Amazon Web Services Beginner

AIF-C01

AWS Certified AI Practitioner

The AWS Certified AI Practitioner (AIF-C01) validates foundational knowledge of artificial intelligence (AI), machine learning (ML), and generative AI concepts and use cases on the AWS platform. Launched in August 2024, this is AWS's newest foundational-level certification designed for anyone in a technical or non-technical role who wants to demonstrate knowledge of AI/ML technologies.

The exam covers five domains: Fundamentals of AI and ML (20%), Fundamentals of Generative AI (24%), Applications of Foundation Models (28%), Guidelines for Responsible AI (14%), and Security, Compliance, and Governance for AI Solutions (14%). Candidates will need to understand core AI/ML concepts, generative AI architectures, prompt engineering, Amazon Bedrock, Amazon SageMaker, and responsible AI practices.

This certification is ideal for business analysts, product managers, marketing professionals, project managers, and early-career IT professionals who work with AI/ML stakeholders and need to understand AI capabilities, limitations, and responsible use within an AWS environment.

Updated Aug 2024 AI & Machine Learning

65

Domande

6

Simulazioni

70%

Punteggio Minimo

70

Visualizzazioni

0

Tentativi Totali

0%

Punteggio Medio

0%

Tasso di Superamento

0

Discussioni

Simulazioni Flash Cards Argomenti dell'Esame Scheda di Riferimento

€5.00

AIF-C01 Practice Exam 1

Complete 65-question practice exam covering all five AIF-C01 domains with detailed explanations for every option.

65 Q 90 minuti 70%

Test Drive

€5.00

AIF-C01 Practice Exam 2

Comprehensive 65-question practice exam covering all five AIF-C01 domains including AI/ML fundamentals, generative AI, foundation model applications, responsible AI, and security governance.

65 Q 90 minuti 70%

Test Drive

€5.00

AIF-C01 Practice Exam 3

Comprehensive 65-question practice exam covering all five AIF-C01 domains including AI/ML fundamentals, generative AI, foundation model applications, responsible AI, and security governance.

65 Q 90 minuti 70%

Test Drive

€5.00

AIF-C01 Practice Exam 4

Comprehensive 65-question practice exam covering all five AIF-C01 domains including AI/ML fundamentals, generative AI, foundation model applications, responsible AI, and security governance.

65 Q 90 minuti 70%

Test Drive

€5.00

AIF-C01 Practice Exam 5

Comprehensive 65-question practice exam covering all five AIF-C01 domains including AI/ML fundamentals, generative AI, foundation model applications, responsible AI, and security governance.

65 Q 90 minuti 70%

Test Drive

€5.00

AIF-C01 Practice Exam 6

Comprehensive 65-question practice exam covering all five AIF-C01 domains including AI/ML fundamentals, generative AI, foundation model applications, responsible AI, and security governance.

65 Q 90 minuti 70%

Test Drive

Sblocca Tutti i Contenuti per AIF-C01

6 Simulazione/i + Flash Cards — accesso per 3 mesi

€39.99 €26.99 Risparmia il 30%

o incluso con abbonamento Mensile / Pacchetto Contenuti

AIF-C01 Cheat Sheet

Guida di riferimento rapido - 6 sezioni

Accesso Gratuito

AWS Certified AI Practitioner (AIF-C01)

The AIF-C01 validates your foundational understanding of artificial intelligence, machine learning, and generative AI concepts, as well as the AWS services used to build and deploy AI/ML solutions. This exam is intended for individuals in any role who want to demonstrate broad AI literacy and knowledge of AWS AI/ML services. No hands-on coding or model building experience is required, but familiarity with AI concepts and AWS cloud fundamentals is expected.

Exam Details

Exam Code	AIF-C01
Duration	120 minutes
Number of Questions	85 questions (65 scored + 20 unscored)
Question Types	Multiple choice and multiple response
Passing Score	700 / 1000 (scaled scoring)
Exam Fee	$150 USD
Delivery Method	Pearson VUE testing center or online proctored
Certification Validity	3 years
Level	Foundational
Recommended Experience	6+ months exposure to AI/ML concepts & AWS Cloud

Domain Breakdown

Domain 1

20%

Fundamentals of AI & ML

Core AI/ML concepts, types of learning, model lifecycle, data types, bias, and responsible AI basics.

