AWS Foundational Certification

AWS Certified AI Practitioner AIF-C01 Study Guide

A comprehensive, exam-focused web app for learning AI/ML fundamentals, generative AI, foundation model applications, responsible AI, and AWS security/governance topics.

Start Study Guide Practice 100 Questions

Exam Snapshot

  • Exam code: AIF-C01
  • Level: Foundational
  • Duration: 90 minutes
  • Questions: 65 total, 50 scored + 15 unscored
  • Passing score: 700 / 1000
  • Question types: Multiple choice, multiple response, ordering, matching
20%Fundamentals of AI and ML
24%Fundamentals of GenAI
28%Applications of Foundation Models
14%Responsible AI
14%Security, Compliance, Governance

How to use this guide

Study with exam logic, not memorization only

What the exam expects

This is not a coding exam. The target candidate is familiar with AI/ML on AWS and can choose appropriate services, explain concepts, understand risks, and apply responsible AI thinking. Focus on identifying the right concept or AWS service for a business scenario.

Best preparation method

Read each domain, memorize the service map, then answer the practice questions. For every wrong answer, return to the matching domain and write a one-line rule. Example: “RAG = trusted retrieval at inference time, not model retraining.”

Detailed Exam Domains

Complete domain-by-domain study guide

Use this section as your main textbook. Each domain includes the concepts, AWS services, exam traps, and memory notes you should know before practicing questions.

Exam Domain • 20% of scored content

Domain 1: Fundamentals of AI and ML

20%

1.1 Explain basic AI concepts and terminologies

  • AI is the broad field of building systems that perform tasks requiring human-like intelligence, such as understanding text, recognizing images, reasoning, planning, and making predictions.
  • Machine learning is a subset of AI where systems learn patterns from data instead of being explicitly programmed for every rule.
  • Deep learning uses neural networks with many layers. It is commonly used for image recognition, speech recognition, language models, and large-scale pattern recognition.
  • Generative AI creates new content such as text, images, audio, video, code, and summaries. Foundation models and large language models are common examples.
  • Agentic AI uses models plus tools, memory, and workflows to take steps toward a goal. Agents can call APIs, retrieve data, and coordinate tasks.
  • Inferencing means using a trained model to produce predictions or outputs. Batch inferencing works on many records later, real-time inferencing responds immediately, asynchronous inferencing handles longer jobs, and serverless inferencing scales without fixed servers.
  • Data types include structured tables, semi-structured JSON/logs, unstructured text/images/audio, time-series data, labeled training data, and unlabeled data.

1.2 Identify practical use cases for AI

  • Regression predicts continuous numeric values, such as sales amount or delivery time.
  • Classification predicts categories, such as fraud/not fraud or high/medium/low risk.
  • Clustering groups similar items without predefined labels, such as customer segmentation.
  • Computer vision identifies objects, detects defects, extracts text from images, and supports safety monitoring.
  • NLP handles text tasks such as sentiment analysis, classification, entity extraction, summarization, and translation.
  • Speech services convert speech to text, text to speech, or support conversational voice interfaces.
  • Traditional ML is often better when the output must be explainable, highly structured, cheaper to run, or built around tabular prediction. Foundation models are better for flexible language, summarization, generation, reasoning, and multi-modal tasks.

1.3 Describe the AI/ML development lifecycle

  • A common lifecycle is business problem definition, data collection, data preparation, model selection, training or customization, evaluation, deployment, monitoring, feedback, and retraining.
  • MLOps focuses on repeatable experiments, versioning, automated deployment, monitoring, governance, and reducing technical debt.
  • Key ML metrics include accuracy, precision, recall, F1 score, false positives, false negatives, latency, and cost per prediction.
  • Business metrics include ROI, user satisfaction, development cost, adoption, task completion rate, conversion rate, and customer feedback.
  • AWS services in this area include Amazon SageMaker AI for ML development, Amazon Bedrock for foundation model applications, Amazon Q for AI assistance, Amazon Transcribe, Translate, Comprehend, Lex, Polly, and Rekognition for managed AI capabilities.
Exam memory:
  • Know the difference between AI, ML, deep learning, GenAI, and agentic AI.
  • Choose regression for numeric prediction, classification for category prediction, and clustering for grouping without labels.
  • The exam is foundational. Expect service matching and concept selection, not coding or math-heavy modeling.

