AWS AIF-C01 AI Practitioner: Exam Domains and AI Literacy

The AWS Certified AI Practitioner (AIF-C01) examination is designed for people who need to understand what artificial intelligence can and cannot accomplish in a business or cloud environment. It does not require the depth of an engineer training large models from scratch. It does require clear distinctions among AI, machine learning and generative AI; knowledge of model selection and evaluation; awareness of security and responsible-use risks; and the ability to recognize where AWS services can support a particular workload.

AIF-C01 is a current foundational AWS Certification as of October 2026. The certification sits within Amazon's AWS offering: Amazon is the vendor, while AWS is the certification and cloud-program designation. AWS lists 65 questions and 90 minutes for the exam. Its official guide was updated in 2026, including version 1.1 published April 30, 2026, so candidates should use the current task statements rather than assume a launch-era summary includes all material now relevant to agentic AI, foundation models and cloud controls.

The five weighted areas of AIF-C01

The current AWS exam guide divides scored content into 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%). The weighting shows that application of foundation models accounts for the largest share of scored material, but candidates should not neglect the governance and security domains. A use case that appears technically feasible may be inappropriate if its data handling, reliability or legal constraints are not addressed.

AWS describes the target candidate as someone demonstrating foundational understanding of AI and AWS AI tools. Many examination scenarios are written from an organizational perspective: a company wants to summarize documents, classify support requests, forecast demand or help employees search internal knowledge. The right response depends on whether the requirement calls for prediction, generation, retrieval, rule-based automation or human judgment. Memorizing service names without identifying the actual business problem produces weak choices.

The difference from AWS Cloud Practitioner (CLF-C02) is one of subject focus. Cloud Practitioner validates broad cloud concepts, shared responsibility, pricing and core services. AI Practitioner concentrates on AI and ML concepts, foundation-model applications and their particular safety and governance implications. Neither certification should be described as a mandatory prerequisite for the other solely because one can provide useful context.

AI, machine learning and generative AI are different levels of description

Artificial intelligence is the broad field of systems that perform tasks associated with intelligent behavior. Machine learning (ML) is a family of techniques in which patterns are learned from data rather than expressed entirely as manually written rules. Deep learning uses multilayer neural-network architectures to learn complex representations. Generative AI (GenAI) focuses on producing content such as text, images, audio or code, often using a foundation model trained on broad data. These categories overlap, but they are not interchangeable labels for every automated system.

A fraud-screening product can use rules, supervised ML or a combination. A demand forecast may use statistical methods or ML regression. A document-summarization assistant typically uses language-generation capabilities. Choosing a generative model for a simple deterministic tax calculation may increase cost and introduce avoidable errors. Conversely, a rigid keyword filter may be inadequate when the task involves paraphrased customer intent. The practical skill is aligning the technique with data, expected outputs, tolerable errors and governance requirements.

AWS includes supervised, unsupervised and reinforcement-learning concepts in its current objectives. Supervised learning uses labeled examples to learn a mapping, such as predicting a category or value. Unsupervised methods explore patterns without predefined target labels, such as clustering similar observations. Reinforcement learning concerns policies learned through interaction and feedback. Candidates should be able to recognize the appropriate family from a scenario without assuming that all AI systems follow the same training process.

The AI lifecycle begins before model selection

A viable AI initiative starts by defining the business decision and success criteria. What should the output accomplish? Who uses it? What data are available and legally usable? What happens when the system is wrong? A support-ticket classifier might be measured by correct routing and the cost of misrouted urgent requests; a content assistant might be measured by factual consistency, appropriate tone and ability to abstain when the source material is missing. “It produces fluent text” is not a sufficient success criterion.

Data preparation has a strong effect on performance. Labeled datasets may contain inconsistencies, missing groups or historic biases. Training and evaluation data should be separated appropriately to avoid measuring performance on examples the model has effectively already seen. In deployment, data may change with customer behavior, products or seasonal effects. Such drift can reduce performance even when the underlying code stays the same. Monitoring and periodic reassessment are part of the lifecycle, not tasks reserved solely for advanced machine-learning engineers.