Domain 2

24%

Fundamentals of Generative AI

Foundation models, LLMs, transformers, prompt engineering, RAG, fine-tuning, and generative AI use cases.

Domain 3

28%

Applications of Foundation Models

Amazon Bedrock, model selection, prompt engineering techniques, agents, RAG implementation, and model customization.

Domain 4

14%

Guidelines for Responsible AI

Fairness, transparency, privacy, safety, governance, toxicity detection, and responsible AI practices on AWS.

Domain 5

14%

Security, Compliance & Governance for AI Solutions

IAM for AI services, data encryption, VPC configurations, compliance frameworks, and securing AI workloads on AWS.

Key Exam Tips

1. Amazon Bedrock is the most heavily tested service — understand its features, model providers, and customization options thoroughly.
2. Generative AI concepts make up 52% of the exam (Domains 2 + 3). Focus heavily on foundation models, prompt engineering, and RAG.
3. Know the difference between traditional ML (supervised, unsupervised, reinforcement) and generative AI approaches.
4. Understand when to use pre-trained AI services (Comprehend, Rekognition, Textract) vs. foundation models via Bedrock vs. custom training via SageMaker.
5. Responsible AI questions appear across all domains — know bias types, mitigation strategies, and AWS guardrails features.
6. This is a foundational-level exam. Questions test conceptual understanding, not hands-on implementation details.

Scoring Note

AWS uses a scaled scoring model from 100 to 1000. The 20 unscored questions are being evaluated for future exams. You will not know which questions are unscored, so answer every question. There is no penalty for guessing. At $150 USD, this is one of the most affordable AWS certifications.

Domain 1: Fundamentals of AI & ML (20%)

This domain tests your understanding of core artificial intelligence and machine learning concepts, including the types of learning, the ML pipeline, data fundamentals, and basic model evaluation. It provides the conceptual foundation for the rest of the exam.

AI vs. ML vs. Deep Learning vs. Generative AI

Concept	Definition	Scope
Artificial Intelligence (AI)	Broad field of creating systems that can perform tasks requiring human-like intelligence	Broadest — includes all subfields
Machine Learning (ML)	Subset of AI where systems learn from data without being explicitly programmed	Subset of AI
Deep Learning (DL)	Subset of ML using neural networks with multiple layers to learn complex patterns	Subset of ML
Generative AI	Subset of DL that creates new content (text, images, code, audio) from learned patterns	Subset of DL

Tip: Think of these as nested circles — AI is the outermost, Generative AI is the innermost. Every generative AI system is also AI, but not every AI system is generative.

Types of Machine Learning

Supervised Learning

Learns from labeled data (input → known output). The algorithm maps inputs to correct outputs using training examples.

• Classification — Predict a category (spam/not spam, fraud/legitimate)
• Regression — Predict a continuous value (house price, temperature)
• Examples: Linear regression, logistic regression, decision trees, random forests, XGBoost

Unsupervised Learning

Learns from unlabeled data. The algorithm discovers hidden patterns and structures without predefined outcomes.

• Clustering — Group similar items (customer segments, anomaly detection)
• Dimensionality Reduction — Reduce features while preserving information (PCA)
• Association — Find relationships (market basket analysis)

Reinforcement Learning (RL)

An agent learns by interacting with an environment, receiving rewards or penalties for actions taken.

• Agent — The learner/decision-maker
• Environment — The world the agent interacts with
• Reward Signal — Feedback that guides learning
• Examples: Game playing, robotics, AWS DeepRacer

Self-Supervised Learning

A form of unsupervised learning where the system generates its own labels from the data. This is how most foundation models (like LLMs) are pre-trained.

• Masked Language Modeling — Predict hidden words (BERT)
• Next Token Prediction — Predict the next word (GPT-style)
• Key to understanding how foundation models learn

The ML Pipeline

Understanding the end-to-end ML workflow is critical. Questions often ask which step addresses a specific problem.

