Exam AI-900: Microsoft Azure AI Fundamentals
Your Quick Reference Guide
This cheat sheet covers the core concepts, terms, and key insights you need to know for the Exam AI-900: Microsoft Azure AI Fundamentals. We've distilled the most important topics, common pitfalls, and critical details to help you prepare efficiently.
💡 Note: While this guide highlights essential concepts, it's designed to complement—not replace—comprehensive study materials. Use it for quick reviews, last-minute prep, or to identify areas that need deeper study.
About This Cheat Sheet: This guide covers core concepts for Exam AI-900: Microsoft Azure AI Fundamentals. It highlights key insights, common mistakes, and frequently confused topics to support your exam preparation.
Use this as a quick reference alongside comprehensive study materials.
Exam AI-900: Microsoft Azure AI Fundamentals
Cheat Sheet •
About This Cheat Sheet: This guide covers core concepts for Exam AI-900: Microsoft Azure AI Fundamentals. It highlights key insights, common mistakes, and frequently confused topics to support your exam preparation.
Use this as a quick reference alongside comprehensive study materials.
Artificial Intelligence Workloads and Considerations
18%Scenario → Workload Map (CV / NLP / GenAI)
Classify a use case as Vision, NLP, or Generative AI and pick the matching Azure service while applying Responsible AI.
Key Insight
Match intent — detect/extract/understand/generate — to workload; don't default to GenAI and always check fairness, safety, transparency.
Often Confused With
Common Mistakes
- Choosing Generative AI for simple template or rule-based tasks
- Confusing model training (learning) with inference (prediction)
- Overlooking speech and translation as part of NLP
Pixels — RGB Grid & Spatial Context
Pixels are numeric channel values (often RGB) in a spatial grid — raw features for vision models that use intensity +位置.
Key Insight
Pixels = numbers per channel + position; spatial arrangement matters (CNNs use local neighborhoods), not just isolated values.
Often Confused With
Common Mistakes
- Thinking models read color names instead of numeric intensity values
- Treating a pixel as a single scalar instead of multi-channel data
- Assuming higher resolution always improves model accuracy
Fairness — Group Parity vs Individual Accuracy
Detect and mitigate dataset/model bias so AI doesn't discriminate across demographic groups.
Key Insight
Removing sensitive fields doesn't remove proxy bias — evaluate group metrics and accept accuracy/privacy trade-offs.
Often Confused With
Common Mistakes
- Deleting race/gender fixes nothing — proxies can reintroduce bias.
- Expecting identical outcomes for everyone; fairness targets comparable treatment or group parity.
- Assuming more data always removes bias; biased data amplifies unfairness.
Accountability — Owners, Artifacts, and Redress
Assign responsibility, keep evidence (model cards, logs), and define remediation for harmful AI outcomes.
Key Insight
Accountability = governance + technical evidence + redress; logs alone don't prove responsible practice.
Often Confused With
Common Mistakes
- Treating governance as only legal compliance, not operational oversight.
- Viewing model cards/datasheets as optional PR instead of audit evidence.
- Believing audit logs alone equal accountability without roles and remediation processes.
Fundamental Principles of Machine Learning on Azure
18%Classification — Discrete Labeling
Supervised task that assigns inputs to discrete labels; use confusion matrix and precision/recall/F1 for decision-making
Key Insight
Confusion matrix -> TP/FP/FN/TN; accuracy can lie on imbalanced data — prefer precision/recall, F1 or AUC; multiclass ≠ multilabel
Often Confused With
Common Mistakes
- Believing class labels must be numeric — encoding is preprocessing, not the concept
- Relying on accuracy for imbalanced classes instead of precision/recall, F1 or AUC
- Mixing up multiclass (one label per instance) with multilabel (multiple labels per instance)
Regression — Predict Continuous Values
Supervised task that predicts continuous numeric targets; evaluate with RMSE/MAE/R² and handle outliers/temporal effects
Key Insight
Regression predicts quantities not classes; choose MAE for outlier-robustness, RMSE penalizes large errors; forecasting requires temporal methods and持
Often Confused With
Common Mistakes
- Expecting regression to output class labels (that's classification)
- Assuming categorical features can't be used — they need appropriate encoding
- Treating time-series forecasting as identical to regression — forecasting needs temporal features and time-aware validation
Train/Validation/Test Split
Partition data into train (fit), validation (tune) and test (final) to avoid leakage and measure generalization.
