AI-900 Practice Test Bootcamp: 300+ MCQs with Explanations

AI-900 validates that you can describe AI workloads and identify Azure services that implement them, along with core concepts like training vs inference and basic evaluation. The exam expects you to recognize when to use classification, regression, clustering, and anomaly detection; to explain the difference between supervised and unsupervised learning at a practical level; and to map real-world prompts (like “detect fraudulent transactions” or “group customers by behavior”) to the right approach.

In Azure terms, you’re also expected to understand the “service menu” for common scenarios: computer vision (image analysis, OCR, customization), NLP (text analytics, language understanding, speech), and generative AI (prompt concepts, copilots, and responsible AI). The exam is less about memorizing every product name and more about selecting the service category that fits the requirement and constraints described.

Common trap: confusing a workload with a service. Workloads are the task type (e.g., OCR, classification). Services are the tools (e.g., Azure AI Vision OCR capabilities). Microsoft-style questions often describe the workload without naming it; your job is to label it correctly first, then choose the service.

Exam Tip: Train your brain to do a two-step mapping: (1) identify the workload (classification/regression/clustering/anomaly detection; OCR vs image analysis; text analytics vs language understanding; speech-to-text vs text-to-speech; generative AI with prompts), then (2) pick the Azure service family that delivers it. If you skip step 1, you’ll fall for distractors that “sound Azure-ish” but solve a different problem.

Registering for AI-900 is straightforward, but exam-day issues often come from preventable scheduling and ID problems. Start with Microsoft Learn (or the certification dashboard) to locate the AI-900 exam page, then schedule through Pearson VUE. You’ll choose either an online proctored exam or an in-person test center appointment. Both are valid; pick the one you can execute with the least risk.

Online proctoring is convenient, but it is strict: stable internet, a private room, a clear desk, and compliance with proctor instructions. In-person testing reduces technical risk, but adds travel and scheduling constraints. If you have any uncertainty about your environment—shared workspace, unpredictable noise, or unreliable connectivity—test centers are often the safer choice.

Accommodations are handled through the official workflow and can take time. If you need accommodations, initiate the request early so it’s approved before you lock your target exam date. Also confirm name matching: the name on your government ID must match your registration details closely to avoid check-in delays.

Exam Tip: Schedule your exam first, then build your study plan backwards from that date. A fixed deadline improves consistency and reduces “one more week” drift. If you’re choosing online delivery, run a system test on the same device and network you’ll use on exam day, not “something similar.”

Microsoft exams use scaled scoring, and AI-900 has a published passing score of 700. Don’t over-interpret raw “percent correct” guesses, because different question sets can vary in difficulty and weighting. Your goal is to consistently answer the mainstream objective areas correctly—especially the high-frequency items around workload identification, service selection, and responsible AI concepts.

Expect multiple-choice questions, including “choose one” and “choose all that apply.” Some items may use scenario wording with constraints (cost, speed, offline processing, need for customization) and then ask which service or approach best matches. Case-study style groupings can appear, where multiple questions share a scenario; the trap there is changing your assumptions from one question to the next. Stay consistent with the scenario facts.

Time management is typically not extreme for AI-900, but it becomes an issue if you reread long prompts repeatedly. Do one deliberate read, underline (mentally) the requirement, then evaluate options. If you’re unsure, eliminate obvious mismatches and flag the item mentally for a quick reconsideration at the end—without turning it into a time sink.

Retake policies exist, but treat your first attempt as the real attempt. A retake should be a contingency, not a plan. The most common reason people need a retake is not lack of intelligence—it’s weak exam reading skills and shallow service mapping.

Exam Tip: On “select all that apply,” assume Microsoft expects you to identify every valid statement—not the “best” one. Read each option as true/false independently. The common trap is picking the single best option and forgetting that multiple can be correct.

A high-yield study strategy for AI-900 uses domain weighting: spend the most time where the exam spends the most questions, but never ignore smaller domains because they can be the difference between 690 and 710. Your approach should cycle through four actions: learn the concept, drill with questions, review explanations, and repeat with a tighter focus on weak areas. Passive reading alone rarely translates into score gains.

For a 2-week plan, prioritize daily mixed practice and rapid feedback. Use short learning blocks to patch gaps, then immediately drill. Example structure: 30–45 minutes learning (service mapping + key concepts), 45–60 minutes practice questions, 20 minutes mistake-log review. For a 4-week plan, add depth: spend the first two weeks building a clean mental map of workloads/services, then increase practice volume and add timed sets in weeks three and four.

• Week 1–2 focus: Workload identification (classification/regression/clustering/anomaly detection), training vs inference, evaluation basics, and “which service for which job.”
• Week 3–4 focus: Higher practice volume, more scenario-based questions, responsible AI, and generative AI concepts (prompting, copilots, safety).

Common trap: studying domains in isolation and then struggling when questions blend them (e.g., vision + responsible AI, or generative AI + governance). Mixed practice is essential because the real exam is mixed by design.

Exam Tip: Track your accuracy by domain and by error type (misread requirement, didn’t know service, confused similar services, changed answer without evidence). Improving “error type” is often faster than learning new content.

Microsoft-style items are engineered to test whether you can interpret requirements, not whether you recognize buzzwords. Start by extracting the task and output type. If the question says “predict the category,” that’s classification. If it says “predict a value,” that’s regression. If it says “group similar items without labels,” that’s clustering. If it says “identify unusual behavior,” that’s anomaly detection. For vision, “extract printed/handwritten text” points to OCR, while “identify objects/tags/captions” points to image analysis. For language, “sentiment/key phrases/entities” suggests text analytics, while “intent and utterances for conversation” suggests language understanding.

Then use constraints to eliminate distractors. Words like “custom model,” “your own data,” or “tailored to your domain” are customization signals. Words like “real-time,” “low latency,” or “on the edge” hint at inference requirements. Responsible AI terms (fairness, reliability, privacy, transparency, accountability, inclusiveness) often appear as “most important consideration” distractors—choose the one that directly addresses the scenario risk described.

Elimination technique: if an option solves a different task than requested, cross it out immediately. Don’t debate “close enough.” The exam typically has one option that exactly matches the requirement and others that are plausible but mismatched (wrong workload, wrong service family, or wrong capability).

Exam Tip: Be wary of answers that are true statements but do not satisfy the question’s requirement. For example, a service may “support AI,” but if the question asks for OCR and the option is about image classification, it’s a distractor even if it’s a legitimate Azure service.

This bootcamp includes heavy MCQ practice because AI-900 is mastered through pattern recognition and correction. Your goal is not to “finish 300 questions.” Your goal is to convert every missed (or guessed) question into a durable rule you won’t miss again. That requires a baseline diagnostic and a disciplined review loop.

Start with a baseline diagnostic quiz early—before you feel “ready.” The diagnostic is not a grade; it’s a map. Categorize every miss into: (1) concept gap (didn’t know), (2) mapping gap (knew concept, chose wrong service), (3) reading error (missed a keyword), (4) overthinking (changed from correct to incorrect). This categorization determines what you do next: study content, build a service comparison table, slow down your reading, or adjust your answer-changing rule.