Step	Activities	Key Considerations
1. Problem Definition	Define business problem, determine if ML is appropriate, set success metrics	Not all problems need ML — rule-based systems may suffice
2. Data Collection	Gather training data from various sources (databases, APIs, files, streams)	Data quality > data quantity; representative data avoids bias
3. Data Preparation	Clean, transform, encode, normalize, split into train/validation/test sets	80% of ML effort typically goes here
4. Model Training	Select algorithm, train on training data, tune hyperparameters	Choose algorithm based on data type and problem type
5. Model Evaluation	Evaluate on test data using metrics (accuracy, precision, recall, F1, AUC)	Use validation set for tuning, test set for final evaluation only
6. Deployment & Monitoring	Deploy model, monitor performance, retrain as needed	Model drift requires retraining; A/B testing validates changes

Model Evaluation Metrics

Metric	Formula / Concept	When to Use
Accuracy	Correct predictions / Total predictions	Balanced datasets only; misleading with imbalanced data
Precision	True Positives / (True Positives + False Positives)	When false positives are costly (spam filtering)
Recall (Sensitivity)	True Positives / (True Positives + False Negatives)	When false negatives are costly (disease detection, fraud)
F1 Score	Harmonic mean of Precision and Recall	Balance between precision and recall; imbalanced datasets
AUC-ROC	Area under the Receiver Operating Characteristic curve	Binary classification; threshold-independent evaluation
RMSE / MAE	Root Mean Squared Error / Mean Absolute Error	Regression problems; measures prediction error magnitude

Overfitting vs. Underfitting

Problem	Symptom	Cause	Solutions
Overfitting	High training accuracy, low test accuracy	Model too complex, memorizes training data	More data, regularization, dropout, cross-validation, simpler model
Underfitting	Low training and test accuracy	Model too simple, cannot capture patterns	More features, more complex model, longer training, less regularization

Tip: The exam often describes a scenario and asks you to identify whether overfitting or underfitting is occurring. Look for the gap between training and validation performance.

Data Concepts

Labeled vs. Unlabeled Data

• Labeled — Data with known outcomes (used in supervised learning)
• Unlabeled — Data without outcomes (used in unsupervised learning)
• Labeling is expensive; Amazon SageMaker Ground Truth helps automate it

Structured vs. Unstructured Data

• Structured — Tabular data in rows/columns (databases, CSV)
• Unstructured — Text, images, audio, video
• Semi-structured — JSON, XML, logs

Exam Tip

Domain 1 questions are conceptual. You will not be asked to write code or configure services. Focus on understanding when to use each type of learning, which metrics suit which problem, and how the ML pipeline stages connect. If a question describes a scenario with labeled data, the answer likely involves supervised learning.

Domain 2: Fundamentals of Generative AI (24%)

This domain covers the core concepts behind generative AI, including foundation models, large language models, transformers, prompt engineering, and the techniques used to customize and improve model outputs. This is one of the two highest-weighted domains.

Foundation Models (FMs)

Foundation models are large AI models pre-trained on massive datasets that can be adapted to a wide range of downstream tasks. They are the backbone of modern generative AI.

Characteristic	Description
Pre-trained	Trained on vast, diverse datasets (internet text, images, code) using self-supervised learning
General-purpose	Can handle many tasks without task-specific training (text generation, summarization, Q&A)
Adaptable	Can be fine-tuned or used with prompts to specialize for specific domains or tasks
Emergent Abilities	Capabilities that appear at scale (reasoning, chain-of-thought, in-context learning)
Multi-modal	Some FMs can process and generate multiple data types (text + images + code)

Large Language Models (LLMs)

LLMs are foundation models specifically trained on text data. They use the transformer architecture and predict the next token in a sequence.

Concept	Explanation
Tokens	The basic units of text processed by LLMs. A token is roughly 3/4 of a word. Models have token limits (context windows).
Context Window	The maximum number of tokens a model can process in a single request (input + output combined)
Temperature	Controls randomness. Low (0) = deterministic/factual. High (1) = creative/diverse. Adjusts probability distribution.
Top-p (Nucleus Sampling)	Limits token selection to the smallest set of tokens whose cumulative probability exceeds p. Lower = more focused.
Hallucination	When a model generates plausible-sounding but factually incorrect or fabricated information. A key risk of LLMs.