Key Insight
Split BEFORE any preprocessing; use stratified sampling for class imbalance and time-ordered splits for temporal data; keep test set untouched.
Often Confused With
Common Mistakes
- Doing feature selection/scaling/encoding before splitting — causes leakage.
- Using a simple random split for time-series or highly imbalanced classes.
- Reusing the same validation set repeatedly for hyperparameter tuning (biases estimates).
Precision — Positive Predictive Value (PPV)
TP/(TP+FP): fraction of predicted positives that are actually positive — vital when false positives are costly.
Key Insight
Precision trades off with recall via the decision threshold: raising threshold usually increases precision but reduces recall.
Often Confused With
Common Mistakes
- Equating precision with overall accuracy.
- Mixing up precision with recall (correctness of positives vs. coverage of positives).
- Assuming precision is unaffected by class imbalance or threshold changes.
Azure AI Translator (NMT)
Cloud neural machine translation, language detection, and transliteration using Transformer models.
Key Insight
Use Translator for text translation/transliteration/language detection (Transformer NMT). Voice and raw images require Speech or OCR first.
Often Confused With
Common Mistakes
- Expecting built‑in real‑time voice‑to‑voice translation (that's Azure AI Speech).
- Treating Translator as a subfeature of Azure AI Language instead of a standalone service.
- Feeding raw images for translation without running OCR first.
NLP — Text Understanding & Generation
ML/DL techniques to analyze, interpret, and generate text — tasks include classification, NER, translation, summarizing.
Key Insight
NLP models learn statistical patterns from data (not human understanding); pick task-specific tooling (NLU vs NLG) on the exam.
Often Confused With
Common Mistakes
- Thinking NLP is only for generating new text rather than for interpreting text.
- Assuming every grammatical rule must be hand-coded instead of learned from data.
- Confusing NLU (meaning extraction) with machine translation (language conversion).
ML Training Lifecycle (Azure ML)
Select algorithm, prepare features, train/tune, validate, and register the trained model artifact.
Key Insight
Training produces a registered model artifact; avoid data leakage — use separate train/validation/test and proper cross‑validation for tuning.
Often Confused With
Common Mistakes
- Deploying before you have a trained, registered model artifact
- Assuming more features always improve performance — may add noise or multicollinearity
- Using validation data to update model weights or skipping a separate holdout test set
Azure AI Services — Prebuilt vs Custom
Map workloads to prebuilt cognitive APIs, managed Azure ML, or custom models by accuracy, cost, latency, and compliance.
Key Insight
Use prebuilt APIs for fast, low‑customization needs; use Azure ML for custom data, tuning, governance and compliance trade‑offs.
Often Confused With
Common Mistakes
- Believing Azure only offers non‑trainable, prebuilt AI APIs
- Treating integration as 'just call an API' — ignoring data prep, latency, security, monitoring
- Assuming self‑hosting is always cheaper and as easy as managed endpoints
Computer Vision Workloads on Azure
18%Image Classification — Labels Only
Assigns one or more labels to an entire image; no object coordinates or bounding boxes.
Key Insight
Image-level output only — choose multi-label if multiple classes can co-occur; use detection for locations.
Often Confused With
Common Mistakes
- Assuming an image can only get a single label (ignore multi-label mode).
- Expecting bounding boxes or coordinates from a classification model.
- Thinking classification also performs OCR or facial recognition.
Azure AI Vision — Prebuilt vs Custom
Microsoft's image-analysis service: use prebuilt APIs for general tasks or Custom Vision to train domain-specific models
Key Insight
Prebuilt = fast, general-purpose results; Custom Vision = trainable for niche classes and works with modest labeled sets.