Exam Domain • 24% of scored content

Domain 2: Fundamentals of Generative AI

24%

2.1 Explain basic concepts of generative AI

  • Tokens are pieces of text processed by a model. Token count affects cost, latency, and how much context a model can handle.
  • Chunking splits large documents into smaller parts for retrieval and embedding. Good chunking improves search and RAG quality.
  • Embeddings convert text, images, or other data into vectors so similarity can be searched mathematically.
  • Vector databases store embeddings and allow similarity search for semantic retrieval.
  • Transformers are a neural network architecture behind modern LLMs. They learn relationships across tokens and context.
  • Foundation models are large models trained on broad data and adapted to many tasks. They can be used through prompting, RAG, fine-tuning, or agents.
  • Multi-modal models can process or generate more than one modality such as text, images, audio, or video.
  • Diffusion models are commonly associated with image generation, where noise is gradually refined into an output.
  • Context engineering is the practice of providing the right instructions, retrieved documents, examples, memory, tools, and constraints to guide an FM application.

2.2 Understand capabilities and limitations of GenAI

  • GenAI is useful for summarization, chat assistants, search, content generation, translation, code assistance, knowledge retrieval, and customer support.
  • Advantages include adaptability, natural conversation, reusable models, rapid prototyping, and the ability to generate or transform content.
  • Limitations include hallucination, nondeterministic responses, bias, security risks, data leakage, latency, cost, and difficult explainability.
  • Model selection depends on task type, modality, performance, latency, context window, cost, regional availability, compliance, customization needs, and safety requirements.
  • Business value should be measured with relevant metrics such as productivity gain, task completion, customer satisfaction, conversion rate, ROI, and cost per interaction.

2.3 Describe AWS infrastructure and technologies for GenAI

  • Amazon Bedrock is a managed service for building GenAI applications with foundation models, model customization, agents, knowledge bases, guardrails, and evaluation features.
  • Amazon SageMaker AI supports ML model building, training, deployment, monitoring, and MLOps workflows.
  • SageMaker JumpStart provides prebuilt models and solution templates that can accelerate ML and GenAI experimentation.
  • Amazon Q is an AI assistant family for business and developer productivity.
  • AWS infrastructure benefits include security, governance, compliance capabilities, scalable compute, managed services, availability, and regional options.
  • Cost tradeoffs include on-demand token pricing, provisioned throughput, model size, output length, latency needs, customization cost, and the cost of retrieval or external tools.
Exam memory:
  • Connect tokens with cost and latency.
  • Connect embeddings and vector databases with semantic search and RAG.
  • Know Bedrock as the main managed AWS service for FM/GenAI applications.

Exam Domain • 28% of scored content

Domain 3: Applications of Foundation Models

28%

3.1 Describe design considerations for FM applications

  • Choose an FM based on cost, latency, supported modality, model size, context window, language support, customization options, quality, and safety requirements.
  • Temperature controls randomness. Lower temperature usually gives more consistent responses; higher temperature gives more creative or varied responses.
  • Input and output length affect cost, speed, and whether the model can use enough context.
  • RAG retrieves relevant information from trusted sources and gives it to the model at inference time. It is useful for company knowledge bases, support bots, policy assistants, and document Q&A.
  • Amazon Bedrock Knowledge Bases helps connect FMs to data sources using embeddings and retrieval.
  • AWS options for vector storage include Amazon OpenSearch Service, Amazon Aurora, Amazon Neptune, and Amazon RDS for PostgreSQL with vector capabilities.
  • Customization options have different costs: prompting and in-context learning are simplest, RAG grounds answers in data, fine-tuning adapts behavior, continued pre-training is heavier, and model distillation can reduce serving cost.
  • AI agents combine models with tools and orchestration so they can complete multi-step business tasks.

3.2 Choose effective prompt engineering techniques

  • A good prompt often includes role, task, context, constraints, examples, output format, and safety boundaries.
  • Zero-shot prompting gives only the instruction. Single-shot and few-shot prompting provide one or more examples.
  • Prompt templates make prompts reusable and consistent across use cases.
  • Negative prompts tell the model what to avoid, especially in generation tasks.
  • Chain-of-thought style prompting can improve reasoning, but exam-safe practice is to ask for structured reasoning or intermediate steps only when appropriate.
  • Risks include prompt injection, prompt hijacking, jailbreaking, data exposure, and poisoned retrieval content.
  • Amazon Bedrock Prompt Management helps version, test, and manage prompts.