Build a mistake log (spreadsheet or notes) with columns: domain, question theme, what you chose, why it was wrong, the rule for next time, and the keywords that should trigger the correct choice. Review the log daily for 10–15 minutes. This is where score gains compound, because the exam repeats the same decision patterns with different wording.

Exam Tip: Don’t just record the correct answer—record the reason the wrong option was tempting. Most repeat mistakes are caused by the same “temptation,” such as confusing OCR with image analysis, mixing up training vs inference, or selecting a generic “AI service” when the question requires a specific workload capability.

Finally, use practice tests in two modes: untimed (learning mode) to build accuracy and timed (performance mode) to build stability. Move to timed sets only when your untimed accuracy is consistently strong; otherwise you’re practicing speed at the expense of correctness.

AI-900 expects you to differentiate AI, machine learning (ML), and deep learning by tying each to realistic workloads. AI is the umbrella term: any system that performs tasks that typically require human intelligence (e.g., understanding language, recognizing objects, making decisions). ML is a subset of AI where systems learn patterns from data rather than being explicitly programmed. Deep learning is a subset of ML that uses multi-layer neural networks and is especially common in vision and speech.

On the exam, workload selection is driven by business outcomes. For example, “scan invoices and capture text” is an OCR outcome (vision workload). “Detect sentiment in reviews” is a text analytics outcome (NLP workload). “Predict next month’s demand” is forecasting (often regression with time-series considerations). “Suggest products” is recommendation (ranking/personalization).

• Vision workloads: image classification, object detection, OCR, image captioning, custom vision.
• NLP workloads: language detection, key phrase extraction, sentiment analysis, named entity recognition, Q&A, translation.
• Speech workloads: speech-to-text, text-to-speech, speaker recognition (conceptual), real-time transcription.
• Generative AI workloads: content generation, summarization, chat-based copilots, retrieval-augmented experiences.

Common exam trap: confusing “computer vision” with “document understanding.” If the scenario emphasizes text in images (forms, receipts, invoices), look for OCR/document extraction rather than generic image classification. Similarly, if it emphasizes conversation or intent, prefer language understanding/Q&A over simple sentiment analysis.

Exam Tip: If the prompt mentions “custom” (custom labels, custom objects, brand-specific defect detection), the intended answer often points to a customization workflow rather than a prebuilt, generic API.

AI-900 frequently checks ML vocabulary: features are input variables (columns) used to make predictions; a label is the target output you want to predict. In supervised learning, you train on labeled data (features + known labels). In unsupervised learning, you typically have features but no labels (e.g., clustering).

Know the ML lifecycle at a high level (not tool-specific): define the problem, collect/prepare data, choose an algorithm/model, train, evaluate, deploy, and perform inference. Training is when the model learns parameters from data; inference is when a trained model is used to make predictions on new data. Many exam questions simply ask you to identify whether a described step is training or inference.

Evaluation is another common target. You are expected to recognize that model performance is measured on data the model did not train on (validation/test) and that “accuracy” is not always sufficient. While AI-900 stays high-level, it may reference basic evaluation ideas: classification often uses accuracy/precision/recall; regression uses error measures like MAE/RMSE; clustering uses cohesion/separation concepts.

Common exam trap: mixing up “feature engineering” with “labeling.” Labeling is assigning the correct outcome for each training example (e.g., “spam” vs “not spam”). Feature engineering is transforming inputs (e.g., scaling, encoding categories, extracting n-grams) to help learning.

Exam Tip: If the scenario says “a model is already trained and is used to score new customer records,” that is inference—even if it sounds like “the model is learning,” it’s not. Training happens when the model updates parameters using known labels (or an unsupervised objective).

This is one of the most testable skills in the chapter: mapping a scenario to classification, regression, or clustering. Use the “output type” rule.

Classification predicts a category/label. Binary classification examples include fraud vs not fraud, churn vs stay, approved vs denied. Multi-class classification includes classifying images into “cat/dog/bird” or routing tickets to “billing/technical/sales.”

Regression predicts a numeric value. Examples: predict house price, forecast energy usage, estimate delivery time, predict temperature. On AI-900, “forecasting” is typically treated as a regression-style problem (numeric prediction over time), even though time series adds domain nuance.

Clustering groups similar items when no labels are provided. Examples: customer segmentation, grouping documents by topic, grouping devices by usage patterns. The output is a cluster assignment or similarity grouping—not a known business label like “high risk.”

• If the output is a class name → classification.
• If the output is a number → regression.
• If the output is a group/segment discovered from data → clustering.

Common exam trap: confusing clustering with classification when the scenario uses business-friendly terms like “segment customers into platinum/gold/silver.” If those segments are predefined and labeled, it’s classification. If the segments are discovered based on similarity without labels, it’s clustering.

Exam Tip: Watch for phrasing like “predict whether” (classification), “predict how much/how many” (regression), and “identify groups” or “find similar” (clustering). These verbs are deliberate cues in AI-900 questions.

Anomaly detection identifies rare or unusual observations compared to expected patterns. AI-900 scenarios include credit card fraud spikes, sensor readings indicating equipment failure, unusual login locations, or network traffic outliers. The key idea is that anomalies are exceptions, and the system often learns “normal” behavior to flag deviations. Outcomes are commonly “anomaly score” or “flag/not flag.”

Anomaly detection is sometimes presented as classification (“fraud/not fraud”), but the exam will hint at limited labeled data or “unknown future anomalies,” which pushes you toward anomaly detection rather than supervised classification. If you have abundant labeled fraud examples, classification can work; if anomalies are rare/novel, anomaly detection is typically the better fit.

Recommendation aims to suggest relevant items to users—products, movies, articles, or next best actions. Conceptually, recommendations can be based on user-item interaction history (collaborative filtering), item attributes/content similarity (content-based), or a hybrid. The business outcome is usually ranked lists (“top 5 products to recommend”).

Common exam trap: confusing recommendation with classification. If the question asks “which product category will the user buy?” that’s classification. If it asks “which products should we suggest?” that’s recommendation/ranking. Also, recommendation scenarios often mention “people who bought X also bought Y” which is a strong cue.

Exam Tip: Look for “rare events,” “outliers,” “deviation from baseline,” or “unusual pattern” to select anomaly detection. Look for “personalized suggestions,” “rank,” or “similar users/items” to select recommendation.

AI-900 includes Responsible AI basics, often as principle matching. You’re not expected to design a full governance program, but you must recognize key principles and how they affect workload selection and deployment decisions.

Fairness is about avoiding unintended bias—ensuring similar individuals are treated similarly across sensitive groups (e.g., gender, race, age). Exam scenarios frequently involve hiring, lending, insurance, or admissions, where biased training data can lead to discriminatory outcomes.