Transformer Architecture

The transformer is the neural network architecture behind all modern LLMs. Understanding its key mechanism — attention — is tested on the exam.

Self-Attention Mechanism

Allows the model to weigh the importance of different words relative to each other in a sequence, regardless of their distance. This is what makes transformers superior to older RNN/LSTM architectures.

Encoder vs. Decoder

• Encoder-only — Understands input (BERT: classification, NER)
• Decoder-only — Generates output (GPT: text generation)
• Encoder-Decoder — Input to output (T5: translation, summarization)

Prompt Engineering Fundamentals

Prompt engineering is the practice of designing effective inputs to guide foundation models toward desired outputs. It is a major exam topic.

Technique	Description	Example
Zero-Shot	Ask the model to perform a task with no examples	"Classify this review as positive or negative: ..."
Few-Shot	Provide a few examples before the actual task	"Review: Great! Sentiment: Positive. Review: Awful! Sentiment: Negative. Review: Decent. Sentiment:"
Chain-of-Thought (CoT)	Ask the model to reason step by step	"Solve this math problem. Think step by step."
System Prompts	Set the model’s persona, constraints, and behavior rules	"You are a helpful customer service agent. Answer only from the provided FAQ."
Negative Prompting	Explicitly tell the model what NOT to do	"Answer the question. Do not make up facts. Do not include personal opinions."

Model Customization Techniques

The exam tests your understanding of how to adapt foundation models for specific tasks. Know the tradeoffs between approaches.

Technique	Data Needed	Cost	When to Use
Prompt Engineering	None	Lowest	Quick experimentation; first approach to try
Retrieval-Augmented Generation (RAG)	Knowledge base / documents	Low–Medium	Need current or proprietary data; reduce hallucination
Fine-Tuning	Labeled task-specific dataset	Medium–High	Domain-specific behavior, style, or terminology
Continued Pre-Training	Large unlabeled domain corpus	Highest	Teach model new domain knowledge (medical, legal, financial)

Tip: The exam frequently asks "what is the MOST cost-effective approach?" Always try prompt engineering first, then RAG, then fine-tuning. Continued pre-training is the last resort.

Retrieval-Augmented Generation (RAG)

RAG is a pattern that combines a foundation model with an external knowledge base to ground responses in factual, up-to-date data. It is one of the most tested concepts on the AIF-C01.

How RAG Works

1. User submits a query
2. Query is converted to an embedding (vector)
3. Similar documents retrieved from a vector database
4. Retrieved context is added to the prompt
5. Foundation model generates a grounded response

RAG Benefits

• Reduces hallucinations with factual grounding
• No model retraining needed
• Knowledge stays current (just update the documents)
• Cost-effective vs. fine-tuning
• Responses can cite sources for transparency

Embeddings & Vector Databases

Embeddings are numerical representations (vectors) of text, images, or other data. They capture semantic meaning, enabling similarity search.

Concept	Description
Embedding Model	Converts text/data into fixed-size vectors (e.g., Amazon Titan Embeddings, Cohere Embed)
Vector Database	Stores and searches embeddings by similarity (Amazon OpenSearch Serverless, Pinecone, FAISS)
Similarity Search	Finds the nearest vectors to a query vector using cosine similarity or Euclidean distance
Chunking	Splitting documents into smaller pieces before embedding. Chunk size affects retrieval quality.

Exam Tip

When a question asks how to make an LLM answer questions using company-specific or up-to-date information without retraining, the answer is RAG. When a question asks about changing the model’s tone, style, or format, the answer is more likely fine-tuning. Temperature and top-p control creativity/randomness, not factual accuracy.

Domain 3: Applications of Foundation Models (28%)

This is the highest-weighted domain. It focuses on AWS services for using, customizing, and deploying foundation models — primarily Amazon Bedrock. You must know the available models, Bedrock features, agents, knowledge bases, and guardrails.

Amazon Bedrock

Amazon Bedrock is a fully managed service that provides access to high-performing foundation models from leading AI companies through a single API. It is the central service for the AIF-C01 exam.