Often Confused With
Common Mistakes
- Assuming prebuilt APIs are 'good enough' for specialized domains.
- Believing custom models require millions of images to be useful.
- Thinking the service only handles static photos and not video frames or temporal analysis.
Binary Classification — Two‑Class Decisions
Supervised model that predicts one of two labels or a probability; choose metrics and thresholds for class balance.
Key Insight
Use precision/recall/F1 or ROC-AUC for imbalanced data and tune the probability threshold, not default 0.5.
Often Confused With
Common Mistakes
- Relying on accuracy with imbalanced classes — it hides poor minority performance.
- Assuming 0.5 is the optimal decision threshold — tune for cost/precision-recall tradeoffs.
- Treating binary like multiclass/multilabel — output format and metrics differ.
Classification vs Detection vs Segmentation
Image-level labels vs bounding boxes vs pixel masks — pick based on whether you need presence, location, or exact shape.
Key Insight
If you need only presence use classification; need bounding locations use detection; need exact object shape use segmentation.
Often Confused With
Common Mistakes
- Expecting object detection to provide pixel-accurate shapes — it returns bounding boxes only.
- Assuming segmentation always separates instances — semantic segmentation groups class labels, not instances.
- Using classification when localization is required — image-level labels give no object coordinates.
Azure AI Vision — Prebuilt Object Detection
Managed prebuilt detector via REST/SDK; finds common objects and returns labeled bounding boxes.
Key Insight
No custom training; returns bounding boxes not segmentation masks, and coordinate formats can differ by endpoint/SDK.
Often Confused With
Common Mistakes
- Trying to train or fine-tune the prebuilt model — use Custom Vision.
- Expecting pixel‑level masks — only bounding boxes and labels are returned.
- Assuming perfect accuracy on niche objects or identical output formats across SDKs.
Object Tracking — Persistent IDs Across Frames
Links detections through time to follow movement and assign persistent track IDs using bounding boxes.
Key Insight
Detection locates per frame; tracking links those detections into tracks (IDs) using motion/appearance — it doesn't provide real-world identity.
Often Confused With
Common Mistakes
- Treating tracking as identical to one-frame detection.
- Expecting track IDs to reveal a person's name or confirmed identity.
- Thinking tracking requires GPS or live capture — it works on stored video/image sequences too.
Object Detection — Labels, Boxes, Scores
Finds and localizes items in images: label, confidence score, and bounding coordinates (origin = top-left).
Key Insight
Coords may be absolute pixels or normalized 0–1 with a top-left origin; score = confidence, not count.
Often Confused With
Common Mistakes
- Interpreting confidence score as the number of objects
- Assuming coordinates are always absolute pixels (they can be normalized 0–1)
- Thinking detection = face recognition or that only one instance can exist
Optical Character Recognition (OCR) — Read API vs Containers
Extracts printed or handwritten text into machine-readable text; choose cloud Read API or containers for on‑prem/edge.
Key Insight
Cloud Read usually has newest models/features; containers give offline/low‑latency and data‑residency but may lag parity.
Often Confused With
Common Mistakes
- Expecting OCR to provide semantic understanding of extracted text
- Assuming container images automatically match the cloud service feature set
- Believing OCR is English-only or that handwriting is never supported
Facial Recognition — Identification vs Verification
Identify (1:N) or verify (1:1) by comparing facial templates; must mitigate bias and follow Limited Access rules.
Key Insight
1:1 = verification, 1:N = identification; matching uses mathematical templates (not raw images) and Microsoft restricts use cases.
Often Confused With
Common Mistakes
- Assuming the service stores raw images instead of mathematical templates.
- Confusing facial recognition with facial analysis (age, emotion, attributes).
- Believing facial recognition is freely usable—Microsoft enforces Limited Access eligibility.
Face System Metrics — FAR, FRR, Precision, IoU
Use detection, localization, and recognition metrics to tune thresholds and balance security vs usability trade-offs.
