AIAI Agent Dev Labaiagentdev-lab.com
Sign in to save progressSign in
Menu

STEP 1 · Study

AIF-C01 StudyAI Practitioner

Read the key points for each exam domain, then use “Practice this domain” to answer questions from that domain only. Domains with a larger weight appear more often on the exam.

01

Fundamentals of AI and ML

20%

Learning goals

  • Explain the relationship between AI, machine learning, deep learning, generative AI and agentic AI, and the difference between training and inference
  • Map supervised, unsupervised and reinforcement learning, and regression, classification and clustering, to business problems
  • Judge situations where AI is not a good fit (a fixed, explainable answer is required, the cost-benefit does not add up, and so on)
  • Explain managed AI services such as Comprehend, Transcribe, Translate, Lex, Polly, Rekognition and Textract, and the ML lifecycle with SageMaker AI (data preparation → training → evaluation → deployment → monitoring)

Key points

  • AI ⊃ machine learning (learning patterns from data) ⊃ deep learning (multi-layer neural networks). Generative AI creates new text, images and so on from learned patterns. Agentic AI plans toward a goal and works autonomously while using tools.
  • Training creates a model from data; inference uses the model to predict or generate from new data. Inference can be real-time, batch, asynchronous or serverless.
  • Supervised learning learns from labeled data (regression = predict a number, classification = predict a category). Unsupervised learning finds structure without labels (clustering, anomaly detection). Reinforcement learning learns actions that maximize reward through trial and error.
  • Data types: labeled/unlabeled, tabular, time series, image, text, structured/unstructured. The type determines which techniques and services to use.
  • Where AI is not a good fit: the same fixed answer is always required (it can be computed by rules), accountability demands rule-based logic, there is not enough data, or the cost-benefit does not add up.
  • Managed AI services: Comprehend for text analysis, Transcribe for speech → text, Translate for translation, Lex for chatbots, Polly for text → speech, Rekognition for image and video analysis, Textract for reading forms, Personalize for recommendations.
  • ML lifecycle: define the problem → collect and prepare data → feature engineering → training and tuning → evaluation → deployment → monitoring (data and model drift). MLOps is the mechanism that runs this continuously through experiment management, reproducibility and automation.
  • Evaluation metrics: for classification, accuracy, precision, recall and F1 score; for regression, RMSE and others. Also look at business metrics (cost per user, development cost, customer response, ROI).

Workplace uses

  • SalesBuild a model that classifies the likelihood of winning a deal from past (labeled) opportunity data, and decide which prospects to visit first.
  • General AffairsTranscribe help-desk calls with Transcribe, classify and run sentiment analysis on them with Comprehend, and identify frequent inquiries.
  • AccountingRead amounts, dates and vendors from invoice PDFs with Textract and register them automatically in the expense system. Approvals that rules can decide are handled by business rules, not AI.

Exam objectives (from the exam guide)

  • 1.1 AI の基本的な概念と用語を説明する。
  • 1.2 AI の実用的なユースケースを特定する。
  • 1.3 AI/ML 開発ライフサイクルについて説明する。

Learning resources

Learn in detail (3 topics with mini quizzes) →Practice this domain (60 questions) →Next domain ↓
02

Fundamentals of Generative AI

24%

Learning goals

  • Explain tokens, chunking, embeddings, vectors, Transformers, LLMs, foundation models (FMs), multimodal models and diffusion models
  • Explain the FM lifecycle, per-token pricing, context engineering, and agent concepts (multi-agent, MCP)
  • Apply the advantages of generative AI (adaptability, conversation, generation) and its drawbacks (hallucinations, poor interpretability, nondeterminism) to business decisions
  • Judge where AWS services such as Bedrock, SageMaker AI, JumpStart, Amazon Q, Amazon Quick, Kiro and Strands Agents fit, and the cost trade-offs (on-demand, Provisioned Throughput, batch)