Reliability and safety concern consistent performance under expected conditions and minimizing harmful failures. For instance, a vision model used in manufacturing quality checks must perform reliably under varying lighting; a medical triage assistant must fail safely and escalate uncertainty.

Privacy and security address protection of personal or sensitive data (PII/PHI), data minimization, proper access controls, and safe handling of prompts and outputs in generative AI settings. You may see cues like “customer addresses,” “medical records,” or “confidential documents,” which should trigger privacy considerations.

Transparency is about explainability and communicating limitations—users should understand when AI is being used and what it can/can’t do. In exam questions, transparency is often the best match when the scenario mentions “explain why the model made a decision” or “provide interpretability.”

Common exam trap: mixing transparency with fairness. “Explain why the loan was denied” maps to transparency/explainability. “Ensure approval rates are equitable across groups” maps to fairness. Another trap is assuming privacy only means encryption; on AI-900 it also includes data governance, consent, and not exposing sensitive information through outputs.

Exam Tip: For generative AI/copilot scenarios, map risks like “hallucinations” primarily to reliability/safety, “leaking customer data” to privacy/security, and “unclear source/citations” to transparency (and sometimes reliability). Use the scenario’s harm type to choose the principle.

This section coaches you through how AI-900 questions are typically constructed without listing the actual items. The exam often provides a short scenario plus four plausible-sounding workloads or services. Your job is to ignore brand names and focus on: (1) input modality (text/image/audio/tabular), (2) desired output (category/number/rank/flag), and (3) availability of labels.

Workload-selection method (use this every time):

• Step 1: Identify the data type. Images/documents → vision/OCR; free-form text → NLP; speech audio → speech; tabular business data → classic ML; mixed + generation → generative AI.
• Step 2: Identify the output type. Category → classification; number → regression; groups → clustering; unusual/rare → anomaly detection; ranked suggestions → recommendation.
• Step 3: Check constraints. Limited labels? Need customization? Sensitive domain? These push you toward unsupervised/anomaly methods, custom models, or stronger Responsible AI controls.

Expect distractors that are “adjacent.” For example, sentiment analysis vs language understanding: sentiment is polarity (positive/negative), while language understanding is intent/entities for commands or routing. In vision, image classification vs object detection vs OCR: classification answers “what is in the image overall,” detection answers “where are the objects,” OCR answers “what text is present.”

Common exam trap: choosing an algorithm name instead of a workload. AI-900 is workload-oriented. If one answer is a general workload category that matches the outcome, and another is a specific technique that doesn’t match the given output, prefer the workload match.

Exam Tip: If the scenario mentions “summarize,” “draft,” “chat,” “generate,” or “rewrite,” you’re in generative AI territory (prompting/copilot patterns). If it mentions “extract,” “detect,” “classify,” or “predict,” it is more likely a traditional ML/vision/NLP analytic workload. Use verbs as your fastest signal under time pressure.

Section 3.1: Describe ML concepts: training, validation, testing, and inference

AI-900 expects you to understand the basic lifecycle of a machine learning model and the purpose of each data split. Training is the process of fitting a model to labeled (or unlabeled) data so it learns patterns. During training, the algorithm adjusts internal parameters to reduce error. Validation data is used during model development to compare candidates (for example, different algorithms or hyperparameters) and make choices without “peeking” at the final test set. Testing is the final, unbiased evaluation on data that was not used to make model decisions.

Inference is when the trained model is used to generate predictions on new data. On Azure, inference typically happens via an endpoint, a deployed service, or a batch scoring job. The exam also tests the difference between real-time (online) scoring and batch scoring. Real-time scoring returns a prediction immediately (e.g., fraud check during checkout). Batch scoring runs predictions over a large dataset on a schedule (e.g., nightly churn risk scores for all customers).

Exam Tip: Watch for prompts that say “must respond in milliseconds” or “interactive application”—that implies real-time. Prompts like “process 10 million records overnight,” “weekly report,” or “score existing storage data” imply batch.

Common trap: confusing validation and test sets. If the question says “you used the validation set to choose the best model,” the test set should be reserved for the final performance estimate. Another trap is assuming retraining happens during inference; it does not. Inference uses fixed model parameters. Retraining is a separate training pipeline step (often triggered by schedule or drift monitoring in real solutions, but AI-900 keeps this conceptual).

Section 3.2: Supervised vs unsupervised learning and when to use each

This exam domain overlaps with the “AI workloads” outcome: classification, regression, clustering, and anomaly detection. Supervised learning uses labeled examples (input features + known target). Two major supervised workloads: classification (predict a category such as “spam/not spam”) and regression (predict a numeric value such as “house price”). The key signal in exam questions is the presence of a target label in historical data: if the prompt says “past outcomes are known,” that’s supervised.

Unsupervised learning uses unlabeled data and tries to discover structure. The most common AI-900 unsupervised workload is clustering (group similar items, such as customer segmentation). Unsupervised techniques can also support anomaly detection when you don’t have many labeled “bad” examples; the system learns what “normal” looks like and flags deviations.

• Classification: categories; usually evaluated with confusion matrix metrics.
• Regression: numeric prediction; evaluated with error metrics like RMSE.
• Clustering: groups without labels; evaluated more qualitatively in AI-900 scenarios.
• Anomaly detection: detect rare or unusual events; can be framed as supervised (if labeled anomalies exist) or unsupervised/semisupervised (if anomalies are rare).

Exam Tip: When you see “predict whether,” “approve/deny,” “yes/no,” or multiple discrete outcomes, it’s classification. When you see “forecast,” “estimate,” “how many,” or “how much,” it’s regression. When you see “segment,” “group,” “cluster,” it’s unsupervised.

Common trap: “prediction” does not automatically mean supervised classification. You can “predict” a continuous value (regression) or assign a cluster ID (unsupervised). Your job is to identify whether a labeled target exists and whether the output is categorical or numeric.

Section 3.3: Model evaluation metrics for classification and regression

AI-900 frequently tests whether you can pick an appropriate metric and interpret it. For classification, start with a confusion matrix: true positives (TP), true negatives (TN), false positives (FP), and false negatives (FN). From that, you get accuracy (overall correctness), precision (how many predicted positives were correct), and recall (how many actual positives were found). Microsoft likes scenarios where “positive” means a rare but important event (fraud, disease, defects), which makes accuracy misleading.

Precision vs recall is a classic exam pivot. If false positives are costly (e.g., blocking legitimate transactions), emphasize precision. If false negatives are costly (e.g., missing a fraud case or missing a cancer diagnosis), emphasize recall. The exam may describe this in business language rather than ML terms, so translate “minimize missed detections” into “maximize recall.”

For regression, AI-900 commonly uses RMSE (Root Mean Squared Error) as a measure of average prediction error magnitude (in the same units as the target). Lower RMSE indicates better fit (all else equal). Be careful: RMSE is sensitive to large errors due to squaring, which is often desirable when big misses are especially bad.