Feature	Description
Model Access	Access FMs from Amazon (Titan), Anthropic (Claude), Meta (Llama), AI21 Labs (Jamba), Cohere, Mistral, Stability AI
Serverless	No infrastructure to manage. Pay per API call (input/output tokens or image generated)
Customization	Fine-tune and continue pre-training models with your own data (data never leaves your account)
Knowledge Bases	Managed RAG implementation — connect FMs to your data sources for grounded responses
Agents	Create autonomous AI agents that can plan, execute multi-step tasks, and call APIs/Lambda functions
Guardrails	Define content filters, denied topics, word filters, and PII redaction for responsible AI
Model Evaluation	Evaluate and compare FM outputs using automatic metrics or human evaluation workflows

Foundation Models on Bedrock

Provider	Model Family	Strengths
Amazon	Titan Text, Titan Embeddings, Titan Image Generator	Cost-effective, built-in watermarking for images, native AWS integration
Anthropic	Claude 3 (Haiku, Sonnet, Opus)	Advanced reasoning, 200K context window, strong safety, vision capabilities
Meta	Llama 3	Open-source, good for fine-tuning, multiple sizes (8B, 70B)
AI21 Labs	Jamba-Instruct	Large context window (256K tokens), multilingual
Cohere	Command, Embed	RAG-optimized, strong embeddings, enterprise search
Mistral AI	Mistral, Mixtral	Efficient, fast inference, good cost-to-performance ratio
Stability AI	Stable Diffusion XL	Image generation, inpainting, outpainting, image-to-image

Bedrock Knowledge Bases (Managed RAG)

Bedrock Knowledge Bases provide a fully managed RAG pipeline. You supply data sources, and Bedrock handles chunking, embedding, indexing, and retrieval automatically.

Supported Data Sources

• Amazon S3 (PDFs, text files, HTML, Markdown, CSV, Word, Excel)
• Confluence
• SharePoint
• Salesforce
• Web Crawler

Vector Store Options

• Amazon OpenSearch Serverless (default)
• Amazon Aurora PostgreSQL (pgvector)
• Pinecone
• Redis Enterprise Cloud
• MongoDB Atlas

Bedrock Agents

Bedrock Agents enable foundation models to autonomously plan and execute multi-step tasks by connecting to company systems and APIs.

Component	Purpose
Instructions	Natural language description of the agent’s role and behavior
Action Groups	Define API operations the agent can invoke (backed by Lambda functions or API schemas)
Knowledge Bases	Connect agents to RAG for information retrieval from your documents
Orchestration	The agent automatically breaks down tasks, calls tools, and synthesizes results

Tip: When a question describes an AI assistant that needs to look up information AND take actions (like booking a flight or updating a database), the answer is Bedrock Agents.

Bedrock Guardrails

Guardrails let you implement safeguards for your generative AI applications, independent of the underlying foundation model.

Content Filters

• Filter categories: Hate, Insults, Sexual, Violence, Misconduct
• Configurable strength: None, Low, Medium, High
• Applied to both inputs and outputs independently

Other Safeguards

• Denied Topics — Block responses about specified topics
• Word Filters — Block specific words or phrases
• PII Redaction — Detect and mask personal data (SSN, emails, phone numbers)
• Contextual Grounding — Check if responses are grounded in the provided source

AWS AI/ML Services (Pre-Trained)

AWS offers fully managed AI services that require no ML expertise. These are pre-trained for specific tasks and exposed via simple APIs.

Service	Category	Capabilities
Amazon Rekognition	Vision	Object detection, facial analysis, content moderation, text in images, celebrity recognition
Amazon Textract	Vision / NLP	Extract text, tables, and forms from scanned documents and PDFs
Amazon Comprehend	NLP	Sentiment analysis, entity recognition, topic modeling, language detection, PII detection
Amazon Translate	NLP	Real-time and batch language translation, custom terminology
Amazon Polly	Speech	Text-to-speech with natural-sounding voices, SSML support, Neural TTS
Amazon Transcribe	Speech	Speech-to-text, real-time and batch, speaker identification, custom vocabulary
Amazon Lex	Conversational AI	Build chatbots and voice bots with NLU, intent recognition, slot filling
Amazon Kendra	Search	Intelligent enterprise search using NLP, semantic ranking, connectors to 40+ data sources
Amazon Personalize	Recommendations	Real-time personalized recommendations (products, content, search results)
Amazon Forecast	Time Series	Time-series forecasting (demand, revenue, capacity) using AutoML

Amazon SageMaker (Key Concepts)

While AIF-C01 is not a deep SageMaker exam, you need to know when SageMaker is the right choice and its high-level capabilities.

