AWS Certified AI Practitioner
(AIF-C01) Study Guide
Free Study Guide — Pattern recognition. Decision matrices. Clear exam-focused explanations.
Independently written. Reviewed against AWS AIF-C01 Exam Guide v1.2, published 29 October 2026.
Use this free AWS AIF-C01 study guide to prepare for the AWS Certified AI Practitioner exam with five-domain coverage, 17 chapters, AWS service comparisons, decision matrices, and practical revision plans.
What This Free AWS AIF-C01 Study Guide Covers
This free AWS Certified AI Practitioner (AIF-C01) study guide is a structured learning resource for candidates preparing for the AWS AIF-C01 certification exam. It explains the exam concepts, compares relevant AWS AI services, and provides decision frameworks and revision plans across all five official exam domains.
- Domain 1 — Fundamentals of AI and Machine Learning: 20% of scored content.
- Domain 2 — Fundamentals of Generative AI: 24% of scored content.
- Domain 3 — Applications of Foundation Models: 28% of scored content.
- Domain 4 — Guidelines for Responsible AI: 14% of scored content.
- Domain 5 — Security, Compliance, and Governance for AI Solutions: 14% of scored content.
The guide contains 17 learning chapters, 13 advanced reference modules, AWS service comparisons, decision matrices, a glossary, and 7-day and 14-day study plans. It was reviewed against the official AWS AIF-C01 exam guide, version 1.2, published on 29 October 2026.
What is not included: this study guide is not a full-length practice exam and does not reproduce the separate diagnostic question set. Use the free 25-question AIF-C01 diagnostic on its dedicated page to check your readiness after studying.
Recommended use: review the exam overview, study each domain, compare services using the decision tables, then follow the 7-day or 14-day revision plan that fits your schedule.
Editorial Scope, Sources, and Exam Alignment
Who this resource is for: learners preparing for the AWS Certified AI Practitioner (AIF-C01) exam who need a structured explanation of AI and machine learning fundamentals, generative AI, foundation model applications, responsible AI, and AI security and governance.
How the material is organized: the five domain names, task objectives, and scored-content weightings follow the AWS exam guide. The explanations, comparison tables, decision frameworks, examples, and revision recommendations are CloudExamPro learning material designed to help readers understand and apply those objectives; they are not official AWS exam questions or an AWS publication.
Source-checking approach: exam alignment is checked against the official AIF-C01 exam guide, its revision history, and relevant in-scope AWS services and AWS product documentation. AWS may revise its objectives and service scope, so candidates should verify the latest official materials before booking the exam.
Last content review: 8 November 2026, against Exam Guide v1.2 (published 29 October 2026). This review date describes the current editorial check; it does not imply that AWS endorsed or certified this guide.
Important distinction: domain percentages represent the share of scored exam content, not a guaranteed number of questions. The official exam guide remains the authority for exam format, policies, and objectives.
How to Use This Guide
This guide is built for the AIF-C01 exam, not for general AI education. Every chapter maps to official exam objectives. Every decision matrix reflects a distinction the exam tests. Every trap is a confusion that costs candidates points.
Read the chapters in order if you are new to AI/ML. If you already work with AWS AI services, start with the Advanced Reference Modules (M1–M8), then read only the chapters where the matrices reveal gaps in your reasoning. The Final-Day Revision Sheet (M10) is designed for the last 24 hours before your exam.
Reading Paths
| If You Are… | Start With | Then | Skip Until Review |
|---|---|---|---|
| New to AI/ML | Terminology Primer → Ch 1–5 | Ch 6–17 in order | Advanced Reference Modules M6–M8 |
| Experienced with AWS but new to AI | Ch 6–9 (GenAI) | Ch 10–13 (FM applications) | Ch 1–5 (skim) |
| Experienced with AI/ML but new to AWS | Ch 5, Ch 10, Ch 11 | Ch 16–17 (security/governance) | Ch 1–4 (skim) |
| Exam in 7 days | M10 (Final-Day Revision) | M6 (Trap Library) → M7 (Trigger Phrases) | Deep chapter reads |
Free readiness check: After reviewing the concepts, use the AIF-C01 free 25-question diagnostic exam to check your current readiness. The diagnostic is kept on its dedicated exam page so this guide can stay focused on teaching and reference material.
Conventions Used in This Guide
- Decision Matrix — a comparison table with a decisive factor column. The exam rewards the candidate who can name the deciding requirement, not the one who memorizes service names.
- Exam Trap — a specific confusion between two or more concepts that the exam exploits in distractor options.
- Memory Anchor — a short, accurate recall phrase. Not a mnemonic gimmick — a compressed distinction.
- Coach's Takeaway — the one-sentence rule to carry into the exam.
AIF-C01 Exam Overview
Verified against the current official AWS Certified AI Practitioner exam guide. This guide follows the official blueprint, not the distribution of any practice question bank.
| Domain | Weight | Focus |
|---|---|---|
| Domain 1: Fundamentals of AI and ML | 20% | AI/ML concepts, learning paradigms, problem types, data, evaluation metrics, AWS AI services |
| Domain 2: Fundamentals of Generative AI | 24% | Foundation models, LLM mechanics, prompt engineering, hallucinations, RAG |
| Domain 3: Applications of Foundation Models | 28% | Amazon Bedrock, SageMaker AI, customization strategy, agents, Guardrails, evaluation |
| Domain 4: Guidelines for Responsible AI | 14% | Fairness, bias, explainability, transparency, human oversight, safety |
| Domain 5: Security, Compliance, and Governance | 14% | IAM, encryption, PrivateLink, CloudTrail, Macie, Config, Inspector, Artifact, Trusted Advisor, data residency, governance |
The percentage of scored exam content is not a guaranteed number of questions on any individual exam. This guide teaches all domains at full depth. Do not skip a domain because its weight is lower.
Exam Logistics — Official Format
- 65 questions: 50 scored and 15 unscored.
- Duration: 90 minutes (about 83 seconds per question on average).
- Question types: Multiple choice, multiple response, ordering, and matching.
- Passing score: 700 on a scaled range of 100–1,000. Scoring is compensatory; only the overall result determines pass/fail.
- Unscored questions: The 15 pilot questions are not identified. Treat every question as scored.
- Answer strategy: Unanswered questions are incorrect; there is no penalty for guessing. Answer every question.
Verified against AWS's official AIF-C01 exam guide and certification overview. Recheck before booking.
AIF-C01 Exam Guide v1.2 — Alignment Update
Reviewed against AWS AIF-C01 Exam Guide v1.2, published 29 October 2026. The five domain weights remain 20%, 24%, 28%, 14%, and 14%. This supplement strengthens the existing 17 chapters with focused revision of current objectives.
Official sources: Exam Guide · Revision history · Exam overview.
Domain 2 — Agentic AI and Context Engineering
- Agentic AI: Goal-directed, multi-step behavior using reasoning, tools, and actions within permissions; distinguish it from one-shot text generation.
- Multi-agent patterns and communication: Specialized agents delegate, collaborate, exchange results, and coordinate complex work.
- MCP: Model Context Protocol standardizes connections between AI applications/agents and external tools or data systems.
- Memory management: Retain relevant state across steps or interactions. Workflow orchestration: Coordinate sequence, handoffs, checkpoints, and completion.
- Context engineering: Design the full context—instructions, retrieved knowledge, history, tools, and constraints—not just prompt wording.
- Token-based pricing: Input/output tokens, model choice, and pricing mode affect cost. Long context and verbose outputs increase cost and may increase latency.
- AWS technologies to recognize: Amazon Bedrock, SageMaker AI, SageMaker JumpStart, Amazon Quick, Kiro, Strands Agents, and Bedrock AgentCore.
Domain 3 — Foundation Model Application Decisions
| Service | Scenario clue |
|---|---|
| Amazon OpenSearch Service | Search plus vector similarity retrieval. |
| Amazon Aurora | Relational application data with vector retrieval. |
| Amazon Neptune | Graph data and relationships. |
| Amazon RDS for PostgreSQL | PostgreSQL with vector extensions. |
| Approach | Best fit | Trade-off |
|---|---|---|
| In-context learning | Prompt examples/instructions guide the model. | Fast, but uses context tokens and window capacity. |
| RAG | Current, private, or source-grounded knowledge. | Depends on retrieval/index quality. |
| Fine-tuning | Consistent task behavior or domain style. | Needs curated data, evaluation, time, and cost. |
| Continued pre-training | Deeper domain language/knowledge exposure. | More compute- and data-intensive. |
| Model distillation | Smaller student model approximates a larger teacher. | Lower latency/cost may mean capability loss. |
Model selection: Compare modality, language support, size/complexity, latency, input/output limits, customization, prompt caching, cost, compliance, and quality. Evaluation: Know ROUGE, BLEU, BERTScore, LLM-as-a-judge, human review, and benchmarks. Evaluate RAG, agents, and workflows using task completion, user satisfaction, and cost per interaction. Fine-tuning data: Know curation, governance, size, labeling, representativeness, and RLHF.
Domain 4 — Responsible AI
- Include environmental impact and sustainability in model selection.
- Assess dataset inclusivity, diversity, balance, and source quality.
- Recognize IP claims, biased outputs, hallucinations, end-user harm, and loss of trust as legal or reputational risks.
- Know subgroup analysis, human audits, label quality, SageMaker Clarify, SageMaker Model Monitor, and human-in-the-loop review workflows.
- Understand Model Cards, open-source model information, licensing, data provenance, explainability, and human feedback.
