AI-901: Azure AI Fundamentals Practice Exam

What the AI-901 exam covers

• Fundamentals of AI and responsible AI135 questions
• Fundamentals of generative AI and language models150 questions
• Build and manage AI solutions with Microsoft Foundry150 questions
• AI agents and agentic solutions135 questions
• Computer vision, natural language, speech, and document intelligence on Azure150 questions

Free AI-901 sample questions

A sample of 10 questions with answers and explanations. Sign up free to practice all 720.

1. Question 1Fundamentals of AI and responsible AI

   In supervised machine learning, what is the role of a label in a training dataset?

   • AThe known output value the model is trained to predictCorrect
   • BThe input variable the model uses to make a prediction
   • CThe numeric weight assigned to each neuron in a network
   • DThe score that measures how accurate a prediction was
   ✓ Correct answer: A

   In supervised learning each training example pairs input features with a label, which is the correct answer for that example. The model adjusts itself to map features to labels, and after training it predicts labels for new, unlabeled data.

   Why the other options are wrong

   • BThat describes a feature, which is an input variable, not the target output.
   • CWeights are internal model parameters learned during training, not labels in the dataset.
   • DThat describes an evaluation metric such as accuracy, not a label.
2. Question 2Fundamentals of AI and responsible AI

   Which of the following is the best example of a scenario suited to clustering rather than classification?

   • ASegmenting shoppers into unknown groups based on behavior, with no predefined labelsCorrect
   • BDeciding whether a transaction is fraudulent using historical fraud labels
   • CPredicting tomorrow's temperature in degrees
   • DSorting emails into predefined folders that were labeled by users
   ✓ Correct answer: A

   Clustering is used when there are no predefined labels and the goal is to discover natural groupings in the data. Segmenting shoppers into unknown groups with no labels fits this because the categories are not defined in advance. Classification, by contrast, needs labeled examples of known categories.

   Why the other options are wrong

   • BDeciding fraud from historical fraud labels uses labeled data, which is classification.
   • CPredicting a temperature in degrees is a continuous numeric output, which is regression.
   • DSorting emails into user-labeled folders relies on predefined labels, which is classification.
3. Question 3Fundamentals of generative AI and language modelsSelect all that apply

   Which of the following are true statements about tokens in a large language model? Each correct answer presents a complete statement.

   • ABoth the input prompt and the generated output consume tokensCorrect
   • BA token can represent a whole word or part of a wordCorrect
   • CA token always corresponds to exactly one character
   • DTokens are irrelevant to how usage of the model is billed
   ✓ Correct answer: A, B

   Language models count tokens for both the prompt and the completion, and total token usage typically determines cost. Each token can be a whole word, a subword fragment, or punctuation depending on how the tokenizer segments the text.

   Why the other options are wrong

   • CA token does not always map to one character; it commonly represents a subword of several characters.
   • DTokens are central to billing because usage is generally metered by the number of input and output tokens.
4. Question 4Fundamentals of generative AI and language modelsSelect all that apply

   Which components are typically part of a RAG solution built on Azure? (Choose three.)

   • AA retrieval source such as an Azure AI Search indexCorrect
   • BA generative model such as an Azure OpenAI chat deploymentCorrect
   • CAn embedding model to vectorize content and queriesCorrect
   • DA dedicated physical GPU that the customer must install on-premises
   ✓ Correct answer: A, B, C

   A typical Azure RAG solution uses a retrieval source such as an Azure AI Search index, an embedding model to vectorize documents and queries, and a generative chat model to produce the grounded answer. These cloud services work together in the retrieve-then-generate flow. No customer-installed on-premises GPU is required.

   Why the other options are wrong

   • DAzure OpenAI runs as a managed cloud service, so customers do not install physical GPUs on-premises.
5. Question 5Build and manage AI solutions with Microsoft Foundry

   A team wants to reuse the same Azure Storage account and Azure AI Search resource across several related AI projects without reconfiguring each one. Where should these shared resources be configured?

   • AAs connections on the hub so all projects under it inherit themCorrect
   • BAs a separate resource group inside each project
   • CAs local files uploaded individually to every project
   • DAs environment variables set on the user's workstation
   ✓ Correct answer: A

   Connections defined at the hub level are available to all projects created under that hub, avoiding repeated setup. This is the intended pattern for sharing resources like Storage or AI Search across related projects. Projects can also define their own project-scoped connections when needed.