Key Insight
AUC/ROC is threshold‑independent; operational error rates (FAR/FRR, precision/recall) depend on the chosen threshold; IoU is for box overlap.
Often Confused With
Common Mistakes
- Relying on AUC alone to claim low operational FAR/FRR at your operating point.
- Treating IoU as interchangeable with precision/recall—IoU measures localization overlap only.
- Assuming lower FAR always improves security—lower FAR usually increases FRR (more rejected genuine users).
Face Tasks — Detect • Verify • Identify
Locate faces (bounding boxes), compare two faces (verification), or match a face against an indexed gallery (identif.).
Key Insight
Different tasks, different needs: detection needs no identity DB; verification gives a similarity score with thresholds; identification queries an ID/
Often Confused With
Common Mistakes
- Thinking detection requires a pre-existing database of known people.
- Treating verification as definitive instead of a thresholded similarity/confidence score.
- Expecting sensitive attributes (age/emotion/gender) are returned by default — often restricted.
Azure AI Foundry (Azure AI Studio)
Unified workspace: model catalog, prompt engineering, deployment, and vision features (OCR) for generative and classicAI
Key Insight
Platform — not a single model. Integrates Microsoft + open models and supports both generative and classic CV/OCR workflows.
Often Confused With
Common Mistakes
- Thinking it only hosts Microsoft-proprietary models like GPT-4.
- Assuming it's only for generative/text apps — it also integrates classic CV and OCR.
- Treating Foundry as an OCR engine or a single AI model instead of a development platform.
Natural Language Processing (NLP) Workloads on Azure
18%Text Classification — Prebuilt vs Custom
Assign labels to text (binary/multiclass/multilabel); use prebuilt for common labels, custom for domain accuracy.
Key Insight
Multi-label allows multiple simultaneous labels; choose prebuilt for generic tasks, custom+transfer learning for domain gaps.
Often Confused With
Common Mistakes
- Assuming prebuilt classifiers always meet production accuracy
- Treating multi-label as multi-class (only one label allowed)
- Believing deep learning or huge datasets are always required
Conversational Language Understanding (CLU)
Maps utterances to intents and extracts entities for dialog apps; needs text input (use Speech-to-Text first for audio).
Key Insight
Intent = action to take; entity = data needed to complete it — CLU interprets text only, so run STT before CLU for audio bots.
Often Confused With
Common Mistakes
- Confusing intents (action) with entities (data points)
- Feeding audio directly into CLU without Speech-to-Text
- Assuming CLU only answers questions — it can trigger workflows
Key Phrase Extraction (Azure Text Analytics)
Pre‑trained Text Analytics feature that extracts important words/phrases from text to surface main talking points.
Key Insight
Not NER or summarization — returns salient words/phrases (including multi-word phrases), no custom training needed but input-size limits apply.
Often Confused With
Common Mistakes
- Expect labeled entity categories (Person/Location) — that's NER.
- Believe it generates summaries or new sentences — that's abstractive summarization.
- Ignore input-size and per-request limits; chunk or batch long documents.
REST APIs (HTTP/JSON) for Azure AI
Platform-agnostic HTTP/JSON endpoints for Azure AI services; SDKs wrap these calls but endpoints, auth, and limits still
Key Insight
SDKs are wrappers over REST — you still supply endpoints/auth and must handle request size, rate limits, and JSON formats.
Often Confused With
Common Mistakes
- Assume a language-specific SDK is required — you can call REST directly.
- Ignore request size or rate limits on endpoints.
- Think REST endpoints exist only for Azure OpenAI — other Azure AI services expose REST too.
Named Entity Recognition (NER) — Slot Extractor
Finds people/places/orgs in text; supplies 'slots' (who/what/where) for intents in Azure AI Language.
Key Insight
NER tags spans and types (slots); it categorizes entities but does not resolve identity — that's NEL's job.
Often Confused With
Common Mistakes
- Confusing NER with Key Phrase Extraction.