Domain 5 — Security, Privacy, and Governance
| Item | Exam association |
|---|---|
| Bedrock AgentCore Identity / Policy | Agent identity, authentication, and permissions. |
| Data lineage/cataloging | Documenting data origin, movement, ownership, and use. |
| Secure data engineering | Data quality, privacy techniques, access control, and integrity. |
| AWS Config / Amazon Inspector | Configuration/compliance evaluation / vulnerability management. |
| AWS CloudTrail / Trusted Advisor | API audit trails / security, reliability, and cost recommendations. |
| Generative AI Security Scoping Matrix | Identify security concerns for different GenAI architectures. |
AI security risks: Prompt injection, data leakage, unsafe tool invocation, toxicity, and poor output validation. Recognize least privilege, encryption at rest/in transit, input/output filtering, audit logs, and human approval for sensitive actions. Grounding: RAG ties answers to retrieved sources; validation checks outputs; confidence scoring can flag uncertainty. Citations improve traceability but do not guarantee correctness.
AWS's v1.2 history records six services removed from the out-of-scope list: Amazon Elastic Transcoder, AWS Elemental MediaStore, AWS IoT 1-Click, AWS IoT Analytics, AWS IoT Events, and Amazon Nimble Studio. This does not mean they are newly emphasized exam topics. Use current task statements and the in-scope list as the primary boundary.
Study Plans
7-Day Aggressive Plan
| Day | Focus | Deliverable |
|---|---|---|
| 1 | Ch 1–3 (AI/ML foundations, lifecycle) | Complete Term Primer + Ch 1–3 Knowledge Checks |
| 2 | Ch 4–5 (Evaluation, AWS services) | Memorize evaluation metrics matrix |
| 3 | Ch 6–7 (GenAI, LLM mechanics) | Complete inference parameter matrix |
| 4 | Ch 8–9 (Prompt engineering, RAG) | Prompt technique comparison table |
| 5 | Ch 10–13 (Bedrock, SageMaker, customization) | Bedrock vs SageMaker matrix |
| 6 | Ch 14–17 (Responsible AI, security) | M6 Trap Library + M7 Trigger Phrases |
| 7 | M9–M10 (Rapid review, final-day sheet) | M11 Self-Assessment + M13 Exam-Day Checklist |
14-Day Recommended Plan
| Days | Focus | Priority |
|---|---|---|
| 1–3 | Ch 1–5 (Domain 1) | Foundations — do not rush |
| 4–6 | Ch 6–9 (Domain 2) | Highest combined weight with Domain 3 |
| 7–10 | Ch 10–13 (Domain 3) | Largest single domain (28%) |
| 11 | Ch 14–15 (Domain 4) | Responsible AI |
| 12 | Ch 16–17 (Domain 5) | Security, compliance, governance |
| 13 | M1–M8 (Advanced Reference Modules) | Decision matrices + trap library |
| 14 | M9–M13 (Review + Exam Day) | Final-day revision + checklist |
AI/ML Terminology Primer
If you are new to AI/ML, learn these 40 terms before Chapter 1. They appear throughout the guide and the exam.
| Term | Plain-English Definition | Exam Context |
|---|---|---|
| Artificial Intelligence (AI) | Broad field of making machines perform tasks requiring human intelligence | The outermost category |
| Machine Learning (ML) | Subset of AI that learns patterns from data | Most exam scenarios sit here |
| Deep Learning | ML using multi-layer neural networks | Images, audio, text at scale |
| Generative AI | Capability to generate new content, commonly using deep-learning foundation models | Text, image, audio, video, and code generation |
| Foundation Model (FM) | Large pre-trained model adaptable to many tasks | Bedrock's core offering |
| Large Language Model (LLM) | Foundation model specialized for text | Chatbots, summarization, translation |
| Token | Unit of text a model processes (word, subword, character) | Context window, billing |
| Embedding | Numeric vector representing meaning | RAG, semantic search |
| Context Window | Maximum tokens a model can process at once | Why long inputs fail |
| Prompt | Input instruction to a model | Prompt engineering |
| Inference | Using a trained model to generate output | Runtime phase |
| Training | Learning model weights from data | Pre-training, fine-tuning |
| Fine-Tuning | Further training a model on labeled task data | Domain adaptation |
| Continued Pre-Training | Further training on unlabeled domain data | Domain knowledge expansion |
| RAG | Retrieval Augmented Generation — grounding responses in external data | Reduces hallucinations |
| Hallucination | Model generates plausible but false content | Key GenAI risk |
| Temperature | Inference parameter that influences sampling variability | Lower = more focused; higher = more varied (not a factuality guarantee) |
| Top K | Limits next-token candidates to K most probable | Diversity control |
| Top P | Nucleus sampling — limits candidates by cumulative probability | Alternative to Top K |
| Guardrails | Safety controls on model input and output | Bedrock Guardrails |
| Agent | AI system that reasons and calls external tools/APIs | Bedrock Agents |
| Vector Database | Database optimized for similarity search on embeddings | OpenSearch, Aurora pgvector |
| Chunking | Splitting documents into smaller units for embedding | RAG pipeline |
| Overfitting | Model memorizes training data, fails on new data | Train-good/test-bad |
| Underfitting | Model too simple to learn the pattern | Train-bad/test-bad |
| Regularization | Technique to reduce overfitting (L1, L2) | SageMaker Linear Learner |
| Precision | Correct positive predictions / all positive predictions | Minimize false positives |
| Recall | Correct positive predictions / all actual positives | Minimize false negatives |
| F1 Score | Harmonic mean of precision and recall | Imbalanced classes |
| ROUGE | Summarization evaluation metric | Compare to references |
| BLEU | Translation evaluation metric | Compare to references |
| BERTScore | Semantic similarity metric using embeddings | Bedrock auto-eval for summarization |
| Data Drift | Input distribution changes over time | Model Monitor detects |
| Concept Drift | Input-output relationship changes | Model Monitor detects |
| Bias | Systematic error producing unfair outcomes | SageMaker Clarify detects |
| Explainability | Understanding why a model made a decision | SageMaker Clarify SHAP values |
| Data Residency | Data must remain in a specific geographic region | Compliance requirement |
| PrivateLink | Private network connection to AWS services | Bedrock without public internet |
| Model Card | Standardized model documentation | Governance and audit |
| Prompt Injection | Malicious input overriding system instructions | Security risk |
Foundations of AI, Machine Learning, Deep Learning, and Generative AI
This chapter establishes the vocabulary you will use throughout the exam. It draws the boundaries between four related but distinct fields and shows where AWS services fit.
1.1 The AI Hierarchy
Think of these four terms as nested circles. Artificial intelligence (AI) is the broad goal of making machines perform tasks that normally require human intelligence. Machine learning (ML) is a subset of AI: instead of programming rules by hand, you let the system learn patterns from data. Deep learning is a subset of ML that uses multi-layer neural networks. Generative AI creates new content — text, images, audio, video, or code — rather than only classifying or predicting. Modern GenAI commonly uses deep-learning foundation models, but treat GenAI as a capability rather than a strict extra layer in every AI hierarchy. Agentic AI describes a system pattern in which an application can select actions, use tools, and coordinate steps; it is not simply another model type.
AI is the broadest field. ML learns patterns from data; deep learning is ML based on multi-layer neural networks. GenAI is a content-generation capability commonly powered by deep-learning foundation models. Agentic AI describes systems that use models, tools, memory, and workflows to pursue goals.
1.2 Neural Networks and Deep Learning Mechanics
A neural network consists of layers of interconnected nodes. Data enters through the input layer, passes through hidden layers, and exits through the output layer. Each connection has a weight that determines influence. During training, the network adjusts these weights to minimize error.
Input Data → Input Layer → Hidden Layer 1 → Hidden Layer 2 → Output Layer
↓ ↓ ↓
Weights Weights Weights
↑ ↑ ↑
Backpropagation ← Loss Function ← Prediction vs Actual
Figure 1.1 — Basic neural network training loop. Data flows forward; error gradients flow backward to update weights.
1.3 AWS AI Service Taxonomy
| Dimension | Managed AI Services | ML Platforms |
|---|---|---|
| Examples | Rekognition, Comprehend, Transcribe, Polly, Translate, Personalize, Lex, Textract | SageMaker AI, Bedrock, SageMaker JumpStart |
| ML expertise required | None | Yes (SageMaker); moderate (Bedrock fine-tuning) |
| Time to value | Minutes | Days to weeks |
| Best for | Standard tasks | Domain-specific tasks, proprietary data |
Start with managed AI services. They solve standard problems with zero ML overhead. Move to platforms only when you need custom models or proprietary data.
Learning Paradigms and Problem Types
This chapter teaches the three learning paradigms and the problem types they solve. By the end, you will match a scenario to the right approach quickly.
2.1 Supervised Learning
Uses labeled training data — each input is paired with the correct output. The model learns the mapping and applies it to new data.
Labeled data + prediction target = supervised learning. Category = classification. Number = regression.
2.2 Unsupervised Learning
Works with unlabeled data. The algorithm discovers structure on its own — groupings, patterns, anomalies.
2.3 Reinforcement Learning
Trains an agent to make decisions by rewarding good actions and penalizing bad ones. The agent learns through trial and error.
2.4 Classification vs. Regression
| Dimension | Classification | Regression |
|---|---|---|
| Output type | Category | Continuous number |
| Example | Will customer churn? (yes/no) | How much will customer spend? ($) |
| Loss function | Cross-entropy | MSE, MAE |
| Evaluation metric | Precision, recall, F1, accuracy | RMSE, MAE, R² |
The ML Development Lifecycle and Data
This chapter walks through the ML lifecycle from business goal identification to deployment and monitoring, then dives into data concepts.
3.1 The ML Lifecycle
flowchart LR
A[Business Goal
Identification] --> B[ML Problem
Framing]
B --> C[Data Processing]
C --> D[Model Development]
D --> E[Evaluation]
E --> F[Deployment]
F --> G[Monitoring]
G --> A
style A fill:#0b2545,color:#fff
style B fill:#13315c,color:#fff
style C fill:#1d4e89,color:#fff
style D fill:#2d6a9f,color:#fff
style E fill:#4a8db7,color:#fff
style F fill:#4a8db7,color:#fff
style G fill:#4a8db7,color:#fff
Figure 3.1 — ML lifecycle. Business goal identification comes first. Compliance and regulatory requirements are established before data collection.
