AIF-C01 Exam Guide: Study for AWS Certified AI Practitioner
AIF-C01 validates a foundational ability to explain AI, ML, and generative AI concepts, choose suitable AWS technologies for business problems, and consider responsible use. It is aimed at people who use AI/ML services on AWS rather than necessarily build models or pipelines. This guide helps you decide whether the exam fits your current experience, identify the concepts that need deliberate study, and build a preparation plan based on the published AWS content outline.
Decide whether AIF-C01 fits your role
AIF-C01 is a foundational AWS certification for people who need to discuss AI and generative AI use cases, service choices, value, risk, and governance without being expected to implement models or production ML infrastructure.
AWS describes the target candidate as having up to 6 months of exposure to AI/ML technologies on AWS. The intended candidate uses AI/ML solutions but does not necessarily build them. That makes the exam a sensible fit for business analysts, product and project roles, technical sales or customer-facing staff, cloud practitioners expanding into AI, and early-career technologists who need a structured AWS AI vocabulary.
The boundary matters when choosing study material. AWS lists developing or coding AI/ML models and algorithms, data or feature engineering, hyperparameter tuning, model optimization, and building or deploying AI/ML pipelines as out of scope. You should understand why an approach may be selected and what its tradeoffs are; you do not need to train a model or write a pipeline to prepare effectively.
Before booking, test your starting point with a few plain-language prompts. Can you distinguish a prediction task from a generative task? Can you explain why a business might prefer a managed service? Can you identify when a fixed rule is more appropriate than an AI prediction? Can you connect IAM, pricing models, and the shared responsibility model to an AI workload? A confident answer to none of these is not a reason to avoid the exam, but it is a reason to schedule study before scheduling the test.
AWS recommends familiarity with core AWS services and their use cases, including Amazon EC2, Amazon S3, AWS Lambda, Amazon Bedrock, and Amazon SageMaker AI. It also recommends familiarity with IAM, AWS pricing models, and the shared responsibility model. Treat those as foundation topics, not optional background reading.
Know what the certification validates
The exam tests decision-making language around AI on AWS: explaining concepts, matching technologies to a business problem, recognizing limitations, and using AI responsibly.
AWS says the exam validates the ability to describe AI, ML, and GenAI concepts, methods, and strategies; select appropriate technologies for business problems; determine suitable types of AI/ML technology for use cases; and use AI, ML, and GenAI technologies responsibly. This is why scenario-based study is more valuable than a disconnected glossary.
A useful way to interpret the objective is to separate four questions a candidate should be able to answer. What kind of problem is this? What AI or GenAI approach is appropriate? What AWS service or capability supports that approach? What cost, safety, privacy, quality, or governance consequence must be considered?
For example, a question about a customer support assistant is not merely a request to name a service. It may require you to recognize a GenAI use case, identify a foundation-model application design consideration, account for inaccurate output, consider grounding or output validation, and connect the solution to a meaningful measure such as task completion or user satisfaction. The official outline explicitly includes business-alignment metrics for AI applications.
Do not prepare as though every business requirement should lead to AI. The published objectives include determining when AI/ML solutions are not appropriate, such as when cost-benefit analysis does not justify it or when a specific outcome is required instead of a prediction. In study notes, record both the recommended approach and the reason an alternative is unsuitable.
Use the published domain weights to allocate effort
Prioritize foundation-model applications first, then build enough AI/ML and GenAI fundamentals to explain the choices made in those applications. The published weights should guide revision time, but they should not become a reason to neglect any domain.
Content Domain 3, Applications of Foundation Models, represents 28% of scored content. This is the largest published domain and includes selecting foundation models, prompt engineering, training and fine-tuning concepts, and performance evaluation. Give it the most repeated practice because it draws on terminology and decision criteria from the other domains.
Content Domain 2, Fundamentals of GenAI, represents 24% of scored content. Its scope covers core GenAI concepts, capabilities and limitations, business value, model-selection factors, and AWS infrastructure and technologies for GenAI applications. Learn its terms well enough to explain them in a business scenario rather than as isolated definitions.
Content Domain 1, Fundamentals of AI and ML, represents 20% of scored content. This is the conceptual base for AI terminology, use cases, the development lifecycle, inference types, data types, learning types, and performance and business metrics.
Content Domain 4, Guidelines for Responsible AI, represents 14% of scored content. Content Domain 5, Security, Compliance, and Governance for AI Solutions, also represents 14% of scored content. Both need focused preparation because responsible design, transparent use, security, privacy, data governance, and compliance considerations can affect the correct choice in a scenario.