- Treating NER as the same as NEL (linking to external records).
- Using entities as intents — entities are parameters/slots, not goals.
Custom NER — Domain Entity Models
Train NER on labeled domain entities when prebuilt models miss terms; trade off accuracy vs time, cost, and maintenance.
Key Insight
Build custom only if you can supply quality labeled examples and ongoing maintenance; otherwise use prebuilt for quick wins.
Often Confused With
Common Mistakes
- Assuming a handful of examples per label is enough.
- Underestimating annotation effort — needs guidelines and review.
- Believing prebuilt models require no tuning or cost oversight.
Sentiment Analysis — Polarity & Confidence
Classifies text polarity (positive/neutral/negative) and confidence at sentence or document level to summarize opinions.
Key Insight
Document labels can mask mixed sentence sentiments — inspect sentence-level scores and choose an aggregation rule (majority, weighted, min/max).
Often Confused With
Common Mistakes
- Treating one document label as the full truth when text contains mixed opinions
- Averaging sentence scores blindly — extreme/polarized sentences get lost
- Interpreting a 0.5 score as 'half-positive' instead of neutral/uncertain
Fine-Tuning Models — Domain Adaptation
Adapt a pre-trained model with domain data to improve task-specific accuracy (e.g., sentiment); requires curation and re
Key Insight
Small, well-curated datasets or parameter-efficient tuning can be effective; always validate iteratively to catch overfitting or bias.
Often Confused With
Common Mistakes
- Expecting fine-tuning to eliminate hallucinations or bias
- Thinking more data always helps — low-quality data can degrade results
- Treating fine-tuning as training from scratch; parameter-efficient options exist
Transformer Decoder (Autoregressive Block)
Autoregressive transformer unit that attends to prior tokens to generate the next token; core of GPT-style models.
Key Insight
Generates tokens iteratively using masked self-attention; decoder-only models (e.g., GPT) don't require encoders.
Often Confused With
Common Mistakes
- Thinking a whole sentence is generated in one step — it's token-by-token.
- Mixing up the decoder's generative role with the encoder's contextual embedding role.
- Assuming a decoder always needs a paired encoder (not true for decoder-only models).
Text Completion (Generative Output)
Produce probable continuations of prompts; primary generative mode for chat, code, summaries, and assistants.
Key Insight
Sampling (temperature/top-p) controls creativity; max tokens and stop sequences control length — outputs are probabilistic, not authoritative.
Often Confused With
Common Mistakes
- Treating completions as factual answers — models can hallucinate plausibly.
- Thinking temperature changes length — it only changes randomness; use max tokens to limit length.
- Assuming longer prompts always improve quality — irrelevant context can hurt results and increase cost.
Azure Speech (STT & TTS)
Managed speech-to-text and text-to-speech for real-time or batch use — cloud, container, or on-device deployments.
Key Insight
STT/TTS are service-hosted — you must pick deployment, model, translation, and privacy settings explicitly.
Often Confused With
Common Mistakes
- Assuming translation is automatic — enable Speech Translation explicitly.
- Treating cloud STT/TTS as privacy-free — check data logging, residency, and opt-out settings.
- Thinking the SDK contains models — the SDK is client code; models run in the service or container.
Phoneme — ASR sound unit
The smallest distinct sound unit ASR acoustic models use to map audio into text via phonetic and language models.
Key Insight
Phonemes are sounds, not letters; ASR typically goes audio → phoneme (acoustic) → grapheme/word (language model).
Often Confused With
Common Mistakes
- Equating a phoneme with a written letter (grapheme).
- Assuming phonemes are identical across languages.
- Believing ASR maps audio directly to text without a phonetic intermediate.
