AI-900 Retired: Free Practice Questions and the AI-901 Shift

What Happened to AI-900?#

Microsoft's AI-900 exam page now carries a notice that the certification requirements have changed: the AI-900 exam was retired on 30 June 2026 and has been replaced by AI-901. To earn the Microsoft Certified: Azure AI Fundamentals certification, candidates must now pass AI-901.

The last AI-900 blueprint had five domains: AI workloads and considerations, fundamental principles of machine learning on Azure, computer vision workloads, natural language processing workloads, and generative AI workloads. AI-901 is organized differently and expects hands-on implementation, so treat AI-900 material as background rather than as a direct match for the new exam.

What Does AI-901 Measure?#

The AI-901 study guide, with skills measured as of 15 April 2026, has two domains.

Identify AI concepts and capabilities (40-45%)#

• Describe principles of responsible AI: fairness, reliability and safety, privacy and security, inclusiveness, transparency, and accountability
• Identify AI model components and configurations: how generative AI models work, choosing an appropriate model based on capabilities, and choosing deployment options and configuration parameters
• Identify AI workloads: scenarios for common workloads including generative and agentic AI, text analysis, speech, computer vision, and information extraction; text analysis techniques such as keyword extraction, entity detection, sentiment analysis, and summarization; speech recognition and synthesis; computer vision and image generation; and extracting information from text, images, audio, and video

Implement AI solutions by using Microsoft Foundry (55-60%)#

• Generative AI apps and agents: writing effective system and user prompts, deploying a model and interacting with it in the Foundry portal, building a lightweight chat client with the Foundry SDK, and creating and testing a single-agent solution
• Text and speech: building a lightweight application that includes text analysis, responding to spoken prompts with a multimodal model, and using Azure Speech in Foundry Tools
• Computer vision and image generation: interpreting visual input in prompts, creating visual outputs with generative models, and building a lightweight vision application
• Information extraction: extracting information from documents, forms, images, audio, and video with Azure Content Understanding in Foundry Tools

Microsoft's audience profile says you should have conceptual knowledge of AI solutions in Azure, foundational technical skills, knowledge of Python coding syntax and programming techniques, and familiarity with Azure resources.

How Much of This AI-900 Practice Still Helps?#

The first AI-901 domain is about AI concepts, such as responsible AI principles, how generative models work, and recognizing which AI workload fits a scenario. Questions on those ideas are a reasonable warm-up for that domain. The second domain, which is the larger one, calls for hands-on work: deploying a model, writing prompts, and building small client applications with the Foundry SDK.

So use the AI-900 questions to check your grasp of concepts and vocabulary, then spend most of your time on the AI-901 objectives above, in the Foundry portal and with the SDK. Compare each question you miss with the AI-901 skills list before you decide how much time it deserves.

Free AI-900 Practice Questions#

10 free AI-900 practice questions#

These are real questions from our 464-question AI-900 bank - the same pool the full practice exam draws from, not simplified samples. Answer before you expand the explanation.

Question 1#

A developer is starting to work with Azure AI Vision Studio for the first time. After successfully logging into the studio portal, what is a fundamental prerequisite or initial task they must complete to begin experimenting with image analysis features within that studio environment?

• A. Input a key and endpoint into the studio
• B. Customize the API request.
• C. Associate a resource with the studio

Show answer and explanation

Correct answer: C. Associate a resource with the studio

Azure AI Studios (like Vision Studio, Language Studio, etc.) generally require you to associate an existing Azure AI resource (e.g. an Azure AI Vision resource or a general Azure AI services resource) with the studio session. This resource provides the backend processing power and is where your usage will be billed. The key and endpoint are used by applications, but the studio needs this resource association. Inputting keys/endpoints is more for SDK/API usage. Customizing API requests happens after a resource is linked and a task is chosen.

Question 2#

A librarian at 'ArchiveCentral' is using an NLP tool to analyze a collection of historical speeches. The tool identifies and counts how often significant words like 'freedom,' 'democracy,' and 'justice' appear in each speech to understand their thematic emphasis. Which text analysis technique is primarily being used to determine the occurrence rates of these words?