Business goal first, always. Compliance, regulatory requirements, and success metrics are defined before any technical work begins.
3.2 Data Types
| Data Type | Example | ML Approach | AWS Service |
|---|---|---|---|
| Structured (tabular) | Customer records, sales transactions | Classical ML | SageMaker AI, Personalize |
| Unstructured (text) | Reviews, emails, documents | NLP, LLMs | Comprehend, Bedrock, Textract |
| Unstructured (image/video) | Photos, surveillance | Computer vision | Rekognition, SageMaker AI |
| Unstructured (audio) | Voice recordings | Speech recognition | Transcribe |
| Time-series | Stock prices, sensor data | Forecasting | SageMaker AI (DeepAR) |
Model Evaluation and Metrics
This chapter teaches the metrics used to evaluate classification, regression, and generative AI models, plus overfitting, underfitting, bias, and drift.
4.1 Classification Metrics
| Metric | Formula | Measures | When to Use |
|---|---|---|---|
| Accuracy | (TP+TN) / Total | Overall correctness | Balanced classes |
| Precision | TP / (TP+FP) | Correctness of positive predictions | Minimize false positives |
| Recall | TP / (TP+FN) | Coverage of actual positives | Minimize false negatives |
| F1 Score | 2 × (P × R) / (P + R) | Balance of precision and recall | Imbalanced classes |
Accuracy is misleading on imbalanced data. If 95% of transactions are legitimate, a model predicting "legitimate" every time achieves 95% accuracy but catches zero fraud. Use F1 score instead.
4.2 Generative AI Evaluation Metrics
| Metric | Purpose | When to Use |
|---|---|---|
| ROUGE | Summarization quality (recall-oriented) | Comparing summaries to references |
| BLEU | Translation quality (precision-oriented) | Comparing translations to references |
| BERTScore | Semantic similarity via embeddings | Bedrock automatic evaluation for summarization |
Amazon Bedrock's automatic evaluation feature for text summarization reports BERTScore using the Gigaword dataset. The academic default for summarization is ROUGE. Both are correct in their contexts. If the question specifies Bedrock's automatic evaluation, choose BERTScore.
4.3 Overfitting and Underfitting
| Signature | Diagnosis | Remedy |
|---|---|---|
| Train-good, test-bad | Overfitting | Increase regularization, reduce complexity, more data |
| Train-bad, test-bad | Underfitting | Increase complexity, add features, reduce regularization |
AWS AI/ML Services Landscape
This chapter organizes AWS AI/ML services by the problem they solve — managed AI services, SageMaker AI, Bedrock, and generative AI developer tools.
5.1 Managed AI Services
| Service | Problem Solved | When to Use | Alternative |
|---|---|---|---|
| Amazon Rekognition | Image and video analysis | Object detection, facial analysis, moderation | SageMaker AI |
| Amazon Comprehend | Text analysis | Sentiment, entities, toxicity | SageMaker AI, Bedrock |
| Amazon Transcribe | Speech to text | Subtitles, call transcription | None |
| Amazon Polly | Text to speech | Voice-overs, accessibility | None |
| Amazon Translate | Language translation | Multilingual content | None |
| Amazon Textract | Document extraction | OCR, forms, tables | Bedrock Data Automation |
| Amazon Lex | Conversational interfaces | Build chatbots and voice assistants | Amazon Bedrock |
| Amazon Personalize | Recommendations | Product/content recommendations | SageMaker AI |
| Amazon SageMaker AI | Custom fraud classification | Build a model for organization-specific fraud patterns | Managed AI services |
5.2 Amazon SageMaker AI Feature Map
| Feature | Purpose |
|---|---|
| SageMaker Studio | Integrated development environment |
| SageMaker Autopilot | Automated model building and tuning |
| SageMaker JumpStart | Pre-trained models and solution templates |
| SageMaker Data Wrangler | Visual data preparation |
| SageMaker Clarify | Bias detection and explainability |
| SageMaker Model Monitor | Production drift detection |
| SageMaker Model Registry | Model versioning and approval |
| SageMaker Model Cards | Standardized model documentation |
| SageMaker Ground Truth | Data labeling with human review |
| SageMaker Canvas | No-code ML for business analysts |
| SageMaker Feature Store | Centralized feature repository |
| SageMaker Pipelines | ML workflow orchestration |
| SageMaker Serverless Inference | Model deployment without managing infrastructure |
| SageMaker Batch Transform | Offline inference for large datasets |
5.3 Amazon Bedrock Feature Map
| Feature | Purpose |
|---|---|
| Model access | Browse and invoke FMs from multiple providers |
| Knowledge Bases | Managed RAG — connect data and retrieve context |
| Agents | Orchestrate multi-step tasks with API calls |
| Guardrails | Content filters, denied topics, PII detection, grounding |
| Prompt Management | Store, version, and reuse prompts |
| Model Evaluation | Automatic and human evaluation of outputs |
| Provisioned Throughput | Dedicated capacity for steady traffic |
| Batch Inference | Bulk generation at discounted rates |
| Prompt Caching | Reuse prompt prefixes to reduce latency and cost |
| Prompt Router | Dynamically select the best model per request |
Match the requirement to the service class: managed AI service for standard tasks, SageMaker AI for custom models, Bedrock for foundation models. When in doubt, start managed.
Domain 1 — v1.2 Focus: Inference, MLOps, and Business Outcomes
Choose the inference pattern from the workload. Batch inference processes collections of inputs together; real-time inference returns a response during a request; asynchronous inference accepts a request and returns its result later; serverless inference reduces the need to manage inference infrastructure.
| Requirement | Best fit |
|---|---|
| Score thousands of records overnight | Batch inference |
| Return a prediction inside a live user request | Real-time inference |
| Accept a large job and retrieve its result later | Asynchronous inference |
| Run inference without managing servers | Serverless inference |
MLOps covers repeatable experimentation, deployment, monitoring, and retraining practices. Connect technical metrics (accuracy, precision, recall, F1) to business measures such as cost per user, customer feedback, development cost, and ROI. A model with strong F1 can still fail the business case if operating cost is too high.
Use a traditional ML model when a structured prediction, lower cost, or more explainable result is the better fit. Do not choose a foundation model simply because it is newer.
When AI/ML Is Not the Right Choice
AI is useful when prediction, classification, recommendation, perception, or language capabilities add measurable value. It is not automatically the best solution for every business problem.
| Situation | Better choice | Why |
|---|---|---|
| A fixed, deterministic rule solves the task | Ordinary software or business rules | Easier to test, explain, and maintain |
| A guaranteed outcome is required | Deterministic logic with validation | ML produces probabilistic predictions, not guarantees |
| Representative data is insufficient | Improve data or use a simpler approach | Weak or biased data makes predictions unreliable |
| Error cost exceeds expected benefit | Retain human decisions or do not automate | Compare model, infrastructure, monitoring, and review costs with measurable benefit |
| Structured historical data supports a clear prediction | Consider traditional ML | Classification or regression may be cheaper and more explainable than an FM |
Do not choose AI because it is fashionable. If a deterministic rule solves the problem reliably, or the cost-benefit case is weak, a non-AI solution may be the better answer.
Generative AI and Foundation Models
This chapter introduces generative AI fundamentals: foundation models, large language models, and the architecture that powers them.
6.1 What Makes Generative AI Different
Traditional ML predicts a label or value. Generative AI creates new content — text, images, audio, video, code. The distinction matters because the exam tests whether you recognize when a scenario calls for generation versus prediction.
Predict a category → classification. Predict a number → regression. Create new content → generative AI.
6.2 Foundation Models and LLMs
A foundation model is a large pre-trained model adaptable to many downstream tasks. A large language model is a foundation model specialized for text. Transformers — the architecture behind modern LLMs — use self-attention to capture contextual relationships across all tokens simultaneously.
flowchart TD
A[Foundation Model] --> B[Text: LLM]
A --> C[Image: Diffusion]
A --> D[Multimodal]
B --> E[Chatbot]
B --> F[Summarization]
B --> G[Code Generation]
C --> H[Image Generation]
C --> I[Image Editing]
D --> J[Text + Image Understanding]
D --> K[Video Generation]
style A fill:#0b2545,color:#fff
style B fill:#1d4e89,color:#fff
style C fill:#1d4e89,color:#fff
style D fill:#1d4e89,color:#fff
Figure 6.1 — Foundation model taxonomy. Foundation models specialize by modality (text, image, multimodal) and application (chat, summarize, generate).
6.3 Model Selection Considerations
| Factor | What It Means | Exam Trigger |
|---|---|---|
| Modality | Input/output types supported | "Text and image inputs" → multimodal |
| Latency | Response time | "Real-time" → prioritize low latency |
| Context window | Maximum input size in tokens | "Long documents fail" → context limit |
| Cost | Per-token pricing | "Minimize cost" → smaller model |
| Customization | Fine-tuning support | "Domain-specific" → fine-tuning |
Select the smallest model that satisfies the performance requirement. Larger models cost more without adding value when a smaller model already meets the task.
LLM Mechanics: Tokens, Embeddings, Context Windows, and Inference Controls
This chapter explains how LLMs process text and how inference parameters control output.
7.1 Tokens and Tokenization
Tokenization splits text into smaller units — words, subwords, or characters — that the model processes numerically. These tokens are the building blocks for all NLP tasks.
7.2 Embeddings
Embeddings map text into a high-dimensional vector space where proximity reflects semantic similarity. That enables mathematical comparison — cosine similarity, nearest-neighbor search — for semantic search, clustering, and retrieval.
Embeddings are not generated text. Embeddings are numeric vectors for comparison. Generated text is the model's output. A text embeddings model converts text to vectors; a text generation model produces new text.