Exam Tip: If a prompt says “the dataset is imbalanced” or the positive class is rare, do not default to accuracy. Look for precision/recall (or F1 in some materials). If the output is numeric, do not pick accuracy—pick RMSE or related regression error metrics.

Common trap: mixing up precision and recall. A memory shortcut: precision is about the purity of what you predicted as positive; recall is about coverage of all actual positives. Another trap is assuming a “high accuracy” model is good even if it never predicts the minority class; the confusion matrix would expose that problem.

Section 3.4: Overfitting, underfitting, and improving generalization

Overfitting and underfitting are foundational concepts that Microsoft uses to test your ability to reason about training vs test performance. Underfitting happens when the model is too simple to capture the true patterns; it performs poorly on both training and test data. Overfitting happens when the model learns noise or overly specific patterns; it performs very well on training data but worse on validation/test data.

Generalization is the goal: strong performance on unseen data. On the exam, cues for overfitting include “excellent training accuracy but low test accuracy” or “model fails when deployed to new data.” Cues for underfitting include “low accuracy everywhere” or “the model cannot capture the relationship.”

• Reduce overfitting: add more training data, reduce model complexity, use regularization, or use early stopping (conceptually).
• Fix underfitting: use a more expressive model, add better features, or train longer (if training was insufficient).

Exam Tip: If training performance is high but test performance is low, think “overfitting” first. If both are low, think “underfitting” or “poor features/data quality.” Microsoft may phrase this as “model does not generalize.”

Common trap: assuming more training always solves issues. More training epochs can worsen overfitting. Another trap is confusing data leakage with overfitting. If the model performs unrealistically well, consider whether the prompt hints that the target label accidentally appears in the input features or that you evaluated on training data instead of a test set.

Section 3.5: Azure ML on Azure: workspaces, compute, datasets, and AutoML (conceptual)

AI-900 does not require you to implement pipelines, but it expects recognition of core Azure Machine Learning concepts and when to use Automated ML vs the designer. The central organizing unit is the Azure Machine Learning workspace, which acts as a management boundary for experiments, models, data connections, compute targets, and deployments.

Compute is where training or inference runs. Conceptually, you may see references to compute instances (interactive development), compute clusters (scalable training), and inference endpoints (deployment targets). Datasets/data assets represent the training/validation/test data sources and metadata, helping you reuse and version data connections in a repeatable way.

Automated ML (AutoML) is designed to help you quickly train and select a model by trying algorithms and hyperparameters for a given labeled dataset and target column. It is a strong fit when the question emphasizes “quickly find the best model,” “no deep ML expertise,” or “compare models automatically.” The Azure ML designer is a drag-and-drop interface for building ML workflows visually; it fits prompts that emphasize “visual interface,” “no-code/low-code,” and “build a training pipeline with modules.”

Exam Tip: If the scenario says “user wants to upload data, choose the target column, and let Azure pick the best algorithm,” pick Automated ML. If it says “create a workflow by dragging modules,” pick the designer. If it says “manage models, experiments, and deployments,” the umbrella answer is typically Azure Machine Learning (workspace-based).

Common trap: confusing Azure Machine Learning (the platform) with Azure AI services (prebuilt). If the prompt requires custom model training on your data, Azure Machine Learning is a likely fit. If the prompt requires prebuilt vision or NLP (OCR, sentiment, key phrases), Azure AI services are more appropriate—but this chapter stays focused on ML fundamentals and Azure ML options.

Section 3.6: Domain practice set: exam-style MCQs with explanations (ML on Azure)

This section prepares you for the “service selection + metrics interpretation” style the AI-900 exam favors. In your practice test set for this chapter, expect items that combine two skills: (1) identify the ML task (classification vs regression vs clustering vs anomaly detection), and (2) choose the metric or Azure ML option that matches the requirement. The exam rarely asks you to compute metrics numerically; instead, it checks whether you know which metric matters and why.

Scoring pattern questions are also common. Prompts about “process all records in storage each night” are testing batch inference, while prompts about “call an endpoint from a mobile app” are testing real-time inference. Another frequent pattern is lifecycle ordering: you train using training data, tune with validation, evaluate with test, then deploy for inference.

Exam Tip: Train yourself to underline (mentally) these keywords in each prompt: label/target (supervised), group/segment (clustering), rare event (anomaly detection), numeric forecast (regression), imbalanced (precision/recall), endpoint/latency (real-time), overnight/schedule (batch), visual drag-and-drop (designer), automatically try algorithms (AutoML).

Mini-review of common pitfalls (and how Microsoft frames them): (1) choosing accuracy on imbalanced data; you should pivot to precision/recall and justify based on FP vs FN cost, (2) mixing up validation and test sets; the test set is for final evaluation, (3) calling training “scoring”; scoring is inference, (4) assuming clustering needs labels; it doesn’t, and (5) choosing Azure ML when a prebuilt AI service would suffice—watch whether the prompt says “custom model with our data” versus “use a prebuilt capability.” Your practice questions will reinforce these patterns so you can answer quickly and consistently under timed conditions.

On AI-900, “computer vision workloads” are less about neural network details and more about recognizing the outcome the user wants from images or video. The exam commonly tests four buckets: (1) image analysis (what’s in the picture), (2) OCR (what text is in the image/document), (3) face-related scenarios (detecting faces or face attributes—subject to policy constraints), and (4) custom vision (training a model to recognize your own categories/objects).

Image analysis use cases include tagging a photo (“outdoor, dog, car”), generating a natural-language caption (“a dog running on grass”), detecting objects with bounding boxes, or identifying content characteristics. OCR use cases include reading a receipt photo, extracting text from a sign, or digitizing scanned documents. Face scenarios show up as “find faces in images,” “verify whether the same person appears,” or “analyze facial features,” but the exam expects you to know that face recognition and biometric identification have responsible AI constraints and may be restricted depending on region/policy.

Custom vision scenarios involve non-generic categories like “detect cracks in a turbine blade,” “classify product SKUs by packaging,” or “identify defects.” Here you need labeled training images and a training lifecycle.

Exam Tip: Translate every question into a noun phrase: “tags/caption,” “printed/handwritten text,” “face detection/verification,” or “custom classification/object detection.” Then select the Azure service that natively supports that noun phrase.

• Common trap: Choosing a general image analysis service when the requirement is “extract fields” (structured key-value pairs). That leans toward document processing rather than generic tags.
• Common trap: Confusing “object detection” (bounding boxes) with “image classification” (single label for the whole image). Custom Vision supports both; know which the scenario describes.

Azure AI Vision (often referenced on the exam as “Azure AI Vision” or “Computer Vision”) provides prebuilt image understanding. In questions, look for tasks like generating tags, a caption, identifying objects, and other general “what’s in this image?” requirements. The exam typically doesn’t require API method names, but it does require you to pick Vision for image analysis scenarios where you are not training a custom model.

