading text characters; facial recognition identifies human faces, irrelevant to land-cover mapping.20. A developer wants to add a feature that reads printed text from photographs of street signs and receipts and converts it into machine-readable text. Which capability should they use?
- A. Optical character recognition (OCR)
- B. Sentiment analysis
- C. Speech synthesis
- D. Anomaly detection
Show answer & explanation
Answer: A
Optical character recognition is designed specifically to detect and extract printed or handwritten text from images and convert it into machine-readable text. Sentiment analysis judges emotional tone in text that already exists digitally, speech synthesis converts text to audio, and anomaly detection flags unusual data patterns, none of which extract text from images.21. A company wants its support app to automatically determine whether a customer review expresses a positive, negative, or neutral opinion. Which natural language processing capability should be implemented?
- A. Sentiment analysis
- B. Named entity recognition
- C. Language translation
- D. Text-to-speech synthesis
Show answer & explanation
Answer: A
Sentiment analysis specifically evaluates text to determine the emotional tone, such as positive, negative, or neutral. Named entity recognition extracts entities like names, places, and organizations rather than tone; language translation converts text between languages; text-to-speech converts text into spoken audio, none of which assess opinion polarity.22. A legal document review tool needs to automatically identify and tag all mentions of people, organizations, dates, and monetary amounts within contracts. Which NLP capability is required?
- A. Named entity recognition
- B. Sentiment analysis
- C. Language detection
- D. Speech-to-text transcription
Show answer & explanation
Answer: A
Named entity recognition identifies and categorizes specific entities in text, such as people, organizations, dates, and monetary values, exactly matching the contract-tagging requirement. Sentiment analysis measures opinion polarity, not entity extraction; language detection identifies which language text is written in; and speech-to-text converts audio to text, unrelated to tagging entities within existing text.23. A global company wants its support ticketing system to automatically convert incoming French customer messages into English for its support agents. Which NLP capability should be used?
- A. Machine translation
- B. Key phrase extraction
- C. Text summarization
- D. Speech synthesis
Show answer & explanation
Answer: A
Machine translation converts text from one language to another, which is exactly what's needed to turn French messages into English. Key phrase extraction pulls out important terms but doesn't translate language; text summarization condenses content length rather than changing language; speech synthesis converts text into audio, not between languages.24. A news aggregator wants to automatically produce a two-sentence summary of each long article for a mobile app preview. Which capability addresses this need?
- A. Text summarization
- B. Named entity recognition
- C. Object detection
- D. Facial recognition
Show answer & explanation
Answer: A
Text summarization condenses longer text into a shorter version while retaining key meaning, which fits generating brief article previews. Named entity recognition extracts entities rather than condensing text; object detection and facial recognition are computer vision tasks applied to images, unrelated to summarizing written articles.25. A team building a customer-facing chatbot with a large language model in Microsoft Foundry is concerned the model might state incorrect facts confidently. What is this phenomenon commonly called, and what technique helps mitigate it?
- A. Hallucination; mitigated by grounding the model's responses in verified source data (such as retrieval-augmented generation)
- B. Overfitting; mitigated by increasing the learning rate during inference
- C. Latency; mitigated by using a smaller context window
- D. Underfitting; mitigated by removing all training data
Show answer & explanation
Answer: A
When a generative model produces plausible-sounding but factually incorrect or fabricated content, it's called hallucination, and grounding techniques like retrieval-augmented generation, which supply verified source documents at inference time, help reduce it. Overfitting and underfitting are training-time model-fit problems unrelated to runtime factual errors, learning rate is a training hyperparameter not adjustable at inference, and latency concerns response speed, not factual accuracy.26. A developer wants a Foundry-based generative AI application to answer questions using the company's internal knowledge base rather than relying solely on the model's pretrained knowledge. Which architecture pattern should they implement?
- A. Retrieval-augmented generation (RAG), retrieving relevant internal documents and injecting them into the model's context
- B. Reinforcement learning from scratch on raw pixel data
- C. A rule-based decision tree with no language model involved
- D. Simple image classification of scanned pages
Show answer & explanation
Answer: A
Retrieval-augmented generation retrieves relevant passages from an external knowledge source and supplies them as context to a language model, grounding its answers in company-specific data rather than only pretrained knowledge. Training reinforcement learning from raw pixels is unrelated to text Q&A, a rule-based decision tree bypasses the language model entirely, and image classification categorizes images rather than answering knowledge questions.27. When configuring a generative AI model deployment in Microsoft Foundry, what is the general effect of lowering the 'temperature' parameter?