7.3 Context Windows
Language models have a maximum context window measured in tokens. If input exceeds that limit, the model cannot process the full input — leading to errors, truncation, or failure.
7.4 Inference Parameters
| Parameter | Effect | Use When |
|---|---|---|
| Temperature | Controls sampling variability; lower values make outputs more focused, higher values allow more variation | Lower for consistency, higher for creative diversity |
| Top K | Limits next-token candidates to K most probable | Balance diversity and focus |
| Top P | Nucleus sampling — cumulative probability cutoff | Alternative to Top K |
| Max tokens | Maximum output length | Control response size |
| Stop sequences | Tokens that end generation | Prevent runaway output |
Lower temperature generally makes sampling more focused; higher temperature generally increases variation. A temperature of 0 does not guarantee identical outputs in every system or context, and low temperature does not guarantee factual accuracy. Use grounding, evaluation, and validation to address hallucinations.
Prompt Engineering Techniques
This chapter teaches the prompting techniques the exam tests: zero-shot, few-shot, chain-of-thought, and their applications.
8.1 Prompt Engineering Fundamentals
A prompt is the input instruction you give the model. Prompt engineering is the practice of designing prompts that produce accurate, relevant, and well-formatted responses.
8.2 Prompting Techniques
| Technique | What It Does | When to Use | Example Trigger |
|---|---|---|---|
| Zero-shot | No examples; task only | Simple, well-defined tasks | "Classify the following text…" |
| Few-shot | A few labeled examples before task | Format or style must be matched | Three examples, then "Now classify this" |
| Single-shot | One worked example before the task | One example clarifies the expected pattern | Show one input/output pair, then ask for the next |
| Negative prompt | States content, style, or elements to avoid | Generation needs explicit exclusions | "Do not include text, logos, or extra objects" |
| Iterative refinement | Uses follow-up feedback or comments to improve an earlier response | The first output needs correction or more detail | Review, give targeted feedback, and regenerate |
| Chain-of-thought | Instructs step-by-step reasoning | Complex multi-step problems | "Think step by step" |
| ReAct | Reasoning + external tool/API calls | Need live data or actions | "Check inventory and respond" |
| Prompt chaining | Output of one prompt feeds next | Multi-stage complex tasks | Sequential subtask execution |
No examples → zero-shot. Few examples → few-shot. Step-by-step → chain-of-thought. Reason + act → ReAct. Chained prompts → prompt chaining.
8.3 Prompt Injection and Security
Prompt injection is an attempt to make a model ignore its intended instructions or misuse the context and tools available to it. It can arrive directly from a user or indirectly through retrieved documents, web pages, and other untrusted content. Retrieved text should be treated as data, not as trusted instructions.
| Risk | Meaning | Example / Response |
|---|---|---|
| Exposure | Sensitive prompts, system instructions, or confidential context are revealed | Minimize secrets in model context; restrict access and redact sensitive data |
| Prompt poisoning | Malicious or misleading content is inserted into prompt inputs or retrieved sources | Validate and curate data sources; treat retrieved content as untrusted |
| Prompt hijacking / injection | Attacker-supplied instructions attempt to override the intended task or system rules | Separate trusted instructions from data; apply input screening and tool authorization |
| Jailbreaking | Prompts attempt to bypass safety restrictions or policy controls | Use safety testing, guardrails, monitoring, and human escalation for sensitive actions |
Guardrails are one layer, not a complete security boundary. Combine them with least-privilege IAM, restricted tool permissions, validation of external content, output checks, logging, and human approval for consequential actions.
8.4 Prompt Management and Versioning
Prompts are application assets. When teams change an instruction, template, example, or model setting, they should be able to identify exactly which version produced a response and compare the effect of the change.
- Use a reusable template: separate stable instructions from variables such as customer question, retrieved context, or product details.
- Store versions: label meaningful prompt revisions and preserve the previous working version.
- Record the configuration: track the model, prompt version, inference parameters, and relevant evaluation dataset together.
- Test before promotion: run representative examples, adversarial inputs, and regression tests against the same evaluation criteria.
- Promote and monitor: release the approved version, monitor quality and cost, and roll back if results deteriorate.
Amazon Bedrock Prompt Management helps create, store, version, and manage reusable prompts. For exam scenarios that mention centralized prompt templates, prompt versions, or controlled prompt iteration on AWS, recognize Prompt Management.
Prompt versioning is not model fine-tuning. Versioning changes and tracks instructions or templates; fine-tuning changes model weights using training data. If a team only needs to improve instructions and compare revisions, start with prompt engineering and management.
Hallucinations, Grounding, and RAG Fundamentals
This chapter teaches why LLMs hallucinate, how RAG grounds responses in external data, and how to evaluate RAG systems.
9.1 Hallucinations
A hallucination is plausible-sounding but factually incorrect content. It is the primary production risk for generative AI in customer-facing and regulated contexts.
9.2 Retrieval Augmented Generation (RAG)
RAG retrieves relevant information from an external knowledge base at inference time and feeds it to the model as context. The model stays fixed; retrieved context changes. That grounds responses in verifiable sources and reduces hallucinations.
flowchart LR
A[User Query] --> B[Query Embedding]
B --> C[Vector Search]
C --> D[Retrieve Relevant Chunks]
D --> E[Augment Prompt with Context]
E --> F[LLM Generates Grounded Response]
F --> G[Response with Source Citations]
style A fill:#0b2545,color:#fff
style F fill:#1d4e89,color:#fff
style G fill:#2d6a9f,color:#fff
Figure 9.1 — RAG workflow. The query is embedded, relevant chunks are retrieved, and the LLM generates a response grounded in those chunks.
9.3 RAG Pipeline: Online vs. Offline
| Stage | When It Runs | Latency Requirement |
|---|---|---|
| Parse documents | Offline (batch) | None |
| Chunk documents | Offline (batch) | None |
| Generate content embeddings | Offline (batch) | None |
| Build search index | Offline (batch) | None |
| Generate query embeddings | Online (query time) | Low |
| Retrieve relevant content | Online (query time) | Low |
| Generate response | Online (query time) | Low |
RAG's core value is reducing hallucinations through grounding. It does not speed up training or inference. Update the knowledge base, not the model.
Domain 2 — v1.2 Focus: Context Engineering, Agents, and Cost
Context engineering designs all information supplied to a foundation model at inference time: instructions, user input, retrieved documents, conversation history, tool descriptions, and constraints. Prompt engineering focuses on wording; context engineering manages the wider information environment.
| Concept | Exam clue |
|---|---|
| Tool use | Agent invokes an external function, API, or service |
| Memory | Relevant state is retained across steps or interactions |
| Multi-agent pattern | Specialist agents delegate, collaborate, and exchange results |
| Workflow orchestration | Coordinates sequence, handoffs, and completion |
| Model Context Protocol (MCP) | Standardized connection to external tools and data |
Token economics: input and output tokens contribute to inference cost. Longer context, larger prompts, and verbose responses can increase cost and may add latency. Compare model capability, token price, throughput, and response-time requirements together.
AWS infrastructure and tool selection: use Amazon Bedrock for managed access to foundation models and application features; use SageMaker AI when broader control is needed to build, train, customize, or host models. SageMaker JumpStart helps discover and deploy pretrained models. Amazon Quick is an in-scope analytics/work-assistance offering, while Kiro and Strands Agents are developer and agent-building technologies; do not confuse them with model-training platforms. Bedrock AgentCore supports building and operating agents. Architecture choices should weigh managed API access against self-hosting, on-demand against provisioned capacity, token cost against throughput, and latency against redundancy and regional availability. Custom models can improve fit but add training, hosting, governance, and operational costs.
An agent is not simply a chatbot that writes a longer answer. Agentic behavior includes selecting tools or actions and coordinating steps within permission boundaries.
Amazon Bedrock: The Managed FM Platform
This chapter covers Amazon Bedrock's capabilities: model access, Knowledge Bases, Agents, Guardrails, and pricing models.
10.1 Bedrock Overview
Amazon Bedrock is a fully managed service that provides access to high-performing foundation models from multiple providers through a single API. You do not manage infrastructure, and you pay for what you use.
10.2 Bedrock Capabilities
| Capability | Purpose | When to Use |
|---|---|---|
| Knowledge Bases | Managed RAG | Ground responses in company data |
| Agents | Multi-step task orchestration | API calls, tool use, transactions |
| Guardrails | Safety controls | Block harmful content, PII, denied topics |
| Prompt Management | Store and reuse prompts | Standardize prompt templates |
| Model Evaluation | Assess model quality | Compare models, measure accuracy |
| Prompt Caching | Reuse prompt prefixes | Reduce latency and cost |
| Prompt Router | Dynamic model selection | Single endpoint, best model per request |
10.3 Pricing Models
| Model | Best For | When to Choose |
|---|---|---|
| On-Demand | Unpredictable or low usage | Pilot phase, variable traffic |
| Provisioned Throughput | Steady, high-volume traffic | Predictable usage, custom models |
| Batch | Periodic bulk generation | Weekly reports, catalog updates |
Steady, predictable traffic → Provisioned Throughput. Unpredictable or low usage → On-Demand. Periodic bulk → Batch.
Amazon SageMaker AI and Custom Model Development
This chapter covers SageMaker AI's platform capabilities: Autopilot, JumpStart, Clarify, Model Monitor, Ground Truth, and deployment options.
11.1 SageMaker AI Overview
SageMaker AI is AWS's end-to-end platform for building, training, and deploying custom ML models. It provides infrastructure, tools, and workflows for every stage of the ML lifecycle.
