Microsoft AI Fundamentals for Non-Technical Pros (AI-900)

AI-900 validates that you can describe common AI workloads and map them to appropriate Azure capabilities—at a conceptual level. The exam targets business stakeholders, project managers, analysts, students, and anyone who needs to evaluate AI solutions responsibly. Expect questions framed as workplace scenarios: a customer support team wants to extract sentiment from reviews, a factory needs to detect defects in images, or leadership wants guidance on when generative AI is appropriate.

The exam is aligned to five outcome areas you will see repeatedly: (1) AI workloads and considerations (including Responsible AI), (2) fundamentals of machine learning on Azure (supervised vs unsupervised, training vs inference, evaluation metrics), (3) computer vision workloads (classification, object detection, OCR, facial analysis concepts), (4) NLP workloads (text analytics, translation, speech, language understanding), and (5) generative AI workloads (foundation models, prompting, copilots, Azure OpenAI, safety/governance).

Exam Tip: AI-900 rewards “workload thinking.” Before choosing an answer, label the task as one of: prediction/classification (ML), perception from images (Vision), understanding text/speech (NLP), or content creation/assistance (Generative AI). Many distractors are plausible services from the wrong workload family.

Common trap: confusing “Azure Machine Learning” (a platform for building/training/managing ML) with “Azure AI services” (prebuilt APIs like Vision or Language). If the scenario is “we don’t have ML expertise and need quick capabilities,” the exam often expects prebuilt Azure AI services. If it emphasizes training a custom model, tracking experiments, or deploying endpoints, Azure Machine Learning is more likely.

Scheduling is simple, but you must choose the right testing provider based on your context. Most candidates schedule through Microsoft’s certification dashboard and test with Pearson VUE (either online proctored or at a test center). Some academic programs and training providers use Certiport, especially in school settings. Confirm which provider your program requires before purchasing an exam attempt to avoid administrative delays.

Typical Pearson VUE path: sign in with your Microsoft account, find “AI-900,” select your country and language, then choose either an online proctored delivery or a test center. Online proctoring requires a compatible device, stable internet, and a clean testing environment. Test centers reduce the risk of connectivity or room-compliance issues but require travel and scheduling around available slots.

Exam Tip: If you select online proctoring, run the system test well before exam day and again the day prior. Many failed check-ins are not knowledge problems—they’re webcam permissions, corporate VPNs, or restricted networks.

Decision guidance: choose online proctoring if you can control your environment (quiet room, no interruptions) and have reliable internet. Choose a test center if your home/office is unpredictable, you share spaces, or your network is locked down. For either option, schedule for a time when you are mentally fresh; fundamentals exams still require careful reading and sustained attention.

AI-900 is organized around objective domains that map closely to the course outcomes: AI workloads and Responsible AI, ML fundamentals on Azure, computer vision, NLP, and generative AI concepts. The exam is not a hands-on lab exam; you are tested through scenario-based questions, conceptual comparisons, and service-selection decisions. You should be comfortable with the “why” behind a choice (for example, why OCR is the right approach for extracting printed text from receipts) rather than implementation details.

Common item types include multiple choice, multiple response (select all that apply), and scenario-driven prompts where you must identify the most appropriate workload or service. Expect “best answer” wording. Two options might be true statements, but only one fits the scenario constraints (cost, speed to implement, data labeling effort, governance, or real-time needs).

Exam Tip: Watch for qualifiers: “real-time,” “no labeled data,” “explainable,” “must detect multiple objects,” “extract text,” “translate,” “summarize,” “generate,” “minimize bias.” These words usually indicate the target domain and eliminate half the options immediately.

Common traps to anticipate: mixing up image classification vs object detection (classification answers “what is in the image,” detection answers “what and where” with bounding boxes), confusing sentiment analysis (opinion polarity) with key phrase extraction (important terms), and treating generative AI as a replacement for deterministic business rules. If a requirement demands strict, repeatable outputs with no creativity (like validating a fixed ID format), traditional software or rules may be preferred over AI.

Microsoft certification exams use scaled scoring. You’ll see a score report and a pass/fail result; the underlying number of questions can vary, and not every question contributes equally. Because scoring is scaled, you should focus on mastering the objective domains rather than trying to “game” the test by predicting how many questions appear from each area.

Policy essentials: arrive early (or check in early online), follow ID requirements, and understand that exam rules are strict—especially for online proctoring. Online sessions may require photos of your testing space and continuous monitoring; looking away frequently, using a phone, or having notes visible can trigger warnings or termination. At a test center, personal items are typically stored in a locker and breaks may be limited by the exam rules.

Exam Tip: For online testing, remove extra monitors, close background apps, and silence notifications. Minor compliance issues can become major disruptions that affect your focus and timing.

If you need accommodations, request them early through the official accommodations process; do not wait until the week of your exam. Retake policies can change over time, so verify current rules at scheduling time. Practically, your plan should assume that if you do not pass, you will review weak domains from the score report and retake after targeted practice—especially focusing on service selection and workload identification, which are common miss areas for non-technical candidates.

AI-900 preparation is most efficient when you treat it like vocabulary plus decision-making patterns. Passive reading creates a false sense of mastery. Use active recall: close your notes and explain, in your own words, the difference between supervised and unsupervised learning; what “training” produces (a model) versus what “inference” does (uses the model to predict); and which Azure service family fits a given scenario (Vision vs Language vs Azure Machine Learning vs Azure OpenAI).

Layer in spaced repetition: revisit key concepts over multiple days. Build a short “concept deck” (digital or paper) with prompts like “When would you use OCR?” or “What is a responsible AI consideration?” Repetition matters because the exam uses similar ideas across different domains (for example, fairness and privacy appear in ML, Vision, NLP, and generative AI scenarios).

A “labs-lite” approach works well even for non-technical learners: you don’t need deep coding, but you should see what the services do. Try guided demos or portal walkthroughs (where available) to connect terms to outcomes: classify an image, run OCR on a document, analyze sentiment, or test a prompt for summarization. This reduces confusion between similarly named capabilities.

Exam Tip: Build a one-page “service map” as you study: list common tasks (translate text, detect objects, extract key phrases, generate a summary) and map them to the correct Azure service category. Many questions are solved by recognizing the category immediately.

Two beginner-friendly schedules:

• 2-week plan: Days 1–2 exam orientation + Responsible AI basics; Days 3–5 ML fundamentals and evaluation; Days 6–8 Vision workloads; Days 9–11 NLP workloads and speech; Days 12–13 Generative AI/Azure OpenAI and governance; Day 14 full review with weak-area focus.
• 4-week plan: Week 1 orientation + Responsible AI + AI vs traditional software; Week 2 ML fundamentals + Azure Machine Learning concepts; Week 3 Vision and NLP with labs-lite demos; Week 4 Generative AI + comprehensive review and timed practice on service selection patterns.