- A. The model switches to a completely different underlying architecture
- B. The model begins ignoring the user's prompt entirely
- C. The model's outputs become more deterministic and focused, favoring the most likely next tokens
- D. The model's response time increases exponentially
Show answer & explanation
Answer: C
Temperature controls the randomness of token selection during generation; lowering it makes the model favor higher-probability tokens, producing more consistent, focused output, while raising it increases variety and creativity. It does not cause the model to ignore prompts, does not directly control response latency, and does not change the underlying model architecture, which is a separate deployment choice.28. A developer is choosing between deploying a small, efficient language model versus a large frontier model for a simple FAQ-answering feature within Microsoft Foundry. Which factor most directly favors choosing the smaller model?
- A. Lower compute cost and faster response times for a relatively simple task
- B. Small models always produce more creative and original content than large models
- C. Small models require no prompt engineering at all
- D. Small models can process unlimited context length compared to large models
Show answer & explanation
Answer: A
Smaller models generally cost less to run and respond faster, making them well suited to simpler tasks like FAQ answering where the full reasoning power of a large frontier model isn't required. Creativity and originality are not inherently tied to model size in the smaller model's favor, prompt engineering is still valuable regardless of model size, and smaller models typically support shorter, not unlimited, context windows compared to larger frontier models.29. A developer deploying an AI solution through Microsoft Foundry wants to apply content filters that automatically block outputs containing hate speech, violence, or self-harm content before they reach end users. What is this capability an example of?
- A. Content safety / harm mitigation controls built into the AI platform
- B. Data encryption at rest
- C. Autoscaling of compute resources
- D. Model version rollback
Show answer & explanation
Answer: A
Automated filters that detect and block categories of harmful content, such as hate speech or self-harm content, before it reaches users are content safety and harm mitigation controls, a key responsible AI implementation practice. Encryption at rest protects stored data, autoscaling manages compute capacity under load, and version rollback reverts to a prior model deployment, none of which filter harmful output content.30. A team wants their AI-powered form-processing app to keep working correctly and produce consistent results even when given slightly noisy or low-quality scanned documents. Which responsible AI principle does this primarily address?
- A. Reliability and safety
- B. Inclusiveness
- C. Accountability
- D. Privacy and security
Show answer & explanation
Answer: A
Reliability and safety concerns an AI system performing consistently and safely, including under less-than-ideal conditions such as noisy input, without unexpected failures. Inclusiveness is about empowering people of diverse abilities and backgrounds, accountability concerns having clear human responsibility for AI system outcomes, and privacy and security concerns protecting data, none of which specifically address robustness to noisy input.31. Which scenario best illustrates the responsible AI principle of 'accountability'?
- A. An AI model is deployed without any human oversight because it is assumed to be fully autonomous
- B. An AI model's training data is deleted immediately after training to save storage costs
- C. An organization designates specific people to be responsible for reviewing and governing the behavior of its deployed AI system
- D. An AI system's internal logic is kept completely secret from regulators
Show answer & explanation
Answer: C
Accountability means people and organizations remain responsible for how an AI system operates and are able to govern, audit, and correct its behavior. Deploying without human oversight, hiding internal logic from regulators, and deleting training data all reduce oversight and traceability, working against accountability rather than demonstrating it.32. A voice-assistant vendor wants its speech recognition system to work equally well for users with different accents, speech patterns, and disabilities affecting speech. Which responsible AI principle does this reflect?
- A. Inclusiveness
- B. Transparency
- C. Privacy and security
- D. Reliability and safety
Show answer & explanation
Answer: A
Inclusiveness means designing AI systems to bring benefits to and work well for people of all backgrounds, abilities, and characteristics, which matches supporting varied accents and speech patterns. Transparency is about explaining how the system works, privacy and security is about protecting data, and reliability and safety is about consistent safe operation, none of which specifically target designing for diverse user populations.33. When Azure AI-901 candidates schedule their certification exam, through which provider is the exam typically scheduled?