Key points

  • A token is the unit of text a model handles; pricing and input/output limits are based on token counts. Chunking splits long documents into sizes suited to search and processing.
  • An embedding turns the meaning of text or images into a numeric vector. Similar meanings produce nearby vectors, which are used for semantic search and RAG.
  • An LLM is a large language model based on the Transformer. A foundation model (FM) is a general-purpose model pre-trained on large amounts of data that can be used for many tasks. Diffusion models are commonly used for image generation. Multimodal models handle multiple types of data such as text and images.
  • FM lifecycle: data selection → model selection → pre-training → fine-tuning → evaluation → deployment → feedback. Most companies do not pre-train; they choose and use an existing FM.
  • Context engineering: designing the information passed to the model (instructions, reference documents, conversation history, tool results) to improve answer quality. Agents call tools to get work done, and MCP (Model Context Protocol) is a common mechanism that connects agents to external tools and data.
  • Drawbacks: hallucinations (plausible but wrong output), outputs that vary for the same input (nondeterminism), and difficulty explaining reasons. Leave calculations and decisions that must be exact to rules or existing systems, and build in grounding checks (RAG, human review).
  • AWS services: Bedrock is a fully managed service for using multiple FMs through an API (knowledge bases, agents, guardrails). SageMaker AI and JumpStart handle model training, customization and hosting. Amazon Q Business is an AI assistant grounded in company data, Q Developer and Kiro support development, and Strands Agents is an open-source SDK for building agents.
  • Cost: on-demand charges only for tokens used; Provisioned Throughput reserves a fixed amount of capacity (needed for steady high-volume use and for custom models); batch inference processes large volumes that need not be immediate at a lower price. Also compare model size, response speed and Regional availability.

Workplace uses

  • SalesGenerate a first draft of a proposal from meeting notes with Bedrock, and have the account rep check the figures and wording before finishing it. Handle confidential information such as customer names according to company rules.
  • General AffairsConnect company policies to Amazon Q Business so that when an employee asks “What is the per diem for business trips?”, it answers while showing the relevant section of the policy.
  • AccountingUse generative AI to draft explanations of monthly expense anomalies, but perform amount calculations and approval decisions with accounting-system rules, not generative AI.

Exam objectives (from the exam guide)

  • 2.1 生成 AI (GenAI) の基本概念を説明する。
  • 2.2 ビジネス上の問題の解決に GenAI を使用する場合の可能性と制限を理解する。
  • 2.3 GenAI アプリケーションを構築するための AWS インフラストラクチャとテクノロジーについて説明する。

Learning resources

Learn in detail (3 topics with mini quizzes) →Practice this domain (72 questions) →Next domain ↓
03

Applications of Foundation Models

28%

Learning goals

  • Choose an FM based on cost, modality, latency, multilingual support, model size, input/output length and so on, and explain the effect of inference parameters such as temperature
  • Explain where RAG (Bedrock Knowledge Bases), vector databases (OpenSearch, Aurora, Neptune, RDS for PostgreSQL, S3 Vectors and others) and agents fit
  • Use zero-shot, few-shot, chain-of-thought, prompt templates and negative prompts appropriately, and explain risks such as prompt injection
  • Explain the cost differences among pre-training, fine-tuning, continued pre-training, distillation, RAG and in-context learning, and when to use ROUGE, BLEU, BERTScore, LLM-as-a-judge and human evaluation

