AI-901 Practice Questions: Free Study Resource for Microsoft Azure AI Fundamentals 2026

Last updated: May 2026

Exam AI-901: Microsoft Azure AI Fundamentals launched as an English beta on 15 April 2026 and is expected to reach general availability around June 2026. It replaces AI-900, which retires on 30 June 2026 [1]. If you are sitting AI-901 in the next eight weeks, structured practice questions are the single fastest readiness check you can run, because the official Microsoft practice assessment has not yet been published [2].

This article gives you five fully worked AI-901 practice questions covering both exam domains, plus a clear method for using practice questions to identify and close skill gaps before exam day. Every question is grounded in the current AI-901 skills outline and the Microsoft Learn study guide. None of these are real exam questions, and you will not find real exam questions on any reputable preparation site, because Microsoft candidates sign an NDA agreeing not to share them.

TL;DR

AI-901 is the refreshed Azure AI Fundamentals exam covering responsible AI, model selection, and implementation through Microsoft Foundry. The two domains weight Foundry implementation more heavily (55-60%) than concepts (40-45%). Use practice questions to confirm 80% readiness across both domains before booking. Five worked examples below, with explanations for every option.

AI-901 exam overview

AI-901 is the Microsoft Certified: Azure AI Fundamentals exam in its 2026 form. It is delivered in beta until general availability, accessible online proctored or at an authorised Pearson VUE test centre, with a passing score of 700 out of 1000 [2]. The exam is currently offered in 13 languages, including English, Japanese, German, French, Spanish, and Portuguese (Brazil) [2].

Microsoft does not publish a fixed question count or duration for fundamentals exams, so any source quoting "exactly 60 questions in 60 minutes" is making an assumption. Plan for roughly the same scope as AI-900: a mix of single-best-answer and multi-select multiple choice, no labs, no case studies. Pricing follows the standard fundamentals tier of approximately $99 USD, with UK candidates paying close to £79 GBP plus VAT depending on Pearson VUE's exchange rate at checkout. Confirm the exact figure when you schedule [2].

The biggest practical implication of the AI-900 retirement on 30 June 2026 is that search results and study resources are still dominated by AI-900 material. Microsoft's official warning banner on the AI-900 exam page is unambiguous:

The requirements for this certification are changing. The related exam for AI-900 will retire on June 30, 2026. It will be replaced with AI-901. You can continue to earn this certification after AI-900 retires by passing AI-901.

— Microsoft Learn, AI-900 exam page [3]

If you are reading AI-900 practice content from before April 2026, treat anything about Form Recognizer or standalone Cognitive Services as outdated for the new exam. Stick to the Microsoft Learn AI-901 study guide and resources updated since 15 April 2026.

AI-901 exam domains and weighting

The 2026 skills outline organises AI-901 into two domains [4]:

Domain 1: Identify AI concepts and capabilities (40-45%)

This domain tests conceptual understanding rather than implementation. The three sub-skill areas are responsible AI principles, AI model components and configuration, and identifying AI workloads. Expect roughly nine to thirteen questions on these topics in a 25-question section sample.

The responsible AI block covers Microsoft's six principles (fairness, reliability and safety, privacy and security, inclusiveness, transparency, and accountability). The model components block covers how generative AI models work, choosing an appropriate model based on capabilities, and identifying deployment options and configuration parameters. The AI workloads block asks you to match scenarios to workload categories: generative and agentic AI, text analysis, speech, computer vision, and information extraction.

Domain 2: Implement AI solutions by using Microsoft Foundry (55-60%)

This is the larger and more practical half. The four sub-skill areas cover generative AI and agents, text and speech, computer vision and image generation, and information extraction, all framed through Microsoft Foundry implementation rather than standalone services.

The Foundry framing is the central change from AI-900. Where AI-900 asked you to identify Azure Cognitive Services capabilities in isolation, AI-901 asks you to deploy a model in the Foundry portal, build a lightweight client with the Foundry SDK, create and test single-agent solutions, and extract information using Azure Content Understanding rather than the legacy Form Recognizer (now rebranded as Azure AI Document Intelligence) [5]. If you have not opened the Foundry portal yet, do that before sitting the exam.

AI-901 practice questions by exam domain

Each of the questions below is followed by a full explanation: why the correct answer is correct, and why each distractor is wrong. The pattern matters because Microsoft exam questions are designed so that more than one answer looks plausible at first glance. Reading the wrong-answer explanations is what trains your eye for the distractor patterns the real exam uses.

