3. Case Studies and AI-900 Exam Preparation
3.1 Agenda
Estimated reading time: ~15 minutes
Learning Outcomes
- Analyze (phân tích) real-world case studies through the lens (lăng kính) of Responsible AI principles
- Identify which principle is most relevant to each scenario
- Review key concepts, service selection patterns, and exam traps (bẫy) across all 8 workshops
- Apply a structured (có cấu trúc) reasoning framework to AI-900 scenario questions
3.2 Quick Glossary
| Term | Quick Explanation |
|---|---|
| Exam scenario question | Dạng câu hỏi AI-900 phổ biến nhất — mô tả (describe) một tình huống kinh doanh (business situation) và hỏi service hoặc principle nào phù hợp. |
| Distractor (Lựa chọn gây nhầm lẫn) | Câu trả lời sai nhưng có vẻ hợp lý (plausible) — được thiết kế để thử thách (test) sự phân biệt chính xác (precise distinction) của bạn. |
4. Real-World Case Studies
4.1 Case Study 1 — Healthcare: Diagnostic AI with Privacy and Fairness Concerns
Context: A hospital in Vietnam deploys an AI model to assist (hỗ trợ) radiologists in detecting lung cancer from chest X-rays (phim X-quang ngực). The model was trained on a US hospital's dataset.
What happened:
- Overall (tổng thể) accuracy: 91%
- Accuracy for Vietnamese patients (unseen (chưa thấy) demographic in training data): 74%
- Patient CT scans (quét cắt lớp) are sent to an Azure endpoint without PII stripping (loại bỏ PII)
- Radiologists trust the AI output without additional review (xem xét thêm)
Responsible AI analysis:
| Principle | Violation | Fix |
|---|---|---|
| Fairness | 74% accuracy for Vietnamese patients vs. 91% overall — representation bias | Evaluate on local demographic (nhân khẩu học địa phương) data; retrain or fine-tune on Vietnamese scans |
| Privacy | Patient scans sent without PII anonymization | Strip (loại bỏ) patient name, ID, and DOB before API call |
| Reliability | Radiologists over-rely (phụ thuộc quá mức) on AI without independent (độc lập) review | Implement mandatory (bắt buộc) human-in-the-loop (con người trong vòng lặp) for all positive flags (cờ dương) |
| Transparency | Model's 74% subgroup accuracy not disclosed (tiết lộ) to clinical (lâm sàng) staff | RAI Dashboard error analysis by demographic group; staff education (giáo dục nhân viên) |
| Accountability | No clear owner (chủ sở hữu) if model misdiagnoses (chẩn đoán sai) a patient | Assign clinical AI governance role; maintain audit logs per scan |
4.2 Case Study 2 — Finance: Credit Scoring Model
Context: A fintech (công nghệ tài chính) company uses Azure ML to build a loan approval model trained on 5 years of historical data. The model uses zip code (mã bưu chính) as a feature.
What happened:
- Model achieves 88% accuracy overall
- Rejection rate (tỷ lệ từ chối) is 40% higher for applicants in historically (lịch sử) low-income zip codes
- Customers who are rejected receive only: "Application declined — does not meet criteria (không đáp ứng tiêu chí)."
- Regulators flag the system as potentially violating (vi phạm) fair lending laws (luật cho vay công bằng)
Responsible AI analysis:
| Principle | Issue | Resolution |
|---|---|---|
| Fairness | Zip code proxies (đại diện) for race — disparate impact on minority (thiểu số) neighborhoods | Remove zip code as feature; test model for demographic parity |
| Transparency | Rejection reason not provided — violates (vi phạm) GDPR "right to explanation" | Implement Local Explainability (SHAP) — generate per-decision (từng quyết định) reasons |
| Accountability | No audit trail (dấu vết kiểm toán) of model versions or decisions | Azure ML Model Registry + Azure Monitor logging for all API calls |
| Privacy | No data minimization — unnecessary personal data collected | Review feature set; collect only what demonstrably (có thể chứng minh) improves accuracy |
4.3 Case Study 3 — Hiring Tool (Công cụ tuyển dụng): NLP for CV Screening
Context: An HR department uses Azure AI Language to extract keywords from CVs and rank candidates. The system is used without human review for the initial screening (sàng lọc) stage.