AI-900 success is largely a reading-and-reasoning skill. Treat each question like a mini-consulting prompt: identify the goal, the data type (text, images, audio, structured data), constraints (real-time, limited data, no ML expertise, compliance), and the desired output (classification label, extracted text, translation, generated content). Then match to the workload and service family before you look at answer choices. This prevents distractors from steering you off course.

Elimination strategy: remove answers from the wrong domain first. If the prompt is about detecting items in an image with locations, eliminate NLP and general ML options that don’t provide bounding boxes. If the prompt is about summarizing or drafting content, eliminate deterministic analytics services and look for generative AI concepts (foundation models, prompting, copilots, Azure OpenAI). If it’s about predicting a numeric value from labeled historical data, favor supervised learning.

Exam Tip: When two answers seem correct, choose the one that best matches the operational requirement in the scenario—speed to implement (prebuilt AI services), need for customization (custom model in Azure ML), or safety/governance expectations (responsible AI controls, content filtering, human review).

Time management: don’t over-invest in any single tricky question. Mark it mentally, choose the best option based on domain match, and move on. Fundamentals exams often include several straightforward items; you want to secure those points. Finally, keep a “trap list” in mind: classification vs detection, training vs inference, supervised vs unsupervised, NLP analytics vs generative AI, and “AI when rules would do.” Avoid these traps and you’ll convert knowledge into a passing score efficiently.

Section 2.1: AI workload categories: vision, language, speech, decision support

AI-900 expects you to recognize the “big buckets” of AI workloads and link them to business value. Four categories show up repeatedly: computer vision (images/video), language (text), speech (audio), and decision support (recommendations, forecasting, anomaly detection). The exam isn’t testing deep implementation details; it’s testing whether you can classify the problem correctly.

Computer vision workloads interpret images or video frames. Common scenarios include: image classification (label the whole image, e.g., “damaged” vs “not damaged”), object detection (find and label items with bounding boxes, e.g., “helmet” and “person”), OCR (extract printed/handwritten text from images like receipts), and facial analysis concepts (detect presence of a face or compare similarity, depending on policy and service availability). Business value often ties to automation: faster inspections, reduced manual data entry, safer workplaces.

Language (NLP) workloads understand or generate meaning from text. Typical scenarios: sentiment analysis, key phrase extraction, entity recognition, summarization, and question answering. Business value is usually scale: analyze thousands of reviews, route support tickets, or search knowledge bases more effectively.

Speech workloads involve converting spoken audio to text (speech-to-text), converting text to natural audio (text-to-speech), and translation. The key clue is that the input/output is audio. Call-center transcription and real-time captions are classic examples.

Decision support workloads help choose actions: recommendations (“customers who bought X also bought Y”), forecasting (predict sales demand), and anomaly detection (unusual transactions or sensor readings). These often rely on machine learning patterns rather than explicit if/then rules.

Exam Tip: If the scenario mentions “bounding boxes,” “counting objects,” or “locating items,” choose object detection, not image classification. If it mentions “extract text from a scanned form,” that’s OCR, not language understanding.

Section 2.2: Key terminology: model, features, labels, training, inference

This section targets the ML fundamentals vocabulary that AI-900 uses in nearly every question. You must know what a model is, what features and labels are, and the difference between training and inference. These terms also help you differentiate AI vs ML vs deep learning vs generative AI.

A model is the learned “function” that maps inputs to outputs. In business terms: it’s the artifact you deploy to make predictions (classify images, predict churn, detect anomalies). Features are the input variables the model uses (age, purchase frequency, transaction amount, image pixels, word tokens). A label is the correct answer used during supervised training (e.g., “fraud” vs “not fraud,” or the correct product category). If labels are present, you’re usually in supervised learning; if not, you may be clustering or anomaly detection.

Training is the process of learning patterns from data—adjusting model parameters to reduce error. Inference is using the trained model to make a prediction on new data. On the exam, a common trap is mixing these up: training happens before deployment; inference happens during usage (often in production).

High-level differentiation: AI is the broad umbrella (systems performing tasks that typically require human intelligence). Machine learning (ML) is a subset where models learn from data rather than being explicitly programmed. Deep learning is a subset of ML using multi-layer neural networks, often strong for vision, speech, and language. Generative AI focuses on producing new content (text, images, code) using foundation models; it still relies on ML/deep learning under the hood, but the user experience is often prompting rather than feature engineering.

Exam Tip: When a question says “use the model to predict,” “score,” or “classify new data,” that’s inference. When it says “use historical data to build,” “fit,” or “learn,” that’s training.

Section 2.3: When to use AI vs rules-based solutions (decision criteria)

AI-900 often tests judgment: not every problem needs AI. You should be able to justify when a rules-based approach is sufficient versus when ML/AI is appropriate. This is especially important for non-technical decision makers because AI introduces costs: data collection, monitoring, governance, and risk management.

Use rules-based logic when the domain is stable, rules are explicit, and edge cases are limited. Examples: calculating tax based on a published rate table, routing requests based on fixed thresholds, validating that a form field contains a 5-digit code. Rules-based solutions are easier to audit and explain, which can be critical for compliance.

Use AI/ML when: (1) rules are hard to define (e.g., “Is this email spam?”), (2) patterns are complex and high-dimensional (images, audio, natural language), (3) you need adaptability (fraud patterns change), or (4) you’re optimizing based on data-driven outcomes (recommendations, forecasting). AI is also appropriate when you can collect representative data and measure success.

Decision criteria the exam loves: data availability (do you have enough quality data?), variability (does the environment change?), explainability requirements (do stakeholders need simple reasons?), and risk tolerance (will errors cause harm?). If the scenario suggests high risk (health, finance, safety), expect responsible AI and human oversight considerations.

Common trap: Assuming “automation” always means “AI.” Many business automations are deterministic workflow tools. Look for cues like “learn from past examples,” “improve over time,” or “unstructured data (images/text/audio).”

Exam Tip: If the question emphasizes “clear business rules” and “no need to learn from data,” the safest answer is usually a traditional rules-based solution, not ML.

Section 2.4: Responsible AI: fairness, reliability, privacy, inclusiveness, transparency, accountability

Responsible AI is explicitly tested in AI-900. You’re expected to recognize the core principles and apply them to scenarios. Microsoft commonly frames these as: fairness; reliability and safety; privacy and security; inclusiveness; transparency; and accountability. The exam typically gives a scenario (loan approvals, hiring screening, patient triage, surveillance) and asks what principle is at risk or what action supports responsible AI.