Cover the answer with your hand or scroll past the explanation before reading on. Five questions is a small sample but you should aim to score at least four out of five before booking, and rework the explanations on any you miss.

Domain 1 sample questions

Question 1 (Responsible AI principles)

A mortgage lender deploys an AI model to automate loan application decisions. After deployment, the lender discovers that the model denies applications from candidates living in certain postcodes at a significantly higher rate, even when other financial indicators are equivalent. Which responsible AI principle is most directly violated?

• A. Privacy and security
• B. Fairness
• C. Transparency
• D. Accountability

Correct answer: B. Fairness.

Microsoft defines fairness as the principle that AI systems should treat all people fairly and allocate opportunities, resources, and information equitably [6]. Differential denial rates correlated with postcode, in the absence of an equivalent financial basis, are the textbook fairness failure pattern.

Why the distractors are wrong:

• A. Privacy and security is about protecting personal data and resisting unauthorised access. The scenario does not describe a data breach or unauthorised access, only an outcome disparity, so privacy and security is not the primary violation.
• C. Transparency is about whether people correctly understand what the system does and why. The scenario tells us what is happening (postcode-correlated denials) but transparency would be the issue if the lender could not explain the model's decisions at all. Here the issue is the decisions themselves, not their explicability.
• D. Accountability is about people maintaining oversight and being held responsible for AI behaviour. The scenario does not describe a missing oversight mechanism, only that the existing model produces biased outcomes. Accountability would be the lead violation if the lender had no review process at all.

This question type appears in almost every AI-901 sitting. Memorise the six principles and at least one canonical failure scenario for each.

Question 2 (Identifying AI workloads)

A logistics company receives several hundred scanned purchase order PDFs per day from suppliers. They need to automatically capture the purchase order number, supplier name, delivery date, and individual line items from each document, then push the structured fields into their order management system. Which AI workload is most appropriate?

• A. Text analysis
• B. Generative AI
• C. Information extraction
• D. Computer vision

Correct answer: C. Information extraction.

The AI-901 skills outline lists information extraction explicitly as a workload category and ties it to scenarios that "extract information from text, images, audio, and videos" [4]. Capturing named fields from a scanned document is the canonical information extraction scenario.

Why the distractors are wrong:

• A. Text analysis covers tasks like keyword extraction, entity detection, sentiment analysis, and summarisation. It works on text that is already digital and focuses on language understanding rather than structured field extraction from a document layout.
• B. Generative AI produces new content based on prompts. Generating a summary of a purchase order would be generative; extracting specific fields with their original values is not.
• D. Computer vision is the workload for identifying objects, classifying images, or detecting features in pictures. It is part of the underlying pipeline that processes scanned documents but is too broad as the named workload. AI-901 distinguishes computer vision (image-level tasks) from information extraction (structured data from documents and other media) [4].

When you see "scanned PDF" plus "named structured fields", reach for information extraction. When you see "image" plus "what is in this picture", reach for computer vision.

Domain 2 sample questions

Question 3 (Foundry deployment options)

A developer needs to make a Meta Llama 3 model available to a customer-facing chatbot. The team wants to start serving traffic within the day, has no infrastructure budget for reserved GPUs, and prefers to pay only for the tokens they actually consume. Which deployment option in Microsoft Foundry best meets these requirements?

• A. Serverless API deployment
• B. Managed compute (managed endpoint) deployment
• C. Self-hosted deployment in Azure Kubernetes Service
• D. Local download and run on a development laptop

Correct answer: A. Serverless API deployment.

Serverless API deployments (sometimes called MaaS or pay-as-you-go) allow developers to access models in the Foundry catalogue without provisioning any GPU compute. Billing is purely token-based, and deployment takes minutes [7]. Llama models are explicitly available through this path.

Why the distractors are wrong:

• B. Managed compute deployment lets you host an open model on dedicated compute that you provision and pay for, whether or not traffic uses it. That conflicts with the "no infrastructure budget for reserved GPUs" constraint.
• C. Self-hosted in Azure Kubernetes Service is a custom infrastructure pattern outside the Foundry portal deployment flows. It requires substantial cluster setup and ongoing management, contradicting the same-day timeline.
• D. Local laptop is not a production deployment pattern and would not serve a customer-facing chatbot reliably.