Feature	Description
SageMaker Studio	Integrated development environment (IDE) for ML — build, train, deploy all in one place
SageMaker JumpStart	Hub with pre-trained models and solution templates; one-click deploy of FMs
SageMaker Canvas	No-code ML for business users; visual drag-and-drop model building
SageMaker Clarify	Detect bias in data and models; explain model predictions (SHAP values)
SageMaker Ground Truth	Managed data labeling with human annotators and active learning

Exam Tip

Domain 3 is the most practical. Questions describe business scenarios and ask you to choose the right service or approach. If the scenario involves using an existing FM with minimal setup, the answer is Bedrock. If it involves custom model training with your own algorithm, the answer is SageMaker. If it involves a specific AI task (like OCR or sentiment), look for the dedicated AI service first.

Domains 4 & 5: Responsible AI (14%) & Security, Compliance & Governance (14%)

These two domains together account for 28% of the exam. They cover the ethical, fairness, privacy, and security dimensions of building AI systems on AWS. Responsible AI principles appear throughout the entire exam, not just in these domains.

Responsible AI Principles

AWS defines responsible AI as the practice of designing, developing, and deploying AI with the intention to empower employees and customers while ensuring fairness, safety, and transparency.

Principle	Description	AWS Implementation
Fairness	AI systems should not discriminate based on protected attributes (race, gender, age)	SageMaker Clarify for bias detection, balanced training data
Explainability	Users should understand why an AI system made a particular decision	SageMaker Clarify (SHAP), Bedrock model evaluation, logging
Transparency	Organizations should be open about how AI systems work and their limitations	Model cards, documentation, disclosure of AI-generated content
Privacy & Security	AI systems must protect personal data and be resilient against attacks	Encryption, VPC, IAM, PII redaction in Guardrails
Safety	AI systems should not cause harm, and there should be human oversight mechanisms	Guardrails content filters, human-in-the-loop, testing
Governance	Policies, processes, and accountability for AI development and deployment	AWS AI Service Cards, audit trails, CloudTrail logging

Types of Bias in AI

Bias Type	Where It Occurs	Example
Sampling / Selection Bias	Data collection	Training data not representative of the real-world population
Measurement Bias	Data collection	Inconsistent or inaccurate data collection across groups
Label Bias	Data labeling	Human annotators introduce their own biases into labels
Algorithmic Bias	Model training	Model amplifies patterns in biased data or overweights certain features
Confirmation Bias	Evaluation	Evaluators interpret results in ways that confirm pre-existing beliefs

Tip: The best way to mitigate bias is to start with diverse, representative training data. SageMaker Clarify can detect bias both before training (pre-training metrics) and after (post-training metrics).

Generative AI Risks & Mitigations

Key Risks

• Hallucinations — Generating false or fabricated information
• Toxicity — Generating harmful, offensive, or inappropriate content
• Prompt Injection — Malicious inputs that override model instructions
• Data Leakage — Model inadvertently revealing training data
• Copyright Issues — Generated content may resemble copyrighted material
• Deepfakes — Realistic fake images, audio, or video

Mitigations

• Bedrock Guardrails — Content filters, denied topics, PII redaction
• RAG — Ground responses in verified data to reduce hallucinations
• Human Review — Human-in-the-loop for high-stakes decisions
• Input Validation — Sanitize and validate user inputs
• Watermarking — Amazon Titan Image Generator adds invisible watermarks
• Model Evaluation — Test models against safety benchmarks before deployment