3.3 Describe training and fine-tuning process for FMs

  • Pre-training creates a base model from broad data. It is expensive and usually not needed for foundational users.
  • Fine-tuning adapts an existing model to specific tasks, tone, formats, or domain patterns.
  • Instruction tuning teaches the model to follow instructions better.
  • Transfer learning reuses a pre-trained model for a related task.
  • Continuous pre-training can update a model with additional domain data but costs more than prompt-based methods.
  • Distillation trains a smaller model to mimic a larger one, often to reduce latency and serving cost.
  • Fine-tuning data should be representative, governed, labeled where needed, high quality, diverse, and reviewed for bias or sensitive content.
  • RLHF uses human feedback to improve model responses and alignment.

3.4 Describe methods to evaluate FM performance

  • Use human evaluation when output quality, tone, usefulness, safety, or business fit needs judgment.
  • Use benchmark datasets to compare models on consistent tasks.
  • Amazon Bedrock Model Evaluation helps compare model performance for selected use cases.
  • Metrics include ROUGE for summarization overlap, BLEU for translation overlap, BERTScore for semantic similarity, and LLM-as-a-judge for model-based evaluation.
  • For applications, evaluate task completion rate, user satisfaction, answer groundedness, retrieval quality, cost per interaction, latency, and escalation rate.
  • For RAG, test retrieval relevance, source citation quality, hallucination rate, and answer accuracy against source documents.
Exam memory:
  • This is the largest domain. Spend the most time on Bedrock, RAG, prompts, FM customization, and model evaluation.
  • RAG is usually preferred when answers must use up-to-date or private company knowledge without training a new model.
  • Fine-tuning changes model behavior; RAG adds external knowledge at response time.

Exam Domain • 14% of scored content

Domain 4: Guidelines for Responsible AI

14%

4.1 Explain responsible AI development

  • Responsible AI includes fairness, inclusivity, safety, robustness, veracity, transparency, privacy, and accountability.
  • Bias can appear in training data, labels, model assumptions, evaluation sets, or user feedback loops.
  • High variance can lead to overfitting, where a model performs well on training data but poorly on new data.
  • High bias can lead to underfitting, where a model is too simple to capture important patterns.
  • Dataset quality matters. Look for diverse, representative, balanced, curated, and legally usable data.
  • Amazon Bedrock Guardrails can help apply safety controls and content filtering for GenAI applications.
  • SageMaker Clarify helps detect bias and explain model behavior. SageMaker Model Monitor helps monitor models in production. Amazon A2I supports human review workflows.
  • Legal risks include IP infringement, biased outputs, hallucinations, harmful content, privacy issues, and loss of customer trust.
  • Sustainability and environmental impact are also model selection considerations.

4.2 Recognize transparent and explainable models

  • Transparent models are easier to inspect directly, such as simple rules or some linear models. Large deep learning models and FMs are usually less transparent.
  • Explainability helps users understand why a system produced a result and when they should trust it.
  • Model Cards document intended use, limitations, evaluation results, ethical considerations, and operational details.
  • SageMaker Clarify, SageMaker Model Cards, Amazon Bedrock Model Evaluation, open-source model documentation, and license information can support transparency.
  • There is often a tradeoff between interpretability, performance, safety, and model complexity.
  • Human-centered design means users should receive useful explanations, feedback channels, clear limitations, and ways to contest or correct results.
Exam memory:
  • Responsible AI questions often ask for the safest, most transparent, or least risky choice.
  • For bias and explainability, remember SageMaker Clarify.
  • For GenAI safety controls, remember Amazon Bedrock Guardrails.