A practical recommendation is to make a coverage tracker with one row for every published task statement and objective cluster. Mark each item as unfamiliar, understood, or ready to apply. Use the domain weights to choose how often topics recur in revision, while using the tracker to prevent a familiar high-weight domain from hiding gaps in responsible AI or governance.
Build the AI and ML foundation before specializing
Start with the distinctions between AI, ML, deep learning, GenAI, foundation models, and agentic AI, then apply those distinctions to data, inference, learning methods, use cases, and lifecycle decisions.
Domain 1 requires you to explain core concepts and terminology. The published examples include AI, ML, deep learning, neural networks, computer vision, natural language processing, training, inferencing, bias, fairness, fit, large language models, GenAI, and agentic AI. A memorized definition is not enough if you cannot identify how a term affects a use case.
Create comparison cards instead of a long glossary. One card can contrast supervised learning, unsupervised learning, and reinforcement learning by the kind of learning described. Another can distinguish labeled and unlabeled data, then add tabular, time-series, image, text, structured, and unstructured data. A third can compare batch, real-time, asynchronous, and serverless inference by the operational need implied in a scenario.
Practice selecting the broad technique, not implementing it. The objectives name regression, classification, and clustering as examples for matching techniques to use cases. In your notes, state the business question first: predict a numeric value, assign a category, or find groups with shared characteristics. That habit reduces confusion when a question includes distracting AWS service names.
The lifecycle topic needs a similarly practical treatment. Be able to recognize stages and the AWS services or features relevant to a stage, while keeping the exam boundary in mind. AWS includes concepts such as experimentation, repeatable processes, scalable systems, managing technical debt, production readiness, model monitoring, and model retraining under fundamental MLOps concepts. It does not require you to build those systems.
Know both technical and business measures. The outline names accuracy, precision, recall, and F1 score as model metrics, as well as cost per user, development costs, customer feedback, and return on investment as business metrics. When reviewing a metric, write the decision it informs. This prevents treating a model-quality measure as a substitute for whether an application meets its business objective.
Learn GenAI concepts through business constraints
Domain 2 rewards candidates who can connect GenAI vocabulary to limits, costs, user needs, and AWS capabilities rather than treating a foundation model as the answer to every content problem.
Study tokens, chunking, embeddings, vectors, prompt engineering, transformer-based large language models, foundation models, multimodal models, and diffusion models as a connected set. The value of this approach is that it explains a workflow: content can be divided into chunks, represented as embeddings, retrieved using vector search, supplied as context, and used by a model to generate a response.
The outline also includes the foundation-model lifecycle: data selection, model selection, pre-training, fine-tuning, evaluation, deployment, and feedback. Make a one-page lifecycle map that identifies the purpose of each stage and the questions a business stakeholder would ask. For instance, model selection concerns capability and constraints; evaluation concerns quality against the intended outcome; feedback supports improvement after deployment.
Do not overlook agentic AI concepts. AWS includes multi-agent system patterns, the Model Context Protocol, multi-agent communication patterns, memory management, tool use, and workflow orchestration within the GenAI fundamentals objectives. Learn the function of each concept at a foundational level and avoid turning it into a coding exercise.
Model choice is a tradeoff exercise. The official objectives identify model types, performance requirements, capabilities, constraints, compliance, cost, latency, and model complexity as selection factors. When you answer practice scenarios, make a small decision table: requirement, relevant factor, preferred capability, and risk or tradeoff. It trains you to identify the fact that actually resolves the scenario.
GenAI limitations are equally important. AWS explicitly identifies hallucinations, interpretability, inaccuracy, and nondeterminism among disadvantages to consider. A strong answer does not assume these concerns make GenAI unusable; it identifies a suitable safeguard or a situation where a different approach is better.
Cost deserves a dedicated review pass. The outline includes token-based pricing and its effect on inference cost and performance, as well as tradeoffs involving responsiveness, availability, redundancy, regional coverage, provisioned throughput, and custom models. Keep the relationship clear: a business requirement may increase cost or complexity, so the appropriate choice is the one that meets the stated need rather than the one with the longest feature list.
Master foundation-model application decisions
Foundation-model application design is the center of AIF-C01 preparation: choose a model and interaction pattern that meet the stated business requirement while accounting for quality, cost, safety, and operational constraints.
AWS lists model-selection criteria such as cost, modality, latency, multilingual support, model size, model complexity, customization, input and output length, and prompt caching. Build scenario drills that force a choice between these criteria. A multilingual requirement points to a different consideration than a short response-time requirement; a need for images or audio points to modality; a specialized organizational task may raise customization questions.