- A. The candidate's own employer's HR department
- B. ProctorU only, with no test center option
- C. Pearson VUE
- D. Prometric
Show answer & explanation
Answer: C
Microsoft certification exams such as AI-901 are scheduled through the exam delivery provider Pearson VUE from the exam or certification details page. Prometric is a different, unrelated testing vendor used by other certification programs; the exam is not exclusively ProctorU-only since a local test center option also exists; and employer HR departments do not schedule Microsoft certification exams.34. A candidate fails their first attempt at the AI-901 exam. How long must they wait before they are permitted to retake it?
- A. 24 hours
- B. 7 days
- C. 30 days
- D. There is no waiting period; they may retake immediately
Show answer & explanation
Answer: A
Candidates can retake the exam 24 hours after their first attempt. Longer waiting periods like 7 or 30 days do not apply to the first retake, and there is in fact a mandatory minimum wait rather than immediate retake eligibility.35. A candidate whose preferred language is not available for the AI-901 exam wants extra time to complete it. How much additional time can they request?
- A. 30 minutes
- B. 15 minutes
- C. 60 minutes
- D. No additional time is available under any circumstance
Show answer & explanation
Answer: A
Candidates whose preferred language is unavailable can request an additional 30 minutes to complete the exam. Shorter or longer time allowances listed are not the documented accommodation, and additional time is in fact available under this specific circumstance.36. A system reads printed text from scanned invoices and returns it as machine-readable characters. Which AI workload category is this?
- A. Computer vision, specifically optical character recognition
- B. Natural language processing
- C. Anomaly detection
- D. Conversational AI
Show answer & explanation
Answer: A
Extracting characters from an image is a vision task, and language processing begins once the text exists and meaning must be derived from it. A document processing pipeline typically chains the two, with vision producing text and language services extracting entities and relationships from it.37. A model determines whether a customer review expresses a positive or negative opinion. Which capability is this?
- A. Sentiment analysis, a natural language processing capability classifying the opinion expressed in text
- B. Entity recognition, which identifies named things in text
- C. Key phrase extraction, which surfaces main topics
- D. Language detection, which identifies the language used
Show answer & explanation
Answer: A
Sentiment analysis classifies expressed opinion, while entity recognition finds people, places and organizations, key phrase extraction surfaces the main topics and language detection identifies which language the text is in. A review analysis pipeline commonly applies several of these together, since each answers a different question about the same text.38. A system predicts a house's sale price from its characteristics. Which type of machine learning problem is this?
- A. Regression, since the predicted value is continuous
- B. Classification, since the model assigns a category
- C. Clustering, since similar houses are grouped
- D. Anomaly detection, since unusual prices are identified
Show answer & explanation
Answer: A
Predicting a numeric quantity is regression, while predicting a category such as whether a house will sell within thirty days is classification. The distinction determines both the algorithm family and the evaluation metrics, since regression uses error measures such as mean absolute error while classification uses accuracy, precision and recall.39. A retailer groups customers with similar purchasing behaviour without predefined categories. Which technique is this?
- A. Clustering, an unsupervised technique that discovers groupings from the data's own structure
- B. Classification, which assigns items to predefined labels
- C. Regression, which predicts a numeric outcome
- D. Forecasting, which projects a series forward in time
Show answer & explanation
Answer: A
Clustering requires no labels because the groupings emerge from similarity in the data, which is why it suits segmentation where the useful categories are not known in advance. The resulting clusters carry no inherent meaning, so interpreting what each represents is a human step after the algorithm runs.40. A monitoring system flags sensor readings that deviate markedly from normal patterns. Which capability is this?
- A. Anomaly detection, which identifies data points departing from an established normal pattern
- B. Object detection, which locates items within an image
- C. Speech recognition, which converts audio to text
- D. Translation, which converts text between languages
Show answer & explanation
Answer: A
Anomaly detection learns what normal looks like and flags departures, which suits equipment monitoring, fraud screening and quality control. Its practical challenge is that anomalies are rare by definition, so a detector tuned too sensitively overwhelms operators with false alerts while one tuned too conservatively misses the events it exists to catch.41. A vision model must locate and draw a box around each vehicle in a photograph. Which capability is required?