Exam Domain • 14% of scored content

Domain 5: Security, Compliance, and Governance for AI Solutions

14%

5.1 Explain methods to secure AI systems

  • Secure AI systems with IAM least privilege, roles, policies, encryption, network controls, logging, monitoring, and data protection.
  • Encrypt data at rest and in transit. Control who can access model inputs, outputs, training data, logs, and vector stores.
  • Use AWS PrivateLink where private connectivity to services is required without exposing traffic to the public internet.
  • Amazon Macie helps discover and protect sensitive data such as personally identifiable information in Amazon S3.
  • Amazon Bedrock Guardrails can filter harmful content and reduce unsafe outputs.
  • Prompt injection is a security risk where user or retrieved content attempts to override instructions or leak data.
  • Mitigations include input validation, output filtering, system instruction separation, retrieval source controls, least privilege tools, logging, and human review for high-risk workflows.
  • Source citation, data lineage, data cataloging, and SageMaker Model Cards help document origins and improve traceability.
  • Hallucination reduction uses RAG grounding, source citations, output validation, confidence scoring, and human review.

5.2 Recognize governance and compliance regulations

  • Governance includes policies, review cadence, training, logging, retention, data residency, auditability, and risk management.
  • AWS CloudTrail records API activity for auditing. AWS Config evaluates resource configuration. AWS Audit Manager helps collect evidence for audits.
  • AWS Artifact provides access to AWS compliance reports and agreements.
  • Amazon Inspector helps identify software vulnerabilities and unintended network exposure.
  • AWS Trusted Advisor gives checks and recommendations across cost, performance, security, fault tolerance, and service limits.
  • Data governance strategies include lifecycle rules, retention, monitoring, ownership, access control, and deletion policies.
  • The Generative AI Security Scoping Matrix helps classify and reason about generative AI security responsibilities and risk boundaries.
Exam memory:
  • Security questions often reduce to IAM least privilege, encryption, private networking, logging, and sensitive data protection.
  • CloudTrail means API audit logs; Config means configuration compliance; Artifact means compliance reports.
  • For hallucinations, think grounding, validation, and source citation.

AWS Service Map

Know which service fits which AI use case

ServiceWhat it doesExam use case
Amazon Bedrock Build GenAI apps with FMs, agents, knowledge bases, guardrails, prompt management, customization, and evaluation. Choose for managed foundation model applications.
Amazon Bedrock Knowledge Bases Connect FMs to enterprise data through retrieval and embeddings. Choose for RAG and grounded Q&A.
Amazon Bedrock Guardrails Apply safety controls and content policies to GenAI interactions. Choose for responsible AI and output filtering.
Amazon Bedrock Agents / AgentCore Create agent workflows that use tools, identity, policies, and orchestration. Choose for multi-step tasks and tool use.
Amazon SageMaker AI Build, train, deploy, monitor, and govern ML models. Choose for ML lifecycle and MLOps.
SageMaker JumpStart Use prebuilt models, notebooks, and solution templates. Choose for faster experimentation.
SageMaker Clarify Detect bias and explain model predictions. Choose for responsible AI and explainability.
SageMaker Model Monitor Monitor model quality and drift in production. Choose for MLOps monitoring.
SageMaker Model Cards Document model purpose, performance, limitations, and governance information. Choose for transparency and lineage.
Amazon A2I Human review workflows for ML predictions. Choose when human-in-the-loop review is needed.
Amazon Q AI assistant for business and developer productivity. Choose for productivity assistants.
Amazon Transcribe Speech to text. Choose for call transcription or subtitles.
Amazon Translate Text translation. Choose for language translation.
Amazon Comprehend NLP entity, sentiment, key phrase, classification. Choose for text analytics without custom ML.
Amazon Lex Conversational chatbots and voice bots. Choose for chatbot interfaces.
Amazon Polly Text to lifelike speech. Choose for text-to-speech.
Amazon Rekognition Image and video analysis. Choose for object, face, label, or moderation tasks.
Amazon Textract Extract text, forms, and tables from documents. Choose for OCR and document extraction.
Amazon OpenSearch Service Search, analytics, and vector search use cases. Choose for semantic search/vector storage.
Amazon Aurora / RDS PostgreSQL Relational databases that can support vector use cases with extensions or features. Choose when vector search must live near relational data.
Amazon Neptune Graph database that can support relationship-heavy knowledge and graph use cases. Choose for connected data.
AWS IAM Manage users, roles, policies, and permissions. Choose for least privilege access control.
AWS KMS Create and manage encryption keys. Choose for encryption key management.
Amazon Macie Discover sensitive data in S3. Choose for PII/sensitive data discovery.
AWS PrivateLink Private connectivity to AWS services. Choose to avoid public internet exposure.
AWS CloudTrail Audit API activity. Choose for who did what and when.
AWS Config Track and evaluate resource configuration compliance. Choose for config rules and drift.
AWS Audit Manager Collect audit evidence. Choose for audit preparation.
AWS Artifact Access compliance reports and agreements. Choose for compliance documentation.
Amazon Inspector Find vulnerabilities and exposure. Choose for vulnerability management.
AWS Trusted Advisor Best-practice checks. Choose for recommendations across cost/security/performance.