Retrieval Augmented Generation, or RAG, is a high-value concept to understand as an application pattern. The outline asks candidates to define RAG and describe business applications, including Amazon Bedrock Knowledge Bases. It also names Amazon OpenSearch Service, Amazon Aurora, Amazon Neptune, and Amazon RDS for PostgreSQL as examples of services that can store embeddings within vector databases.
Study RAG as a response to a specific problem: an application needs relevant contextual information when generating an answer. Do not reduce it to a service-name association. The official security objectives also include RAG grounding as an example of a technique that can improve output accuracy, which helps connect application architecture with safe and reliable output handling.
Prompt engineering requires both construction and risk awareness. The exam outline includes context, instructions, negative prompts, chain-of-thought, zero-shot, single-shot, few-shot, and prompt templates. It also includes specificity, concision, experimentation, guardrails, discovery, and prompt versioning and management through Amazon Bedrock Prompt Management.
A practical study exercise is to take one ambiguous request and rewrite it with a role or context, a precise task, constraints, and a desired response format. Then ask what could go wrong if the prompt contains sensitive content or is manipulated by a hostile input. AWS lists exposure, poisoning, hijacking, and jailbreaking among prompt-engineering risks and limitations, so the prompt itself must be treated as part of the application’s risk surface.
Understand customization as a selection decision, not a recipe. The published objectives cover pre-training, fine-tuning, continuous pre-training, distillation, instruction tuning, domain adaptation, transfer learning, and preparing data through curation, governance, size, labeling, representativeness, and reinforcement learning from human feedback. They also require the cost tradeoffs among pre-training, fine-tuning, in-context learning, RAG, and model distillation. For revision, create a matrix with columns for goal, data implication, cost implication, and evaluation need.
Evaluation should connect output quality to business success. AWS includes human-in-the-loop evaluation, benchmark datasets, Amazon Bedrock Model Evaluation, ROUGE, BLEU, BERTScore, and LLM-as-a-judge. It also includes evaluating applications built with RAG, agents, and workflows, alongside business measures such as productivity, user engagement, task completion rate, user satisfaction, and cost per interaction. Learn what an evaluation approach is intended to reveal before trying to memorize its name.
Treat responsible AI as a design requirement
Responsible AI questions ask you to recognize that model and dataset decisions can affect people, output quality, trust, and legal risk; they are not a separate afterthought after a solution has been chosen.
The responsible-AI domain includes bias, fairness, inclusivity, robustness, safety, and veracity. AWS also includes dataset characteristics such as inclusivity, diversity, curated data sources, and balanced datasets. Make the connection explicit in your notes: data choices can affect the behavior and equity of an AI system, and evaluation must be able to reveal relevant problems.
The outline addresses bias and variance, including effects on demographic groups, inaccuracy, overfitting, and underfitting. It names analyzing label quality, human audits, and subgroup analysis as examples of methods to detect and monitor bias, trustworthiness, and truthfulness. Focus on the purpose of each activity rather than assuming one check proves a system is fair.
Transparency and explainability have their own task statement. AWS expects candidates to recognize differences between models that are transparent and explainable and those that are not, and to understand tradeoffs between model safety and transparency. Amazon SageMaker Model Cards, Amazon Bedrock Model Evaluations, open-source models, data, and licensing are cited as relevant considerations or tools.
Human-centered design is part of this domain. The official examples include user-feedback mechanisms and AI decision transparency. When a scenario names a user who must understand, assess, or challenge an AI-supported decision, look for choices that make the system’s use and limits clearer rather than choices that merely increase automation.
Review legal and trust risks alongside model behavior. AWS lists intellectual property infringement claims, biased model outputs, loss of customer trust, end-user risk, and hallucinations as legal risks of working with GenAI. The practical preparation decision is to ask, for every use case you study, what harm could arise and what design, review, or communication measure addresses it.
Connect security, governance, and compliance to AI workloads
Security and governance preparation should focus on protecting data and AI interactions, controlling access, documenting origins, and following governance practices—not on implementing every control yourself.
Domain 5 includes methods to secure AI systems and recognizing governance and compliance regulations for AI systems. AWS names IAM roles, policies, and permissions; encryption; Amazon Macie; AWS PrivateLink; the shared responsibility model; Amazon Bedrock AgentCore Identity; Policy in AgentCore; and Amazon Bedrock Guardrails among examples relevant to securing AI systems.