11.2 Key SageMaker Features
| Lifecycle Stage | Feature | Purpose |
|---|---|---|
| Data preparation | Data Wrangler | Visual, low-code data preparation |
| Data labeling | Ground Truth | Human review and labeling |
| Model building | Autopilot | Automated model building and tuning |
| Model building | Canvas | No-code ML for business analysts |
| Model building | JumpStart | Pre-trained models and templates |
| Training | SageMaker Training | Managed training jobs |
| Bias/explainability | Clarify | Bias detection and SHAP explanations |
| Governance | Model Cards | Standardized model documentation |
| Governance | Model Registry | Versioning and approval workflows |
| Deployment | Endpoints | Real-time inference |
| Deployment | Serverless Inference | No infrastructure management |
| Deployment | Batch Transform | Offline inference for large datasets |
| Monitoring | Model Monitor | Drift detection |
Model Monitor vs. Clarify. Model Monitor detects production drift. Clarify detects bias and explains predictions. Both are governance features, but they solve different problems.
Customization Strategy: Prompt Engineering vs. RAG vs. Fine-Tuning vs. Continued Pre-Training
This chapter teaches the decision framework for customizing foundation models.
12.1 The Customization Decision
| Approach | Data Required | Effort | When to Use |
|---|---|---|---|
| Prompt engineering | None | Lowest | Style, format, simple task adaptation |
| RAG | Documents (unlabeled) | Moderate | Frequently changing data, grounded answers |
| Fine-tuning | Labeled input-output pairs | High | Task-specific behavior, style adaptation |
| Continued pre-training | Unlabeled domain text | Highest | Domain knowledge expansion |
Least to most effort: prompt → RAG → fine-tune → continued pre-training. Prompt engineering first. RAG for changing data. Fine-tuning for task behavior. Pre-training for domain knowledge.
12.2 Training and Fine-Tuning: What Happens to the Data?
For the exam, distinguish the purpose of each training stage and the kind of data it uses. You are expected to recognize the approach—not implement a training pipeline.
| Method | What Changes? | Typical Data | Exam Clue |
|---|---|---|---|
| Pre-training | Builds broad language or multimodal capabilities from large datasets | Large-scale, often unlabeled data | Creating a foundation model from the beginning |
| Continued pre-training | Extends a pretrained model's knowledge in a domain | Additional domain-specific, usually unlabeled data | Specialized vocabulary or domain knowledge |
| Instruction fine-tuning | Improves following instructions and producing desired responses | Instruction-and-response examples | Consistent task behavior or response format |
| Transfer learning | Adapts knowledge learned on one task or dataset to another | Task- or domain-specific examples | Reuse a pretrained model rather than start from zero |
| Model distillation | Transfers useful behavior from a larger teacher model to a smaller student model | Teacher outputs and/or curated training examples | Lower inference cost or latency with a smaller model |
| RLHF | Uses human preferences to guide model behavior | Human rankings or preference comparisons | People compare responses and indicate which is better |
Fine-Tuning Data Preparation
- Define target behavior: specify the task, audience, expected output, and success criteria.
- Curate representative examples: cover normal cases, edge cases, and the real distribution of expected inputs.
- Check quality and labels: remove duplicates, correct errors, standardize formats, and verify labels.
- Apply governance: confirm rights to use the data, minimize sensitive information, and apply access controls.
- Separate evaluation data: keep a held-out set not used for training, so evaluation measures generalization.
More training is not always the answer. If facts change frequently, RAG is often more practical than fine-tuning. If the goal is consistent output behavior, instruction fine-tuning may fit. If the goal is lower serving cost, consider distillation. Choose by the stated business need.
A company needs an assistant to use a frequently updated internal policy library but does not need to change the model's writing style. Best fit: RAG, because the knowledge changes independently of model weights.
Agents, Guardrails, and Model Evaluation on Bedrock
This chapter covers Bedrock Agents, Guardrails, and model evaluation capabilities.
13.1 Bedrock Agents
Amazon Bedrock Agents orchestrate multi-step tasks: interpreting user input, calling external APIs, and producing outputs. They are the right choice when the application needs to reason and act.
| Pattern | Main job | Example | What it does not do by itself |
|---|---|---|---|
| RAG | Find relevant information and provide it as evidence for a generated answer | Answer a policy question using the latest approved policy documents | It does not normally approve requests, update records, or execute a business process |
| AI agent | Coordinate steps, select tools, and take permitted actions toward a goal | Check an order, verify eligibility, create a support ticket, and report the result | It does not automatically have reliable, current knowledge unless connected to trusted data sources |
| RAG + agent | Ground decisions in retrieved information, then orchestrate authorized actions | Retrieve the current refund policy, check the order, and submit an eligible refund request for approval | Requires access controls, action validation, monitoring, and human approval where appropriate |
Need to answer from changing documents? Use RAG. Need to perform a multi-step task with tools? Use an agent. Need both grounded information and actions? Combine RAG with an agent, while enforcing permissions outside the model.
13.2 Bedrock Guardrails
| Policy | What It Does | When to Use |
|---|---|---|
| Content filters | Block harmful categories (hate, violence, sexual, insults) | General safety |
| Denied topics | Block custom subject areas | Politics, competitor products, legal advice |
| Word filters | Block specific words or phrases | Profanity, brand names |
| Sensitive information filters | Block or mask PII | Customer data protection |
| Contextual grounding check | Verify responses are supported by source context | Reduce hallucinations |
| Prompt attacks | Detect jailbreak attempts | Public-facing chatbots |
Content filters vs. denied topics. Content filters handle predefined harmful categories (hate, violence). Denied topics handle custom subject areas you define (politics, competitor products).
13.3 Bedrock Model Evaluation
| Metric | What It Measures | Task Type |
|---|---|---|
| BERTScore | Semantic similarity to reference | Summarization (automatic) |
| ROUGE | N-gram overlap with reference | Summarization (academic) |
| BLEU | Translation quality | Translation |
| Faithfulness | Response grounded in retrieved context | RAG |
| Completeness | All parts of prompt addressed | Agent responses |
| Correctness | Factual accuracy | General |
| Harmfulness | Toxic content detection | Safety |
| Refusal | Model declines to answer | Guardrail testing |
13.4 Choosing the Right Evaluation Method
No single metric proves that a foundation model is good. Match the evaluation method to the task, compare outputs against a suitable reference or rubric, and use human review when quality requires judgment.
| Method / Metric | What It Helps Assess | Useful When | Important Limitation |
|---|---|---|---|
| Benchmark dataset | Performance on repeatable representative tasks | Comparing models or versions | May not represent real users or production data |
| Human evaluation | Usefulness, correctness, tone, relevance, nuanced quality | Subjective or high-impact tasks | Needs a clear rubric and reviewer consistency |
| ROUGE | Word or sequence overlap with reference text | Summarization comparisons | Overlap does not prove factual correctness |
| BLEU | n-gram precision against reference translations | Machine translation evaluation | May miss meaning and valid alternative wording |
| BERTScore | Semantic similarity using contextual token representations | Text generation where meaning matters beyond exact wording | Similarity is not proof that claims are true |
| LLM-as-a-judge | Scores or compares responses against a rubric | Scaling structured evaluation | Judge bias requires calibration and human spot checks |
13.5 Evaluate the Application, Not Only the Model
A model can produce fluent text while the complete application still fails. Evaluate the end-to-end system using measures that match its architecture.
| Application | What to Test | Useful Measures |
|---|---|---|
| RAG application | Whether retrieval finds the right passages and the answer is supported by them | Retrieval relevance/recall, faithfulness, answer relevance, citation correctness |
| AI agent | Whether it chooses the right tool, follows required steps, and completes the task safely | Task success, tool-call accuracy, completion rate, policy violations |
| Workflow | Whether handoffs, steps, and final outputs work reliably end to end | End-to-end success rate, error rate, latency, escalation rate |
Check two stages. Retrieval: did the system find the right evidence? Generation: did the model answer accurately using that evidence? Relevant retrieval does not guarantee a faithful answer.
13.6 Connect AI Performance to Business Objectives
Technical quality matters when it improves the intended user or business outcome. Choose the metric that measures the outcome named in the scenario.
| Metric | Meaning | Example |
|---|---|---|
| Task completion rate | Percentage of user tasks completed successfully | Support requests resolved without handoff |
| User satisfaction | User-reported experience or satisfaction | Post-interaction survey score |
| Cost per interaction | Average cost to serve one interaction | Model, retrieval, and infrastructure cost per request |
| Productivity | Time or effort saved on a process | Average handling time before and after deployment |
| Engagement | Whether users meaningfully use the application | Repeat usage or successful feature adoption |
Do not confuse model quality with business success. BLEU or ROUGE may compare generated text, but neither tells a business whether users completed tasks or whether interactions were cost-effective.
13.7 Applied Decision Examples
| Application situation | Appropriate approach | Reasoning and limitation |
|---|---|---|
| Comparing summaries against human-written reference summaries | ROUGE | Measures overlap with reference text; it does not independently establish factual accuracy |
| Knowledge assistant produces fluent but unsupported answers | Evaluate retrieval relevance and answer faithfulness | Check both the evidence retrieved and whether generated claims are supported by that evidence |
| Customer-service agent must resolve requests successfully | Task completion rate | Measure completed user goals; pair with cost per interaction and user satisfaction |
| Large model is too expensive for a narrower, repeatable task | Model distillation | A smaller student model can approximate useful behavior from a larger teacher; validate quality and safety after distillation |
Domain 3 — v1.2 Focus: Application Design and Evaluation
For vector storage questions, identify the data model and surrounding application requirement. The exam expects recognition of common AWS options, not database administration.
| Service | Scenario clue |
|---|---|
| Amazon OpenSearch Service | Search workloads combined with vector similarity retrieval |
| Amazon Aurora | Relational application data with vector retrieval |
| Amazon Neptune | Knowledge graphs and relationship-rich data |
| Amazon RDS for PostgreSQL | PostgreSQL application with vector extensions |
Evaluate the whole application, not only the foundation model. For RAG, check retrieval relevance, answer faithfulness to retrieved evidence, and citation correctness. For agents, check tool-call accuracy, task completion, policy violations, and safe escalation. For workflows, check end-to-end completion, latency, and failure rates.