• A. Tokenization, to split the speeches into sentences.
• B. Frequency analysis, to count word occurrences.
• C. Stemming, to reduce words to their root form.
• D. Entity recognition, to identify proper nouns.

Show answer and explanation

Correct answer: B. Frequency analysis, to count word occurrences.

Frequency analysis involves counting the occurrences of words (or tokens) within a document or corpus. This helps in understanding which words are most commonly used and can be an indicator of the main themes or topics, as in the librarian's case. Tokenization is a prerequisite. Stemming normalizes words. Entity recognition identifies specific categories of words.

Question 3#

An online retail platform, 'FashionFind,' allows users to upload a photo of a clothing item. The platform's AI then tries to identify the main item in the photo (e.g., 'summer dress,' 'leather jacket,' 'running shoes') to suggest similar products. What is this AI process of assigning a category or label to an image based on its visual content called?

• A. A mechanism for visual perception
• B. An array of pixel values
• C. A process of naming images
• D. A method for classifying images based on their contents (Image Classification)

Show answer and explanation

Correct answer: D. A method for classifying images based on their contents (Image Classification)

Image classification is a computer vision task where an AI model is trained to assign a predefined category or label (e.g., 'summer dress,' 'running shoes') to an image based on its visual features. This accurately describes what FashionFind's AI is doing.

The other approaches are insufficient for this use case. Representing images as arrays of pixel values describes the raw technical format of image data, but doesn't address the categorization process itself. Simply naming images is too general and doesn't capture the AI-driven analysis of visual content. Visual perception mechanisms describe the broader concept of how systems interpret visual information, but don't specifically define the classification process that assigns categorical labels based on image content.

Question 4#

A travel tech company, 'LinguaTrip,' is developing a mobile app feature that provides real-time voice translation for travelers. When a user speaks a phrase in their native language (e.g., English), the app should translate it into the local language (e.g., Japanese) and speak the translation aloud. For the translation to be natural and accurate, it's important that the AI considers the full meaning and context of the spoken phrase, not just a literal word-for-word conversion. What key aspect of advanced AI translation, as opposed to older, simpler methods, is crucial for achieving this high-quality, contextually aware translation?

• A. The use of comprehensive profanity filtering to remove offensive words from the translation output.
• B. The ability to translate into only one predefined destination language at a time to maintain focus and accuracy.
• C. The selective translation of only proper nouns (like names of people or places) while leaving common words untranslated to preserve some original flavor.
• D. Semantic translation, where the AI model considers the overall context and meaning of the source phrase to produce a more natural and accurate translation in the target language, rather than a purely literal, word-for-word conversion.

Show answer and explanation

Correct answer: D. Semantic translation, where the AI model considers the overall context and meaning of the source phrase to produce a more natural and accurate translation in the target language, rather than a purely literal, word-for-word conversion.

Modern AI translation systems, especially those based on neural machine translation (NMT), strive for semantic translation. This means they analyze the meaning and context of the source text or speech as a whole, rather than just translating words or short phrases in isolation. This contextual understanding allows them to produce translations that are more fluent, natural-sounding, and accurate in conveying the intended meaning in the target language. Literal translation often results in awkward or incorrect translations. Profanity filtering and selective translation are supplementary features, not the core principle for contextual accuracy. Limiting to one language is an operational choice, not a quality aspect.

Question 5#

A corporation is building an AI system to make hiring recommendations. To ensure the system does not unfairly favor candidates from certain universities or backgrounds due to biases in the historical hiring data it was trained on, which Responsible AI principle must be rigorously applied and monitored?

• A. Privacy and Security
• B. Fairness
• C. Reliability and Safety
• D. Transparency

Show answer and explanation

Correct answer: B. Fairness

Fairness is the Responsible AI principle that aims to avoid bias in AI solutions and ensure equitable treatment. If an AI hiring system shows bias due to skewed training data, it violates this principle. Rigorous application and monitoring for fairness are crucial. Reliability ensures it works as intended. Privacy protects data. Transparency explains how it works.