Conceptually, know the difference between outputs: tags are keyword-like labels, captions are short natural-language descriptions, and detection implies locating items (bounding boxes) rather than just describing the image. If the scenario mentions “highlight where the item appears,” that is detection. If it just says “categorize photo as beach/mountain/city,” that’s classification-like behavior and may be handled by prebuilt tagging/categorization, unless the categories are business-specific.

AI-900 also tests your ability to avoid over-engineering. If a solution can be done with prebuilt image analysis, don’t pick Custom Vision. If the scenario wants a quick proof-of-concept or standard labels, Vision is the likely answer.

Exam Tip: “No training data + wants tags/caption/objects” is the signature for Azure AI Vision image analysis. If you see “bounding boxes,” read carefully: prebuilt detection can help for common objects; if the objects are unique to your domain (factory parts, specialized medical imagery), expect Custom Vision object detection instead.

• Common trap: Selecting OCR when the question says “analyze the image content” but never mentions text extraction.
• Common trap: Selecting a generative AI service for captioning when the exam is clearly targeting the dedicated vision service’s caption capability.

OCR questions are extremely common because they look deceptively similar to generic image analysis. OCR is specifically about extracting text—printed or handwritten—from images or scanned documents. On Azure, many AI-900 questions reference the idea of a “Read” capability (extract text lines/words) versus “Analyze document” capabilities (extract structured data such as fields, tables, and key-value pairs). Even when the exam uses simplified wording, you must distinguish between “just get the text” and “understand the document’s structure.”

Use the following mental split. If the scenario says: “extract all text from a photo of a receipt” or “read text from street signs,” that is OCR/Read. If it says: “extract invoice number, total amount, vendor name, and line items,” that is a forms-like/document processing scenario—structured extraction—commonly mapped to Azure AI Document Intelligence (formerly Form Recognizer) rather than generic OCR alone.

Exam Tip: The keyword “fields” is your giveaway. OCR gives you words; document intelligence gives you meaningful labeled fields and tables. Don’t let a distractor push you to simple OCR when the scenario demands structured outputs.

• Common trap: Picking image analysis tags/captions when the user requirement is “digitize text.” Tags describe; OCR transcribes.
• Common trap: Picking OCR for invoices when the question expects extraction of specific fields (invoice ID, totals). That’s document analysis.

Finally, remember that “training vs inference” still applies: many document scenarios can use prebuilt models (receipts, invoices) without custom training. If the scenario asks for unusual document types (custom forms), then custom document models may be needed—but AI-900 often stays at the recognition level: Read vs structured document extraction.

Face-related questions test both capability recognition and responsible AI awareness. Technically, face workloads include detecting faces in an image (finding face regions) and analyzing attributes (for example, locating facial landmarks). Some scenarios describe verifying whether two faces match (verification) or identifying a person from a known set (identification). Conceptually, verification is a “one-to-one” check, while identification is “one-to-many.” This distinction often appears as a subtle wording change in multiple-choice options.

However, the exam also expects you to recognize that biometric identification and sensitive attribute inference are high-risk. Azure’s face capabilities are governed by strict requirements, limited access in some cases, and responsible AI policies. If a scenario implies surveillance, mass identification, or inferring sensitive traits, that should raise a red flag. AI-900 may not ask you to memorize policy documents, but it will reward selecting answers that align with responsible use and choosing non-biometric alternatives when appropriate.

Exam Tip: When options include “Face” versus a more general “Vision,” choose Face only when the scenario explicitly requires face detection/verification-like functionality. If the task is “count people” or “detect persons,” that might be object/person detection rather than face recognition.

• Common trap: Assuming any “people in images” scenario is Face. Many are just object detection (“person” class) in image analysis.
• Common trap: Missing the verification vs identification cue words. “Is this the same person?” (verification) vs “Who is this person among employees?” (identification).

In your exam reasoning, include governance: if the scenario is questionable, the safest answer is often the one that avoids biometric identification or adds human review and consent-oriented controls.

Custom Vision is tested as the “when prebuilt isn’t enough” option. The exam focuses on the workflow, not deep ML theory: collect images, label them, train, evaluate, iterate, and then use the trained model for prediction (inference). Two major custom tasks show up: image classification (assign a label to the whole image) and object detection (assign labels plus bounding boxes). The question stem usually tells you which one you need by mentioning either “categorize each image” (classification) or “locate each defect/object” (detection).

Training concepts that appear on the exam include the need for representative images (lighting, angles, backgrounds), sufficient examples per label, and consistent labeling. AI-900 is also likely to mention “iterations” or “improving accuracy,” where the correct action is to add more labeled data, fix label quality, and retrain—rather than changing unrelated services.

Evaluation is another testable area. You don’t need to compute metrics, but you should recognize that models are evaluated using measures like precision/recall or overall performance, and that you validate before deploying. If the scenario says “too many false positives,” you should think “precision issues” and consider adjusting thresholds or improving training data; if it says “missing many true defects,” think “recall issues.”

Exam Tip: The fastest way to spot Custom Vision is the presence of business-specific categories or unique objects plus an explicit willingness to provide labeled training images. Prebuilt services rarely learn your proprietary SKUs without custom training.

• Common trap: Choosing Custom Vision just because the word “classify” appears. If it’s generic categories (dog/cat/car), prebuilt Vision can suffice.
• Common trap: Forgetting that object detection requires bounding boxes during labeling; classification does not.

This section is your “practice set without the questions.” Use it as a checklist to self-grade your reasoning on any MCQ you encounter. AI-900 vision items are usually answerable by identifying two things: (1) the output type (tags/caption, text, structured fields, faces, custom labels/boxes) and (2) whether training is implied or not.

How to identify the correct answer on test day: First underline requirement verbs: “describe,” “detect,” “read,” “extract fields,” “verify,” “train.” Then map to service families: Azure AI Vision for general image understanding; OCR/Read for text transcription; Azure AI Document Intelligence for forms-like extraction; Face for face-specific operations (with policy awareness); Custom Vision for bespoke classification/detection with your own labeled images.

Common distractor patterns: (a) “OCR” offered when the scenario wants key-value pairs (should be document analysis). (b) “Custom Vision” offered when no training data is available (prebuilt is intended). (c) “Face” offered for “people counting” (object detection is sufficient). (d) “Image classification” offered when the scenario needs bounding boxes (object detection).

Exam Tip: When two options both seem plausible, pick the one that satisfies the requirement with the least customization and fewest moving parts. AI-900 emphasizes choosing the correct managed service, not building pipelines from scratch.

Rapid recap (capabilities + limitations): Prebuilt Vision is great for generic concepts and common objects; OCR reads text but doesn’t inherently understand “invoice total” as a field; document analysis is for structured extraction; face/biometric scenarios require extra governance and may be restricted; Custom Vision requires labeled training images and supports classification and object detection with an iterative improvement loop.

