Azure AI Fundamentals (AI-901): 5 Practice Questions on Identify AI concepts and capabilities

How well do you know Identify AI concepts and capabilities? This Azure AI Fundamentals (AI-901) set has 5 questions to find out. It carries about 42.5% of the exam score. Answers and full explanations sit under each question.

Azure AI Fundamentals (AI-901) · Identify AI concepts and capabilities · 5 questions with answers and explanations

🧭 Before you start

This fundamentals domain gives short business cases and asks you to name the concept behind them: a responsible AI principle, an AI workload, a model type or an extraction feature. Few items need product depth. What decides the answer is reading what the system actually produces or what went wrong, then matching that to the closest category rather than a neighbouring one.

Watch for these while you work

• Classify the workload by its output: a new image made from a prompt is generative AI, while describing or labelling an existing image is computer vision.
• Learn the responsible AI principles by their typical failure: missing owners and approval records point to accountability, uneven outcomes to fairness.
• For model choice, weigh accuracy against speed. Multi-step math and planning favour a reasoning model; simple, high-volume language tasks favour a small model.

📊 This set at a glance

Exam: AI-901 · 45 questions · 45 minutes · passing score 700
This set: 3 applied, 2 advanced · 2 scenario, 1 concept check, 1 trade-off, 1 troubleshooting · 1 with more than one answer
Topics: Azure Content Understanding in Foundry Tools, Microsoft Foundry, Foundry Models, Azure Speech in Foundry Tools

Question 1 of 5

An insurer is automating claim form processing. It wants values the AI is unsure about to go to a human reviewer, and reviewers must be able to see exactly where on the page each value came from. Which TWO extraction features meet these needs? (Select TWO)

AMarkdown output, which returns the whole form as formatted text that a search index can use for later retrieval.

BImage tagging, which returns keywords about each scanned page so that reviewers can find forms by their content.

CSegmentation, which splits a long file into logical sections so that each section can go to a different analyzer.

DGrounding, which identifies the specific region of the source content where each field value was extracted from.

EConfidence scores, which give each extracted field value a reliability estimate between 0 and 1 for routing decisions.

Show answer

✅ Answer: D, E

Difficulty: advanced · Topic: Azure Content Understanding in Foundry Tools

Trusting extracted data
Extraction is only useful in automation if you know how much to trust each value and can check it quickly. Azure Content Understanding in Foundry Tools provides two features for this:
• Confidence scores: a value from 0 to 1 for each extracted field. High scores allow straight-through processing; low scores can be routed to a person. This is how you reduce manual review without blindly trusting every value.
• Grounding: the specific region of the content where each value was found or generated, so a reviewer can trace a value back to its origin on the page.

Other features in the same pipeline
• Segmentation divides a document or video into logical parts (for example by document type, or a video into scenes).
• Markdown output suits search and retrieval-augmented generation (RAG).
Both are useful, but neither tells you how reliable a value is or where it came from.

Example
A rule sends any claim whose PolicyNumber confidence is below 0.8 to review, and the reviewer's screen highlights the grounded region on the scanned form.

Why each option works or fails
• ✗ A. Markdown output: it is a text representation for search and RAG; it gives no reliability estimate and no link between a field and its location.
• ✗ B. Image tagging: keywords describe what a picture shows and say nothing about the reliability or origin of extracted field values.
• ✗ C. Segmentation: it splits content into sections for targeted processing, which neither scores values nor shows their source.
• ✓ D. Grounding (correct): it shows the region of the source where each value came from, so reviewers can verify a value at a glance.
• ✓ E. Confidence scores (correct): a per-field reliability estimate lets the workflow send uncertain values to a human and pass confident ones straight through.

🎯 Key idea: Use confidence scores to decide which values need human review and grounding to show reviewers where each value came from.

Question 2 of 5

After a pricing model produced unfair quotes for a week, an online retailer investigates. Nobody can say who owned the model, who approved the version that went live, or when it was deployed. Which responsible AI principle did the retailer's process mainly fail?

AAccountability, because the organization had no governance record of who owned and approved the model and why.

BTransparency, because customers were not told that an AI system set the prices they were offered for products.

CReliability and safety, because the model should have been tested against a wider range of unexpected inputs.