Business alignment metrics include task completion rate, user satisfaction, and cost per interaction. ROUGE, BLEU, and BERTScore assess aspects of generated text; human evaluation and benchmark datasets provide complementary evidence. No single metric proves business success.
Responsible AI: Fairness, Bias, and Inclusivity
This chapter covers the responsible AI dimensions the exam tests: fairness, bias, inclusivity, and mitigation strategies.
14.1 Responsible AI Dimensions
| Dimension | Definition | Exam Trigger |
|---|---|---|
| Fairness | Equitable outcomes across demographic groups | Diverse data, protected groups |
| Explainability | Understanding why a model made a decision | Rationale, SHAP, feature importance |
| Transparency | Visibility into how AI systems make decisions | Model Cards, documentation |
| Privacy and Security | Data protection and access control | Encryption, IAM, PII |
| Safety | Preventing harmful output and misuse | Guardrails, content filters |
| Robustness | Resilience to unexpected inputs | Adversarial testing |
| Controllability | Human oversight and intervention | Human-in-the-loop review workflows |
14.2 Bias Detection and Mitigation
Bias enters through unrepresentative training data, historical bias in labels, proxy features, or sample selection. SageMaker Clarify computes pre-training and post-training bias metrics and generates reports on feature importance.
Bias from training data → fix the data. Bias from model design → use fairness-aware techniques. Diverse, balanced data is the foundation of fair AI.
14.3 Dataset Quality, Robustness, and Veracity
Responsible AI begins before model training. Review whether the dataset represents the people and situations affected by the system, whether labels are reliable, and whether important groups are missing or underrepresented.
| Check | What to Examine | Why It Matters |
|---|---|---|
| Diversity and inclusivity | Coverage across relevant demographic groups, languages, locations, and use cases | Reduces the risk that some groups are poorly represented |
| Balance | Whether classes or outcomes are disproportionately represented | Skewed data can distort model behavior and evaluation |
| Label quality | Accuracy, consistency, and agreement of labels | Incorrect labels can teach the model incorrect patterns |
| Curated sources | Origin, permission, relevance, and quality of collected data | Supports reliability, traceability, and responsible use |
| Robustness | Performance on unusual, noisy, shifted, or adversarial inputs | Helps the system behave reliably outside ideal test cases |
| Veracity | Whether generated statements are accurate and supported by evidence | Reduces misleading or fabricated outputs |
Bias and Variance: Recognize the Symptoms
- High bias / underfitting: the model is too simple or constrained to learn important patterns; it performs poorly on both training and unseen data.
- High variance / overfitting: the model fits training examples too closely and performs worse on new data.
- Demographic performance gaps: overall accuracy can hide poor results for a subgroup. Compare relevant metrics across groups rather than relying only on an aggregate score.
Use label-quality reviews, human audits, subgroup analysis, and ongoing monitoring. Amazon SageMaker Clarify can help examine bias and explainability; SageMaker Model Monitor can help detect data and model-quality changes over time. Select the tool based on the issue being investigated.
14.4 Legal, Trust, and Sustainability Risks
- Intellectual property: generated content may create infringement or licensing concerns; review data rights and intended use.
- Biased outputs: unfair recommendations can harm affected users and damage trust.
- Hallucinations: unsupported statements can create user, financial, legal, or safety risks.
- Loss of customer trust: opaque or harmful behavior can damage confidence in the organization.
- Environmental impact: consider energy use, model size, compute demand, and whether a smaller or more efficient model can meet the requirement.
A model with the highest benchmark score is not automatically the most responsible choice. Consider fairness, robustness, safety, licensing, user impact, and environmental cost alongside performance.
Explainability, Transparency, and Human Oversight
This chapter covers explainability techniques, transparency documentation, and human-in-the-loop workflows.
15.1 Explainability Techniques
SHAP values show how much each feature contributed to a prediction. Feature importance ranks inputs by influence. Counterfactuals show what would need to change for a different outcome. SageMaker Clarify provides these capabilities.
15.2 Transparency Documentation
| Tool | What It Documents | When to Use |
|---|---|---|
| SageMaker Model Cards | Purpose, performance, limitations, risks | Custom model governance |
| AWS AI Service Cards | AWS service capabilities and limitations | Understanding managed services |
| SageMaker Model Registry | Version tracking and approval status | Model lifecycle management |
15.3 Human Oversight
Human-in-the-loop (HITL) workflows route low-confidence or flagged predictions to human reviewers. The exam focus is the purpose of human review, the escalation trigger, and how feedback improves oversight—not implementation of a specific A2I service.
High-stakes decisions about people (hiring, lending, healthcare) require appropriate human review, clear escalation paths, and documented accountability.
15.4 Transparency vs. Explainability
These ideas are related but not identical. Transparency makes relevant information about a model and its use available—for example, its purpose, training-data characteristics, intended use, limitations, and evaluation results. Explainability helps a person understand why a particular output or prediction occurred.
| Question | Transparency | Explainability |
|---|---|---|
| What is disclosed? | Model purpose, data, limits, intended use, documentation | Factors or evidence behind a specific output |
| Typical artifact | Model Card, service documentation, licensing information | Feature attribution, explanation, counterfactual |
| Who benefits? | Users, reviewers, auditors, developers, decision-makers | People affected by or reviewing a decision |
| Key limitation | Disclosure alone does not make a model understandable | An explanation may be approximate and does not prove the model is correct |
15.5 Model Evaluation and Open-Source Considerations
Amazon Bedrock Model Evaluations can help assess foundation-model performance using supported evaluation approaches and metrics. When comparing open-weight or open-source models, review more than model availability: check the license, permitted uses, training-data information where available, evaluation evidence, known limitations, and the organization's ability to operate and secure the model.
More disclosure can help users and reviewers understand a system, but exposing sensitive implementation details, personal data, or abuse-enabling information can create risk. Aim for meaningful, audience-appropriate transparency while protecting security and privacy. Do not assume that maximum disclosure is always safest.
15.6 Human-Centered Explanations and Feedback
- Tell users when AI is being used and what role it plays in the decision.
- Explain important outcomes in language appropriate to the audience.
- Provide a way to ask questions, correct information, or challenge an outcome.
- Route uncertain, harmful, or high-impact cases to qualified human review.
- Use feedback and review findings to improve the system and its documentation.
A Model Card improves transparency; it does not automatically explain an individual prediction. SHAP or feature attribution can help explain a prediction; it does not replace documentation, fairness testing, or human accountability.
15.7 Transparency, Interpretability, and Safety Trade-offs
Transparency and explainability improve trust and review, but disclosure must be balanced with security, privacy, intellectual-property protection, and misuse risk. A highly interpretable model may be less capable for a particular task; a more capable complex model may require stronger documentation, evaluation, monitoring, and human oversight. Select the level of explanation that fits the decision’s impact and the audience.
| Situation | Useful approach | Important limitation |
|---|---|---|
| Need to document intended use and known limits | Model Card and service documentation | Does not explain every individual prediction |
| Need to understand feature contribution | Feature importance or SHAP-style attribution | Attributions can be approximate and do not prove causation |
| Need to explain a high-impact outcome to a person | Audience-appropriate explanation plus human review and appeal path | Must preserve privacy and avoid overstating certainty |
Domain 4 — v1.2 Focus: Responsible Selection and Human-Centred Explanations
Responsible model selection is a multi-factor decision. Consider fairness across relevant groups, robustness to unusual inputs, safety, veracity, licensing and intellectual-property risks, user impact, and environmental cost. A larger or more accurate model is not automatically the most responsible choice.
Transparency helps users understand what a system is intended to do, its limitations, and how it was evaluated. Explainability helps make a particular output or decision understandable. Model Cards can document intended use, evaluation, limitations, and risks; user-feedback mechanisms and clear explanations support human-centred design.
When overall performance hides unequal outcomes, examine subgroup results and label quality. When users need to understand a decision, prioritize appropriate explanations and transparent documentation.
Securing AI Workloads: IAM, Encryption, and Private Connectivity
This chapter covers IAM, KMS, PrivateLink, and the shared responsibility model for AI services.
16.1 Access Control with IAM
AWS Identity and Access Management (IAM) controls who can access which AWS resources. IAM policies can restrict access to specific Bedrock models, SageMaker endpoints, or datasets.
16.2 Encryption with KMS
AWS Key Management Service (KMS) manages encryption keys. For Bedrock, you can use customer-managed keys to encrypt custom model artifacts. The API parameter is modelKmsKeyArn.
16.3 Private Connectivity with PrivateLink
AWS PrivateLink creates a private connection between your VPC and Amazon Bedrock. Traffic stays on the AWS private network and never traverses the public internet.
flowchart LR
A[VPC] --> B[VPC Endpoint]
B --> C[AWS PrivateLink]
C --> D[Amazon Bedrock]
A --> E[No Internet Gateway]
style A fill:#0b2545,color:#fff
style C fill:#1d4e89,color:#fff
style D fill:#2d6a9f,color:#fff
style E fill:#c53030,color:#fff
Figure 16.1 — Private connectivity to Bedrock via AWS PrivateLink. Traffic never traverses the public internet.
16.4 Shared Responsibility for AI
| Layer | AWS Responsibility | Your Responsibility |
|---|---|---|
| Infrastructure | Data centers, hardware, networking | None |
| Managed service endpoints | API availability, patching | Configure VPC endpoints |
| Foundation models (Bedrock) | Model hosting, base model integrity | Model access policies, guardrails |
| Training data | Storage durability | Data classification, encryption, access control |
| Application logic | None | Input validation, output filtering, logging |
Amazon Bedrock does not share your inputs or outputs with third-party model providers, and it does not use them to train base models unless you explicitly opt in.