Section 5.1: Describe NLP workloads: text classification, extraction, summarization

NLP workloads on AI-900 are best understood by the “job” being done on text. The three most tested categories are: classification (assign labels), extraction (pull structured information from unstructured text), and summarization (condense or rewrite content). When the exam says “categorize emails into Billing/Support/Sales,” that’s classification. When it says “find people, organizations, locations, or invoice numbers,” that’s extraction. When it says “create a brief summary of a long call transcript,” that’s summarization.

On Azure, these map broadly to Azure AI Language (for analysis) and generative AI (for free-form summarization and content creation). AI-900 questions often blur summarization: classic extractive summarization can be treated as an NLP analysis feature, while generative summarization is a generation workload. If the prompt emphasizes “use an LLM” or “draft a new summary in natural language,” you should lean toward a generative AI workload. If it emphasizes “extract important sentences” or “analyze documents,” that leans toward language analytics features.

Exam Tip: Classification and extraction typically produce structured outputs (labels, entities, key phrases). Generative summarization typically produces new text. When an option mentions “tokens,” “prompts,” or “chat,” it’s pointing you to generative AI.

• Text classification: spam detection, topic labeling, routing tickets.
• Extraction: key phrases, named entities, PII detection, language detection.
• Summarization: meeting notes, executive summaries, shortened descriptions.

Common trap: choosing a generative AI service for simple extraction tasks. If the goal is to identify entities or sentiment reliably at scale, text analytics is usually the intended answer because it’s deterministic in output shape and easier to validate. Conversely, choosing text analytics when the requirement is to “write” or “compose” is also a trap—analytics won’t generate a new email reply or create a marketing paragraph.

Section 5.2: Text analytics concepts: sentiment, key phrase extraction, entity recognition

Text analytics questions are common because they map cleanly to business outcomes. The AI-900 exam expects you to know what each analysis returns and when you’d use it. Sentiment analysis predicts the emotional tone (for example, positive/negative/neutral). The exam often frames this as “measure customer satisfaction from reviews” or “monitor social media sentiment.” The output is usually a label plus scores (confidence).

Key phrase extraction identifies the main talking points in a document (for example, “late delivery,” “refund policy,” “battery life”). It’s not summarization; it’s a list of important phrases. This distinction is tested: key phrases give “topics,” whereas summarization gives “sentences/paragraphs.”

Named entity recognition (NER) extracts entities such as people, locations, organizations, dates, and sometimes domain-specific entities depending on the feature set. Many exam questions describe “extract company names and product names from news articles” or “find patient names and dates in clinical notes.” You should associate that with entity recognition/extraction.

Exam Tip: Look for the noun in the requirement: if they want “a score” (sentiment), “a list of phrases” (key phrase extraction), or “structured fields like names/places” (NER). Don’t overthink model training—AI-900 questions frequently assume you can use prebuilt capabilities without custom ML.

• Sentiment: trend dashboards, alerts on negative feedback spikes.
• Key phrases: topic indexing, tag generation, search facets.
• Entity recognition: compliance checks, CRM enrichment, document indexing.

Common trap: confusing NER with OCR. OCR extracts text from images; NER extracts entities from text. If the scenario starts with scanned PDFs or photos, OCR is needed first (computer vision domain), then language analytics can run on the extracted text. Another trap is mixing translation with sentiment: if the data is multilingual, the exam may hint you should detect language or translate before downstream analytics.

Section 5.3: Speech and conversational concepts: speech-to-text, text-to-speech, bots

Speech workloads are tested as “audio in, text out” or “text in, audio out,” plus conversational experiences like bots and virtual agents. Speech-to-text is transcription: converting call center recordings, meetings, or dictated notes into text. Text-to-speech is voice synthesis: reading responses aloud for accessibility, IVR systems, or voice assistants.

AI-900 questions often provide an accessibility requirement (“read product descriptions aloud”) or a contact center requirement (“transcribe calls for QA review”). Map these directly to Azure AI Speech capabilities. If the question mentions “real-time captions,” “transcription,” or “speaker diarization,” it’s in speech territory. If it mentions “translate speech,” it still indicates Speech plus translation features (the key is that the input is audio).

Conversational solutions combine NLP and orchestration. The exam may describe a bot that answers FAQs, hands off to an agent, or triggers workflows. The key concept: bots need understanding (intent/entities) and dialog management (conversation flow). In many modern designs, a generative model can help produce fluent responses, but classic bots rely on structured intents, predefined responses, or grounded knowledge sources.

Exam Tip: Identify the modality first. If the source is audio, start with Speech. If it’s typed text, start with Language. Then decide if the output is analysis (labels/entities) or generation (free-form text).

• Speech-to-text: compliance, searchable transcripts, meeting minutes.
• Text-to-speech: kiosks, accessibility, voice UI.
• Bots: customer support triage, HR self-service, IT helpdesk.

Common trap: selecting a chatbot/generative AI option when the requirement is only transcription. Another trap is forgetting that many conversational systems require grounding to avoid making up answers—if the scenario emphasizes “must answer only from approved documentation,” you should think beyond a generic bot and toward a grounded or retrieval-augmented approach.

Section 5.4: Generative AI workloads: content generation, Q&A, and retrieval-augmented patterns

Generative AI questions in AI-900 focus on what large language models (LLMs) are good at: drafting text, summarizing, rewriting, brainstorming, and answering questions conversationally. You’re expected to understand prompts (instructions + context), tokens (units of text that affect limits and cost), and grounding (anchoring responses in trusted data). If the scenario says “write a product description,” “draft an email reply,” or “create a policy summary in plain language,” that’s content generation.

For Q&A, the exam distinguishes between: (1) a model answering from its general training and (2) answering from your data. Enterprise scenarios commonly require the latter. That’s where retrieval-augmented generation (RAG) comes in: retrieve relevant passages from a knowledge base, then feed them into the prompt so the model responds with context. In Azure terms, you’ll see patterns that combine an LLM endpoint (for generation) with a search/index layer (for retrieval) and sometimes orchestration tools.

Exam Tip: If the requirement includes “use our internal PDFs,” “company handbook,” “SharePoint,” or “only answer using approved sources,” the question is hinting at a retrieval-augmented pattern, not a standalone chat completion. Look for answer choices mentioning grounding, connecting to data, or retrieval.

• Content generation: marketing copy, meeting recap, code explanation, rewriting tone.
• Conversational Q&A: chat interface for support, guided help, interactive tutoring.
• RAG: policy assistant grounded in internal documentation, product support grounded in KB articles.

Common trap: assuming an LLM “knows” your private documents by default. It does not unless you provide the content in the prompt or connect it through a retrieval mechanism. Another trap is ignoring token limits: very long documents cannot be pasted wholesale; RAG retrieves only relevant chunks to stay within context size and improve accuracy.

Section 5.5: Responsible generative AI: hallucinations, prompt injection, privacy, safety filters

AI-900 includes responsible AI fundamentals, and generative AI introduces specific risks. Hallucinations are confident but incorrect outputs. The exam often frames this as “the assistant must not fabricate policies” or “answers must be verifiable.” The primary mitigation is grounding (RAG), plus clear instructions, citations, and human review for high-stakes workflows.