Security for AI on AWS

Security Layer	AWS Service / Feature	Purpose
Identity & Access	IAM roles, policies, SCPs	Control who can access AI services, models, and data
Encryption at Rest	AWS KMS, SSE-S3	Encrypt training data, model artifacts, and inference data stored in S3
Encryption in Transit	TLS/SSL	All AWS API calls and Bedrock requests use HTTPS (TLS 1.2+)
Network Isolation	VPC, PrivateLink, Security Groups	Run SageMaker in private VPC; use VPC endpoints for Bedrock
Auditing	AWS CloudTrail, CloudWatch	Log all API calls to AI services, monitor for anomalies
Data Privacy	Bedrock data isolation, Amazon Macie	Bedrock does not use your data to train base models; Macie detects sensitive data in S3

Bedrock Data Privacy & Compliance

Data Protection in Bedrock

• Your data is NOT used to train or improve the base foundation models
• Your data is NOT shared with model providers
• Fine-tuning data and custom models stay in your AWS account
• All data encrypted at rest and in transit by default
• Optional: Use customer-managed KMS keys for additional control

Compliance Frameworks

• SOC 1, SOC 2, SOC 3 reports
• HIPAA eligible (with BAA)
• GDPR compliant
• ISO 27001, 27017, 27018
• FedRAMP (select services)
• AWS Artifact for compliance reports

AI Governance Best Practices

•Model Cards — Document model purpose, training data, performance metrics, limitations, and intended use cases
•AWS AI Service Cards — Published by AWS for each AI service; describe intended use, limitations, and responsible use guidance
•Audit Trails — Use CloudTrail to log all interactions with AI services for accountability
•Model Monitoring — Continuously monitor deployed models for drift, bias, and degradation
•Human Oversight — Maintain human-in-the-loop for critical decisions (healthcare, finance, legal)

Exam Tip

Questions about data privacy in Bedrock almost always highlight that your data is NOT used to train base models and is NOT shared with third-party model providers. For security questions, remember the defense-in-depth approach: IAM for access control, KMS for encryption, VPC for network isolation, CloudTrail for auditing. When asked about detecting bias, the answer is SageMaker Clarify. When asked about filtering harmful content, the answer is Bedrock Guardrails.

Service Comparisons & Decision Guide

The AIF-C01 exam frequently presents scenarios where you must choose between similar services or approaches. These comparison tables will help you make the right choice quickly.

Bedrock vs. SageMaker vs. AI Services

Criteria	Amazon Bedrock	Amazon SageMaker	AWS AI Services
Best For	Using foundation models (LLMs, image gen) via API	Custom ML model training and deployment	Specific AI tasks with no ML expertise
ML Expertise	Low — use FMs via prompts	High — build/train custom models	None — simple API calls
Customization	Prompt engineering, RAG, fine-tuning	Full control over algorithm, data, hyperparameters	Limited (custom labels in Rekognition, custom vocabulary in Transcribe)
Infrastructure	Fully serverless	Managed instances (you choose type/size)	Fully serverless
Example Use Case	AI chatbot, content generation, document Q&A	Fraud detection model, demand forecasting	Extract text from receipts (Textract), translate text (Translate)

Tip: If the question mentions "foundation model," "LLM," or "generative AI" → Bedrock. If it mentions "custom training" or "own algorithm" → SageMaker. If it mentions a specific task like OCR or translation → AI Service.

RAG vs. Fine-Tuning vs. Continued Pre-Training

Approach	What It Changes	Data Required	Cost	Best For
RAG	Context provided to the model (not the model itself)	Documents in a knowledge base	Low	Current or proprietary information; factual accuracy
Fine-Tuning	Model weights are updated	Labeled prompt-completion pairs	Medium	Domain tone/style, specific output format, task specialization
Continued Pre-Training	Model learns new domain knowledge	Large unlabeled domain corpus	High	Teach model entirely new domain vocabulary/concepts

Bedrock Knowledge Bases vs. Amazon Kendra

Feature	Bedrock Knowledge Bases	Amazon Kendra
Primary Purpose	RAG — provide context to FMs for generation	Enterprise search — find and rank documents
Output	Generated natural language answer from FM	Ranked list of relevant documents/passages
Uses FM?	Yes (Bedrock FM generates the response)	No (traditional NLP-based search ranking)
Best For	Q&A chatbots, document summarization with your data	Enterprise document search, intranet search portal