Flashcards

Tap/hover to revise key concepts

AI vs MLTap or hover
AI is the broad goal of intelligent systems; ML is a subset where systems learn from data.
Deep learningTap or hover
ML using multi-layer neural networks, often used for images, speech, and large language models.
GenAITap or hover
AI that creates new content such as text, images, audio, code, or summaries.
Agentic AITap or hover
AI that can plan steps, use tools, maintain memory, and orchestrate workflows.
RegressionTap or hover
Predicts a continuous number.
ClassificationTap or hover
Predicts a category or class.
ClusteringTap or hover
Groups similar records without predefined labels.
Batch inferenceTap or hover
Runs predictions on many records at a scheduled or offline time.
Real-time inferenceTap or hover
Returns predictions immediately for interactive applications.
TokenTap or hover
A piece of text processed by a model; affects cost and context usage.
EmbeddingTap or hover
A vector representation of data used for similarity search.
Vector databaseTap or hover
Stores embeddings and supports semantic retrieval.
RAGTap or hover
Retrieval Augmented Generation: retrieve trusted context, then generate an answer.
Fine-tuningTap or hover
Adapts an existing model with additional training examples.
Prompt templateTap or hover
Reusable prompt pattern with variables.
TemperatureTap or hover
Controls response randomness.
HallucinationTap or hover
Confident but incorrect or unsupported model output.
GroundingTap or hover
Connecting model output to trusted sources.
SageMaker ClarifyTap or hover
Bias detection and explainability.
Bedrock GuardrailsTap or hover
Safety controls for GenAI applications.
CloudTrailTap or hover
API audit logs.
AWS ConfigTap or hover
Configuration compliance.
AWS ArtifactTap or hover
Compliance reports and agreements.
Amazon MacieTap or hover
Sensitive data discovery in S3.
AWS PrivateLinkTap or hover
Private connectivity to AWS services.

Practice Exam

100 original sample questions with explanations

Use the filter to focus on a domain. Click an answer to reveal the correct option and explanation.

10-Day Plan

Fast but complete revision schedule

Day 1
Exam overview + Domain 1 concepts

Learn AI, ML, DL, GenAI, agentic AI, learning types, inference types, and model metrics.

Day 2
Domain 1 AWS managed AI services

Map Transcribe, Translate, Comprehend, Lex, Polly, Rekognition, Textract, SageMaker AI, and Bedrock to use cases.

Day 3
Domain 2 GenAI basics

Study tokens, embeddings, vectors, chunking, FMs, LLMs, multi-modal models, context engineering, and cost factors.

Day 4
Domain 2 AWS GenAI services

Focus on Bedrock, SageMaker AI, JumpStart, Amazon Q, infrastructure benefits, and token/provisioned throughput tradeoffs.

Day 5
Domain 3 FM application design

Study model selection, latency, modality, context window, RAG, vector stores, agents, and customization tradeoffs.

Day 6
Domain 3 prompt engineering

Practice zero-shot, few-shot, templates, negative prompts, prompt risks, prompt management, and evaluation basics.

Day 7
Domain 3 training + evaluation

Review pre-training, fine-tuning, RLHF, distillation, ROUGE, BLEU, BERTScore, LLM-as-a-judge, and business metrics.

Day 8
Domain 4 Responsible AI

Study fairness, bias, variance, transparency, explainability, Guardrails, Clarify, Model Cards, Model Monitor, and A2I.

Day 9
Domain 5 Security and governance

Review IAM, encryption, Macie, PrivateLink, CloudTrail, Config, Artifact, Audit Manager, Inspector, Trusted Advisor.

Day 10
Mock exam day

Answer all 100 questions, review every explanation, then revise weak domains using the left menu.