For AI-901, remember the deployment-option decision tree: serverless for fastest path and token-based pricing, managed compute when you need a dedicated endpoint, and Azure OpenAI deployments for GPT-family models routed through the Azure OpenAI sub-service.

Question 4 (Content Understanding)

A mortgage lender wants to use Azure Content Understanding to extract loan amounts from incoming applications. Their compliance team requires that any extracted value where the model is uncertain must be flagged for human review before it is written to the system of record. Which Content Understanding capability should the lender configure?

• A. Set estimateFieldSourceAndConfidence to true in the analyser configuration
• B. Increase the content safety filter threshold on the analyser to "high"
• C. Switch from Content Understanding to the legacy Form Recognizer custom model
• D. Enable Prompt Shields in the analyser configuration

Correct answer: A. Set estimateFieldSourceAndConfidence to true in the analyser configuration.

Content Understanding analysers support an opt-in flag that, when enabled, returns a per-field confidence score and grounding metadata alongside each extracted value [5]. The lender's downstream system can then route any field with a confidence score below a chosen threshold to a human reviewer. This is the documented pattern for human-in-the-loop information extraction.

Why the distractors are wrong:

• B. Content safety filter governs whether generated or extracted content is blocked for hate, sexual, violence, or self-harm severity. It has nothing to do with extraction confidence and would not flag uncertain numerical extractions.
• C. Switching to Form Recognizer is the wrong direction. Form Recognizer was rebranded to Azure AI Document Intelligence and is the older service. Content Understanding is the higher-level service named in the AI-901 skills outline, and it natively supports per-field confidence scoring.
• D. Prompt Shields is a content safety feature that protects models from jailbreak attempts and prompt injection. It does not produce confidence scores for extracted fields.

This question is a good test of whether you have actually worked with Content Understanding analysers rather than just read about them. If you have not built one, the Content Understanding analyser quickstart is worth an hour of your study time.

Question 5 (Agent core components)

A developer creates a customer service agent in Microsoft Foundry. They configure it to use the GPT-4o model, provide the system instruction "You are a helpful customer service assistant for Contoso", and connect a file search tool that can read the company knowledge base. Which three components are combined in this agent?

• A. Endpoint, API key, deployment name
• B. Trigger, action, condition
• C. Embedding, retriever, generator
• D. Model, instructions, tools

Correct answer: D. Model, instructions, tools.

The Microsoft Foundry Agent Service documentation defines an agent as the combination of three core components: a model from the Foundry catalogue that provides reasoning capabilities, instructions that define the agent's goals and behaviour, and tools that give the agent access to data or actions such as file search, code interpreter, or external APIs [8]. The scenario maps exactly: GPT-4o is the model, the customer service prompt is the instruction, the file search tool is the tool.

Why the distractors are wrong:

• A. Endpoint, API key, deployment name are properties needed to authenticate to a deployed model, not the conceptual components of an agent.
• B. Trigger, action, condition belong to workflow automation platforms like Power Automate or Logic Apps, not to the Foundry Agent Service definition.
• C. Embedding, retriever, generator describes a retrieval-augmented generation (RAG) pipeline architecture. RAG is one pattern an agent can use, but it is not the definition of an agent. An agent might use RAG via its tools, but the components of the agent itself are still model, instructions, and tools.

If you can describe an agent in those three words in your sleep, you have the AI-901 agentic AI sub-skill covered.

How to use AI-901 practice questions effectively

Working through practice questions is not the same as studying them. Three practices separate effective preparation from time-wasting:

Time yourself per question. AI-901 is not lab-heavy, so each question should take 60 to 90 seconds. If you find yourself reading a practice question for three minutes, you do not know the underlying concept well enough yet. Mark the question, move on, and return to the underlying Microsoft Learn module before retrying.

Score the wrong-answer explanations, not just the right ones. Microsoft writes distractor answers deliberately. A "fairness" question with three plausible-sounding distractors trains you to spot the specific cue (postcode correlation, demographic disparity, opportunity allocation) that signals fairness rather than another principle. If you got the question right but cannot explain why each wrong answer was wrong, you got it right by elimination and might miss the same pattern next time.