Inference is the stage at which a model processes new input and produces an output. Depending on the business need, inference can be real-time, asynchronous, batched or served through managed infrastructure. That decision changes latency, throughput and cost. A company generating overnight document classifications can optimize differently from one that needs a low-latency customer-facing response. AIF-C01 tests comprehension of these trade-offs more than deployment commands.

Foundation models, prompts and retrieval

A foundation model is trained on broad data and can support multiple downstream tasks. Language models predict or generate token sequences in response to prompts and context; multimodal models can work with additional types of input or output. Models differ in capabilities, supported modalities, context windows, latency, cost and safety features. A sensible selection begins with use-case tests and operational constraints, not an assumption that the largest available model is always the best.

Prompting directs the model's behavior through instructions, examples and context. Effective prompting clarifies the role, task, format, constraints and available evidence. However, a prompt cannot guarantee factual accuracy or prevent all unsafe behavior. One-shot or few-shot examples may improve consistency, but poor examples can also steer outputs in the wrong direction. Evaluation is required to determine whether an improvement is reliable across representative inputs.

Retrieval-augmented generation (RAG) supplies relevant external information to a model at inference time, often through a retrieval layer that searches indexed documents or vector representations. Embeddings map data to numeric representations that support similarity comparison. RAG can help answer questions about a private and frequently changing knowledge base without retraining the entire foundation model. It is not automatic fact checking: retrieval can return irrelevant, stale, confidential or manipulated passages, and the model can still fabricate unsupported statements. Foundation models and generative AI are therefore best understood alongside retrieval quality and evidence-grounded evaluation.

Consider an employee assistant that answers questions about an organization's internal policies. Fine-tuning might change style or specialized behavior, but updating the policy corpus through controlled retrieval may be more appropriate for frequently changing factual material. The solution should track document provenance, access permissions and effective dates. Users should see uncertainty when the indexed evidence does not support an answer. Those requirements matter more than the label of a fashionable model technique.

Choosing AWS services for the actual workload

Amazon Bedrock provides managed access to foundation models and capabilities for building generative AI applications. It can support prompting, model invocation, knowledge-base patterns and other application components. Amazon SageMaker AI supports a broader set of machine-learning development, training and deployment workflows. The services can coexist in an organization's architecture, but neither should be selected merely because it contains the term “AI.” Sound AI service selection for business use cases asks who owns the model-development process, how much customization is required, where data reside and what operational controls the team can maintain.

For a business team wanting to evaluate foundation-model summaries, a managed foundation-model service may reduce the infrastructure work required to start. A specialist team developing a predictive model from structured proprietary data may need a different ML workflow. Candidates should understand this distinction without being expected to implement production inference pipelines. Those deeper engineering responsibilities are more closely associated with AWS Machine Learning Engineer – Associate (MLA-C01).

AWS also offers purpose-built AI services for tasks such as extracting document information, understanding language or processing speech. Choosing a managed specialized service instead of a general-purpose foundation model may improve predictability and simplify integration for constrained tasks. Conversely, a flexible foundation model may be useful where inputs and outputs vary substantially. Compare options against accuracy, operational effort, cost, privacy, integration and the need for human review.

Evaluating outputs and managing hallucinations

Generative models can produce plausible language that is false or not supported by the prompt and retrieved evidence. This phenomenon is often called hallucination. The risk is consequential because a confident response can be mistaken for a verified answer. An assistant that invents a product warranty condition or cites a nonexistent policy section can cause real business harm even if most of its output sounds useful.

Evaluation should therefore use representative data and metrics suited to the task. For classification, precision and recall may matter, especially if missing a harmful case is more costly than creating a false alert. For generation, human or model-assisted assessments may examine factual grounding, relevance, completeness, safety and consistency. No single automated score establishes all aspects of quality. A model can be fluent yet unfaithful to its source, or factually accurate but too slow and expensive to meet business requirements.