What happened:
- Model learned that top-performing past hires (nhân viên tốt) had keywords like "led," "managed," "architected" — all historically more common in male candidate language
- Female candidates using "collaborated (hợp tác)," "supported (hỗ trợ)," "contributed (đóng góp)" were systematically (có hệ thống) ranked lower
- No human reviewed the ranking logic
Responsible AI analysis:
| Principle | Issue |
|---|---|
| Fairness | Gender-coded language (ngôn ngữ liên quan đến giới tính) bias in training data → disparate impact |
| Accountability | No human in the loop — AI made hiring decisions autonomously (tự chủ) |
| Transparency | Candidates had no visibility (khả năng thấy được) into why they were screened out |
| Inclusiveness | System disadvantaged (gây bất lợi) candidates whose language reflected cultural or gender norms (chuẩn mực) |
Fix: Human-in-the-loop for all screening; blind (ẩn danh) CV processing; fairness audit (kiểm toán) against gender; regular bias testing.
5. AI-900 Exam Preparation
5.1 How AI-900 Scenario Questions Work
AI-900 scenario questions follow a predictable (có thể dự đoán) pattern:
- Describe a business problem — usually 2–4 sentences
- Ask which Azure AI service / principle / tool applies
- Include 2–3 distractors — services that almost fit but miss a key detail
The winning framework (khung chiến thắng):
- Identify the data type (text / image / audio / document / multimodal)
- Identify the task type (classify / extract / generate / detect / translate / transcribe)
- Identify any special constraints (custom domain / regulated industry / privacy / real-time)
- Match to the correct service using the decision tables from each workshop
5.2 Master Service Selection Reference
| Workload | Task | Correct Azure Service |
|---|---|---|
| Language | Sentiment, NER, key phrases, language detection | Azure AI Language (prebuilt) |
| Language | Custom intent + entity recognition | Azure AI Language — CLU |
| Language | FAQ chatbot from documents | Azure AI Language — Question Answering |
| Language | Custom text categories | Azure AI Language — Custom Classification |
| Language | Real-time text translation | Azure AI Translator |
| Speech | Audio → Text (general) | Azure AI Speech — STT |
| Speech | Audio → Text (domain-specific (chuyên biệt) vocabulary) | Azure AI Speech — Custom Speech |
| Speech | Identify WHO is speaking in a recording | Azure AI Speech — Speaker Diarization |
| Speech | Verify if a voice matches a registered user | Azure AI Speech — Speaker Verification |
| Vision | Describe, tag, detect objects in general images | Azure AI Vision — Image Analysis |
| Vision | Extract raw text from any image | Azure AI Vision — Read API (OCR) |
| Vision | Extract structured fields (trường có cấu trúc) from invoices/forms | Azure AI Document Intelligence |
| Vision | Train your own image classifier | Azure AI Custom Vision |
| Vision | Count people in a video stream (luồng video) | Azure AI Vision — Spatial Analysis |
| Vision | Verify two face photos are the same person | Azure AI Face — Face Verification |
| GenAI | Generate new text, summarize, answer conversationally | Azure OpenAI — GPT-4o |
| GenAI | Generate images from text prompt | Azure OpenAI — DALL-E 3 |
| GenAI | Transcribe audio with high multilingual accuracy | Azure OpenAI — Whisper |
| GenAI | Q&A grounded (có cơ sở) in private documents | Azure OpenAI + RAG (AI Search) |
| Safety | Detect harmful text/image content | Azure AI Content Safety |
| Custom ML | Domain-specific model, full data control | Azure Machine Learning |
5.3 Key Distinctions That AI-900 Tests Most
| Question Type | The Trap | The Answer |
|---|---|---|
| Vision Read API vs. Document Intelligence | Both do OCR | Read API = raw text; Document Intelligence = structured fields |
| Azure AI Language vs. Azure OpenAI | Both process text | Language = deterministic (tất định), structured; OpenAI = generative (tạo sinh), flexible |
| CLU vs. Question Answering | Both are conversational | CLU = intent/entity for action; QA = retrieves answers from documents |
| Speaker Diarization vs. Speaker Verification | Both involve identifying speakers | Diarization = separate (phân biệt) unknown speakers in a recording; Verification = match against a known (đã biết) voice |
| Azure AI Vision vs. Custom Vision | Both analyze images | Azure AI Vision = general tasks; Custom Vision = domain-specific model you train |