- A. Object detection, which returns both a class and a bounding box location for each instance
- B. Image classification, which assigns a single label to the whole image
- C. Optical character recognition, which extracts text
- D. Face detection, which locates human faces specifically
Show answer & explanation
Answer: A
Classification answers what the image contains overall while detection answers what is present and where each instance is, which is why detection requires bounding box annotations in its training data. Semantic segmentation goes further still, classifying every pixel rather than drawing boxes, and each level of detail raises the labelling cost.42. A conversational assistant must recognize what a user wants to accomplish and extract the details needed to act. What are these two elements called?
- A. Intent, the user's goal, and entities, the parameters extracted from the utterance
- B. Entities, the user's goal, and intents, the parameters
- C. Prompts and completions
- D. Features and labels
Show answer & explanation
Answer: A
An utterance such as booking a flight to Lisbon on Friday carries a booking intent with destination and date entities, and the assistant needs both to act. Where a required entity is absent the assistant prompts for it, which is why designing the slot-filling conversation is as important as recognizing the intent.43. A team must decide whether to use a prebuilt AI service or train a custom model for image recognition. What favours the prebuilt service?
- A. The objects to be recognized are common ones the service already covers, so no training data or ML expertise is required
- B. The objects are highly specific to the organization's domain
- C. A large labelled dataset is already available
- D. Complete control over the model architecture is required
Show answer & explanation
Answer: A
Prebuilt services cover general categories well and remove the data and expertise barrier entirely, so they suit anything a general model already recognizes. Custom training earns its cost where the categories are organization-specific, such as identifying a manufacturer's own components, which a general model has never seen.44. A custom vision model is trained on 40 images per class and performs poorly on new photographs. What is the most likely issue?
- A. Insufficient and insufficiently varied training data, so the model has not learned to generalize across lighting, angle and background
- B. The images were stored in the wrong file format
- C. The model was trained for too many iterations
- D. The service does not support custom training
Show answer & explanation
Answer: A
Small training sets that do not span the variation present at inference produce models that work on images resembling the training set and fail otherwise. Variety across conditions matters as much as raw count, and data augmentation can extend a limited set by generating rotated, cropped and recoloured variants.45. A document processing service extracts fields from invoices with varying layouts. What capability makes this possible without a template per vendor?
- A. A prebuilt model trained to recognize common document types and their typical fields regardless of specific layout
- B. A fixed coordinate map applied to every document
- C. Manual data entry with spell checking
- D. Optical character recognition alone with no structural understanding
Show answer & explanation
Answer: A
Prebuilt document models understand that an invoice has a total, a due date and line items wherever they appear, which coordinate-based extraction cannot handle across varying layouts. Custom models remain useful for organization-specific forms the prebuilt models do not cover, and they can be trained on a modest number of labelled samples.46. A generative AI application must answer questions using the organization's internal documents. What approach supplies that knowledge without retraining the model?
- A. Retrieval augmented generation, where relevant document passages are retrieved and included in the prompt at query time
- B. Increasing the model's temperature setting
- C. Selecting a model with more parameters
- D. Asking the model to recall the documents from training
Show answer & explanation
Answer: A
Retrieval places the organization's content in the model's context at query time, so answers reflect current documents and can cite them. A larger model brings more general capability but no knowledge of private documents it never saw, which is why capability and knowledge are separate concerns.47. A generative model returns an answer that sounds authoritative but is not supported by the provided source documents. What should the application do?
- A. Require and display citations to the source passages, so unsupported claims are visible and verifiable by the reader
- B. Increase the response length to add detail
- C. Remove the source documents from the prompt
- D. Trust the answer, since the model was given sources
Show answer & explanation
Answer: A
Providing sources reduces but does not eliminate unsupported generation, so citation is what lets a reader check whether a claim actually appears in the cited passage. Instructing the model to answer only from the provided context and to state when the context is insufficient further reduces the problem without removing the need to verify.48. A responsible AI framework lists fairness as a principle. What does applying it require in practice?
- A. Evaluating whether the system's outcomes differ across groups and addressing disparities found, rather than assuming neutrality
- B. Removing all demographic attributes from the training data
- C. Using the largest available model
- D. Requiring users to accept terms of service
Show answer & explanation
Answer: A
Fairness is assessed by measuring outcomes across groups rather than assumed from the absence of explicit demographic features, because correlated proxies carry the same signal. Deciding which fairness definition applies is itself a judgment, since several reasonable definitions can conflict and cannot all be satisfied simultaneously.49. A responsible AI principle of transparency is applied to a deployed model. What does it call for?
