AI-901 (Microsoft Azure AI Fundamentals with Azure AI Foundry) goes beyond conceptual AI knowledge — 55–60% of the exam tests your ability to implement AI solutions in Azure AI Foundry, including reading Python SDK snippets, deploying models, configuring agents, and using Azure Content Understanding.
These 15 questions mirror the real exam's format and difficulty. Each question includes a detailed explanation covering both the correct answer and why each distractor is wrong.
What you'll get:
- ✓15 questions across both AI-901 domains
- ✓Real exam difficulty — Foundry SDK snippets, system prompts, Content Understanding vs. Document Intelligence
- ✓Detailed explanations for every answer with full reasoning
- ✓Coverage of 2026 Foundry topics — agents with tools, Content Understanding, responsible AI in Foundry
📝 Practice Test Instructions
- • Each question has ONE best answer
- • Note your answer before reading the explanation
- • Aim for 20 minutes total (real exam: 60 minutes for 40–60 questions)
- • Target 85%+ on practice — Fundamentals scoring is forgiving but a comfortable margin helps
Identify AI Concepts and Capabilities
Questions 1–6
Your Azure AI model produces outputs that are technically correct but consistently harder to interpret for users with disabilities. You need to address this gap using a specific Responsible AI principle as your framing.
Which Microsoft Responsible AI principle is most directly applicable?
- A.Fairness — addressing potential bias in outputs
- B.Inclusiveness — ensuring AI works for people of all abilities
- C.Accountability — being responsible for AI outcomes
- D.Transparency — making AI decisions understandable
Correct: B. Inclusiveness is the Responsible AI principle that specifically addresses accessibility and ensuring AI works for people with diverse abilities. Fairness covers bias across demographic groups. Accountability covers governance and oversight. Transparency covers explainability of AI decisions.
You have historical sales data with daily revenue figures and want to predict tomorrow's revenue.
Which type of machine learning problem is this?
- A.Classification
- B.Regression
- C.Clustering
- D.Reinforcement learning
Correct: B. Predicting a continuous numerical value (revenue) is a regression problem. Classification predicts a discrete category. Clustering finds groupings in unlabeled data. Reinforcement learning trains agents through reward signals.
A customer-service application needs to automatically detect the language of incoming written tickets and translate them into English for the support team.
Which two Azure AI services together provide this capability?
- A.Azure AI Vision and Azure AI Speech
- B.Azure AI Language (language detection) and Azure AI Translator
- C.Azure AI Document Intelligence and Azure AI Search
- D.Azure OpenAI and Azure AI Speech
Correct: B. Azure AI Language includes a prebuilt language detection capability and Azure AI Translator handles translation. This is the canonical combination for the scenario. Vision and Speech address image and audio, not text. Document Intelligence extracts structured data from documents but does not translate. Azure OpenAI can translate but is overkill for deterministic translation needs.
Your generative AI application has been producing factually incorrect responses, asserting events that did not happen as if they were established facts.
What is this risk called, and what is the primary mitigation?
- A.Bias; rebalance training data
- B.Prompt injection; enable prompt shields
- C.Hallucination; use Retrieval-Augmented Generation (RAG) and groundedness checks
- D.Copyrighted output; enable protected material detection
Correct: C. Hallucination is when a model produces plausible but factually incorrect content. The primary mitigation is grounding the model in retrieved authoritative sources (RAG) and using groundedness evaluators to detect ungrounded outputs. Bias addresses unfair treatment of demographic groups. Prompt injection is malicious instruction insertion. Copyrighted output protection is unrelated to factual hallucination.
You need to extract people, organizations, and locations from a large set of news articles into a structured database.
Which NLP capability does this require?
- A.Sentiment analysis
- B.Key phrase extraction
- C.Named entity recognition (NER)
- D.Translation
Correct: C. Named entity recognition identifies specific entities like people, organizations, and locations in text and tags them with their type — exactly the requirement. Sentiment analysis returns polarity. Key phrase extraction returns important terms without entity typing. Translation converts between languages.
Your Azure OpenAI deployment must block any model output that contains hate speech at medium or higher severity, while leaving the other content categories at default settings.
Which Azure feature provides this capability?
- A.Azure RBAC role assignments
- B.Content filters in the Azure OpenAI deployment, configured per category and severity level
- C.Prompt shields for jailbreak detection
- D.Azure AI Search semantic ranker
Correct: B. Azure OpenAI content filters allow per-category (hate, self-harm, sexual, violence) and per-severity (safe, low, medium, high) configuration for both prompts and completions. RBAC controls access, not content. Prompt shields detect jailbreak attempts, not severity-based content. Semantic ranker re-ranks search results.
Implement AI Solutions Using Microsoft Foundry
Questions 7–15
You are building a customer-support chat application with Azure OpenAI. The model must always respond formally and never discuss topics outside the company's product line.
Where should you specify these behavioral constraints?
- A.In the user prompt for every message
- B.In the deployment configuration
- C.In a system prompt at the start of the conversation
- D.In the Azure OpenAI resource's networking settings
Correct: C. System prompts define the model's persona, tone, and operating constraints for the entire conversation — applied globally before any user input. This is the right place for "respond formally" and "stay on-topic". Repeating instructions in every user prompt is wasteful and inconsistent. Deployment configuration sets capacity and content filters, not behavior. Networking is unrelated.
You see this Python code: from openai import AzureOpenAI client = AzureOpenAI(api_version="2024-10-21", azure_endpoint="https://my-aoai.openai.azure.com") response = client.chat.completions.create( model="gpt-4o-mini", messages=[ {"role": "system", "content": "You are a helpful assistant."}, {"role": "user", "content": "Hi"} ] ) print(response.choices[0].message.content)
What does this code do?