DInclusiveness, because the pricing model was not designed together with customers from a variety of backgrounds.

Show answer

✅ Answer: A

Difficulty: advanced

Accountability needs records
Accountability is about people and organizations owning how their AI systems operate. In practice that requires governance: clear owners, approval steps, and records that show who did what, when and why.

Lineage and audit trails
Machine learning operations (MLOps) tools help here. Azure Machine Learning, for example, can register models with their metadata and capture lineage information such as who published a model, why changes were made, and when a model was deployed or used in production. It can also send alerts on events like model deployment or data drift.

Why the other principles are not the main failure
The unfair quotes themselves are a fairness problem, and better testing would have helped reliability. But the question asks about the process failure revealed by the investigation: nobody can trace ownership or approval, which means nobody can be held answerable or learn from what happened.

Example
A simple record such as pricing-model v7 | trained on orders 2025-01..2026-06 | approved by J. Park | deployed 2026-09-14 would have let the retailer find the cause and the owner within minutes.

Why each option works or fails
• ✓ A. Accountability (correct): without records of ownership, approval and deployment time, no one can be held answerable for the model's behavior, which is a failure of the governance that accountability requires.
• ✗ B. Transparency: disclosing AI pricing to customers would be good practice, but the investigation's problem is internal: the retailer cannot trace its own decisions.
• ✗ C. Reliability and safety: wider testing might have caught the problem earlier, but it would not tell the retailer who approved the model or when it went live.
• ✗ D. Inclusiveness: co-designing with diverse customers is useful, but the scenario's failure is missing ownership and approval records.

🎯 Key idea: Accountability depends on governance: named owners, approvals and lineage records that show who deployed which model, when and why.

Question 3 of 5

A university builds a tutoring app that must solve multi-step physics and math problems and work through complex scheduling puzzles, where accuracy matters more than an instant reply. Which type of model is the best fit?

AAn embedding model, because comparing the vector of each problem with stored solutions always produces the correct worked answer.

BA text-to-speech model, because reading each problem aloud lets the app check every step before it gives the student a final answer.

CAn image generation model, because drawing a diagram of each problem lets the model calculate the result from the generated image.

DA reasoning model, which spends more processing on the request and is strong at science, coding, math and complex problem solving.

Show answer

✅ Answer: D

Difficulty: applied

What reasoning models are
Reasoning models are designed for reasoning and problem-solving tasks. They spend more time processing and understanding a request before answering, which makes them especially strong in science, coding and math, and in complex problem solving, strategy and logistics.
• They generate internal reasoning tokens to break the problem apart and weigh approaches before producing the visible answer.
• Many let you tune how much they think with a reasoning_effort setting (low, medium, high and other values, depending on the model).
• The extra thinking adds latency and output tokens, so they are not the default for simple chat.

Choosing between chat and reasoning models
• Chat completion models suit conversation and general content generation where a fast reply matters.
• Reasoning models suit multi-step problems where accuracy matters more than speed.

Example
"A train leaves at 9:40 at 80 km/h..." The reasoning model works through the distances and times internally, then returns a short, correct step-by-step explanation.

Why each option works or fails
• ✗ A. Embedding model: embeddings measure similarity between texts; a nearby stored solution is not guaranteed to be the right answer for a new problem, and it cannot generate new working.
• ✗ B. Text-to-speech model: it converts text into spoken audio; reading a problem aloud does not solve it or check any steps.
• ✗ C. Image generation model: it creates pictures from descriptions; generated diagrams are not a calculation method and add no reasoning ability.
• ✓ D. Reasoning model (correct): these models spend more processing on a request and are strongest at multi-step science, math and complex problem solving, which matches a tutor where accuracy beats instant replies.

🎯 Key idea: For multi-step math, science and complex problem solving where accuracy matters more than speed, choose a reasoning model.

Question 4 of 5

An online home-goods store wants lifestyle pictures of its sofas placed in different room styles for social media ads. A designer will type descriptions such as "grey sofa in a sunny Scandinavian living room" and pick from the results. Which AI workload is this?

AText analysis, which extracts key phrases such as "grey sofa" and "living room" from the designer's descriptions.

BInformation extraction, which reads the sofa's dimensions and fabric details from the supplier's spec sheets.