Fairness means AI decisions should not discriminate against protected groups. Watch for biased training data (historical hiring decisions) or proxy variables (zip code acting as a proxy for socioeconomic status). Reliability and safety means the system performs consistently and doesn’t create unacceptable harm—think of monitoring, testing under edge conditions, and fallback behaviors. Privacy and security includes minimizing sensitive data collection, protecting data in transit/at rest, and controlling access, especially for PII.

Inclusiveness means solutions work for people with different abilities, languages, accents, and contexts. Speech recognition that fails for certain accents is an inclusiveness issue (and potentially fairness). Transparency means stakeholders understand when AI is being used and, when appropriate, how outputs are produced (explanations, confidence scores, limitations). Accountability means humans remain responsible: clear governance, audit trails, and escalation paths.

Common trap: Confusing transparency with privacy. Transparency is about clarity and explainability; privacy is about protecting personal data. Another trap is treating “accuracy” as the only goal; responsible AI demands trade-offs (e.g., high accuracy but unfair outcomes is not acceptable).

Exam Tip: If the scenario mentions demographic disparities in outcomes, select fairness. If it mentions model failures in unusual conditions (nighttime images, noisy audio), select reliability and safety. If it mentions user consent, sensitive identifiers, or data leaks, select privacy and security.

Section 2.5: Azure AI services overview: what they are and when they fit

AI-900 does not require you to memorize every SKU, but it does expect you to recognize the main Azure “families” for AI workloads and when to use each: prebuilt Azure AI services, Azure Machine Learning for custom ML, and Azure OpenAI for generative AI. The exam frequently tests whether you should choose a prebuilt service versus building/training your own model.

Azure AI services (often referred to as Cognitive Services) provide APIs for vision, language, speech, and decision-related tasks. Choose these when you want fast time-to-value and the task matches a common pattern (OCR, speech-to-text, translation, sentiment analysis). They are ideal for non-technical teams integrating AI into apps without building models from scratch.

Azure Machine Learning is used when you need a custom model trained on your data—such as predicting equipment failure for a specific factory or churn for your subscription business. It supports training, evaluation, deployment, and MLOps practices. Exam wording to watch for: “train a model using historical data,” “track experiments,” “deploy and manage models.” Those usually point to Azure Machine Learning rather than a prebuilt API.

Azure OpenAI supports generative AI workloads using foundation models (for example, producing summaries, drafting text, extracting structured info with prompting, building copilots). Scenarios mentioning “prompt,” “chat interface,” “generate,” “summarize,” or “ground responses on company documents” are strong signals. Governance and safety are part of the fit: content filters, access controls, and responsible usage patterns.

Common trap: Picking Azure Machine Learning when the scenario is simply OCR or translation. If no custom training is needed, the prebuilt service is usually the better exam answer. Another trap is assuming generative AI is the right tool for deterministic extraction; sometimes classic OCR + rules or text analytics is more reliable.

Exam Tip: Look for the phrase “prebuilt model” or “API” (Azure AI services) versus “train and deploy” (Azure Machine Learning) versus “prompt and generate” (Azure OpenAI).

Section 2.6: Domain practice set: scenario mapping and concept checks

AI-900 questions often look easy but are designed to test whether you can map keywords to the correct workload and avoid “near-miss” distractors. Use a two-step approach: (1) identify the input/output data type (image, text, audio, tabular), then (2) identify the task type (classify, detect, extract, translate, generate, recommend, forecast). This quickly narrows the options.

Concept checks you should be able to do mentally during the exam: If labels are provided (known outcomes), you are in supervised learning and can talk about training data and evaluation. If the scenario emphasizes grouping without labels (segment customers by behavior), that’s unsupervised learning (clustering). If it’s “spot unusual behavior,” think anomaly detection (often unsupervised or semi-supervised). If it’s “create new text/images,” that’s generative AI using a foundation model plus prompting.

Responsible AI mapping is also a scenario skill. If a company uses AI to screen resumes, consider fairness (bias), transparency (disclosing AI use), and accountability (human review). If a hospital uses AI for alerts, reliability and safety dominate. If a chatbot uses internal documents, privacy and security plus transparency (source citations, limitations) often appear.

Common trap: Over-indexing on a single word. For example, “analyze text from scanned invoices” is not only language; it starts with vision/OCR to extract text, then language to interpret. On AI-900, pick the primary workload the question is asking about (often OCR if extraction is the core requirement).

Exam Tip: When multiple answers seem plausible, choose the one that matches the minimum necessary capability described. The exam frequently rewards the simplest correct fit (prebuilt service or clear workload category) over a more complex “you could build it” option.

Supervised learning means you train with labeled examples: each row (or image, audio clip, etc.) includes the “correct answer” (the label). AI-900 frequently frames this as “historical data with outcomes.” Your first exam move is to ask: is the label a category or a number?

Classification predicts a discrete class (category). Business examples: approving a loan (approve/deny), flagging email (spam/not spam), predicting customer churn (churn/no churn), or classifying an image as “damaged” vs “not damaged.” Multi-class classification expands to many categories, such as routing support tickets to “billing,” “technical,” or “sales.”

Regression predicts a continuous numeric value. Business examples: forecasting next month’s sales dollars, predicting time-to-delivery in days, estimating house price, or predicting energy consumption. On the exam, words like “estimate,” “forecast,” “how much,” “how many,” and “what value” usually signal regression.

Exam Tip: If the scenario output can be sorted (low/medium/high) but is still a set of labels, it’s classification—not regression—unless the output is explicitly a measured numeric value.

Common trap: “risk score.” If the output is a score on a continuous scale (e.g., 0–1 probability or 0–100 score), that can be treated as regression even though it might later be thresholded into categories. AI-900 questions typically make the intent clear: “predict probability” leans regression; “classify as high/medium/low risk” leans classification.

Another trap is confusing supervised classification with “grouping similar customers.” Grouping without labels is unsupervised clustering (Section 3.2), even if the groups later receive business names.

Unsupervised learning uses unlabeled data: you do not have the “correct answer” column. AI-900 commonly tests whether you can recognize the absence of labels and choose methods that discover structure in the data.

Clustering groups similar items based on features. Business examples include segmenting customers by purchasing behavior, grouping documents by topic similarity, or identifying usage patterns in telemetry. The key phrase is “find groups,” “segment,” “organize into clusters,” or “discover categories.” No pre-defined label exists at training time.