Question 6#

A data science team has finalized a training pipeline in Azure ML Designer for a regression model. To enable other applications to get real-time predictions from this trained model, the team needs to convert this training pipeline into a different type of pipeline suitable for live inferencing. What is a key transformation that occurs during this conversion process from a training pipeline to a real-time inference pipeline?

• A. It primarily involves adding more sophisticated and computationally intensive data pre-processing components to handle raw input data.
• B. It automatically configures the pipeline to retrain the underlying model with new incoming data each time a prediction request is received, ensuring continuous learning.
• C. It focuses on reconfiguring the pipeline for efficient batch scoring of very large datasets in an offline manner, rather than for handling individual real-time prediction requests.
• D. It typically removes components specific to model training and evaluation (like 'Train Model' or 'Evaluate Model') and adds web service input and output points to accept prediction requests and return results.

Show answer and explanation

Correct answer: D. It typically removes components specific to model training and evaluation (like 'Train Model' or 'Evaluate Model') and adds web service input and output points to accept prediction requests and return results.

Converting a training pipeline to a real-time inference pipeline in Azure ML Designer involves adapting it for prediction. This means removing modules that are only relevant for training (like model training algorithms, data splitting for training/validation, and evaluation modules) and adding specific web service input and output components. These allow the deployed pipeline to receive data for a new prediction request and return the model's prediction. It doesn't necessarily add more preprocessing. It doesn't retrain on each call. It's specifically for real-time, not batch.

Question 7#

CodeGenius Software is developing a new mobile application that will integrate features from Azure Computer Vision for image analysis and Azure Speech service for voice commands. Their lead developer needs to ensure the application can securely and correctly connect to these Azure services. What are the two primary pieces of information the mobile application will require from each provisioned Cognitive Service resource in Azure to authenticate and make API requests?

• A. An API Key (Subscription Key) and the service Endpoint URL
• B. A User ID and Password for an Azure Active Directory account
• C. The Resource Group name and the Azure Region
• D. The Azure Subscription ID and the Cognitive Service's deployment name

Show answer and explanation

Correct answer: A. An API Key (Subscription Key) and the service Endpoint URL

To enable client applications, like CodeGenius Software's mobile app, to use Azure Cognitive Services, they primarily need two critical pieces of information for each service:

1. An API Key (often referred to as a Subscription Key): This key is used to authenticate requests to the service.
2. The service Endpoint URL: This is the specific web address to which the application sends its API requests for that particular Cognitive Service. These are fundamental for accessing the functionalities of services like Computer Vision and Speech. Resource Group and Region are important for management but not directly used in the API call from the client for authentication/connection in this primary way. User ID/Password for Azure AD is an alternative authentication mechanism but key-based is standard. Subscription ID and deployment name are for management and identification within Azure.

Question 8#

CloudCompute Inc wants to use Azure Machine Learning to automatically provision and manage computing resources that can scale up for training large models and scale down when not in use. What type of Azure ML capability does this describe?

• A. Model management services
• B. Data and compute services
• C. Model deployment services
• D. Automated machine learning

Show answer and explanation

Correct answer: B. Data and compute services

Data and compute services is correct because this describes the infrastructure and computational resources needed for machine learning workloads. Azure ML provides managed compute instances, compute clusters, and other scalable computing resources that can automatically adjust based on workload demands for training and inference.

Question 9#

A global marketing firm, 'MarketPulse,' analyzes social media posts from around the world to understand public opinion on new product launches. Their AI system needs to interpret the underlying meaning and sentiment of these posts, which are often short, informal, and use colloquial language. Which branch of AI is specifically focused on enabling computers to understand and process human language in this manner?

• A. To compress text files for efficient storage on servers.
• B. To convert textual content directly into spoken audio.
• C. To interpret the subject of text and its contextual meaning, similar to how humans do (Natural Language Processing).
• D. To encrypt text for secure transmission across networks.

Show answer and explanation

Correct answer: C. To interpret the subject of text and its contextual meaning, similar to how humans do (Natural Language Processing).