CComputer vision object detection, which draws a box around each sofa found in the existing product photos.

DGenerative AI, using an image-generation model that creates new pictures from the designer's text prompt.

Show answer

✅ Answer: D

Difficulty: applied · Topic: Microsoft Foundry, Foundry Models

Generative AI is not only text
Generative AI enables applications to generate new content: natural language dialogs, but also images, video, code and other formats. Image-generation models take a natural language prompt and produce a new picture that did not exist before.

Analyzing images versus creating them
• Computer vision analysis takes an existing image as input and returns information about it: a label, object locations, a caption.
• Image generation takes text (and sometimes an image) as input and returns a new image as output.
A useful test is to look at what the business receives at the end. If it is a fact about existing content, it is analysis. If it is new content, it is generative AI.

Example
Prompt: "grey three-seat sofa in a sunny Scandinavian living room, light wood floor, plants". The model returns several candidate images; the designer picks one for the ad.

Why each option works or fails
• ✗ A. Key phrase extraction: listing the main terms in the designer's prompt returns text about the prompt, but it never produces any image at all.
• ✗ B. Information extraction: reading spec sheets returns dimensions and fabric as data, but the requirement is new marketing images, not product fields.
• ✗ C. Object detection: it analyzes photos the store already has and returns locations of sofas; it cannot place a sofa into a new room scene.
• ✓ D. Generative AI image generation (correct): the store wants new pictures that do not exist yet, created from a text description, which is what image-generation models do.

🎯 Key idea: When the output is a new image created from a text prompt, the workload is generative AI (image generation), not computer vision analysis.

Question 5 of 5

A team is comparing older, concatenated synthetic voices with the standard neural voices offered by Azure Speech in Foundry Tools. What best describes how neural text to speech produces more natural speech?

AIt uses deep neural networks that predict prosody, such as stress and intonation, together with the voice, giving fluid, humanlike output.

BIt records a human speaker reading every word in the dictionary and then plays back the matching recordings in the order that the text requires.

CIt converts the text into a written transcript first and then applies sentiment analysis to choose a happy or sad recording for each sentence.

DIt translates the text into a phonetic alphabet that a separate language model reads aloud word by word at a single fixed pitch and pace.

Show answer

✅ Answer: A

Difficulty: applied · Topic: Azure Speech in Foundry Tools

What makes a voice sound natural
Human speech is not just the right sounds: it has prosody, the rhythm, stress and intonation that change with meaning. Older synthesis approaches often sounded flat because they struggled with prosody.
Azure Speech's standard voices are neural voices: deep neural networks predict prosody and synthesize the voice at the same time, which Learn says gives more fluid and natural-sounding output and reduces listening fatigue.

Related terms
• Standard voice: an out-of-the-box neural voice (called Neural on the pricing page).
• HD voices: higher-quality voices for more versatile scenarios.
• Custom voice: a neural voice trained for one brand from a voice talent's recordings (Limited Access).
• SSML: markup to adjust pitch, rate, pauses and pronunciation of any of these voices.

Why each option works or fails
• ✓ A. Deep neural networks predicting prosody (correct): neural voices model stress and intonation together with the audio itself, which is why they sound close to a human speaker.
• ✗ B. Playing back word recordings: stitching stored recordings together describes older concatenative ideas and could never cover every word or sentence naturally.
• ✗ C. Sentiment analysis choosing recordings: sentiment analysis is a text analysis technique; neural synthesis generates audio rather than picking prerecorded clips by mood.
• ✗ D. Phonetic text at a fixed pitch: a fixed pitch and pace is exactly the flat, robotic delivery neural voices avoid by predicting prosody.

🎯 Key idea: Neural text to speech uses deep neural networks to predict prosody and voice together, which makes synthesized speech sound humanlike.

🔗 More from this exam

• All AI-901 practice sets
• Every exam on this blog

📚 Beyond this set

More AI-901 questions: https://hyunhp.github.io/cert-bank/#exam=AI-901
Study guide: https://hyunhp.github.io/cert-bank/study.html#exam=AI-901

These are original practice questions, not real exam questions, and cert-bank is not affiliated with any certification vendor.

Original practice questions written for this blog, not exam dumps.