   Why the other options are wrong

   • BResource groups organize Azure resources, but they do not automatically share connections across Foundry projects.
   • CUploading files to each project does not create reusable connections to external services.
   • DWorkstation environment variables are local to one machine and do not share resources across cloud projects.
6. Question 6Build and manage AI solutions with Microsoft Foundry

   Which of the following is a benefit of consuming a deployed model through Microsoft Foundry rather than embedding a raw model file in your app?

   • AThe endpoint handles hosting, scaling, and authentication so the app only sends requestsCorrect
   • BThe app no longer needs any network connectivity
   • CThe model automatically retrains on every request
   • DThe app gains ownership of the underlying GPU hardware
   ✓ Correct answer: A

   By calling a Foundry endpoint, the application offloads model hosting, scaling, and credential-based access to the platform. The app simply sends requests and processes responses without managing infrastructure. This simplifies development and lets the service manage capacity.

   Why the other options are wrong

   • BConsuming a hosted endpoint requires network connectivity to reach it.
   • CModels do not retrain on every request; inference and training are separate.
   • DThe app does not take ownership of GPU hardware when calling a managed endpoint.
7. Question 7Build and manage AI solutions with Microsoft Foundry

   For Azure OpenAI models deployed through Microsoft Foundry, what does a tokens-per-minute (TPM) quota primarily control?

   • AThe rate of token throughput allowed for a model deploymentCorrect
   • BThe total number of documents that can be stored in an index
   • CThe maximum size of a single uploaded data file
   • DThe number of users who can sign in to the Foundry portal
   ✓ Correct answer: A

   A tokens-per-minute quota sets the maximum rate of tokens a model deployment can process, effectively controlling throughput and helping manage capacity and cost. When you deploy a model in Microsoft Foundry, you allocate TPM from the available quota. Requests beyond the quota can be throttled with rate-limit responses.

   Why the other options are wrong

   • BIndex document counts are governed by the search resource, not TPM quota.
   • CFile size limits are separate storage constraints, not TPM.
   • DPortal sign-ins are governed by identity and access, not TPM.
8. Question 8AI agents and agentic solutionsSelect all that apply

   Which of the following are valid ways to give an agent access to your organization's own data for grounding? (Select all that apply.)

   • AAttach an existing Azure AI Search index with the Azure AI Search toolCorrect
   • BSet a higher model temperature
   • CIncrease the number of threads
   • DUpload documents and enable the File Search toolCorrect
   ✓ Correct answer: A, D

   You can ground an agent in your own data by uploading files for the File Search tool or by connecting an existing Azure AI Search index through the Azure AI Search tool. Both provide retrieval over your content during a run.

   Why the other options are wrong

   • BTemperature affects output randomness, not access to your data.
   • CAdding threads creates more conversations and does not provide grounding data.
9. Question 9Computer vision, natural language, speech, and document intelligence on Azure

   A Custom Vision object detection model returns which additional piece of information that an image classification model does not?

   • AThe bounding box coordinates for each detected itemCorrect
   • BThe dominant color of the image
   • CA natural-language caption of the scene
   • DThe extracted text from the image
   ✓ Correct answer: A

   An object detection model not only labels items but also returns the bounding box coordinates that show where each item is located in the image. Image classification returns only labels for the whole image without location information.

   Why the other options are wrong

   • BDominant color is a feature of Azure AI Vision Image Analysis, not object detection labels.
   • CCaptions come from Image Analysis, not from Custom Vision object detection.
   • DExtracting text is an OCR task, not part of object detection output.
10. Question 10Computer vision, natural language, speech, and document intelligence on Azure

    An organization wants to translate content but ensure that specific product names are never translated. Which Azure AI Translator feature supports this?

    • ACustom Translator or markup to preserve termsCorrect
    • BSentiment analysis
    • CLanguage detection
    • DKey phrase extraction
    ✓ Correct answer: A

    Azure AI Translator supports customization through Custom Translator for domain-specific models and markup such as the notranslate directive to protect specific terms like brand names from being translated. This ensures product names remain unchanged in the output. These features let organizations control translation behavior for specialized content.

    Why the other options are wrong

    • BSentiment analysis scores opinion and cannot preserve terms during translation.
    • CLanguage detection identifies the input language and does not control term translation.
    • DKey phrase extraction lists topics and is unrelated to preserving terms in translation.

AI-901 practice exam FAQ