Key points

  • Choosing an FM: input/output types (text, image), response speed, Japanese support, input length (context window), customizability and cost. Bigger models are not always better; a smaller model suited to the task can be cheaper and faster.
  • Inference parameters: lowering temperature makes output more stable; raising it makes output more varied and creative. Top P also adjusts diversity. Maximum tokens caps output length.
  • RAG searches for internal documents relevant to the question, adds them to the prompt, and has the model answer based on them. It uses up-to-date and internal information without retraining the model, and can show the basis for answers. Bedrock Knowledge Bases manage ingestion, chunking, embedding and retrieval together.
  • Customization cost generally increases in this order: prompt engineering (in-context learning) < RAG < fine-tuning < continued pre-training < pre-training from scratch. Distillation transfers a large model's knowledge to a smaller model to reduce inference cost and latency.
  • Prompting techniques: zero-shot (no examples), one-shot/few-shot (show examples), chain-of-thought (have the model think step by step) and templating. Make the role, context, instructions and output format clear, and state what to avoid with negative prompts. Bedrock Prompt Management handles versioning.
  • Prompt risks: prompt injection and hijacking, jailbreaks, leakage of confidential information, and data poisoning. Mitigate them with guardrails and input validation.
  • Fine-tuning adapts a model to a specific task or domain with labeled data (examples of instructions and answers). Continued pre-training adds knowledge with unlabeled domain documents. Data quality, quantity, representativeness and governance determine the result.
  • Evaluation: ROUGE for summarization, BLEU for translation, BERTScore for semantic similarity, LLM-as-a-judge where an LLM scores outputs, and human evaluation. Bedrock model evaluation supports both automatic and human evaluation. Also evaluate against business goals (task completion rate, satisfaction, cost per interaction).

Workplace uses

  • General AffairsIngest work rules and procedure manuals into a Bedrock knowledge base and build an internal help desk that answers employee questions while citing the relevant policy sections.
  • SalesTemplatize the prompt for summarizing proposals and specify the output format (three lines each for issue, proposal and benefit). Lower the temperature so quality is consistent every time.
  • AccountingFor generating journal-entry descriptions, first try prompt engineering and RAG, and consider fine-tuning only if the output still does not match in-house terminology.

Exam objectives (from the exam guide)

  • 3.1 基盤モデル (FM) を使用するアプリケーションの設計上の考慮事項を説明する。
  • 3.2 効果的なプロンプトエンジニアリング手法を選択する。
  • 3.3 FM のトレーニングとファインチューニングのプロセスを説明する。
  • 3.4 FM のパフォーマンスを評価する方法を説明する。

Learning resources

Learn in detail (4 topics with mini quizzes) →Practice this domain (84 questions) →Next domain ↓
04

Guidelines for Responsible AI

14%

Learning goals

  • Explain the elements of responsible AI (bias, fairness, inclusivity, robustness, safety, accuracy) and tools such as Amazon Bedrock Guardrails
  • Explain the legal risks of generative AI (IP infringement, biased output, loss of customer trust) and dataset characteristics (diversity, balance, provenance)
  • Explain the effects of bias and variance, overfitting and underfitting, and detection methods such as SageMaker Clarify, human audits and subgroup analysis
  • Explain the difference between transparent, explainable models and those that are not, tools such as SageMaker Model Cards, and the principles of human-centered design

Key points

  • Dimensions of responsible AI: fairness (not disadvantaging particular groups), explainability, privacy and security, safety, controllability, veracity and robustness, governance and transparency.
  • Bedrock Guardrails can be applied to an app's inputs and outputs to block harmful content, deny topics, mask personal information and detect ungrounded answers (contextual grounding checks).
  • Datasets: diverse, unbiased data with known provenance and high-quality labels. If a group has little data, prediction accuracy for that group drops (bias).
  • Overfitting (high variance) fits the training data but misses on new data. Underfitting (high bias) does not even fit the training data. Address them with regularization, more data and adjusting model complexity.
  • SageMaker Clarify measures bias in pre-training data and in trained models, and helps explain predictions with feature importance (SHAP). Combine it with human audits and per-subgroup analysis.
  • Explainability and performance can be a trade-off. Decision trees and linear models are easy to explain; deep neural networks perform well but are hard to explain. Choose based on how much accountability the use case demands.
  • Record and share a model's purpose, training data, evaluation results and limitations with SageMaker Model Cards. Make it clear to users when a decision comes from AI, and provide a feedback mechanism (human-centered design).
  • Legal risks: outputs that infringe others' copyrights, biased outputs that lead to discrimination, and wrong answers that lose trust. Manage them with terms of use, provenance checks and human review. Environmental impact (power for training and inference of large models) is also a selection factor.