- A. Making the system's purpose, capabilities and limitations understandable to those affected by it and to those operating it
- B. Publishing the model's complete source code
- C. Sharing the training data publicly
- D. Removing all logging to protect privacy
Show answer & explanation
Answer: A
Transparency is about comprehensibility to affected parties rather than disclosure of proprietary artefacts, so a clear statement of intended use, known limitations and evaluation results serves it better than source code would. Documenting what the system was not designed to do is often the most useful part, since misuse frequently comes from applying a system outside its evaluated scope.50. A facial analysis capability is proposed for a public space. Which responsible AI consideration is most acute?
- A. Privacy and consent, alongside the disproportionate consequences of errors for individuals who cannot contest them
- B. The compute cost of running inference
- C. The programming language used to call the service
- D. The file format of the captured images
Show answer & explanation
Answer: A
Biometric identification of people who have not consented and may not know they are being processed raises privacy questions that technical accuracy does not answer. Providers have restricted access to some facial recognition capabilities for this reason, and lawful basis, notice and a contestation route are prerequisites rather than refinements.51. A speech capability converts spoken audio into written text. What is this called, and what is its counterpart?
- A. Speech to text, with text to speech synthesizing spoken audio from written input
- B. Text to speech, with speech to text synthesizing audio
- C. Speech translation, with transcription as its counterpart
- D. Speaker recognition, with sentiment analysis as its counterpart
Show answer & explanation
Answer: A
Recognition converts audio to text and synthesis converts text to audio, while speech translation combines recognition, translation and optionally synthesis into a pipeline. Speaker recognition is the separate capability of identifying or verifying who is speaking rather than what was said.52. A knowledge base answers common customer questions from documented question-and-answer pairs. Which capability underpins this?
- A. Question answering over a curated knowledge base, matching a user's phrasing to the closest documented question
- B. Object detection over document images
- C. Time series forecasting
- D. Anomaly detection over query logs
Show answer & explanation
Answer: A
Question answering maps varied user phrasings to curated answers, which gives predictable and reviewable responses because the answer text is authored rather than generated. That predictability is why it remains preferred over free generation for regulated or high-stakes support content.53. A content safety capability is applied to a generative application. What does it evaluate?
- A. Both user input and model output against harm categories, so harmful content is blocked in either direction
- B. Only the model's output, since input comes from trusted users
- C. Only the model's input, since output derives from it
- D. Neither, since the model's training handles safety entirely
Show answer & explanation
Answer: A
Screening in both directions catches harmful user submissions and harmful generations independently, and severity thresholds per category let an application tune strictness to its audience. Relying on training alone leaves no organization-specific policy enforcement and no record of what was blocked.54. An AI service is called from an application. What credential practice is recommended?
- A. Publishing the key in the application's documentation
- B. Embedding the service key in the client-side application code
- C. Authenticating with a managed identity where supported, and otherwise storing keys in a secrets service rather than in code or configuration files
- D. Sharing one key across all environments for consistency
Show answer & explanation
Answer: C
Keys embedded in client-side code are visible to every user, and a key shared across environments means a development leak compromises production. Managed identity removes the credential entirely where supported, and regenerating keys periodically limits the window of exposure when one does leak.55. A team measures a classification model with a confusion matrix. What does it display?
- A. Counts of true positives, true negatives, false positives and false negatives, from which accuracy, precision and recall are derived
- B. The model's training loss over each epoch
- C. The distribution of feature values in the dataset
- D. The compute time consumed per prediction
Show answer & explanation
Answer: A
The matrix shows where predictions agreed and disagreed with reality in each direction, which is what makes the two error types visible separately. Aggregate accuracy hides that distinction, and the relative cost of a false positive against a false negative is usually what determines whether a model is fit for its purpose.56. A machine learning workflow includes feature engineering. What does this step accomplish?
- A. Transforming raw data into representations that expose the signal a model can learn from, such as deriving age from a date of birth
- B. Selecting which algorithm to use
- C. Deploying the trained model to production
- D. Monitoring the model after deployment
Show answer & explanation
Answer: A
Feature engineering frequently contributes more to model performance than algorithm selection, since a well-chosen representation makes the pattern learnable. Care is needed to avoid deriving features from information unavailable at prediction time, which produces leakage and evaluation results that cannot be reproduced in production.57. A deployed model must be monitored after release. What should be watched beyond service availability?
