Content domains
The 5 domains of AIF-C01, with every task statement and its objectives from the official guide. Study a whole domain, or drill a single task.
1Fundamentals of AI and ML
20% of examStudy domain
1.1Explain basic AI concepts and terminologies23 q>
Knowledge of
- Core vocabulary of the field: AI, ML, deep learning, neural networks, computer vision, natural language processing (NLP), models, algorithms, training, and inferencing
- What separates AI, ML, deep learning, generative AI, and agentic AI from one another, and how they nest inside each other
- Bias, fairness, and model fit (including overfitting and underfitting) as descriptions of model behavior
- Large language models (LLMs) and where they sit within the broader family of foundation models
- The inferencing styles a deployed model can use: batch, real-time, asynchronous, and serverless
- Data shapes that feed AI systems: labeled compared with unlabeled, structured compared with unstructured, tabular, time-series, image, and text
- The three broad learning paradigms: supervised, unsupervised, and reinforcement learning
Skills in
- Using AI/ML terminology precisely and correctly in conversation with technical teams
- Placing a given technology in the right category (for example, recognizing an LLM chatbot as generative AI rather than classical ML)
- Matching an inference style to a workload's latency, throughput, and cost profile
- Classifying a dataset by type and by whether it carries labels
- Picking the learning paradigm that fits the data and the problem at hand
1.2Identify practical use cases for AI23 q>
Knowledge of
- Where AI/ML pays off: augmenting human judgment, scaling a solution, and automating repetitive work
- Where AI/ML is the wrong tool, such as when a deterministic answer is required or when the cost outweighs the benefit
- The classic technique families and what each one predicts: regression, classification, and clustering
- Real deployments of AI in the wild, including computer vision, NLP, speech recognition, recommendation engines, fraud detection, forecasting, knowledge bases, and agentic AI
- What the AWS managed AI services do, including Amazon SageMaker AI, Amazon Transcribe, Amazon Translate, Amazon Comprehend, Amazon Lex, and Amazon Polly
- The tradeoffs that push a use case toward a traditional ML model instead of a foundation model, such as regulatory pressure, explainability needs, and operational limits
Skills in
- Judging whether a business problem is genuinely a good fit for AI/ML
- Making the case against AI/ML when a rules engine or a simple calculation would serve better
- Selecting regression, classification, or clustering to match a stated outcome
- Mapping a business requirement onto the AWS managed AI service that already solves it
- Weighing a foundation model against a purpose-built traditional model for a specific set of constraints
1.3Describe the AI/ML development lifecycle23 q>
Knowledge of
- The stages of an AI/ML pipeline, from data collection and preparation through training, evaluation, deployment, and monitoring
- Where foundation models come from: open source pre-trained checkpoints compared with custom-trained models
- Ways to serve a model in production, including a managed API service and a self-hosted API
- The AWS services that cover each pipeline stage, including Amazon Bedrock, Amazon Q, Amazon Quick, Kiro, and SageMaker AI
- MLOps fundamentals: experimentation, repeatable processes, scalable systems, technical debt, production readiness, model monitoring, and retraining
- Model quality metrics (accuracy, precision, recall, F1 score) and business metrics (cost per user, development cost, customer feedback, ROI)
Skills in
- Walking through the stages of an ML pipeline and explaining what each one contributes
- Choosing between a pre-trained model and a custom-trained one for a given need
- Deciding whether a managed API or a self-hosted endpoint better serves a deployment
- Assigning the right AWS service to each stage of a proposed pipeline
- Reading model metrics and business metrics together to judge whether a model is worth shipping
2Fundamentals of GenAI
24% of examStudy domain
2.1Explain the basic concepts of generative AI (GenAI)23 q>
Knowledge of
- The building blocks of generative AI: tokens, chunking, embeddings, vectors, and prompt engineering
- Model architectures behind GenAI, including transformer-based LLMs, foundation models, multi-modal models, and diffusion models
- What GenAI gets used for: image, video, and audio generation, summarization, AI assistants, translation, code generation, customer service agents, search, and recommendations
- The foundation model lifecycle: data selection, model selection, pre-training, fine-tuning, evaluation, deployment, and feedback
- Token-based pricing and how token counts drive both cost and inference performance
- Context engineering and the part it plays in getting useful output from a foundation model
- Agentic AI concepts: multi-agent system patterns, the Model Context Protocol (MCP) for wiring agents to external systems, inter-agent communication, memory management, tool usage, and workflow orchestration
Skills in
- Explaining how text becomes tokens and embeddings before a model ever sees it
- Spotting workloads where a generative model is the natural fit
- Tracing a foundation model from raw data through to a deployed, feedback-driven system
- Estimating how prompt and response length will move the cost of a token-priced workload
- Describing how an agent uses memory, tools, and MCP connections to finish a multi-step task