Track your domain split. If you are scoring 90% on Domain 1 and 60% on Domain 2, you do not need more responsible-AI practice. You need to open the Foundry portal, deploy a model, and build a single-agent solution. AI-901 weights Domain 2 at 55-60%, so a deficit there matters disproportionately for your final score. Aim to hit 80% on the smaller domain and 75% on the larger before booking, then book with a few days of buffer in case real-test conditions drag your scores down 5%.

The 80% threshold is a heuristic, not a guarantee. It accounts for the gap between practice conditions (your desk, a cup of tea, reference material a tab away) and proctored conditions (silence, time pressure, no reference material). Aim higher than the 700 of 1000 passing score in practice.

AI-901 vs AI-900: what changed in the practice questions

If you are migrating from AI-900 study material to AI-901, three categories of question change in ways that affect your practice strategy.

Foundry replaces standalone Cognitive Services framing. AI-900 asked questions like "which Azure Cognitive Services service would you use to transcribe speech?" Answer: Speech Service. AI-901 reframes the same workload around Foundry: "build a lightweight application by using Azure Speech in Foundry Tools" [4]. You still need to know what the workload is, but the deployment path is now Foundry-centric. Practice questions that name standalone Cognitive Services endpoints are out of scope.

Agentic AI is genuinely new. AI-900 has no sub-skill mentioning agents. AI-901 explicitly tests "Create and test a single-agent solution in the Foundry portal" and "Create a lightweight client application for an agent" [4]. The Microsoft Foundry Agent Service was introduced at Ignite 2025 [9]. If your study materials predate November 2025, the agent content is missing entirely and you should bridge that gap with the Microsoft Foundry Agent Service overview.

Content Understanding replaces Form Recognizer scenarios. AI-900 tested invoice and receipt extraction using Form Recognizer (since rebranded to Azure AI Document Intelligence). AI-901 tests the same workloads but using Azure Content Understanding, which extends the same idea across documents, images, audio, and video [5]. The conceptual model also shifts: you configure analysers rather than custom models, fields rather than form regions, and confidence scores are opt-in rather than always returned.

For migrating candidates, the safest assumption is that any AI-900 practice question grounded in pre-2025 Azure services is suspect. Build your AI-901 study list from material updated since 15 April 2026.

Free AI-901 practice resources beyond Examinotion

Practice questions are only one preparation tool. Combine them with these free resources for stronger coverage:

• Microsoft Learn AI-901 study guide. The official skills outline at learn.microsoft.com/credentials/certifications/resources/study-guides/ai-901 is the single most important free resource. Skills outlined here are testable; skills not outlined here are not. Read it twice.
• Microsoft Learn modules. Each AI-901 sub-skill maps to a free interactive learning module under Microsoft Learn AI training. Modules include sandbox environments where you can deploy a Foundry model without spending Azure credits.
• Microsoft Q&A. The AI-901 tag on Microsoft Q&A surfaces questions from other candidates and Microsoft employee responses. Useful for clarifying ambiguous skills outline points.
• Microsoft Foundry portal sandbox. A free Azure account gives you a sandbox in the Foundry portal where you can deploy small models, build a single-agent solution, and run the Content Understanding quickstarts at no cost.
• Examinotion AI-901 practice tests. When you are ready for an expanded pool with full exam-format practice, start practising AI-901 questions using Examinotion's paid practice pool, which mirrors AI-901 question format and weighting.

The official Microsoft practice assessment for AI-901 has not yet been published [2], so the practice question gap is genuinely larger for AI-901 than it was for AI-900 at the equivalent moment in its lifecycle.

Tips for answering AI-901 exam questions

Five tactical habits help on AI-901 specifically:

Read the question stem twice before looking at the answers. Microsoft writes long stems for fundamentals exams. The setup often includes a constraint near the end ("they want to pay only for tokens used", "compliance requires human review") that changes the right answer. Read fully before scanning options.

Eliminate deprecated services first. Any answer that names "Cognitive Services" as a standalone, or "Form Recognizer" without the Document Intelligence rebrand, is more likely to be a distractor than a correct answer in AI-901. The current names are Foundry, Content Understanding, and Document Intelligence.

For agent questions, count to three. Model, instructions, tools. If you see an answer that lists those three, it is usually the right one for any "what are the components of an agent" question.

Match workload questions to the AI-901 list, not the AI-900 list. AI-901 lists generative and agentic AI, text analysis, speech, computer vision, and information extraction [4]. AI-900 listed slightly different categories. Use the current list when matching scenarios.

