AZ-305 Cheat Sheet
Workload-to-service mappings, decision flows, and exam traps — optimised for last-minute revision. Scan fast.
Section 1 — Workload-to-Service Mapping
| Workload | Primary Azure Service | Keyword Cues |
|---|---|---|
| Image classification / tagging | Azure AI Vision (Tags) · Custom Vision | label whole image, no boxes, categories |
| Object detection | Azure AI Vision (Objects) · Custom Vision | bounding box, count items, locate |
| OCR — printed/handwritten text | Azure AI Vision (Read API) | extract text from image, receipts, signs |
| Forms with key/value + tables | Azure AI Document Intelligence | invoice, claim form, structured fields |
| Face detection / attributes | Azure AI Face (Detect) | find faces, count, age/emotion attributes |
| Face identification / verification | Azure AI Face (Limited Access) | who is this person, 1:1 / 1:N match |
| Sentiment / key phrase / NER / language | Azure AI Language | tone, named entities, key topics, language ID |
| Speech-to-text / text-to-speech | Azure AI Speech | transcribe audio, voice synthesis, captions |
| Text translation | Azure AI Translator | translate text between languages |
| Speech translation (audio in) | Azure AI Speech (translation) | audio input → translated output |
| Generative chat / completions | Azure OpenAI in Azure AI Foundry | GPT, summarisation, draft email, code-gen |
| RAG over enterprise docs | Azure AI Search + Azure OpenAI (Prompt Flow) | ground answers in company data |
| Train tabular regression/classification | Azure Machine Learning (AutoML / Designer) | predict price, classify churn, no-code/SDK |
| Visual ML pipeline (drag-and-drop) | Azure ML Designer | no-code, components, canvas |
KEY RULE: "OpenAI" in a scenario almost always means Azure OpenAI inside Azure AI Foundry — never the public OpenAI API.
Section 2 — The Six Responsible AI Principles
Fairness
Tested concern: Equal outcomes across demographic groups
Mitigation: Diverse data; fairness assessment in Responsible AI dashboard; demographic parity / equalised odds checks
Reliability and safety
Tested concern: Works on edge cases; human-in-the-loop on high-stakes calls
Mitigation: Red-teaming; monitoring; HITL approval gates; grounding for GenAI
Privacy and security
Tested concern: Personal data protected; resists data exfiltration / prompt injection
Mitigation: Data minimisation; encryption; RBAC; Content Safety; Prompt Shields
Inclusiveness
Tested concern: Everyone can USE the system (accents, abilities, languages)
Mitigation: Diverse data; alternative modalities (captions, voice, alt text); WCAG testing
Transparency
Tested concern: Users understand what the AI does, its limits, and that it IS AI
Mitigation: Transparency Notes; explainability dashboard; AI disclosure to users
Accountability
Tested concern: Humans (not AI) own the system's behaviour
Mitigation: Defined ownership; governance reviews; audit logs; escalation paths
TRAP: Fairness is about EQUAL OUTCOMES. Inclusiveness is about EQUAL ACCESS. If the scenario emphasises participation or accessibility → inclusiveness.
Section 3 — ML Technique Decision Flow
| Technique | Output | Canonical Scenario | Eval Metrics |
|---|---|---|---|
| Regression (supervised) | Continuous number | Predict house price, sales next month, temperature | MAE, RMSE, R² |
| Classification (supervised) | Category / class | Spam vs ham, fraud vs not fraud, cat/dog/rabbit | Accuracy, Precision, Recall, F1, ROC-AUC |
| Clustering (unsupervised) | Group assignments (no labels) | Customer segmentation, anomaly grouping | Silhouette, inertia |
| Deep learning (neural nets) | Anything — usually images/audio/text | Image classification, speech recognition | Task-specific (accuracy, BLEU, etc.) |
| Transformer (attention) | Sequence (usually text) | LLMs, translation, summarisation | Perplexity, groundedness, BLEU, ROUGE |
MEMORISE: "predict a number" → regression. "predict a category" → classification. "group similar with no labels" → clustering.
Section 4 — Computer Vision Cheat Card
| Capability | Returns | Pick When | Service |
|---|---|---|---|
| Image classification | Label(s) for whole image (no box) | Need WHAT is in the picture, not where | Azure AI Vision Tags · Custom Vision |
| Object detection | Labels + bounding boxes | Need WHERE / count / locate items | Azure AI Vision Objects · Custom Vision |
| OCR (Read API) | Text + line/word coordinates | Plain printed/handwritten text from images | Azure AI Vision (Read) |
| Document Intelligence | Key-value pairs, tables, fields | Forms, invoices, claims, structured docs | Azure AI Document Intelligence |
| Face Detection | Bounding box + attributes | Find faces / count / age / emotion | Azure AI Face (Detect) |
| Face Verification (1:1) | Same person? Yes/No + score | Confirm a specific person — Limited Access | Azure AI Face (Verify) |
| Face Identification (1:N) | Person ID from PersonGroup | Who is this? — Limited Access | Azure AI Face (Identify) |
TRAP: Classification = WHOLE image label. Detection = label + bounding box. If the question mentions counting or locating → detection.