2.2Understand the capabilities and limitations of GenAI for solving business problems23 q>
Knowledge of
- What GenAI is good at: adapting to new tasks, responding conversationally, and producing original content
- Where GenAI falls down: hallucinations, weak interpretability, factual errors, and nondeterministic output
- The criteria that drive model selection, including model type, performance targets, capabilities, constraints, compliance, cost, latency, and complexity
- Metrics that show whether a GenAI application earns its keep: cross-domain performance, ROI, efficiency, conversion rate, average revenue per user, accuracy, and customer lifetime value
Skills in
- Setting realistic expectations with stakeholders about what a generative model can and cannot deliver
- Recognizing hallucination and nondeterminism as risks that need mitigation rather than surprises
- Comparing candidate models against cost, latency, and compliance requirements at once
- Defining the business metrics that will prove or disprove the value of a GenAI project
2.3Describe AWS infrastructure and technologies for building GenAI applications23 q>
Knowledge of
- The AWS building blocks for GenAI applications: Amazon Bedrock, Amazon SageMaker AI, SageMaker JumpStart, Amazon Quick, Kiro, Strands Agents, and Amazon Bedrock AgentCore
- Why teams reach for managed AWS GenAI services: accessibility, a lower barrier to entry, efficiency, cost-effectiveness, and speed to market
- What the AWS platform contributes underneath a GenAI workload in terms of security, compliance, responsibility, and safety
- The cost levers and tradeoffs of AWS GenAI services, including responsiveness, availability, redundancy, performance, Regional coverage, token-based pricing, provisioned throughput, and custom models
Skills in
- Choosing the AWS service that matches a team's skill level and delivery timeline
- Explaining the advantage of a managed foundation model API over self-hosting
- Deciding between on-demand token pricing and provisioned throughput for a given traffic pattern
- Accounting for Regional availability and redundancy when planning a GenAI deployment
3Applications of Foundation Models
28% of examStudy domain
3.1Describe design considerations for applications that use foundation models (FMs)23 q>
Knowledge of
- The criteria for picking a foundation model: cost, modality, latency, multi-lingual support, model size, complexity, customization options, input and output length, and prompt caching
- How inference parameters such as temperature and input/output length reshape a model's responses
- Retrieval Augmented Generation (RAG) and the business problems it solves, including via Amazon Bedrock Knowledge Bases
- AWS options for storing embeddings in a vector database: Amazon OpenSearch Service, Amazon Aurora, Amazon Neptune, and Amazon RDS for PostgreSQL
- The cost profile of each customization approach: pre-training, fine-tuning, in-context learning, RAG, and model distillation
- What AI agents are and the business applications they unlock
Skills in
- Matching a foundation model to an application's modality, latency, and budget requirements
- Tuning temperature and length parameters to get more deterministic or more creative output
- Designing a RAG workflow that grounds model answers in an organization's own documents
- Selecting a vector store that fits an existing data platform
- Ranking customization approaches by cost and effort before committing to fine-tuning
3.2Choose effective prompt engineering techniques23 q>
Knowledge of
- The parts of a prompt: context, instruction, and negative prompts
- Prompting techniques including chain-of-thought, zero-shot, single-shot, few-shot, and prompt templates
- Best practices that lift response quality: experimentation, guardrails, discovery, specificity, concision, and layering multiple comments
- Prompt-level attack surface: exposure, poisoning, hijacking, and jailbreaking
- Versioning and managing prompts as artifacts with Amazon Bedrock Prompt Management
Skills in
- Writing prompts that state context and instructions clearly enough to get consistent output
- Reaching for few-shot examples or chain-of-thought when a task needs reasoning or a specific format
- Iterating on a prompt and measuring whether each change actually improved the response
- Recognizing prompt injection and jailbreak attempts and applying guardrails against them
- Keeping prompts versioned and reviewable instead of scattered through application code
3.3Describe the training and fine-tuning process for FMs23 q>
Knowledge of
- The stages of building a foundation model: pre-training, fine-tuning, continuous pre-training, and distillation
- Fine-tuning methods including instruction tuning, domain adaptation, transfer learning, and continuous pre-training
- What good fine-tuning data looks like: curated, governed, appropriately sized, labeled, and representative
- Reinforcement learning from human feedback (RLHF) and how human preference shapes model behavior
Skills in
- Distinguishing pre-training from fine-tuning in terms of purpose, data volume, and cost
- Selecting a fine-tuning method that suits a domain-specific goal
- Assembling a fine-tuning dataset that is representative and free of obvious gaps
- Explaining where human feedback enters the loop and why it matters for alignment
3.4Describe methods to evaluate FM performance3 q>
Knowledge of
- Evaluation approaches: human-in-the-loop review, benchmark datasets, and Amazon Bedrock Model Evaluation