Anomaly detection identifies unusual observations that deviate from normal patterns, often using unsupervised or semi-supervised techniques. Business examples: detecting fraudulent transactions, spotting unusual IoT sensor readings, or flagging abnormal login behavior. The exam often describes this as “rare events,” “outliers,” or “unexpected spikes.”

Exam Tip: If the question says “we don’t know what fraud looks like yet” or “there are very few examples of failures,” anomaly detection is a strong match. If it says “we have many examples labeled fraud/not fraud,” that becomes supervised classification.

Common trap: thinking anomaly detection always requires labeled anomalies. Many real-world anomaly methods learn “normal” patterns from mostly normal data, then flag deviations. On AI-900, you’re mainly tested on recognizing the use case and choosing “anomaly detection” when labels are missing or rare.

Also note the “cluster naming trap”: after clustering, humans might label clusters (“budget shoppers,” “premium buyers”). That does not retroactively make it supervised learning; the algorithm still learned without labels.

AI-900 expects you to distinguish the training lifecycle (building a model) from inference (using a model to make predictions). Training is typically compute-intensive and done periodically; inference is performed repeatedly in production, either in real time or batch.

Training steps often include: collecting data, cleaning/transforming features, splitting into train/validation/test sets, selecting an algorithm, training on compute, and evaluating metrics. In Azure contexts, these steps are commonly orchestrated in a pipeline so they are repeatable.

Inference occurs when you apply the trained model to new data. On the exam, inference is hinted by language like “users submit a form and get an instant decision,” “score new transactions,” or “predict for tomorrow’s orders.” A deployed model is often exposed through an endpoint (a callable interface) that applications can use.

Exam Tip: If the scenario mentions “deploy,” “consume,” “API,” “endpoint,” or “integrate into an app,” you are in inference/serving territory. If it mentions “train,” “experiment,” “tune,” “compare models,” or “evaluate,” you are in training territory.

Common trap: assuming a pipeline is only for training. In Azure ML, pipelines can support both training workflows and batch scoring workflows. However, the exam’s simplest mapping is: pipeline = repeatable process; endpoint = serving predictions. Another trap is confusing “model registration” with “deployment.” Registering stores/version-controls a model artifact; deployment makes it accessible for inference.

Evaluation on AI-900 is about choosing the right metric for the right problem and recognizing overfitting. You’re not expected to compute metrics by hand, but you should know what each metric indicates and when it is misleading.

For classification, accuracy is the percentage of correct predictions. It’s tempting—and often wrong in imbalanced datasets (e.g., fraud where 99.9% are legitimate). That’s where precision and recall matter. Precision answers: “Of the items predicted positive, how many were truly positive?” Recall answers: “Of the truly positive items, how many did we catch?”

A confusion matrix summarizes predictions vs actuals: true positives, false positives, true negatives, false negatives. AI-900 questions often implicitly test whether you understand the cost of false positives vs false negatives. Example: in fraud detection, false negatives (missed fraud) can be costly; in spam filtering, false positives (good email marked spam) can be costly.

For regression, common metrics include RMSE (root mean squared error), which measures average prediction error magnitude (in the same units as the target). Lower RMSE generally indicates better fit.

Overfitting occurs when a model performs very well on training data but poorly on new/unseen data. You’ll see this described as “high training accuracy but low test accuracy.” The fix is usually not “train longer”; rather, consider more data, simpler models, regularization, or better validation.

Exam Tip: When the dataset is imbalanced, eliminate answers that rely only on accuracy. Look for precision/recall language, and tie it to the business cost of mistakes (false positives vs false negatives).

Azure Machine Learning (Azure ML) is the core Azure service for building, training, tracking, and deploying machine learning models. AI-900 stays conceptual: you should know the main components and what problems they solve.

An Azure ML workspace is the top-level container that organizes assets: experiments, models, datasets/data assets, compute targets, endpoints, and monitoring configurations. Think “project space with governance.” Questions often ask where you manage experiments and artifacts—workspace is the best match.

Compute provides processing power for training and sometimes batch inference. On the exam, compute is a scalable resource you attach to runs; it is not the same thing as the model. Data (datasets/data assets) represent training and evaluation inputs; good questions will mention storing/curating data used across experiments.

Models are the trained artifacts. You can register a model to version and track it, then deploy it to an endpoint for inference.

Automated ML (AutoML) helps you automatically try multiple algorithms and feature preprocessing choices to find a strong model for tabular tasks like classification and regression. AI-900 likes AutoML as the answer when: you have labeled data, want a baseline quickly, and don’t want to manually select algorithms.

Exam Tip: If the scenario says “we need the best model but lack ML expertise,” AutoML is a strong indicator. If it says “we need full control over code and architecture,” Azure ML with custom training is a better conceptual fit.

Common trap: picking Azure ML for every AI task. Azure ML is for building/operationalizing ML models; prebuilt AI capabilities (vision, language, etc.) often live in other Azure services. In this chapter, stay focused: for the ML domain, Azure ML + AutoML are the key service concepts.

This section prepares you for the exam’s “identify the right approach” style without turning into a quiz. When you read a scenario, apply a consistent decision process aligned to AI-900 objectives: (1) What is the learning type? (2) What is the output? (3) Is the task training or inference? (4) Which Azure ML capability fits?

Approach selection signals:

• “Predict a category” (approve/deny, churn/no churn) → supervised classification.
• “Predict a number” (price, demand, duration) → supervised regression.
• “Find groups/segments” with no labels → clustering (unsupervised).
• “Detect unusual/outliers” with rare/unknown patterns → anomaly detection.
• “Learn by trial and reward” (optimize decisions over time) → reinforcement learning (less common on AI-900 but must be recognized).

Then map to lifecycle language. If the scenario emphasizes building and improving the model—data preparation, experiment runs, comparing metrics—think training in Azure ML. If it emphasizes integrating into an app for predictions—web service, API calls, real-time scoring—think deployment to an endpoint and inference.

Exam Tip: Watch for “we have labeled historical outcomes.” That single phrase usually eliminates clustering/anomaly answers and points you to supervised learning plus either classification or regression. Conversely, “we don’t have labels” should immediately move you away from supervised learning.

Finally, consider when AutoML is the best conceptual tool: tabular labeled data, time constraints, and the need to automatically evaluate candidate models. A common trap is choosing AutoML for unlabeled clustering/anomaly scenarios; AutoML is typically presented for classification/regression tasks in AI-900 framing.

AI-900 frequently tests whether you can distinguish three foundational computer vision tasks: image classification, object detection, and segmentation. These are “what problem are we solving?” labels—each produces a different type of output, and that output usually dictates the service selection in later sections.