16.5 Securing Agent Identity and Tool Access
AI agents may call tools, access APIs, and interact with business systems. Treat an agent's identity and permissions as security boundaries: grant only the access needed for its task, scope credentials to the intended resources, and monitor its actions.
| Concept | Purpose | Exam Clue |
|---|---|---|
| IAM role and policy | Defines which AWS actions a workload identity may perform | Least privilege for AWS resources |
| Amazon Bedrock AgentCore Identity | Supports identity and authentication for agents accessing tools and resources | Secure agent identity and credential handling |
| Policy in AgentCore | Provides policy-based controls over agent actions and tool use | Restrict what an agent is permitted to do |
| Amazon Bedrock Guardrails | Applies configured safeguards to model inputs and outputs | Filter harmful content or sensitive information |
16.6 Data Lineage and Secure Data Engineering
- Data lineage: record where data originated, how it was transformed, and where it was used.
- Data cataloging: maintain searchable descriptions, ownership, classification, and context for datasets.
- Data integrity: protect data from unauthorized or accidental alteration.
- Access control: restrict datasets to approved users, roles, services, and workloads.
- Privacy-enhancing practices: minimize collected data, mask or redact sensitive fields where appropriate, and consider de-identification or synthetic data when suitable.
Amazon Macie helps discover and assess sensitive data in Amazon S3. Amazon SageMaker Model Cards can document model details and intended use. Use lineage and catalog information to understand the origin and handling of data used by an AI solution.
16.7 Protect Inputs, Outputs, and AI Interaction Logs
- Validate and constrain user inputs; do not treat retrieved documents or tool output as trusted instructions.
- Defend against prompt injection and unauthorized tool requests with layered controls.
- Filter and validate model outputs before displaying them or using them to trigger actions.
- Use grounding, source checks, and confidence thresholds to identify answers that need review.
- Log relevant prompts, model versions, tool calls, decisions, and outcomes while applying data minimization, retention limits, and access controls.
- Monitor for toxicity, sensitive-data leakage, anomalous activity, and policy violations.
Encryption protects data at rest or in transit; it does not prevent a model from generating harmful or incorrect text. Guardrails and output validation address different risks. Logging supports investigation and audit, but logs themselves must be protected and governed.
16.8 Grounding, Hallucination Detection, and Output Validation
Grounding connects a generated response to trusted evidence, such as approved documents retrieved through RAG. It can reduce unsupported claims, but it does not guarantee correctness: retrieval can return irrelevant or outdated material, and a model can still misinterpret evidence.
| Control | What it checks | What it cannot guarantee |
|---|---|---|
| Retrieval relevance | Whether retrieved passages match the user’s request | That the model will use them correctly |
| Grounding and source citations | Whether claims are supported by retrieved sources | That every cited source is complete or current |
| Output validation | Required fields, format, policy rules, and factual constraints | Full factual truth unless the validation has reliable evidence |
| Confidence scoring and thresholds | Whether uncertain cases should be escalated | A universally calibrated probability of correctness |
| Human review | High-impact, ambiguous, or flagged outputs | Freedom from human error; reviewers need guidance and evidence |
For high-impact workflows, combine retrieval evaluation, citation checks, deterministic validation where possible, monitoring, and human escalation. Treat user-supplied documents and tool outputs as data, not as trusted instructions; this also reduces prompt-injection risk.
Data Privacy, Compliance, and Governance for AI
This chapter covers data residency, lifecycle management, compliance services, and governance frameworks.
17.1 Data Privacy and Residency
Data residency is the governance principle that data must be stored and processed within a specific geographic boundary. Data residency requirements depend on the applicable law, contract, sector, and organization policy. GDPR does not impose a blanket requirement to store all personal data in the EU, and HIPAA does not create a general U.S.-only data-residency rule. Organizations must assess cross-border transfer rules, safeguards, contractual commitments, and the locations where data is stored and processed.
17.2 Compliance Services
| Service | Purpose | When to Use |
|---|---|---|
| AWS Config | Resource configuration tracking and compliance | Policy adherence monitoring |
| AWS CloudTrail | API activity logging | Audit trails |
| AWS CloudTrail Lake | Queryable environment for collecting and analyzing audit events | Search and analyze activity over time; apply appropriate retention and access controls |
| AWS Artifact | AWS compliance reports and certifications | Demonstrating AWS compliance |
| Amazon Macie | Sensitive data discovery in S3 | PII detection |
| Amazon Inspector | Vulnerability scanning | EC2, container, Lambda security |
17.3 Governance Frameworks
AI governance frameworks define policies and guidelines for data, transparency, responsible AI, and compliance. AWS CAF for AI provides guidance across six perspectives: Business, People, Governance, Platform, Security, and Operations.
Data lifecycle management covers creation, updates, retention, and scheduled deletion. It is a governance component, not a security service.
17.4 Governance Operations and Review Cadence
Governance is an ongoing process, not a one-time approval. Establish ownership and repeatable reviews throughout the AI lifecycle.
- Set policies: define acceptable use, data classification, model approval, human oversight, and incident escalation.
- Assign owners: identify who is accountable for the data, model, application, security controls, and user impact.
- Review before release: assess intended use, data rights, privacy, fairness, security, evaluation results, and limitations.
- Monitor after release: review incidents, drift, user feedback, access logs, performance, and policy violations.
- Reassess periodically: repeat reviews when the model, data, use case, regulation, or risk level changes.
- Train teams: ensure developers, reviewers, operators, and business owners understand their responsibilities.
17.5 Generative AI Security Scoping Matrix
The Generative AI Security Scoping Matrix is a way to reason about security according to the architecture and level of control in a generative AI solution. Identify where the model runs, who controls the model and infrastructure, what data is supplied, which external tools are connected, and where the application boundary lies. Security responsibilities and attack surfaces change with the deployment pattern.
| Area | Questions to Ask | Evidence or Control |
|---|---|---|
| Data | Where did data come from? Is its use permitted? How long is it retained? | Data catalog, lineage, classification, retention policy |
| Model | Who provides it? What are its limitations and intended uses? | Model Card, evaluation results, license review |
| Application | Can users inject instructions or extract sensitive data? | Threat modeling, guardrails, input/output validation |
| Identity and tools | What can the agent access or execute? | Least-privilege IAM, AgentCore identity and policy controls |
| Operations | Can activity be investigated and controls rechecked? | CloudTrail logs, monitoring, incident and review procedures |
17.6 Governance and Compliance Service Selection
| Service | Primary Use | Scenario Clue |
|---|---|---|
| AWS Config | Records and evaluates resource configurations | Check whether resources comply with configuration rules |
| Amazon Inspector | Finds software vulnerabilities and exposure risks in supported workloads | Assess vulnerabilities in compute workloads |
| AWS CloudTrail | Records AWS API activity | Investigate who performed an AWS action and when |
| AWS Artifact | Provides access to AWS compliance reports and agreements | Obtain AWS audit and compliance documentation |
| AWS Trusted Advisor | Provides recommendations across areas such as security, fault tolerance, performance, and service limits | Identify opportunities to improve AWS environment posture |
Do not mix up the services: CloudTrail records API activity; Config tracks resource configuration; Artifact provides AWS compliance documents; Inspector identifies vulnerabilities; Trusted Advisor provides recommendations.
Data Residency: A Practical Decision Framework
Do not assume that a regulation automatically requires every copy of data to remain in one country. Identify the data subjects and data types, the organization’s role, the applicable legal and contractual requirements, cross-border transfer mechanisms, and the AWS Regions used by each service. Also check where backups, logs, support data, and model-related processing occur. A residency decision is a documented legal, contractual, and architecture assessment—not simply a Region selection.
Data residency is about where data is stored or processed. Data sovereignty concerns the laws and governmental authority that may apply to data. Data localization is a specific requirement to keep certain data within a jurisdiction. These ideas overlap but are not identical.
Domain 5 — v1.2 Focus: Security and Governance Service Map
The current AWS exam guide groups this domain into two tasks: securing AI systems, and recognizing governance and compliance requirements. The examples below follow the official task statements; the official in-scope service list is broader and non-exhaustive.
| Capability | What it protects or controls | Practical interpretation |
|---|---|---|
| IAM roles, policies, and permissions | Identity and authorization | Grant each user, workload, and agent only the actions and resources it needs |
| Encryption | Data at rest and in transit | Protect stored datasets, model artifacts, and network communication; manage keys appropriately |
| Amazon Macie | Sensitive data in Amazon S3 | Discover and help assess sensitive information stored in S3 |
| AWS PrivateLink | Private service connectivity | Reduce exposure to the public internet when connecting to supported AWS services |
| AWS shared responsibility model | Security ownership | Separate AWS responsibilities for the cloud from customer responsibilities in the cloud; the split changes by service type |
| Amazon Bedrock AgentCore Identity | Agent identity and authentication | Establish and manage identity and credentials for agents accessing tools and resources |
| Policy in AgentCore | Agent action authorization | Constrain which actions an agent is allowed to perform |
| Amazon Bedrock Guardrails | Model inputs and outputs | Apply configured safeguards such as content filtering and sensitive-information controls |
| AWS service | Primary purpose | Choose it when the requirement is… |
|---|---|---|
| AWS Config | Resource configuration tracking and compliance evaluation | Check whether AWS resources meet defined configuration rules |
| Amazon Inspector | Vulnerability management | Identify software vulnerabilities and exposure risks in supported workloads |
| AWS Artifact | Compliance reports and agreements | Obtain AWS audit and compliance documentation |
| AWS CloudTrail | Account activity and API event history | Investigate who performed an AWS action and when |
| AWS Trusted Advisor | Recommendations across AWS environment categories | Find opportunities to improve security, reliability, performance, and cost posture |
Related in-scope capabilities: Amazon CloudWatch supports monitoring and operational logs; AWS Secrets Manager helps manage secrets such as credentials and API keys; Amazon S3 is a common storage layer whose access, encryption, retention, and logging must be governed. These services complement the task-statement examples but do not replace them.