A sensible evaluation design includes difficult edge cases and records why an output passes or fails. For a RAG assistant, measure retrieval relevance separately from generation quality. If the answer is wrong because no appropriate document was retrieved, rewriting the prompt may not fix the problem. If the correct document was retrieved but the model ignored it, a different intervention is needed. Understanding the failure layer is more valuable than memorizing model names or copying a successful demonstration prompt.

Responsible AI is a design requirement

Responsible AI includes fairness, transparency, explainability, privacy, safety, accountability and meaningful human oversight. Bias may arise from unrepresentative data, biased labels, deployment context or feedback loops, not only from intentionally discriminatory code. An AI-assisted process affecting people should be assessed for disparate error patterns and whether affected users can understand or challenge consequential outcomes where appropriate.

Transparency helps users know that a system is AI-assisted and what it can reliably do. Explainability concerns why particular outputs or recommendations were made, but the attainable level differs by model and task. A concise cited explanation may support a document assistant; a high-stakes risk model may need a more rigorous validation and governance framework. Explanations should not be fabricated after the fact merely to create an appearance of certainty.

The site's responsible AI concepts for practitioners provide context for these decisions. Candidates should recognize that fairness and safety cannot be treated as post-launch public relations exercises. They influence data selection, model evaluation, guardrail placement, escalation design and acceptable deployment contexts. A use case can be rejected or narrowed because it cannot meet those requirements, even if the model's technical output looks impressive.

Security, governance and the boundaries of guardrails

AI applications inherit familiar cloud-security concerns—identity, access control, encryption, logging, data classification and retention—while introducing new AI security risks. Prompts may carry sensitive data; retrieved documents may be visible to users who are not entitled to them; model responses may reveal confidential information; and external content can contain prompt injection attempts that try to override trusted instructions. A model receiving text from a web page or document should not treat that text as an authoritative system instruction.

Amazon Bedrock Guardrails can help apply content and safety controls to supported model workflows. These features are useful but do not replace correct authorization, secure data retrieval, sound application logic or human oversight. A team building RAG must enforce the requesting user's access permissions when selecting documents, rather than retrieving an entire internal repository and hoping the language model keeps secrets. The relationship between Bedrock guardrails and content safety and ordinary security controls is an important architectural boundary.

The AWS shared-responsibility model remains relevant, but candidates should distinguish the cloud provider's infrastructure responsibilities from the customer's responsibilities for data, identities, application behavior, configurations and legal obligations. AI governance should assign model ownership, approval standards, data-use rules, incident procedures and monitoring. It should also record which controls apply to a managed service and which must be implemented in the customer's application. No service-selection decision transfers all accountability to the provider.

A practical path through AIF-C01 scenarios

The best study scenarios start with a business need, not with the answer already encoded in a service name. A retailer wants to forecast product demand: is a predictive ML approach more appropriate than unconstrained text generation? A law firm wants to find information in authorized internal documents: what retrieval and access controls are necessary? A support team wants consistent message summaries: how should the company test omissions, incorrect details and privacy exposure? An organization wants to automate a consequential decision: which human review and fairness concerns change the design?

For each case, identify data type, task type, acceptable error, evaluation measure, cost or latency constraint, security boundary and responsible owner. Then select the suitable category of AWS capability. This produces a mental model that remains useful even when service offerings change. The distinction between conceptual use-case analysis and production-level coding can be explored further through AWS Generative AI Developer – Professional (AIP-C01), a more implementation-focused certification rather than a required follow-on to AIF-C01.

Use the current AWS AIF-C01 guide—particularly its April 2026 revision notes—to review domain coverage before scheduling. Prioritize the skills with the greatest conceptual gaps: model-family distinctions, retrieval and prompting, evaluation, responsible AI and security boundaries. The essential exam competency is the ability to make a proportionate decision about AI use: identify what will solve the problem, describe why it is appropriate, and recognize the conditions under which it should not be deployed.