Image classification assigns one or more labels to an entire image (or to a cropped region you provide). The output is typically a set of class names with confidence scores (e.g., “dog: 0.92, cat: 0.04”). Use classification when the question implies one primary subject or category for the whole image (product type, scene category, defect vs no-defect).

Object detection finds multiple instances of objects and returns their locations. The hallmark output is bounding boxes (x/y coordinates, width/height) plus a label and confidence per detected object (e.g., “car at box A, person at box B”). Detection is the right task when the scenario asks “where” something is, “count” items, or “highlight” objects in an image.

Segmentation goes further than detection by assigning a label to pixels, producing a mask (outline/shape) rather than a rectangle. You’ll hear “semantic segmentation” (class per pixel) and “instance segmentation” (separate mask per object instance). AI-900 treats segmentation at a high level; focus on recognizing that the output is a mask/region map rather than a simple label or a box.

Common exam trap: Don’t confuse “classification with multiple labels” (e.g., tags) with “detection.” If the scenario needs location data (boxes/masks), it’s not pure classification—even if the question mentions labels.

Exam Tip: If you see verbs like “locate,” “draw a box,” “track,” “count,” or “highlight,” think object detection. If you see “outline,” “area,” “pixel-level,” think segmentation.

Beyond core tasks, many AI-900 questions describe “analyze an image and return insights.” In Azure’s vision stack, these insights often include tags, captions, and content moderation signals. These are typically prebuilt capabilities that summarize what is in the image rather than training a custom model.

Tags are keyword-like labels that describe objects, actions, or concepts present (e.g., “outdoor,” “person,” “bicycle”). Tags are useful when the scenario needs searchable metadata for indexing and retrieval (think “make photos searchable”). The output is a list of tags with confidence scores—no location required.

Captions are natural-language descriptions (e.g., “A person riding a bike on a city street”). Captions are used for accessibility (screen readers) and content summarization. The key clue is that the scenario wants a sentence, not a set of categories.

Content moderation concepts refer to detecting potentially unsafe or inappropriate content (adult, violent, or otherwise sensitive categories), or enforcing policy constraints on what can be shown. Exam questions may describe filtering user-generated images, preventing unsafe thumbnails, or flagging content for human review.

Common exam trap: Don’t assume that “describe the image” means OCR. OCR extracts text that is in the image. Captions describe the visual scene. If the prompt says “extract words from a sign,” that’s OCR; if it says “generate a sentence about the photo,” that’s captioning.

Exam Tip: Look for the noun in the deliverable: “keywords/tags,” “a caption/sentence,” or “a moderation score/flag.” AI-900 questions often include these exact deliverables to steer you.

OCR (Optical Character Recognition) is a staple AI-900 topic because it is easy to describe in business terms and it cleanly differs from image labeling. OCR’s core output is recognized text, often with coordinates (where the text appears). Modern services can also detect reading order and group text into lines and blocks.

AI-900 commonly extends OCR into document processing scenarios: invoices, receipts, forms, IDs, and contracts. Here the question is not only “what text is present,” but “what does it mean in a business structure?” This is often called document understanding—extracting key-value pairs (e.g., VendorName, InvoiceTotal), tables (line items), and sometimes signatures or selection marks.

When you see a scenario like “extract the total amount from thousands of receipts,” pure OCR is usually insufficient because OCR returns raw text; you still must interpret which number is the total. Document-focused models/services add this structure, returning fields with confidence scores and normalized values.

Common exam trap: OCR is not translation. If the scenario says “recognize text in a photo,” it’s OCR. If it says “convert recognized Spanish text into English,” that’s OCR plus translation (an NLP workload). AI-900 may test your ability to see multi-service pipelines.

Exam Tip: A fast clue is whether the output is expected to be free text (OCR) or named fields/tables (document understanding). If the question mentions “invoice number,” “purchase order,” “line items,” or “form fields,” think structured extraction rather than raw OCR.

Face-related questions on AI-900 aim to ensure you understand capability boundaries and responsible AI implications. At a high level, separate face detection from face identification/verification.

Face detection answers: “Is there a face, and where is it?” The output is commonly a bounding box and optional face attributes (depending on service/policy), such as head pose or blur. Detection is often used to crop faces, autofocus, or count people.

Face verification (sometimes described as “same person?”) compares two faces and returns a similarity/confidence that they belong to the same person. Face identification (sometimes described as “who is this?”) matches a detected face to a known set of enrolled identities. The exam may use wording like “match against a database,” “authenticate a user,” or “identify employees.”

Ethics and policy matter here. Face identification can be sensitive and is subject to stricter requirements, limited use cases, and governance controls. AI-900 expects conceptual awareness: biometric use can increase privacy risk, bias risk, and harm if misapplied. You should be ready to recommend using the least intrusive capability that satisfies the need (e.g., detection for counting rather than identification).

Common exam trap: “Detect faces for login” is ambiguous. If the scenario says “confirm the user is the same as their profile photo,” that’s verification. If it says “find which employee this is,” that’s identification. If it just says “find faces in the image,” it’s detection.

Exam Tip: Key verbs: “locate” → detection; “compare” → verification; “match to a list/name” → identification. When in doubt, choose the option that best matches the required business decision.

AI-900 is not a product deep-dive, but it does test your ability to map scenarios to Azure services at the “right bucket” level. The safest way to answer is to first name the workload (classification, detection, OCR, document extraction, face tasks, image analysis) and then select the Azure service family that delivers it.

As of current Azure terminology, common buckets include: Azure AI Vision for general image analysis (tags/captions), object detection capabilities, and OCR-style read operations; Azure AI Document Intelligence (formerly Form Recognizer) for extracting structured fields/tables from documents; and Azure AI Face for face detection and comparison/identity-style scenarios (subject to policy constraints). In some cases, a custom model approach using Azure Machine Learning is appropriate when prebuilt features don’t cover a specialized domain (unique defects, proprietary product categories) and you have labeled data.

Service-fit reasoning examples (conceptual): If the scenario is “auto-generate alt-text for product images,” choose image captioning/image analysis. If it is “extract invoice total and vendor name,” choose document intelligence. If it is “find all safety helmets and draw boxes,” choose object detection rather than image tags.

Common exam trap: Over-selecting “Azure Machine Learning” because it sounds powerful. AI-900 often rewards choosing a prebuilt vision service when the requirement is standard (OCR, captions, basic detection). Custom ML is a better answer when the scenario explicitly mentions “custom classes,” “train,” “labeled images,” or highly specialized recognition.

Exam Tip: Look for signals of prebuilt vs custom: “quickly,” “without training,” “no ML expertise” → prebuilt service; “domain-specific,” “train a model,” “your own labels” → custom ML approach.