- A. Whether the distribution of incoming data and the model's prediction quality have shifted from the conditions it was trained and validated on
- B. Only the number of requests served per hour
- C. Only the size of the deployed model file
- D. Only the programming language of the client
Show answer & explanation
Answer: A
A model can be fully available while producing steadily worse predictions as the world changes, which uptime monitoring cannot detect. Tracking input distributions gives early warning without waiting for outcomes, while prediction quality requires ground-truth labels that may arrive with substantial delay.58. A team uses an automated machine learning capability. What does it do?
- A. Trials multiple algorithms and preprocessing combinations automatically and ranks them by a chosen metric, producing a candidate model
- B. Removes the need to prepare or label data
- C. Guarantees the resulting model is fit for production use
- D. Automatically collects the training data required
Show answer & explanation
Answer: A
Automated machine learning accelerates the search over algorithms and hyperparameters, which compresses experimentation that would otherwise take considerable time. The data preparation, the choice of evaluation metric and the judgment about fitness for purpose all remain human responsibilities, and the metric chosen determines what the search optimizes toward.59. A model is deployed as a real-time endpoint rather than for batch scoring. What determines this choice?
- A. Whether predictions are needed immediately in response to individual requests, rather than periodically over a large set of records
- B. The number of features the model uses
- C. The programming language used for training
- D. The size of the training dataset
Show answer & explanation
Answer: A
Real-time endpoints hold capacity ready to answer individual requests with low latency, while batch scoring processes accumulated records efficiently on a schedule and costs less per prediction. Deploying a real-time endpoint for a workload that only needs nightly scoring pays for idle capacity continuously.60. A generative model's response varies between identical requests. What setting most directly controls this?
- A. Temperature, which governs how much randomness is applied when selecting each next token
- B. The maximum response length
- C. The number of documents retrieved
- D. The deployment region of the service
Show answer & explanation
Answer: A
Temperature scales the sampling distribution, so lowering it toward zero makes output far more consistent while raising it increases variety. Consistency is desirable for extraction and classification and undesirable for creative generation, which is why the setting is chosen per task rather than set once globally.61. An organization must decide whether a proposed AI use case is appropriate. Which consideration is the most fundamental?
- A. Whether the implementation can be completed this quarter
- B. Whether competitors have deployed something similar
- C. Whether the newest available model is being used
- D. Whether the consequences of the system being wrong are acceptable and can be detected and corrected, given that no model is perfectly accurate
Show answer & explanation
Answer: D
Every model errs at some rate, so the decisive question is what happens when it does and whether that failure is detectable and reversible. Use cases where errors are invisible, irreversible or fall on people who cannot contest them demand far stronger safeguards than those where a mistake is noticed and corrected cheaply.62. A translation capability is applied to user-generated content. What limitation should be communicated?
- A. Machine translation quality varies by language pair and domain, and idiom, ambiguity and specialized terminology are common failure points
- B. Machine translation is equally accurate for every language pair
- C. Machine translation cannot handle any technical content
- D. Machine translation requires the user to specify the source language always
Show answer & explanation
Answer: A
Quality follows the volume of available training data for a pair, so widely spoken pairs perform far better than low-resource ones, and domain-specific terminology often needs a custom glossary. Automatic language detection removes the need to specify the source in most cases, which is a separate matter from output quality.63. A team evaluates whether to fine-tune a language model. What does fine-tuning best address?
- A. Teaching a consistent style, format or specialized task behaviour, rather than supplying facts that change frequently
- B. Keeping the model informed about daily changing data
- C. Reducing the model's inference latency to zero
- D. Removing the need for any evaluation of outputs
Show answer & explanation
Answer: A
Fine-tuning shapes how a model responds, which suits enforcing an output format or a domain style that prompting cannot achieve reliably. Facts that change belong in retrieval, since retraining every time a document is updated is impractical and the tuned weights cannot be selectively corrected.64. An AI solution processes personal data. What privacy practice applies at the design stage?