- The standard scoring metrics: ROUGE, BLEU, BERTScore, and LLM-as-a-judge
- How to tell whether a model actually advances a business objective such as productivity, user engagement, or task completion
- Evaluating whole FM-backed applications, not just the model: RAG pipelines, agents, and workflows
- Business alignment metrics including task completion rate, user satisfaction, and cost per interaction
Skills in
- Designing an evaluation that mixes automated scoring with human review
- Picking the metric that matches the task, such as ROUGE for summarization or BLEU for translation
- Using an LLM as a judge while staying aware of its blind spots
- Evaluating a RAG or agent application end to end rather than testing the model alone
- Tying evaluation results back to the business objective that funded the project
4Guidelines for Responsible AI
14% of examStudy domain
4.1Explain the development of AI systems that are responsible23 q>
Knowledge of
- The dimensions of responsible AI: bias, fairness, inclusivity, robustness, safety, and veracity
- Tooling that enforces responsible behavior at runtime, such as Amazon Bedrock Guardrails
- Responsible model selection, including environmental footprint and sustainability
- Legal exposure from generative AI: intellectual property claims, biased outputs, loss of customer trust, end user risk, and hallucinations
- What healthy datasets look like: inclusive, diverse, curated, and balanced
- How bias and variance surface as harm to demographic groups, inaccuracy, overfitting, and underfitting
- Tools for detecting and monitoring bias and truthfulness: label quality analysis, human audits, subgroup analysis, Amazon SageMaker Clarify, SageMaker Model Monitor, and Amazon Augmented AI (Amazon A2I)
Skills in
- Auditing a dataset for gaps and imbalances that would produce biased predictions
- Applying guardrails to block harmful or off-topic model output
- Identifying the legal and reputational risks a GenAI feature introduces
- Reading bias and variance symptoms in a model's behavior and naming the likely cause
- Choosing the AWS tool that fits a given bias-detection or monitoring need
4.2Recognize the importance of transparent and explainable models3 q>
Knowledge of
- What sets a transparent, explainable model apart from an opaque one
- Tools that surface model behavior and provenance: Amazon SageMaker Model Cards, SageMaker Clarify, Amazon Bedrock Model Evaluations, plus open source models, data, and licensing
- The tension between model safety and transparency, and how interpretability trades against raw performance
- Human-centered design principles for explainable AI, including user-feedback mechanisms and transparency about AI decisions
Skills in
- Judging when a use case demands an interpretable model instead of the highest-scoring one
- Documenting a model's intent, data, and limitations for downstream reviewers
- Explaining an automated decision in terms a non-technical user can act on
- Building feedback paths that let users challenge or correct AI output
5Security, Compliance, and Governance for AI Solutions
14% of examStudy domain
5.1Explain methods to secure AI systems4 q>
Knowledge of
- AWS services and features that secure AI workloads: IAM roles, policies, and permissions; encryption; Amazon Macie; AWS PrivateLink; the AWS shared responsibility model; Amazon Bedrock AgentCore Identity; Policy in AgentCore; and Amazon Bedrock Guardrails
- Source citation and provenance practices, including data lineage, data cataloging, and Amazon SageMaker Model Cards
- Secure data engineering practices: data quality assessment, privacy-enhancing technologies, data access control, and data integrity
- Security and privacy risks specific to AI systems, including prompt injection, data leakage, toxicity, encryption at rest and in transit, output filtering and validation, threat detection, vulnerability management, infrastructure protection, and audit logging of AI interactions
- Techniques that keep output honest: RAG grounding, output validation, confidence scoring, and hallucination detection
Skills in
- Applying least-privilege IAM to the services and data an AI workload touches
- Keeping model traffic off the public internet with private connectivity
- Recognizing prompt injection and data leakage as first-class threats and mitigating them
- Recording where training and retrieval data came from so answers can be traced
- Grounding responses in retrieved sources and validating output before it reaches a user
5.2Recognize governance and compliance regulations for AI systems3 q>
Knowledge of
- AWS services that support governance and regulatory compliance: AWS Config, Amazon Inspector, AWS Audit Manager, AWS Artifact, AWS CloudTrail, and AWS Trusted Advisor
- Data governance strategies covering data lifecycles, logging, residency, monitoring, observation, and retention
- Processes that keep governance real: policies, review cadence, review strategies, transparency standards, and team training requirements
- Governance frameworks for generative AI, notably the Generative AI Security Scoping Matrix
Skills in
- Selecting the AWS service that produces the audit evidence a given regulation demands
- Defining retention and residency rules for the data an AI system consumes and generates
- Setting a review cadence that keeps deployed models under ongoing scrutiny
- Placing a GenAI workload in the right scope of the Generative AI Security Scoping Matrix
- Identifying the training a team needs before it can operate an AI system responsibly