- A.Trains a new fine-tuned model called gpt-4o-mini
- B.Sends a chat completion request to an Azure OpenAI deployment named gpt-4o-mini and prints the model's response
- C.Creates a new Azure OpenAI resource via the Azure Management API
- D.Lists all available models in the Azure OpenAI model catalog
Correct: B. The code instantiates an AzureOpenAI client, calls chat.completions.create with a system prompt and user message, and prints the assistant's response from response.choices[0].message.content. This is the canonical Foundry SDK pattern for a chat completion. Training, resource creation, and catalog listing use different methods.
You need an AI assistant that can hold multi-turn conversations with a user, search a folder of PDFs for relevant content, and execute Python code to compute results when needed.
Which Azure AI Foundry primitive should you use?
- A.A single chat completions call
- B.A prompt flow with an LLM node
- C.An Azure AI Foundry agent with file search and code interpreter tools
- D.Azure AI Search by itself
Correct: C. Foundry agents are the right primitive when you need tool use (file search, code interpreter), persistent multi-turn conversation state, and the ability to chain tool calls. Single chat completions force you to hand-roll state and tool orchestration. Prompt flow is for deployable pipelines, not interactive multi-turn assistants. AI Search alone retrieves documents but does not converse or execute code.
You receive PDF invoices from hundreds of vendors with consistent fields (vendor name, invoice number, date, line items, total). You need a service that extracts these fields with high accuracy out of the box.
Which Azure AI service is most appropriate?
- A.Azure AI Vision Image Analysis with OCR
- B.Azure AI Document Intelligence prebuilt Invoice model
- C.Azure AI Custom Vision with object detection
- D.Azure OpenAI with a free-form prompt
Correct: B. Document Intelligence prebuilt Invoice is purpose-trained on millions of global invoices and returns structured vendor, invoice number, date, line items, and total fields with confidence scores. Image Analysis OCR returns raw text only. Custom Vision is for image classification/detection. Azure OpenAI works but is slower, more expensive, and less reliable than the purpose-built model.
You need to extract insights from a mix of customer-call audio recordings, screenshots of dashboards, scanned customer letters, and embedded video testimonials — all into a single structured JSON schema.
Which Azure AI service is the best fit?
- A.Azure AI Document Intelligence prebuilt Layout
- B.Azure AI Content Understanding with custom analyzer schemas
- C.Azure AI Vision Image Analysis 4.0
- D.Azure AI Speech batch transcription only
Correct: B. Content Understanding is the multimodal extractor designed for mixed sources — documents, images, audio, and video — using natural-language-described schemas to produce structured JSON output. Document Intelligence Layout is for documents only. Image Analysis covers images only. Batch transcription covers audio only.
You want to deploy GPT-4o-mini in Azure AI Foundry so your application can call it.
Which portal location lists deployable models and lets you initiate a new deployment?
- A.The Resource Group blade in the Azure portal
- B.The Model catalog in the Azure AI Foundry portal
- C.The Azure Marketplace under "AI + machine learning"
- D.The Azure Active Directory enterprise applications blade
Correct: B. The Azure AI Foundry portal Model catalog lists every deployable model (Azure OpenAI, open-weight, Microsoft-published) and provides the workflow to create a new deployment with capacity, region, and content-filter selections. The Azure portal Resource Group blade does not list model catalog items. Marketplace lists third-party apps, not Foundry models. Entra ID Enterprise Applications is for app registrations.
A background service needs to call your Azure OpenAI deployment to generate completions, but it must not be able to view keys, modify deployments, or delete the resource. You want the least-privilege role.
Which Azure RBAC role should the service's managed identity have?
- A.Owner
- B.Contributor
- C.Cognitive Services OpenAI User
- D.Cognitive Services OpenAI Contributor
Correct: C. Cognitive Services OpenAI User is the least-privilege role for inference: it permits chat completions and embeddings calls but does not allow viewing keys, managing deployments, or modifying the resource. Cognitive Services OpenAI Contributor adds deployment management — over-privileged. Contributor and Owner are far broader than required.
Your Python application needs to authenticate to an Azure OpenAI deployment without storing API keys in source control or environment variables.
Which approach should you use?
- A.Hard-code the API key in the source file
- B.Use a managed identity with the Cognitive Services OpenAI User role, authenticated via DefaultAzureCredential
- C.Store the API key in an Excel spreadsheet shared with the dev team
- D.Pass the API key as a command-line argument
Correct: B. Managed identities authenticated via DefaultAzureCredential are the recommended secret-less authentication pattern for Azure-hosted applications. The managed identity holds the Cognitive Services OpenAI User role on the Azure OpenAI resource and the SDK obtains tokens via Entra ID — no keys to leak. The other options all involve insecure secret handling.
You need to add real-time speech-to-text transcription to a meeting application, with sub-second latency and the ability to identify who is speaking.
Which Azure AI Speech capability should you use?
- A.Batch transcription with a SAS-URL input
- B.Real-time speech-to-text with speaker diarization enabled
- C.Text-to-speech with a neural voice
- D.Speech translation only
Correct: B. Real-time speech-to-text streams audio and returns transcription with low latency, and speaker diarization identifies different speakers. This satisfies both requirements. Batch transcription is asynchronous and unsuitable for real-time meetings. Text-to-speech generates audio from text — opposite direction. Speech translation translates but does not transcribe with diarization as the primary purpose.
Want 485 More AI-901 Questions?
Full coverage of both domains — Responsible AI, Azure AI Foundry portal and SDK, agents, Content Understanding, and Foundry Python snippets. Start with 40 questions free.
Start Free AI-901 Practice →