- A. Collecting and retaining only the data necessary for the stated purpose, and establishing the lawful basis before processing begins
- B. Relying on the AI provider's terms to satisfy all obligations
- C. Collecting as much data as possible in case it proves useful later
- D. Deferring privacy assessment until after deployment
Show answer & explanation
Answer: A
Data minimization and purpose limitation constrain what may be collected and how long it may be kept, and both are design-time decisions that become expensive to retrofit. The provider's terms address the provider's obligations rather than the organization's own lawful basis for processing the data in the first place.65. A team wants to compare two prompt variants for a generative task. What approach produces a defensible comparison?
- A. Evaluating both against the same representative test set using consistent criteria, rather than comparing impressions from a handful of examples
- B. Choosing whichever variant produced a better first response
- C. Selecting the longer prompt, since it provides more context
- D. Selecting the shorter prompt, since it costs less per request
Show answer & explanation
Answer: A
Generative output varies between runs, so a small number of impressions cannot distinguish a genuinely better prompt from favourable sampling. A fixed test set with defined criteria makes the comparison repeatable and lets a change be verified rather than assumed, which is the same discipline applied to model changes.
More in this family
Explore more Technology & IT Certifications
In the same family
More in this category
- Cisco Certified Network Associate (200-301 CCNA)Practice questions →
- Cisco Certified Support Technician (CCST) NetworkingPractice questions →
- Certified Information Systems Security Professional (CISSP)Practice questions →
- Certified Kubernetes Administrator (CKA)Practice questions →
- Claude Certified Associate – FoundationsPractice questions →
- CompTIA A+ (Core 1: 220-1201 and Core 2: 220-1202)Practice questions →
- CompTIA Cybersecurity Analyst+ (CySA+)Practice questions →
- CompTIA Network+Practice questions →
- CompTIA Security+ (SY0-701)Practice questions →
- Google Cloud Certified - Associate Cloud EngineerPractice questions →
2026 statistics
Key facts: AI-901 exam
Every free resource for this exam
Get a free AI-901 study plan
A week-by-week plan plus new practice questions, straight to your inbox.
Official sources
Primary documents used to verify the exam details shown on this page.
- Study guide for Exam AI-901: Microsoft Azure AI FundamentalsMicrosoftlearn.microsoft.com
- Exam AI-901: Microsoft Azure AI FundamentalsMicrosoftlearn.microsoft.com
- Register and schedule an exam (Pearson VUE)Microsoftlearn.microsoft.com
- Exam scoring and score reportsMicrosoftlearn.microsoft.com
- Microsoft Certified: Azure AI FundamentalsMicrosoftlearn.microsoft.com
Last verified against the official exam content outline:
Frequently asked questions
What score do I need to pass the Azure AI Fundamentals (AI-901) exam?
You need a score of 700 or greater to pass AI-901. Scores are reported on a scale of 1 to 1,000, and there is no penalty for guessing — no points are deducted for incorrect answers. That means you should answer every question on the exam, even when unsure, and on practice tests you should aim to score consistently above the 700 threshold before booking your exam date.
Which topics should I spend the most practice time on for AI-901?
AI-901 is assessed across two skill areas: Identify AI concepts and capabilities (40-45%) and Implement AI solutions by using Microsoft Foundry (55-60%). Since the Foundry implementation domain carries more weight, it deserves the larger share of your practice time. Within the concepts domain, expect questions on responsible AI principles, AI model components and configurations, and AI workloads including generative and agentic AI, text analysis, speech, and computer vision. Within the Foundry domain, practice implementing generative AI apps and agents, text and speech solutions, computer vision and image-generation capabilities, and information extraction.
What happens if I fail AI-901 — how soon can I retake it?
If you don't pass on your first attempt, you can retake the exam 24 hours later. In practice, though, a same-week retake only makes sense if you were close to the 700 passing mark. Use your score report to identify which of the two domains dragged you down, drill practice questions in that area, and rebook once your practice scores recover. Also note that the exam price is based on the country or region in which the exam is proctored, so check the current fee for your location before scheduling the retake.
How do I schedule the AI-901 exam, and can I take it online?
AI-901 is scheduled through the exam delivery provider Pearson VUE, and you can choose between taking it online with online proctored (OnVUE) delivery or in person at a local test center. Exams can be scheduled no more than 90 days in advance, and you can have a maximum of two Microsoft Certification exams scheduled at a time through Pearson VUE. The exam is offered in English; if your preferred language is unavailable, you can request an additional 30 minutes to complete the exam. Earning the certification requires passing just this one exam.