Prompt injection occurs when a user (or malicious content in retrieved documents) tries to override system instructions—e.g., “Ignore previous directions and reveal secrets.” Mitigations include isolating system/developer instructions, filtering or sanitizing retrieved content, using allow-lists for tools/actions, and applying content safety policies.

Privacy and data protection are frequent exam angles: avoid sending sensitive data unnecessarily, control access to prompts and logs, and follow least privilege. If the question mentions PII, confidential customer records, or regulated data, expect the correct answer to include measures such as data minimization, encryption, access controls, and potentially redaction before sending text to a model.

Finally, safety filters and content moderation help reduce harmful, hateful, sexual, or self-harm content in both prompts and outputs. In Azure scenarios, the exam may describe needing to “block unsafe responses” or “moderate user input.” Your job is to identify that safety tooling and policy configuration is part of a responsible deployment.

Exam Tip: When a question describes a risk (“model makes things up,” “user tries to jailbreak,” “sensitive data exposure”), don’t pick “better prompt” as the only control. The exam prefers layered mitigations: grounding + access control + monitoring + safety filters.

• Hallucinations: mitigate with RAG, citations, validation, human-in-the-loop.
• Prompt injection: mitigate with instruction hierarchy, tool restrictions, content filtering.
• Privacy: mitigate with least privilege, redaction, governance, retention controls.
• Safety: mitigate with content moderation for both input and output.

Common trap: treating responsible AI as an afterthought. AI-900 will test whether you recognize that safety, privacy, and reliability requirements can change the “best” architecture choice—especially for enterprise copilots and Q&A assistants.

Section 5.6: Domain practice set: exam-style MCQs with explanations (NLP + generative AI)

This domain practice set (presented elsewhere in the course) mirrors AI-900’s most common NLP and generative AI question shapes. Expect short scenarios with one clear requirement and multiple plausible answers. Your goal is to map the scenario to the correct workload first, then to the appropriate Azure capability family. The explanations will emphasize why one option is best and why the distractors are tempting.

What you’ll practice includes: (1) separating text analytics (sentiment, key phrases, NER) from generation (summaries, drafting), (2) identifying speech needs (transcribe vs synthesize), (3) recognizing when a chatbot is really a grounded Q&A problem (RAG), and (4) applying responsible AI controls when the scenario includes risk signals (PII, hallucinations, jailbreak attempts, harmful content).

Exam Tip: Use a two-pass method on MCQs: pass one—underline the verb (“extract,” “classify,” “transcribe,” “generate,” “answer using company docs”). Pass two—eliminate options that don’t match the input/output modality (audio vs text) and output type (structured labels vs free-form text). This quickly removes 2–3 distractors.

• High-frequency scenario cues: “reviews” → sentiment; “tags/topics” → key phrases; “names/addresses” → entities/PII; “audio recordings” → speech-to-text; “read aloud” → text-to-speech; “write a draft” → generative; “must use internal docs” → RAG/grounding.
• Trick cues: “scanned documents” implies OCR first; “multilingual” implies language detection/translation; “compliance” implies privacy controls and auditing.

Common trap: choosing the most “advanced” option (LLM/chat) when a simpler analytics feature directly satisfies the requirement. The exam rewards selecting the appropriate capability, not the fanciest. In the practice set, review explanations for eliminated options—those distractors are often reused across questions with only minor wording changes.

Section 6.1: Mock exam instructions: pacing, marking, and review strategy

Run both mock exams as if they are the real AI-900: timed, uninterrupted, and with the same device setup you plan to use on exam day. Your pacing target should be “steady, not perfect”—you’re training decision-making under time constraints. Use a simple three-pass approach: Pass 1 answers what you know immediately; Pass 2 revisits marked items; Pass 3 is a final sweep for misreads and requirement mismatches.

Marking strategy is critical because AI-900 distractors are often plausible. Mark any question where you (a) are torn between two service names, (b) see negation words (NOT, EXCEPT, LEAST), or (c) must infer the workload type. Your goal is not to “sit and think” for a full minute; it’s to capture the uncertainty and move on, preserving time for higher-value review later.

Exam Tip: Treat every question as “workload → service → feature.” Example mental chain: “extract printed text from images” → OCR workload → Azure AI Vision (Read). If you jump straight to a service, you’ll get tricked by sibling services that sound right.

• Time box: If you cannot justify your choice in one sentence, mark it and move on.
• Keyword scan: Circle mentally: “predict,” “classify,” “group,” “detect anomalies,” “generate,” “summarize,” “translate,” “OCR,” “custom,” “training.”
• Review discipline: On review, first confirm you understood the ask. Many misses come from answering a different question than the one asked.

Finally, use a scratch method for mapping: after the mock exam, tag each miss with a domain label (Workloads, ML fundamentals, Vision, NLP/Speech, GenAI/RAI). That becomes your Weak Spot Analysis in Section 6.4.

Section 6.2: Full mock exam set A (timed) and answer rationales

Mock Exam Set A is designed to stress “service selection under similar-sounding options.” When you review rationales, focus on the decisive clue the scenario gives you. AI-900 commonly tests whether you can distinguish workloads: classification (categorical label), regression (numeric value), clustering (grouping without labels), and anomaly detection (outliers). If the rationale says “the output is a number,” that’s not trivia—it’s the key to the exam objective “Describe AI workloads and when to use each.”

Set A also reinforces Azure ML fundamentals. Expect rationales to highlight training vs inference: training is where the model learns patterns from historical labeled data; inference is where the deployed model produces predictions on new data. Many candidates lose points by mixing up model evaluation (validation/test metrics) with inference (live scoring). The rationales should repeatedly connect “evaluate” to metrics such as accuracy, precision/recall, or mean absolute error—chosen based on classification vs regression.

Exam Tip: If a scenario mentions “deploy,” “endpoint,” “real-time predictions,” or “integrate into an app,” you are in inference territory. If it mentions “experiment,” “labeled dataset,” “split data,” or “improve performance,” you are in training/evaluation territory.

• Common trap: Choosing a vision service for text-only sentiment tasks. Text sentiment belongs to Azure AI Language (Text Analytics) workloads.
• Common trap: Confusing OCR with object detection. OCR extracts text; object detection locates objects with bounding boxes.
• Common trap: Overusing “Azure Machine Learning” as a catch-all. The exam expects you to pick purpose-built Azure AI services when the scenario is a standard vision/NLP/speech task.

When rationales refer to customization, lock onto the word “custom.” “Custom Vision” is for custom image classification/object detection; “Language understanding” maps to building intent/entity models (e.g., conversational command interpretation); “Speech” maps to speech-to-text, text-to-speech, and translation. In Set A, prioritize learning the boundary lines between these families—most distractors live on those borders.