Speech Translation — Real-time vs Batch
Convert spoken language live or from recordings via pipeline (STT→MT→TTS) or end-to-end models; choose by latency and QA
Key Insight
Pipeline gives best tuning and language coverage; end-to-end lowers latency but limits customization and supported languages
Often Confused With
Common Mistakes
- Assuming STT→MT→TTS is always required — end-to-end speech models exist
- Expecting zero latency in real-time — real-time systems trade latency for accuracy
- Believing speech translation equals text translation accuracy — audio quality, accent, and noise reduce accuracy
Text Translation (Machine Translation)
Automatic text-to-text language conversion; on Azure consider model type, domain tuning, latency, and data privacy
Key Insight
Out-of-the-box MT is fast but needs domain adaptation for specialist vocab; language detection is a separate step
Often Confused With
Common Mistakes
- Confusing translation with transcription (speech→text) on exam questions
- Assuming machine translation is always human-quality without domain tuning
- Thinking MT always requires cloud calls — on-device/offline MT exists for privacy/latency
Azure AI Language Service
Unified text NLP on Azure — sentiment, entities, Q&A and pre-built PII/key-phrase features.
Key Insight
Q&A is a feature (not a separate service); audio requires Speech transcription; Translator is separate.
Often Confused With
Common Mistakes
- Treat Question Answering as a separate service — it's a Language feature.
- Send raw audio expecting text results — transcribe with Azure AI Speech first.
- Assume Language handles translations or that PII detection needs custom training.
Language Detection
Identifies text language; returns ISO 639-1 code and a 0.0–1.0 confidence score.
Key Insight
Confidence is a 0–1 probability; undetectable/unsupported inputs return '(Unknown)' with confidence 0.0.
Often Confused With
Common Mistakes
- Expect 'NaN' for unsupported languages — API returns '(Unknown)' and confidence 0.0.
- Treat the confidence score as a percent — it's 0.0–1.0, not 0–100.
- Confuse ISO 639-1 codes ('en') with BCP-47 regional tags ('en-US').
Generative AI Workloads on Azure
28%Real-time Speech-to-Text (ASR)
Converts live spoken audio into text (captions, notes) with optional punctuation/diarization; not semantic understanding
Key Insight
ASR outputs words, not meaning — accuracy depends on audio quality, accents, model tuning, and feature settings
Often Confused With
Common Mistakes
- Expecting perfect transcripts regardless of accent, noise, or mic quality
- Thinking real-time transcription automatically translates speech to another language
- Believing punctuation, capitalization, and speaker labels are always provided by default
Grounding & RAG (Retrieval-Augmented Generation)
Retrieve authoritative documents to condition LLM outputs so answers reflect sources — reduces hallucinations but needs驗
Key Insight
RAG supplies context at query time (no weight changes); source quality and human validation still determine trustworthiness
Often Confused With
Common Mistakes
- Assuming RAG guarantees fully factual answers without human review
- Confusing grounding with fine-tuning (it adds context, it doesn't change model weights)
- Treating any retrieved document as safe grounding regardless of provenance
Embeddings — Semantic Vectors (word/sentence)
Dense float vectors that encode text meaning; power similarity, clustering, retrieval and RAG vector search.
Key Insight
Embeddings are dense floating‑point vectors used with nearest‑neighbor (cosine/dot/L2) in vector stores — pick metric by model & task.
Often Confused With
Common Mistakes
- Assuming cosine is always best — metric choice depends on model and use case.
- Treating embeddings as discrete IDs or integers — they are floating‑point vectors.
- Expecting exact keyword matches — embeddings capture semantic similarity, not token equality.
Privacy — Responsible AI (PII & controls)
Controls and processes to protect personal/sensitive data in AI: PII handling, minimization, encryption, access and data
Key Insight
Privacy ≠ security: privacy is about rights/purpose and risk reduction — anonymization can fail and cloud defaults don't equal compliance.
Often Confused With
Common Mistakes
- Relying on cloud default settings as proof of regulatory compliance.
- Assuming anonymization always eliminates re‑identification risk.
- Treating privacy and security as the same — privacy focuses on rights and purpose limitations.
Platform Safety Layer — Runtime Filters
Centralized runtime filters that detect, tag, and block risky model outputs to reduce abuse at response time.