Data provenance: data lineage records where data originated and how it changed; data cataloging makes datasets discoverable and understandable; source citation helps users trace generated claims to supporting material. Amazon SageMaker Model Cards can document model details, intended use, evaluation, and limitations.
Encryption protects confidentiality of data at rest or in transit. IAM and AgentCore Policy control who or what can act. Guardrails and output validation address model-content risks. CloudTrail supports investigation. These controls solve different problems and should not be treated as interchangeable.
Older study materials may contain services that are not in the current AIF-C01 in-scope service list, including AWS Audit Manager, Amazon Kendra, Amazon Fraud Detector, and Amazon Augmented AI (A2I). Focus first on the current Domain 5 task statements and the official in-scope list. AWS states that the in-scope list is non-exhaustive and may change.
AIF-C01 Study Guide: Frequently Asked Questions
Is this AWS AIF-C01 study guide free?
Yes. This study guide is available free on CloudExamPro and is designed to help candidates understand the AWS Certified AI Practitioner (AIF-C01) exam objectives.
What topics does the guide cover?
It covers all five AIF-C01 domains: AI and machine learning (20%), generative AI (24%), foundation model applications (28%), responsible AI (14%), and security, compliance, and governance (14%).
Is the study guide aligned with the current AWS exam guide?
The guide was reviewed against AWS AIF-C01 Exam Guide version 1.2, published on 29 October 2026. Always check the official AWS certification page for the latest exam information.
Does this study guide include a full practice exam?
No. This page focuses on explanations, reference tables, decision frameworks, and revision planning. A separate free 25-question AIF-C01 diagnostic is available to check your readiness.
How should I use the guide before the exam?
Start with the exam overview, study the five domains, and use the 7-day or 14-day plan that matches your available time. Finish with the revision sheet and the separate diagnostic exam.
Related Free AIF-C01 Study Resources
- AIF-C01 Cheat Sheet — a concise revision reference for key AI, ML, generative AI, and AWS service concepts.
- AIF-C01 14-Day Study Plan — a structured schedule for covering all five exam domains.
- Amazon Bedrock vs. SageMaker AI — compare two important AWS services using exam-focused scenarios.
- AI vs. ML vs. Deep Learning vs. Generative AI — review the foundational terminology and distinctions.
Ready to Check Your AIF-C01 Readiness?
This guide is designed for learning and reference. After reviewing the five domains, use the separate free 25-question AIF-C01 diagnostic to check your current readiness.
Start the Free 25-Question Diagnostic →Final-Day Revision Sheet
A compact, high-value summary of the entire blueprint. Review this the day before your exam.
Domain 1 — AI & ML (20%)
- AI includes ML; deep learning is a subset of ML. GenAI is a content-generation capability commonly powered by deep-learning FMs; agentic AI is a tool-using system pattern.
- Supervised (labeled) / Unsupervised (unlabeled) / Reinforcement (rewards)
- Classification (category) vs. Regression (number)
- Accuracy misleading on imbalanced data → use F1
- Train-good/test-bad = overfitting → regularization
- Business goal first → compliance before data collection
Domain 2 — GenAI (24%)
- Tokens, embeddings, context window
- Lower temperature = more focused sampling; it does not guarantee determinism or factuality
- Zero-shot, few-shot, chain-of-thought, ReAct
- RAG reduces hallucinations by grounding in sources
- Prompt injection = override instructions; jailbreak = bypass safety
Domain 3 — FM Applications (28%)
- Bedrock = managed FMs; SageMaker AI = custom ML
- Bedrock Knowledge Bases = managed RAG
- Bedrock Agents = multi-step API orchestration
- Bedrock Guardrails = safety policies
- Prompt engineering → RAG → fine-tuning → continued pre-training (increasing effort)
- On-Demand (unpredictable) vs. Provisioned Throughput (steady)
Domain 4 — Responsible AI (14%)
- Fairness, explainability, transparency, privacy, safety, robustness, controllability
- SageMaker Clarify = bias detection + SHAP explainability
- Model Cards = standardized documentation
- Human-in-the-loop review = human oversight for low-confidence or high-impact outputs
Domain 5 — Security (14%)
- IAM = access control; KMS = encryption; PrivateLink = private network
- Amazon Bedrock customer prompts and outputs are not used to train base FMs by default; data handling and retention vary by feature, model, and Region
- Macie = sensitive-data discovery in S3; Inspector = vulnerability findings
- CloudTrail = API activity logs; CloudTrail Lake = queryable event analysis with configured retention and access controls
- Artifact = AWS compliance reports; Config = resource configuration/compliance evaluation
- Data residency = where data lives; lifecycle = how long
Read every question for the decisive requirement. The exam rewards the candidate who identifies the constraint, not the one who memorizes the most services.
Practice-Based Self-Assessment Checklist
Mark each item you can explain confidently. Unchecked items are your review priorities.
Domain 1
Domain 2
Domain 3
Domain 4
Domain 5
Acronym and Glossary Reference
Searchable or clearly organized terminology for quick reference.
| Term/Acronym | Expansion | Definition |
|---|---|---|
| AI | Artificial Intelligence | Broad field of simulating human intelligence |
| HITL | Human-in-the-loop | Human review of low-confidence or high-impact AI outputs |
| BERTScore | Bidirectional Encoder Representations from Transformers Score | Semantic similarity metric |
| BLEU | Bilingual Evaluation Understudy | Translation quality metric |
| CTR | Click-Through Rate | Engagement metric |
| FM | Foundation Model | Large pre-trained adaptable model |
| IAM | Identity and Access Management | AWS access control service |
| KMS | Key Management Service | AWS encryption key management |
| LLM | Large Language Model | Text-specialized foundation model |
| MCP | Model Context Protocol | Standard for connecting AI to external tools |
| ML | Machine Learning | Subset of AI that learns from data |
| PII | Personally Identifiable Information | Sensitive personal data |
| RAG | Retrieval Augmented Generation | Grounding LLM responses in external data |
| RLHF | Reinforcement Learning from Human Feedback | Aligning models with human preferences |
| ROUGE | Recall-Oriented Understudy for Gisting Evaluation | Summarization quality metric |
| SHAP | SHapley Additive exPlanations | Feature importance for explainability |
Exam-Day Preparation Checklist
Practical preparation advice, with logistics verified against current official guidance where applicable.
Before the Exam
During the Exam
After the Exam
References
All references are official AWS documentation, AWS Builder Center articles, or AWS whitepapers. Time-sensitive details were verified against the current documentation as of this guide's publication.
- AWS Certified AI Practitioner (AIF-C01) Exam Guide — Official exam domains, task statements, and weights.
- AIF-C01 Domain 1 Objectives — AI/ML fundamentals, practical use cases, and lifecycle.
- AIF-C01 Domain 2 Objectives — GenAI concepts, agentic AI, and AWS infrastructure.
- AIF-C01 Domain 3 Objectives — Foundation model applications, prompting, customization, and evaluation.
- AIF-C01 Exam Guide Revision History — Version changes and publication dates.
- AIF-C01 In-Scope AWS Services — Official list of in-scope services.
- Amazon Bedrock User Guide — Service overview, models, capabilities.
- Amazon Bedrock Knowledge Bases — Managed RAG implementation.
- Amazon Bedrock Guardrails — Content filters, denied topics, PII filters, contextual grounding.
- Amazon Bedrock Agents — Multi-step orchestration and API calls.
- Amazon Bedrock Model Evaluation — Automatic and human evaluation metrics.
- Amazon Bedrock Data Protection — Data-handling guarantees and third-party sharing policy.
- AWS PrivateLink for Amazon Bedrock — Private VPC connectivity.
- SageMaker Model Cards — Standardized model documentation.
- SageMaker Clarify — Bias detection and explainability.
- SageMaker Model Monitor — Production drift detection.
- SageMaker Automatic Model Tuning — Hyperparameter optimization.
- SageMaker Canvas — No-code ML.
- SageMaker JumpStart — Pre-trained models and templates.
- SageMaker Data Wrangler — Visual data preparation.
- SageMaker Ground Truth — Data labeling.
- AWS Trusted Advisor — Recommendations for security, reliability, performance, and cost.
- AWS Config — Resource configuration and compliance.
- AWS CloudTrail — API activity logging.
- AWS CloudTrail Lake — Queryable event data store for audit analysis.
- Amazon Macie — Sensitive data discovery.
- Amazon Inspector — Vulnerability scanning.
- AWS Artifact — AWS compliance reports.
- Amazon Q Business — Enterprise AI assistant.
- Amazon Q Developer — Developer productivity.
- Amazon Nova — Multimodal foundation models.
- Amazon Nova Canvas — Image generation.
- Amazon Nova Reel — Video generation.
- AWS — What is RAG? — RAG explanation.
- AWS — What is Prompt Engineering? — Prompting techniques.
- AWS — What are Embeddings? — Embeddings explanation.
- AWS — What is Overfitting? — Overfitting explanation.
- AWS Responsible AI — Responsible AI dimensions and resources.
- AWS Well-Architected Machine Learning Lens — ML lifecycle phases.
- AWS CAF for AI — AI governance framework.
- AWS Data Residency — Data residency compliance.
- AWS Prescriptive Guidance — Common LLM Attacks — Prompt injection and jailbreak.
- Amazon Bedrock Prompt Management — Store and reuse prompts.
- Amazon Bedrock Provisioned Throughput — Dedicated capacity.
- Amazon Bedrock Batch Inference — Bulk generation.
All references were verified against current official AWS documentation at the time of writing. Time-sensitive details (service features, pricing models, metric names) should be re-verified against the latest AWS documentation before relying on them in production.