Section 6.3: Full mock exam set B (timed) and answer rationales

Mock Exam Set B is tuned to newer AI-900 emphasis: generative AI workloads on Azure and responsible AI considerations. Your rationales should repeatedly connect “generate/summarize/compose” to large language models and Azure OpenAI, and connect “prompt” to concepts like instructions, context, examples (few-shot), and grounding data sources. The exam is not asking you to be a prompt engineer; it is asking you to recognize prompt components and apply safe, effective usage patterns.

Set B also blends classic workloads with GenAI to test whether you can separate “predictive ML” from “generative AI.” If the scenario asks to forecast a numeric demand value, that’s regression; if it asks to draft an email from bullet points, that’s generative. Rationales will often explain that generative models output new content rather than a label/number, and that evaluation differs (human review, relevance, safety) even though you still care about quality measures.

Exam Tip: Watch for “copilot” language. Copilots are embedded assistance experiences that orchestrate prompts, tools, and data. The exam may describe the experience without naming the product; your job is to identify it as a generative AI workload with governance needs.

• Responsible AI trap: Selecting “accuracy” as the main mitigation for harmful outputs. Responsible AI controls also include content filters, grounding, monitoring, human-in-the-loop, and limiting data exposure.
• Data privacy trap: Assuming you should paste sensitive data into prompts. Rationales should stress minimizing sensitive input and using approved enterprise patterns.
• Service-family trap: Using Azure AI Language for open-ended generation. Language services handle analysis (sentiment, key phrases, entity extraction, translation), while Azure OpenAI handles generative text/code/image depending on the model.

Finally, Set B should reinforce speech and multimodal boundaries: speech-to-text and text-to-speech are Azure AI Speech; OCR and image tagging are Vision; intent/entity extraction is Language understanding; and “chat completion with prompts” belongs to Azure OpenAI. Read each rationale as a map: scenario verbs (transcribe, detect, extract, generate) point to the correct domain.

Section 6.4: Score report: mapping misses to the official exam domains

Your weak spot analysis must be objective-driven, not emotion-driven. After each mock exam, build a simple score report table with five rows: (1) AI workloads, (2) ML fundamentals on Azure, (3) Computer vision, (4) NLP & speech, (5) Generative AI & responsible AI. For every miss, tag it to exactly one row and write a one-line “miss reason” such as: “mixed up classification vs regression,” “confused training vs inference,” or “picked wrong service family.” This is how you turn mistakes into guaranteed points.

Patterns matter more than individual questions. If you miss several items due to the same confusion (e.g., OCR vs text analytics), that is a high-ROI fix. If misses are scattered, your issue is likely reading precision or rushing, not knowledge. The exam is designed so that careful readers can eliminate 2 choices quickly—your report should record whether you were able to eliminate distractors and why you didn’t.

Exam Tip: Track “distractor type.” Common distractor types include (a) correct workload but wrong service, (b) correct service family but wrong feature, and (c) plausible buzzword unrelated to the scenario. Naming the distractor type makes your next practice session targeted.

• Workloads misses: Re-drill the output type (label/number/group/outlier) and typical examples.
• ML fundamentals misses: Re-drill training vs inference, dataset splits, and metric selection by task.
• Vision misses: Re-drill image analysis vs OCR vs custom vision, and what “bounding boxes” implies.
• NLP/Speech misses: Re-drill sentiment/key phrases/entities vs intent understanding, plus speech transcription/synthesis.
• GenAI/RAI misses: Re-drill prompt parts, copilots, and mitigations (filters, grounding, monitoring, human review).

Use this report to decide what to review in the final 48 hours. Don’t re-read everything. Fix the top two miss clusters first, then retake a small set of questions focused on those domains to confirm the gap is closed.

Section 6.5: Final review: high-frequency concepts and common distractors

Your final review sprint is a “recognition drill.” AI-900 frequently tests the same conceptual pivots: workload selection, service selection, and responsible usage. Build a one-page mental sheet of high-frequency cues. For workloads: classification predicts categories (spam/not spam), regression predicts numbers (price), clustering finds groups (customer segments without labels), anomaly detection flags outliers (fraud). For ML fundamentals: training builds the model; inference uses the trained model; evaluation selects metrics aligned to the task.

For Azure services, focus on the core match-ups the exam expects: Azure Machine Learning for building/training/deploying custom ML models; Azure AI Vision for image analysis and OCR; Custom Vision for customized image classification/object detection; Azure AI Language for text analytics tasks like sentiment/key phrases/entities and for intent/entity understanding; Azure AI Speech for speech-to-text and text-to-speech; Azure OpenAI for generative text/code/image scenarios driven by prompts.

Exam Tip: When two answers are both “true statements,” the exam wants the statement that best satisfies the scenario constraint (e.g., “needs custom labels,” “must read text from images,” “must generate a summary”). Choose the option that is most specific to the requirement.

• Distractor: “Translator” vs “Speech translation.” If the scenario is spoken audio, Speech is the anchor service family.
• Distractor: “Computer Vision” vs “Custom Vision.” If the scenario needs a custom model trained on your images, choose Custom Vision.
• Distractor: “NLP analysis” vs “GenAI generation.” If the task is extraction (entities, sentiment), choose Language; if it is creation (draft, summarize, answer), choose OpenAI.
• Distractor: Responsible AI framed as only fairness. Also include privacy, safety, transparency, reliability, and accountability.

Last-48-hours plan: Day -2, do a targeted review of your two weakest domains, then do a light timed set to practice pacing. Day -1, stop heavy new learning; do quick service/workload flash review and read your miss-reason notes. Sleep is a performance tool—treat it like one.

Section 6.6: Exam day checklist: setup, rules, and confidence framework

Exam day performance is process plus calm. Your checklist should cover environment, identity requirements, and a mental framework for decision-making. If you’re taking the exam online, prepare a quiet room, stable internet, and a clear desk. Have your ID ready and ensure your webcam and microphone work. If you’re testing at a center, arrive early to buffer for check-in and to settle your nerves.

Rules are where candidates accidentally create stress. Don’t assume you can use your phone, notes, or extra monitors. Remove anything that looks like study material from view. Close background apps and notifications. Most importantly, rehearse your timing plan: three-pass approach, marking uncertain items, and leaving time to review negation/exception questions.

Exam Tip: Your confidence framework should be “eliminate, match, verify.” Eliminate obviously wrong workload types, match the remaining choices to the scenario verb (detect/extract/generate/transcribe), and verify with one detail (custom vs prebuilt, image vs text vs audio, training vs inference).

• Before starting: Read the first question slowly to set your pace; don’t sprint.
• During the exam: If you feel stuck, mark and move; momentum prevents time panic.
• Final five minutes: Re-check questions with NOT/EXCEPT and any you answered while rushed.

Walk in expecting a few unfamiliar phrasings. AI-900 is designed so that if you can map a scenario to the correct workload and service family, you can succeed even when the wording is new. Trust your process: you trained with full mock exams, diagnosed weak spots by domain, and sharpened high-frequency concepts for the final sprint.