Flag and move. Microsoft fundamentals exams let you flag and revisit questions. If a question takes more than 90 seconds, flag it, pick the best-feeling answer, and move on. You can return with the time you save on faster questions. For more exam-day patterns to watch for, see our guide to common AI-901 mistakes to avoid.

Frequently asked questions

Are there free AI-901 practice questions available?

Yes. Five worked examples appear in the sections above, grounded in the current Microsoft Learn AI-901 skills outline. The Microsoft Foundry portal also hosts sandbox modules with practice exercises. The official Microsoft practice assessment for AI-901 has not been published yet, so independent practice pools are currently the main free source.

How many questions are on the AI-901 exam?

Microsoft does not publish a fixed question count for AI-901. Other fundamentals exams typically run between 40 and 60 questions. The official exam page does not state a number, and the duration is also not published on the Microsoft Learn page. Plan your time around the Pearson VUE booking confirmation, which gives the scheduled length.

What is the passing score for AI-901?

The passing score is 700 out of 1000, as confirmed on the official Microsoft Learn AI-901 exam page. Scores are scaled rather than raw percentages, so a 70% raw score does not necessarily equal 700 scaled. Use the 80% target in practice to give yourself a margin for the gap between practice and proctored conditions.

Are AI-900 practice questions still useful for AI-901?

Partly. Responsible AI principles, workload identification, and high-level model concepts overlap. Anything grounded in standalone Cognitive Services, Form Recognizer scenarios, or pre-Foundry deployment patterns is outdated. Use AI-900 material only for Domain 1 conceptual practice, and rely on AI-901-specific resources for the entire Domain 2 Foundry implementation block.

Does Microsoft provide an official AI-901 practice assessment?

Not yet. The official AI-901 exam page states that the practice assessment is not currently available, and Microsoft typically releases practice assessments within eight weeks of an exam leaving beta. With AI-901 expected to reach general availability around June 2026, the official practice assessment is unlikely before August 2026 at the earliest.

How hard are AI-901 practice questions compared to the real exam?

Practice questions written from the published skills outline match the exam well at the concept and recognition level. The real exam adds time pressure, no reference material, and occasional scenario depth that practice questions struggle to replicate. Score 80% on a varied practice pool to give yourself a buffer against those proctored-condition effects.

What topics carry the most weight on AI-901?

Domain 2 (Implement AI solutions by using Microsoft Foundry) is weighted at 55-60% and covers generative AI and agents, text and speech, computer vision and image generation, and information extraction, all framed through Foundry. Domain 1 (Identify AI concepts and capabilities) is 40-45% and covers responsible AI, model components, and workload identification. Prioritise hands-on Foundry practice.

Can I pass AI-901 with practice questions alone?

Practice questions are necessary but not sufficient. AI-901 includes scenario-based questions about deploying models, configuring analysers, and building lightweight agent clients, which test recognition of patterns you have actually used. Pair practice questions with a few hours in the Foundry portal building a deployed model and a single-agent solution end-to-end before booking the exam.

---

Ready to test yourself against a fuller question pool? Start practising for AI-901 with Examinotion's full AI-901 practice tests, mirrored to the current skills outline and updated as Microsoft refreshes AI-901 through its beta period. For broader context on where AI-901 sits in the Microsoft AI certification track, see our Microsoft AI certification roadmap.

Sources

1. Microsoft Learn, Exam AI-900: Microsoft Azure AI Fundamentals (retiring 30 June 2026), updated 11 May 2026.
2. Microsoft Learn, Exam AI-901: Microsoft Azure AI Fundamentals (beta), updated 21 April 2026.
3. Microsoft Learn, Exam AI-900 retirement warning banner, accessed 28 May 2026.
4. Microsoft Learn, Study guide for Exam AI-901: Microsoft Azure AI Fundamentals, updated 23 April 2026.
5. Microsoft Learn, Azure AI Content Understanding overview, updated 27 March 2026.
6. Microsoft, Responsible AI principles and approach, accessed 28 May 2026.
7. Microsoft Learn, Foundry model inference and deployment options, updated 19 May 2026.
8. Microsoft Learn, Microsoft Foundry Agent Service overview, updated 19 May 2026.
9. Microsoft Tech Community, Foundry Agent Service at Ignite 2025: simple to build, powerful to deploy, November 2025.