Natural Language Processing (NLP) is the area of AI dedicated to enabling computers to process, understand, interpret, and generate human language (both written and spoken). Analyzing social media posts for meaning and sentiment is a core NLP task. Encryption is for security. Compression is for storage. Text-to-speech is a specific NLP output task.

Question 10#

A developer is building a sophisticated Conversational Language Understanding (CLU) model. They anticipate users might ask questions or make statements completely unrelated to the application's defined functions (e.g., asking a smart light control app about the weather). How does CLU typically handle such utterances that don't align with any specifically defined purpose?

• A. It attempts to find the closest matching pre-built intent from common scenarios.
• B. It requires the developer to explicitly train the language model to ignore such phrases.
• C. It maps them to the 'None' intent, which serves as a fallback for unrecognized utterances.
• D. It automatically creates a new entity type for these unclassifiable utterances.

Show answer and explanation

Correct answer: C. It maps them to the 'None' intent, which serves as a fallback for unrecognized utterances.

The 'None' intent in Conversational Language Understanding is specifically designed as a fallback mechanism. It is used to handle utterances that do not match any of the other intents explicitly defined and trained by the developer. This helps the application gracefully manage input that is outside its designed scope. Pre-built intents are for specific domains. Explicit training to ignore is less scalable than a 'None' intent. Creating new entities is not the standard handling for unmatched intents.

How Should I Use These Questions?#

Work through them without checking the answers, then score yourself. For every question, including the ones you got right, read the explanation for each option and make sure you can say why the others are wrong. A question you answered correctly for the wrong reason is a gap.

Treat your score as a sign of how well you know the concepts, not as a prediction of how you will do on AI-901. For the exam itself, use Microsoft's materials: the AI-901 study guide, the exam sandbox for the interface, and the practice assessment on AI Skills Navigator, which you need to sign in to launch.

How Do I Prepare for AI-901?#

Start from the study guide and work down the objectives, rating yourself on each. Then get hands-on, because more than half of the exam calls for implementing solutions: deploy a model in the Foundry portal and chat with it, write and refine system and user prompts, build a small client with the Foundry SDK, create and test a single agent, and try text analysis, speech, vision, and Content Understanding.

Because AI-901 expects some Python, make sure you can read short Python snippets that call an SDK. Then take the practice assessment, review the objectives you missed, and repeat. Microsoft's guide notes that most questions cover generally available features, and that the exam may contain questions on preview features if they are commonly used, so keep an eye on the documentation for the services you practise with.

What Should I Know About the Certification Itself?#

The certification is Microsoft Certified: Azure AI Fundamentals, and the exam that earns it is now AI-901. Microsoft notes that the fundamentals credential can help you prepare for role-based certifications, but it is not a prerequisite for any of them. The passing score is 700. If the exam is not offered in your preferred language, you can ask for an additional 30 minutes.

For more on the other Microsoft exams, see the Microsoft certification practice exams page. For the domains of the retired exam, the earlier AI-900 exam guide has the detail.

Official domain weights#

Frequently Asked Questions#

Is AI-900 still available? No. Microsoft retired AI-900 on 30 June 2026.

What replaced AI-900? AI-901. Candidates must now pass AI-901 to earn Microsoft Certified: Azure AI Fundamentals.

How is AI-901 organized? Two domains: Identify AI concepts and capabilities (40-45%) and Implement AI solutions by using Microsoft Foundry (55-60%).

What is the passing score for AI-901? 700.

Does AI-901 require coding? Microsoft's audience profile says you need knowledge of Python coding syntax and programming techniques, and familiarity with Azure resources.

Are these AI-900 questions still worth doing? They are useful for checking concepts and vocabulary, especially for the AI concepts domain. Use the AI-901 study guide to decide how much time each topic deserves, and put most of your time into hands-on work with Microsoft Foundry.

Ready to try the questions? Start with free AI-900 practice questions at CertCompanion.

Checked against official Microsoft AI-900 exam documentation and current candidate reports. Last verified 2026-10-30.