For AI-900 readiness, practice is less about memorizing service names and more about consistently extracting the “required output” from a scenario. In this domain, the exam tends to describe business outcomes and expects you to infer the underlying vision task and select the right capability.

When you review practice items, train yourself to classify each scenario into one of these outputs: (1) label(s) for an image (classification/tags), (2) boxes around items (object detection), (3) masks (segmentation), (4) recognized text with layout (OCR), (5) structured fields/tables (document understanding), (6) face box/compare/match (face detection/verification/identification), or (7) caption/safety flag (captioning/moderation).

Also rehearse elimination strategies. If an option is an NLP service but the scenario has no language transformation (only reading text from an image), it’s likely wrong. If the scenario demands bounding boxes and the option only produces tags, it’s likely wrong. If the scenario discusses extracting invoice line items, OCR-only is likely incomplete because it doesn’t return business fields.

Common exam trap: Multi-step scenarios can tempt you to pick the “last step” only. For example, “extract text from a receipt and store totals in a database” still hinges on OCR/document extraction as the AI component; the database is not the AI service. Focus on the part that requires perception or prediction.

Exam Tip: Before looking at answer choices, write (mentally) the output you expect. Then select the choice that explicitly returns that output type. This prevents being distracted by familiar service names and keeps you aligned with what the exam is testing.

Section 5.1: NLP fundamentals: tokens, intents, entities, classification (conceptual)

NLP workloads deal with human language as data. The exam commonly checks whether you understand the “units” and “goals” of language processing. A token is a chunk of text (a word, subword, or symbol) used by models to process language. You don’t need to calculate tokens, but you should recognize that token limits affect summarization and prompting in generative AI scenarios (longer inputs can be truncated or cost more).

Classification is assigning text to a label: spam vs non-spam, complaint vs compliment, or urgency levels. This appears in both classic NLP (text analytics classification) and ML. On AI-900, classification is often positioned as an NLP workload when the input is text and the output is a category.

Intents and entities are central to language understanding for conversational experiences. Intent answers “what does the user want to do?” (e.g., “reset password”), while entities are the key data points in the utterance (e.g., account type, date, product name). When you see “build a chatbot that identifies what the user is asking,” think intent recognition plus entity extraction rather than sentiment analysis.

Exam Tip: If the scenario says “route a request,” “trigger an action,” or “fill a form,” it’s usually intent + entity extraction. If it says “analyze customer feedback,” it’s usually text analytics tasks like sentiment, key phrases, or entity recognition (insight extraction rather than action execution).

Common trap: confusing “entities” (names/places/organizations) with “key phrases” (important multi-word concepts). Entities tend to be canonical categories (Person, Location, Organization) while key phrases are often domain-specific topics (e.g., “battery life,” “refund policy”).

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

Text analytics is about extracting structured insights from unstructured text. For AI-900, your primary mapping is to Azure AI Language capabilities (often referred to as “Azure AI Language” / “Language service”), which includes features like sentiment analysis, named entity recognition (NER), key phrase extraction, and summarization.

Sentiment analysis classifies opinion polarity (positive/negative/neutral) and can include confidence scores. Exam scenarios: “monitor social media brand sentiment,” “detect unhappy customers,” or “prioritize negative support tickets.”

Entity recognition finds and classifies entities such as people, places, organizations, and sometimes custom domain entities. Exam scenarios: “extract product names from reviews,” “find company names in contracts,” or “identify locations mentioned in incident reports.”

Key phrase extraction identifies the main talking points without requiring a predefined taxonomy. Exam scenarios: “discover common themes in survey responses,” “summarize topics from meeting notes,” or “tag tickets with major issues.”

Summarization reduces long text into a concise version. The exam may present it as “generate an abstract” or “create bullet highlights.” Conceptually, summarization can be extractive (selects existing sentences) or abstractive (rewrites). You don’t need to implement it, but you should recognize it as an NLP workload, sometimes available in language analytics, and also achievable via generative AI depending on the requirement.

Exam Tip: Look for the verb: “analyze,” “extract,” “detect,” “classify,” and “summarize” typically point to text analytics. “Generate,” “draft,” “rewrite,” and “compose” usually point to generative AI. The exam likes mixing these, so pick the option that best matches whether you’re extracting facts from existing text or creating new language.

Common trap: choosing translation when the scenario is actually summarization (both “transform text”), or choosing sentiment when the scenario needs topic extraction (key phrases). Use output type as your anchor: sentiment outputs polarity; key phrases output topical terms; entities output structured named items; summarization outputs shortened text.

Section 5.3: Speech concepts: recognition, synthesis, and real-time vs batch patterns

Speech workloads convert between audio and text, and they commonly appear on AI-900 as part of the Azure AI Speech service. Know the three core concepts: speech-to-text (recognition), text-to-speech (synthesis), and speech translation.

Speech-to-text turns spoken audio into text. Exam scenarios include “transcribe call center recordings,” “create meeting captions,” or “voice-enable note taking.” Text-to-speech produces natural-sounding audio from text, used for “read messages aloud,” “interactive voice response,” or “accessibility narration.”

Speech translation converts spoken language to another language, often producing text and/or synthesized audio. This is different from text translation: the input is audio, not text. Watch for wording like “live interpreter,” “translate a speaker,” or “real-time multilingual meeting.”

The exam also tests pattern selection: real-time vs batch. Real-time is low-latency streaming (captions during a meeting). Batch is asynchronous processing of stored audio (overnight transcription of recordings). You won’t be tested on APIs, but you should match the operational need: immediate feedback implies real-time; large backlogs imply batch.

Exam Tip: When you see “recordings” or “uploaded audio files,” assume batch. When you see “during the call,” “live captions,” or “in the meeting,” assume real-time streaming.

Common trap: picking a language text service for an audio requirement. If the input is audio, speech is the front door. Another trap is assuming speech-to-text automatically implies sentiment or summarization; those are downstream NLP steps you apply to the transcript.

Section 5.4: Generative AI fundamentals: foundation models, embeddings, and typical use cases

Generative AI creates new content—text, images, or code—based on patterns learned from vast datasets. AI-900 focuses on conceptual understanding and Azure alignment: “foundation models,” “copilots,” and the Azure OpenAI Service as a key offering for enterprise generative AI.

A foundation model is a large, pretrained model that can be adapted to many tasks (summarize, draft, classify, answer questions) primarily through prompting and sometimes through fine-tuning. The exam expects you to understand that these models are not trained from scratch for each business case; they’re reused and configured.