Key Insight
Runtime controls cut downstream risk but complement — not replace — model-level fixes; expect false positives and latency.
Often Confused With
Common Mistakes
- Treating platform controls as a substitute for fine-tuning or prompt work
- Expecting filters to guarantee 100% elimination of harmful outputs
- Assuming filters always interpret sarcasm, nuance, or domain terms correctly
Content Safety (Language Models)
Policies and tools (filters, classifiers, guardrails, red‑teaming) used to prevent, detect, and mitigate harmful or dis‑
Key Insight
Layer defenses — system messages, red‑teaming, and filters together; safety decisions trade off utility and UX.
Often Confused With
Common Mistakes
- Relying only on an output filter to make a model safe
- Assuming safety controls have no impact on user experience or utility
- Believing prompt engineering alone stops adversarial inputs
Security & Access Control — RBAC, Managed Identity, VNet
Protect models, data and endpoints using authentication, RBAC, managed identities, VNets and secrets.
Key Insight
Defense-in-depth: RBAC, network isolation, encryption and secrets cover different risks — use all together.
Often Confused With
Common Mistakes
- Thinking RBAC alone secures endpoints — it doesn't replace network or secrets controls.
- Assuming managed identities remove all secret needs — external APIs often still require secrets.
- Believing VNet equals no data exfiltration — you still need encryption, logging and DLP.
Knowledge Mining — Enrich, Index & Search Unstructured Data
End-to-end pipeline that ingests, extracts, enriches (OCR/NLP) and semantically indexes multi‑modal data for search.
Key Insight
Not just keyword search: combines OCR, NLP, enrichment skills and semantic indexing to surface insights.
Often Confused With
Common Mistakes
- Mistaking knowledge mining for only building a knowledge graph.
- Assuming it's just keyword search — it uses OCR, extraction and semantic layers.
- Believing it handles only text — it also extracts from images and other media.
Azure OpenAI — Managed GPTs (secure, enterprise)
Managed Azure access to OpenAI models with enterprise security, endpoints, and resource credentials.
Key Insight
You must provision an Azure OpenAI resource to get endpoints/keys; control access via Azure RBAC/managed identities; Azure doesn't use your data to re
Often Confused With
Common Mistakes
- Thinking Azure OpenAI uses your data to train public OpenAI models
- Trying to authenticate with openai.com API keys
- Expecting immediate ChatGPT-like access; forgetting resource provisioning and Azure RBAC
Tokenization — Tokens, Limits & Cost
Splits text into model tokens; token counts determine limits, truncation risk, and billing.
Key Insight
Tokenizers vary by model — tokens ≠ characters; token counts drive prompt/response limits and cost, so craft prompts to avoid truncation.
Often Confused With
Common Mistakes
- Counting characters as tokens
- Confusing tokenization with lemmatization or stemming
- Assuming max_tokens always caps the entire conversation (prompt + response)
GenAI Capabilities: Text, Image, Code
Map model capability (chat, completion, embeddings, image/audio/OCR) to the task to balance accuracy, cost, and latency.
Key Insight
Chat ≠ completion ≠ embeddings: use chat for dialog/state, completion for single-turn outputs, embeddings for semantic search/similarity.
Often Confused With
Common Mistakes
- Assuming a 'chat' label equals production-ready performance without evaluation
- Always picking the largest foundation model regardless of cost, latency, or task fit
- Using embeddings as if they produce generative text or code
Content Safety: Filtering & Risk Triage
Detect, score, and route harmful content with automated classifiers, severity signals, and human-in-loop review viaAzure
Key Insight
Raising sensitivity raises false positives — combine safety signals, severity thresholds, and human review to balance recall and UX.
Often Confused With
Common Mistakes
- Believing higher sensitivity always improves outcomes
- Treating automated moderation as unbiased or perfectly accurate
- Relying only on automated flags instead of severity-based routing and human review
Ace the Exam AI-900: Microsoft Azure AI Fundamentals Exam
Unlock the full Practice Questions Bank and simulations. Get certified with confidence.