Workplace uses

  • General AffairsBefore using AI for resume screening, measure bias in screening results by gender and age with Clarify, and revisit the training data and features if bias is found. A person makes the final decision.
  • SalesConfigure Bedrock Guardrails on the customer chat to restrict topics so it does not disparage competitors or promise prices, and mask personal information.
  • AccountingRecord the purpose, training data, evaluation results and limitations of the credit-scoring model in Model Cards so it can be explained during audits.

Exam objectives (from the exam guide)

  • 4.1 責任ある AI システムの開発について説明する。
  • 4.2 透明性の高い説明可能なモデルの重要性を認識する。

Learning resources

Learn in detail (2 topics with mini quizzes) →Practice this domain (42 questions) →Next domain ↓
05

Security, Compliance, and Governance for AI Solutions

14%

Learning goals

  • Explain how to protect AI systems with IAM roles and policies, encryption, Macie, PrivateLink and the shared responsibility model
  • Explain data lineage and catalogs, documentation with Model Cards, and privacy-preserving data engineering
  • Explain hallucination detection and grounding (RAG, output validation) and defenses against prompt injection
  • Explain Config, Inspector, Artifact, CloudTrail, Trusted Advisor and Audit Manager, data retention, residency and logging policies, and frameworks such as the Generative AI Security Scoping Matrix

Key points

  • Shared responsibility model: AWS is responsible for security of the cloud infrastructure; customers are responsible for their data, access permissions, app configuration and handling of prompts and outputs. The more managed the service, the more AWS covers.
  • In Bedrock, prompts and outputs are not used to improve the base models and are not shared with model providers (stated on Bedrock's official security and compliance page). Data is encrypted at rest with KMS and in transit with TLS, and PrivateLink (VPC endpoints) lets you call it without traversing the internet.
  • Least privilege: restrict which models and actions can be used with IAM policies. Record who did what and when with model invocation logging (CloudWatch Logs, S3) and CloudTrail.
  • Detect personal information in training data or RAG documents with Macie, and mask or anonymize it as needed. Manage data access with Lake Formation, and catalogs and lineage with the Glue Data Catalog and similar.
  • Hallucination countermeasures: provide supporting documents with RAG (grounding), have the model cite sources, guardrail grounding checks, output validation and human review.
  • Governance services: AWS Config (recording configuration and evaluating compliance), Inspector (vulnerabilities), Artifact (obtaining AWS third-party audit reports and compliance documents), CloudTrail (recording actions), Trusted Advisor (recommendations) and Audit Manager (collecting audit evidence).
  • Data governance: retention periods and deletion (lifecycle), the Region where data is stored (data residency), logging and monitoring, and periodic reviews. The Generative AI Security Scoping Matrix organizes the controls needed for each usage pattern (from using off-the-shelf apps to training your own models).

Workplace uses

  • AccountingFor RAG over accounting documents, call Bedrock through a VPC endpoint, encrypt knowledge base data with a KMS key, and keep model invocation logs for auditing.
  • General AffairsCreate company AI usage guidelines and organize controls per usage pattern with the Generative AI Security Scoping Matrix. For audits, obtain AWS certification reports from Artifact.
  • SalesBefore feeding customer data to generative AI, detect personal information with Macie and mask only the necessary fields.

Exam objectives (from the exam guide)

  • 5.1 AI システムを保護する方法を説明する。
  • 5.2 AI システムのガバナンスとコンプライアンス規制を認識する。

Learning resources

Learn in detail (2 topics with mini quizzes) →Practice this domain (42 questions) →Try the mock exam →

When you finish reading

Practice all domains (300 questions)Mock exam (65 questions · 90 min)AIF-C01 home
Notes (unofficial, original questions, no pass guarantee)