AI Practitioner (AIF-C01)

Domain 1: Fundamentals of AI and ML — 20%

This domain establishes the vocabulary and mechanics behind modern AI and machine learning. You’ll work with core AI concepts, machine learning approaches, common data types, model training and inference, neural networks, deep learning, and the business considerations involved in selecting an AI or ML approach.

The important distinction is knowing what problem you are solving before choosing the technology. Supervised learning uses labeled data, while unsupervised learning looks for patterns without predefined labels. Classification predicts categories; regression predicts numerical values; clustering identifies natural groupings. You’ll also need to understand the difference between training and inference, and how factors such as accuracy, precision, recall, F1 score, latency, cost, and scalability influence whether a model is appropriate for production.

Don’t approach this domain as a collection of definitions. Think about the relationship between the data, the learning approach, the model, and the business outcome. A technically accurate model can still be the wrong solution if its performance characteristics, operating cost, explainability requirements, or deployment model do not fit the use case. Expect questions that test whether you can recognize the appropriate ML approach and understand the tradeoffs involved rather than simply identify terminology.

Domain 2: Fundamentals of Generative AI — 24%

Generative AI introduces a different set of concepts built around foundation models and their ability to generate new content. You’ll encounter foundation models, large language models, tokens, embeddings, vectors, transformers, multimodal models, prompt engineering, context, and the factors that influence model selection.

The key is understanding how a generative AI system turns an input into an output. A prompt provides instructions and context, tokens represent the information processed by the model, and inference parameters can influence how the response is generated. Embeddings represent information numerically so that semantic relationships can be measured and searched. These concepts become especially important when generative AI applications need to work with information that was not part of the model's original training data.

You’ll also need to recognize where foundation models are useful and where they introduce risk. Generative AI can summarize, classify, transform, generate, and reason over information, but it can also produce hallucinations, expose sensitive information, or behave unpredictably when given ambiguous or adversarial instructions. Model selection is therefore more than choosing the largest or most capable model. Consider the required modality, quality, latency, context length, cost, availability, and intended use. The strongest answer is usually the one that balances capability with the requirements of the application.

Domain 3: Applications of Foundation Models — 28%

This is the largest domain on the AIF-C01 exam and focuses on putting foundation models to work. You’ll work with prompt engineering, inference parameters, retrieval-augmented generation, vector stores, knowledge bases, model customization, agents, evaluation, and the AWS services used to build generative AI applications.

RAG is particularly important. A foundation model does not automatically know the current contents of an organization's documents or databases. A RAG architecture retrieves relevant information from an external source and provides that information as context during inference. Understand the flow: source data is prepared and chunked, embeddings are generated, content is stored for retrieval, relevant information is returned for a query, and that information is incorporated into the model's prompt. RAG improves access to external knowledge without requiring the model itself to be retrained.

You’ll also need to distinguish prompting, RAG, fine-tuning, continued pre-training, and other customization approaches. Prompt engineering changes the instructions and context supplied to the model. RAG supplies external information at inference time. Fine-tuning uses additional training data to adapt model behavior for a particular task or domain. These approaches solve different problems and have different cost, maintenance, and data requirements.

AWS services are a major part of this domain. Know the roles of Amazon Bedrock, Knowledge Bases, Agents, Guardrails, Prompt Management, and model evaluation capabilities, as well as the role of Amazon SageMaker AI when building and operating machine learning solutions. Newer agentic-AI concepts are also important, including tool use, memory, orchestration, multi-agent patterns, and the Model Context Protocol. Focus on what each capability is designed to accomplish rather than memorizing service names in isolation.

Domain 4: Guidelines for Responsible AI — 14%

Responsible AI is about ensuring that AI systems are useful without creating unacceptable risks for the people and organizations that depend on them. You’ll need to understand fairness, inclusivity, robustness, safety, veracity, transparency, explainability, and the role of humans in the design and operation of AI systems.

Bias can enter an AI system through training data, labels, model design, evaluation methods, or the way outputs are ultimately used. A dataset that does not adequately represent the population can produce systematically different outcomes for different groups. Strong responsible-AI practices therefore include representative data, appropriate evaluation, bias detection and mitigation, continuous monitoring, and human review where the consequences justify it.

Generative AI adds its own challenges. Hallucinations can produce convincing but incorrect information. Toxicity can result in harmful or inappropriate content. Models may reveal sensitive information or respond to attempts to bypass their safeguards. Grounding, guardrails, output validation, human oversight, and appropriate evaluation all help reduce these risks.

Don’t treat responsible AI as a separate activity performed after the model has been built. It is a lifecycle concern. Consider who is affected by the system, what could go wrong, how those risks can be measured, and what controls are available before the system is placed into production. AWS capabilities such as Amazon SageMaker Clarify, SageMaker Model Monitor, Amazon A2I, Model Cards, and Amazon Bedrock Guardrails support different parts of this process.

Domain 5: Security, Compliance, and Governance for AI Solutions — 14%

This domain applies familiar AWS security and governance principles to AI workloads while introducing threats and controls that are particularly relevant to generative AI. You’ll work with identity and access management, encryption, data protection, privacy, logging, compliance, governance, and AI-specific attacks such as prompt injection, jailbreaking, data poisoning, adversarial inputs, and sensitive-information disclosure.

The fundamental security questions remain the same: Who can access the system? What can they access? What data are they allowed to use? How is that data protected? What happened when something went wrong? IAM provides identity and permissions, KMS manages cryptographic keys, and least privilege limits access to what is actually required. Encryption protects data at rest and in transit, while services such as CloudTrail, CloudWatch, AWS Config, and Audit Manager provide different forms of monitoring, configuration assessment, and audit support.

Generative AI changes the threat surface. A prompt may contain sensitive information. Retrieved documents may contain malicious instructions. A user may deliberately attempt to bypass model restrictions. Model outputs may expose information that should never have been returned. Understand the difference between threats such as prompt injection, jailbreaking, data poisoning, model theft, and adversarial attacks, along with the controls used to reduce their impact.

Governance provides the framework that keeps these controls sustainable. Establish appropriate data handling rules, access controls, logging, monitoring, audit processes, ownership, and review procedures. AWS Artifact provides access to AWS compliance documentation, while AWS Audit Manager helps collect evidence for audits. Amazon Macie helps discover sensitive data in Amazon S3. Amazon Bedrock Guardrails can apply configurable safeguards to generative AI inputs and outputs. The goal is not simply to secure an AI model; it is to operate an AI solution that remains secure, compliant, observable, and accountable throughout its lifecycle.