Comprehend vs. Bedrock for NLP Tasks

Task	Amazon Comprehend	Amazon Bedrock (FM)
Sentiment Analysis	Purpose-built, returns structured score (positive/negative/neutral/mixed)	Can do via prompting, but less structured output
Entity Recognition	Purpose-built NER with standard entity types	Can extract entities via prompting
Text Summarization	Not supported natively	Excellent — FMs excel at summarization
Content Generation	Not supported	Primary use case for generative FMs
PII Detection	Purpose-built PII detection and redaction	Bedrock Guardrails can filter PII in FM inputs/outputs

SageMaker Canvas vs. SageMaker Studio vs. SageMaker JumpStart

Service	Target User	Coding Required	Best For
SageMaker Canvas	Business analysts	No code	Visual drag-and-drop ML, quick predictions
SageMaker JumpStart	Developers / Data Scientists	Minimal	Deploy pre-trained models, fine-tune FMs, solution templates
SageMaker Studio	Data Scientists / ML Engineers	Full coding	End-to-end custom ML development

Bias Detection & Content Safety Tools

Tool	Purpose	When to Use
SageMaker Clarify	Detect bias in training data and model predictions; explain model decisions (SHAP)	Custom ML models on SageMaker; pre- and post-training bias analysis
Bedrock Guardrails	Filter harmful content, block topics, redact PII in FM inputs/outputs	Any Bedrock-based generative AI application
Bedrock Model Evaluation	Compare FM performance across models using automatic or human evaluation	Choosing between FMs or validating model quality
Amazon Rekognition (Content Moderation)	Detect inappropriate or explicit content in images and video	User-uploaded image/video content moderation
Amazon Comprehend (Toxicity)	Detect toxic content in text (insults, threats, profanity)	Text moderation in forums, reviews, user-generated content

Inference Parameters Quick Reference

Parameter	Effect When Low	Effect When High	Use Case
Temperature	Deterministic, focused, factual	Creative, diverse, more random	Low for facts/code, high for creative writing
Top-p	Fewer candidate tokens, more focused	More candidate tokens, more diverse	Lower for precise tasks, higher for brainstorming
Top-k	Consider fewer tokens at each step	Consider more tokens at each step	Fine-grained control over vocabulary diversity
Max Tokens	Shorter responses	Longer responses (up to model limit)	Control output length and cost

Quick Decision Flowchart

→ Need to generate text, images, or code? Use Amazon Bedrock
→ Need to answer questions from company documents? Use Bedrock Knowledge Bases (RAG)
→ Need an AI agent that takes actions? Use Bedrock Agents
→ Need to detect sentiment or extract entities? Use Amazon Comprehend
→ Need to extract text from documents? Use Amazon Textract
→ Need to analyze images or moderate visual content? Use Amazon Rekognition
→ Need to build a chatbot with intents and slots? Use Amazon Lex
→ Need to convert text to speech? Use Amazon Polly
→ Need to convert speech to text? Use Amazon Transcribe
→ Need to translate between languages? Use Amazon Translate
→ Need product/content recommendations? Use Amazon Personalize
→ Need to search enterprise documents? Use Amazon Kendra
→ Need to train a custom model with your own algorithm? Use Amazon SageMaker
→ Need no-code ML for business analysts? Use SageMaker Canvas
→ Need to filter harmful FM outputs? Use Bedrock Guardrails
→ Need to detect bias in ML models? Use SageMaker Clarify

Exam Tip

Many AIF-C01 questions are scenario-based: they describe a business need and ask which AWS service or approach to use. Read the scenario carefully for keywords. "Foundation model" or "generative AI" points to Bedrock. "No ML expertise" points to AI Services or SageMaker Canvas. "Custom training data" with labeled examples points to SageMaker. "Up-to-date company data" without retraining points to RAG. "Prevent harmful content" points to Bedrock Guardrails. Always pick the simplest, most cost-effective solution that meets all requirements.