Embeddings are numeric vector representations of text (and sometimes other modalities) that capture semantic similarity. Embeddings power search and retrieval experiences: “find documents similar to this question,” “recommend related articles,” or “semantic search over policy PDFs.” In generative AI solutions, embeddings are often used in retrieval-augmented generation (RAG) patterns, where the system fetches relevant content and supplies it to the model for grounded answers.

Typical use cases tested at AI-900 level include drafting emails, summarizing documents, generating marketing copy, creating chat assistants, and answering questions over internal knowledge bases. You may also see “code assistance” mentioned conceptually as a copilot-style capability.

Exam Tip: If the scenario emphasizes “searching internal documents” or “use our company knowledge,” think embeddings + retrieval (RAG) as the enabling concept, commonly implemented with Azure OpenAI plus a search index. If the scenario is purely creative (new text from scratch), embeddings may not be required.

Common trap: assuming generative AI is the best answer for any language task. If the requirement is strict extraction of entities or a consistent sentiment score at scale, classic text analytics is often a better fit (more deterministic outputs, simpler governance).

Section 5.5: Prompting basics and safety: grounding, hallucinations, content filters, governance concepts

Prompting is how you instruct a generative model to behave. For AI-900, know the building blocks: role/context, task instructions, input data, constraints (tone, length, format), and examples. Good prompts reduce ambiguity and improve consistency, which matters in business copilots that must produce predictable outputs.

Hallucinations occur when a model produces plausible-sounding but incorrect information. The exam often frames this as a risk that must be mitigated. One of the most important mitigation concepts is grounding: providing the model with trusted, relevant source material (for example, retrieved company policy excerpts) and instructing it to answer only from that content.

Safety and governance show up in AI-900 as “responsible AI” applied to generative scenarios. Expect mentions of content filters (blocking unsafe outputs), access controls, auditability, privacy, and policy enforcement. For Azure OpenAI concepts, you should know that enterprise deployments emphasize security boundaries, monitoring, and safety layers rather than unrestricted public model usage.

Exam Tip: If an answer choice mentions “reduce hallucinations,” “use internal data,” or “provide citations,” the best conceptual match is grounding via retrieval (RAG) and clear prompt constraints (e.g., “If not found, say you don’t know”). If it mentions “prevent harmful content,” look for content filtering and governance controls.

Common trap: confusing “fine-tuning” with “grounding.” Fine-tuning changes model behavior by training on examples; grounding supplies relevant facts at runtime. For many enterprise Q&A scenarios, grounding is preferred because it keeps answers aligned to the latest documents without retraining.

Section 5.6: Domain practice set: service selection and scenario-based items (NLP + GenAI)

AI-900 scenario items frequently test “service selection.” Your approach should be consistent: identify the input modality (text vs audio), identify the output (insights vs generated content), then select the Azure capability family that matches. For this chapter, you’re mainly choosing among Azure AI Language (text analytics and language understanding), Azure AI Speech (audio in/out), Translator (language translation scenarios), and Azure OpenAI (generative text and copilots).

Use these decision cues. If the scenario asks to detect sentiment, extract key phrases, recognize entities, or summarize a document without emphasizing creative generation, it maps cleanly to Azure AI Language text analytics capabilities. If the scenario starts with audio—calls, meetings, dictation—begin with Azure AI Speech for speech-to-text, then optionally add language analytics on the transcript.

If the scenario asks for “drafting,” “rewriting,” “creating a response,” “chatting with a knowledge base,” or “copilot experiences,” think Azure OpenAI concepts. If the scenario is multilingual, decide whether it is text translation (Translator/Language) or speech translation (Speech). Many items are hybrid: transcribe audio (Speech), translate transcript (Translator), then summarize (Language or Azure OpenAI depending on whether the output must be strictly extractive vs more narrative).

Exam Tip: When two options both seem possible (e.g., summarization via Language vs summarization via generative AI), use the requirement’s risk tolerance: regulated, audit-heavy, or extraction-focused scenarios usually favor text analytics; open-ended drafting and conversational assistants favor generative AI with safety controls and grounding.

Common trap: picking generative AI as a “universal translator” or “universal classifier.” The exam often rewards the simplest fit: translation service for translation, speech service for audio, language analytics for extraction. Generative AI is powerful, but it introduces variability and governance considerations that the question may hint at through words like “must be consistent,” “must be auditable,” or “no fabricated content.”

Section 6.1: Mock exam instructions, timing strategy, and how to review answers

Run the mock exam in two blocks to simulate the mental shifts you’ll do on test day. Block 1 focuses on AI workloads and fundamental ML concepts; Block 2 focuses on vision, NLP, and generative AI. Set a timer and commit to a steady pace rather than perfection on the first pass.

A practical timing strategy is “two-pass”: (1) answer what you know quickly, (2) return to flagged items. Don’t reread stems repeatedly; instead, underline (mentally) the constraint words: “best,” “most cost-effective,” “requires labeled data,” “real time,” “explainability,” or “responsible.” These are often the decisive hints.

When reviewing, do not ask “Why is my choice wrong?” first. Ask “What is the exam writer trying to test?” This shifts you from guessing to pattern recognition. Track misses by domain and by confusion type: service mismatch (e.g., choosing an ML platform when a prebuilt AI service is implied), learning type mismatch (supervised vs unsupervised), lifecycle mismatch (training vs inference), or governance mismatch (privacy, fairness, transparency).

Exam Tip: If two answers both sound plausible, look for the one that matches the minimum required capability. AI-900 commonly rewards the simplest correct service (e.g., a prebuilt Azure AI service) over a custom build (Azure Machine Learning) unless customization is explicitly required.

Section 6.2: Mock Exam Part 1: mixed-domain question set (AI workloads + ML)

In Part 1, you’ll see scenarios that test whether you can (a) recognize an AI workload, (b) decide whether AI is appropriate at all, and (c) identify core ML concepts used on Azure. Expect wording that tempts you into overengineering. For non-technical roles, the exam often frames decisions as business requirements: reduce manual effort, predict outcomes, detect anomalies, or segment customers.

Key concepts to rehearse as you answer: classification vs regression vs clustering; supervised vs unsupervised learning; training vs inference; and evaluation metrics at a high level. If a scenario mentions historical labeled outcomes (for example, “approved/denied,” “fraud/not fraud,” “churned/did not churn”), that points to supervised learning. If it mentions “group similar items with no labels,” that points to clustering.

Watch for the “AI vs rules” trap. If the logic is deterministic (for example, “if the total is over $X, apply discount Y”), traditional software rules are the right fit. If the inputs are messy, probabilistic, or pattern-based (images, text, voice, ambiguous customer behavior), AI becomes justified.