| Supervised vs. Unsupervised | Problem type | Supervised = labeled data; Unsupervised = no labels → clustering / anomaly detection |
| Overfitting vs. Underfitting | Model behavior | Overfitting = great training, poor test; Underfitting = poor both |
| RAG vs. Fine-tuning | Customization approach | RAG = external knowledge at query time; Fine-tuning = knowledge baked (mã hóa) into model weights |
| Hallucination source | Why it happens | Next-token prediction, NOT a retrieval failure — structural, not a bug |
| Content Safety vs. Prompt Shields | Two different defenses | Content Safety = filters output; Prompt Shields = blocks jailbreak/injection attempts at input |
5.4 Responsible AI Quick Reference
| Principle | One-Line Definition | Azure Tool |
|---|---|---|
| Fairness | Equal outcomes across demographic groups | Fairlearn, RAI Dashboard — Fairness Analysis |
| Reliability & Safety | Performs correctly under all conditions | Error Analysis, Model Monitoring, Content Safety |
| Privacy & Security | Protects personal data throughout lifecycle | PII Detection, Private Endpoints, Encryption |
| Inclusiveness | Works for all people, languages, and abilities | Multilingual testing, Custom Speech for accents |
| Transparency | Understandable decisions and disclosed limitations | InterpretML, Counterfactual Analysis, Transparency Notes |
| Accountability | Clear human ownership and audit trails | Model Registry, Audit Logs, RBAC, Human-in-the-loop |
5.5 The Five AI-900 Exam Domain Weights
| Domain | Weight | Key Topics |
|---|---|---|
| AI workloads and considerations | ~15% | ML types, AI workloads, Responsible AI principles |
| Fundamental principles of ML on Azure | ~20% | ML concepts, Azure ML, AutoML, Designer |
| Computer Vision on Azure | ~15% | CV tasks, Azure AI Vision, Custom Vision, Document Intelligence |
| NLP on Azure | ~15% | NLP pipeline, Azure AI Language, Azure AI Speech |
| Generative AI on Azure | ~35% | LLMs, Azure OpenAI, RAG, Responsible GenAI |
Note: Generative AI is now the highest-weighted (trọng số cao nhất) domain in AI-900 (updated 2024). Master Workshop 7 content thoroughly (kỹ lưỡng).
5.6 Common Exam Mistakes to Avoid
- Confusing Azure AI Vision and Azure AI Document Intelligence — always ask: "Does the output need to be raw text OR structured fields?"
- Applying Azure OpenAI when Azure AI Language suffices — if the task is deterministic (classify, extract, detect), prefer the specialized (chuyên biệt) Language service
- Thinking hallucination can be "fixed" by better prompting — it's structural; mitigation (giảm thiểu) requires RAG + human review
- Confusing Precision and Recall — always tie (kết nối) to the business cost: "Which is worse: False Positive or False Negative?"
- Treating Responsible AI as only about bias — it covers all 6 principles; security, privacy, reliability, transparency, and accountability are equally (bình đẳng) important
6. Course Capstone Discussion
Q1 — Cross-Workshop Integration: A hospital deploys a system that: (1) transcribes doctor-patient conversations with Azure AI Speech, (2) extracts symptoms and medications from transcripts with Azure AI Language, (3) uses GPT-4o to generate a patient care summary, (4) flags the summary for a doctor's review before sending to the patient. Identify which Responsible AI principle applies most critically (quan trọng nhất) to each of the four steps, and which Azure safety tool mitigates (giảm thiểu) the primary risk at each step.
Q2 — Build vs. Buy — Full Analysis: A Vietnamese e-commerce company wants to implement all of the following: (a) detect fake product reviews (đánh giá sản phẩm giả) in Vietnamese, (b) classify products into 200 custom categories from photos, (c) answer customer questions about return policies conversationally, (d) generate personalized product descriptions. For each, recommend (đề xuất) the appropriate Azure service — and explicitly (rõ ràng) state whether it requires training (cần training) or not.
Q3 — Responsible AI as Competitive Advantage (Lợi thế cạnh tranh): Many students view Responsible AI as a compliance (tuân thủ) cost — something required (bắt buộc) by regulation rather than desired (mong muốn). Construct (xây dựng) a business argument (lập luận kinh doanh) for why an organization that invests proactively (chủ động) in Responsible AI — beyond (vượt ra ngoài) minimum compliance — gains measurable (có thể đo lường) competitive advantage (lợi thế cạnh tranh) in three specific dimensions (chiều).
Made by Anh Tu - Share to be share