Section 5 — NLP Cheat Card
| Capability | What It Does | Service |
|---|---|---|
| Key phrase extraction | Surface main topics in text | Azure AI Language |
| Named Entity Recognition (NER) | Tag people, places, orgs, dates, money | Azure AI Language |
| Sentiment analysis | Positive / neutral / negative + opinion mining | Azure AI Language |
| Language detection | Identify which language a text is in | Azure AI Language |
| Summarisation / Q&A / CLU | Summarise docs; answer over KB; intent + entities | Azure AI Language |
| Speech-to-text / TTS / Speech translation | Transcribe audio; synthesise voice; speech in → translated text out | Azure AI Speech |
| Text translation | Translate text between 100+ languages | Azure AI Translator |
TRAP: Audio in → Azure AI Speech (even if output is translated text). Translator is text-in, text-out only.
Section 6 — Generative AI Stack
Azure AI Foundry
Unified portal for projects, model catalogue, deployments, Prompt Flow, evaluations, Content Safety. Supersedes the old Azure OpenAI Studio.
Azure OpenAI Service
OpenAI's models (GPT-4o, gpt-4.1, o-series, embeddings, DALL·E, Sora) inside Azure with privacy + regional residency. Runs INSIDE Foundry.
Foundry Model Catalog
1,800+ models — Azure OpenAI, Microsoft Phi, Meta Llama, Mistral, Cohere, Hugging Face, Nvidia NIM, DeepSeek, Stability AI.
Serverless API vs Managed Compute
Serverless API = pay per token, Microsoft hosts. Managed compute = your dedicated GPU VMs, more control, you pay for the VM regardless of traffic.
Prompt Flow
Visual / code authoring for chaining LLM calls, Python steps, prompts, and connections (e.g. Azure AI Search) into a runnable evaluatable pipeline. Standard way to build RAG.
Azure AI Content Safety
Four harm categories: Hate · Sexual · Violence · Self-harm (severity 0–6). Plus Prompt Shields, Protected Material detection, Groundedness detection. Default-on for Azure OpenAI deployments.
Content Safety extras to memorise
- • Prompt Shields — blocks direct & indirect prompt injection / jailbreak (User Prompt Attacks).
- • Groundedness detection — flags hallucinated / ungrounded responses in RAG.
- • Protected Material detection — flags copyrighted text / code in completions; required for Customer Copyright Commitment.
Section 7 — Top Exam Traps
Read each one carefully — these are the most commonly missed question patterns on AZ-305.
RAG vs fine-tuning
RAG injects fresh, citable context at runtime — pick it for changing enterprise docs (HR handbook, policies). Fine-tuning bakes style/format into model weights — pick it for a tone or output schema, not for facts.
Image classification vs object detection
Classification labels the WHOLE image (no boxes). Detection adds bounding boxes and counts. If the question mentions WHERE, COUNT, or LOCATE → detection.
Hallucination ≠ prompt injection ≠ bias
Hallucination = confidently wrong (Reliability/Safety, mitigate with grounding). Prompt injection = adversarial override (Privacy/Security, mitigate with Prompt Shields). Bias = unequal outcomes (Fairness).
Fairness vs Inclusiveness
Fairness = equal OUTCOMES across groups. Inclusiveness = equal ACCESS / participation (accents, abilities, modalities). Speech-to-text struggling with non-native accents → inclusiveness, NOT fairness.
Sentiment ≠ language detection
Sentiment returns positive/neutral/negative tone. Language detection returns the language code (en, fr, ja). If the scenario says 'we don't know what language' — that is language detection.
Translator vs Speech Translation
Translator = text-in, text-out. If the input is AUDIO, the answer is Azure AI Speech (speech translation), even if the output is text in another language.
Azure OpenAI vs public OpenAI API
If a scenario mentions Azure data residency, enterprise SLA, or RBAC + Microsoft privacy commitments — the answer is Azure OpenAI in Azure AI Foundry, never 'call OpenAI's public API.'
Limited Access features
Face Identify, Face Verify, the face ID property, Custom Neural Voice Pro, Speaker Recognition, and Celebrity Recognition are gated behind Microsoft's Limited Access policy. Detection and attribute analysis are NOT.
OCR vs Document Intelligence
Plain text from photos / signs / handwriting → OCR (Read). Forms with fields, tables, key-value pairs (invoices, claims, tax forms) → Azure AI Document Intelligence.
AutoML vs Designer
AutoML = wizard that auto-tries algorithms on tabular data (pick a target column and go). Designer = drag-and-drop pipeline canvas where you wire components yourself. Both are no-code; AutoML is auto, Designer is visual.
Cheat sheet last reviewed 2026-04-27. Based on AZ-305 exam objectives current as of May 2, 2025.