Azure framing: Azure Machine Learning is the platform for building, training, and deploying custom models when you need control over features, algorithms, and evaluation. Prebuilt Azure AI services are for common tasks without needing to train a model from scratch. In Part 1, if the scenario emphasizes model training, experiments, pipelines, or deploying an endpoint, it’s often pointing you toward Azure Machine Learning concepts.

Exam Tip: “Training” produces a model; “inference” uses a model. Many candidates mix these up when a question mentions “real-time scoring” (inference) versus “improve accuracy using more data” (training).

Section 6.3: Mock Exam Part 2: mixed-domain question set (vision + NLP + GenAI)

Part 2 shifts to recognition workloads: seeing (vision), understanding language (NLP), and generating content (GenAI). The most frequent exam move here is giving you a description of the user goal and expecting you to map it to the right capability category: image classification (what is it?), object detection (where is it?), OCR (what text is in it?), or analysis of text and speech for sentiment, entities, translation, or transcription.

For vision scenarios, focus on the verb: “classify” (label the image), “detect” (bounding boxes for items), and “read” (extract printed or handwritten text). A common trap is confusing OCR with document understanding. OCR is about text extraction; more complex document processing implies forms/fields and structured output, which may be beyond basic OCR framing.

For NLP, identify whether the task is analytics (extract entities, key phrases, sentiment) versus interaction (conversational language understanding). Translation is its own category. Speech scenarios often pivot on “speech to text” (transcription) versus “text to speech” (synthesis). If the scenario describes turning call recordings into searchable text, think transcription; if it describes generating an audio voice from text, think synthesis.

For generative AI, the exam expects baseline literacy: foundation models, prompting basics, copilots, and Azure OpenAI concepts including safety/governance. If the scenario asks for drafting, summarizing, rewriting, or Q&A over content, it’s likely GenAI. The trap is choosing GenAI when a deterministic extractive approach is asked (e.g., “extract invoice number”), which is usually better served by OCR/document extraction rather than free-form generation.

Exam Tip: If a question mentions “grounding” responses in company data, think retrieval-augmented patterns (using your data to inform outputs) and governance considerations—don’t answer as if the model “already knows” internal content.

Section 6.4: Answer key with rationales mapped to official domain names

Use this section as your scoring rubric. For each missed item, write down (1) the official domain it belongs to, (2) the keyword that revealed the domain, and (3) the rule that would have gotten you to the correct choice. Your objective is to build repeatable decision rules, not isolated facts.

Official domain mapping to use in your rationales:

• Describe AI workloads and considerations: selecting AI vs traditional solutions; common AI scenarios; Responsible AI principles (fairness, reliability/safety, privacy/security, inclusiveness, transparency, accountability).
• Describe fundamental principles of machine learning on Azure: supervised vs unsupervised; training vs inference; evaluation; how Azure Machine Learning supports the lifecycle (experiments, pipelines, deployment endpoints conceptually).
• Describe features of computer vision workloads on Azure: classification vs detection vs OCR; typical use cases and outputs.
• Describe features of Natural Language Processing (NLP) workloads on Azure: sentiment/key phrases/entities; translation; speech to text and text to speech; conversational understanding at a conceptual level.
• Describe features of generative AI workloads on Azure: foundation models; prompting; copilots; Azure OpenAI positioning; safety filters, content moderation, and governance basics.

Common rationale patterns that score well: “The stem describes labeled historical outcomes, so this is supervised learning,” or “The requirement is to locate objects in an image, so object detection fits better than classification.” Avoid vague rationales like “it’s AI,” which don’t prove you can map scenario-to-capability.

Exam Tip: When two service options appear, pick the one aligned to the workload family first (vision vs language vs ML platform). AI-900 questions often reward recognizing the workload type more than knowing deep configuration details.

Section 6.5: Final review: domain-by-domain checklist and last-minute memory anchors

Use this final checklist to tighten recall and reduce second-guessing. Read it once, then explain it back to yourself in your own words—teaching is a fast way to expose gaps.

• AI workloads & considerations: Can you tell when rules-based software is enough? Can you name Responsible AI principles and match them to scenarios (bias/fairness, transparency/explainability, privacy, reliability/safety, inclusiveness, accountability)?
• ML on Azure: Can you distinguish classification vs regression vs clustering? Supervised vs unsupervised? Training vs inference? Can you describe that Azure Machine Learning supports end-to-end model development and deployment?
• Computer vision: Do you map “what is it” to classification, “where is it” to object detection, and “what does it say” to OCR? Can you avoid confusing image labeling with OCR?
• NLP: Can you separate text analytics (sentiment/entities/key phrases) from translation and from speech tasks (STT vs TTS)?
• Generative AI: Can you define foundation models and prompting basics? Do you understand copilots as app experiences built on models? Can you describe why safety/governance (filters, monitoring, data handling) matters?

Memory anchors that help under pressure: “Labels = supervised,” “Groups with no labels = clustering,” “Training builds, inference uses,” “Classification answers what, detection answers where, OCR answers what text,” “Speech-to-text transcribes, text-to-speech narrates,” “GenAI generates; analytics extracts.”

Exam Tip: If you feel stuck, return to outputs: Does the scenario need a label, a number, a group, a bounding box, extracted text, a translation, a transcript, or generated prose? Output type is often the fastest path to the correct answer.

Section 6.6: Exam day readiness: environment setup, pacing, and common traps to avoid

On exam day, reduce preventable stress. Use a quiet location, stable internet, and a single monitor if required by your test provider. Close notifications and background apps. Have your ID ready and complete check-in steps early so you don’t start flustered.

Pacing: start with a quick scan mindset. Don’t let one difficult question consume your time budget. If you can eliminate to two options, pick the best match to the domain and move on; you can revisit if time remains. Keep your energy steady by taking micro-pauses between questions—one deep breath can prevent careless misreads.

Common traps to avoid:

• Overbuilding: choosing Azure Machine Learning when a prebuilt Azure AI service is sufficient (or choosing GenAI when extraction is needed).
• Lifecycle confusion: mixing training activities (improving a model) with inference activities (using a deployed model to score new data).
• Workload mismatch: selecting image classification when the scenario needs object detection or OCR; selecting sentiment analysis when the scenario is translation; selecting text analytics when the scenario is conversational understanding.
• Responsible AI blind spots: ignoring privacy/security, transparency, or fairness requirements embedded in the scenario.

Exam Tip: Read the last line of the question stem twice. AI-900 often places the true requirement there (for example, “identify the location,” “extract text,” “generate a summary,” or “ensure fairness/privacy”). That single phrase should dictate your answer more than the story around it.