Organize security review by the asset being protected. Data needs quality assessment, access control, integrity protection, privacy considerations, and appropriate handling. Model and application interactions need protections against prompt injection, data leakage, unsafe outputs, and other misuse. Operational records may require audit trails and logging. AWS explicitly includes encryption at rest and in transit, threat detection, vulnerability management, output filtering and validation, and logging requirements for AI interactions in the security objectives.
Source citation and data origin are also tested concepts. The outline includes data lineage, data cataloging, and Amazon SageMaker Model Cards as examples for documenting data origins. Do not confuse a citation with a guarantee that content is correct. It is evidence about origin; quality evaluation and output validation remain separate concerns.
For inaccurate or fabricated output, AWS lists hallucination detection methods and grounding techniques, including RAG grounding, output validation, and confidence scoring. Review each as a way to improve output accuracy while recognizing that the business context determines the needed level of review.
Governance questions can involve services and processes. AWS names AWS Config, Amazon Inspector, AWS Artifact, AWS CloudTrail, and AWS Trusted Advisor as examples of services or features that can assist with governance and regulation compliance. It also includes data lifecycle, logging, residency, monitoring, observation, retention, policies, review cadence, review strategies, transparency standards, and team training requirements.
A frequent study mistake is treating responsible AI and security as identical. Responsible AI emphasizes issues such as fairness, safety, transparency, and human impact. Security, compliance, and governance emphasize protection, control, data handling, auditability, and adherence to defined practices. They overlap in real solutions, but the distinction helps identify what a question is asking.
Learn AWS services by job, not by list
Use the in-scope service list as a vocabulary boundary, then learn a service’s primary role and the reason it fits a scenario instead of attempting to memorize every available AWS feature.
AWS publishes a non-exhaustive in-scope list that is subject to change. The list spans machine learning, analytics, compute, storage, databases, management and governance, security, networking, and other categories. Its purpose is not to require deep implementation knowledge of each service; it tells you which service names can reasonably appear in exam scenarios.
Start with the services specifically named across the task objectives. Amazon Bedrock and Amazon SageMaker AI recur in the AI and GenAI objectives. Amazon Bedrock Knowledge Bases is associated with RAG. Amazon OpenSearch Service, Amazon Aurora, Amazon Neptune, and Amazon RDS for PostgreSQL are cited as examples for storing embeddings in vector databases. Amazon Bedrock Prompt Management is relevant to prompt versioning and management.
Then reinforce the surrounding AWS foundation. Amazon S3 relates to storage; Amazon EC2 and AWS Lambda are core compute examples; IAM controls access; AWS KMS is in scope for key management; AWS CloudTrail supports audit-related visibility; AWS Config, Amazon Inspector, AWS Artifact, and AWS Trusted Advisor appear in governance and compliance contexts. The key is to identify the relationship between the requirement and the service capability named in the question.
The Domain 1 objectives also name managed AI/ML services including Amazon Transcribe, Amazon Translate, Amazon Comprehend, Amazon Lex, and Amazon Polly. Study the type of business function implied by each name and the general service category. Avoid inventing architecture details that the question does not provide.
Make a two-column service sheet. In the first column, write the service or feature. In the second, write a short job description grounded in the official outline, such as ‘foundation-model application capability,’ ‘managed ML capability,’ ‘vector embedding storage example,’ ‘access control,’ or ‘governance support.’ Add a third column only for common confusion pairs that you personally need to resolve. This keeps service study focused and manageable.
Plan around the official exam format and delivery options
Plan for the published question types and scoring rules, then select the test delivery option that best supports a calm, compliant testing environment.
AWS lists a 90 minutes exam duration and a format of 65 questions. The exam guide states that 50 questions affect the score and 15 questions are unscored. Candidates cannot rely on identifying unscored items, so every question should receive the same careful treatment.
Question formats may include multiple choice, multiple response, ordering, and matching. A multiple-choice question has one correct response and three distractors. A multiple-response question has two or more correct responses from five or more options, and all correct responses must be selected to receive credit. Ordering questions require correct responses to be selected and placed in the correct order; matching questions require all pairs to be correct.
The exam guide says unanswered questions are scored as incorrect and that there is no penalty for guessing. A practical approach is to mark difficult questions for review where the interface permits, but do not leave a question blank merely because you are uncertain. Use the scenario’s business objective and constraints to eliminate answers that address the wrong problem.
Results are reported on a scaled score of 100–1,000, and AWS states that the minimum passing score is 700. Do not interpret the scaled score as a simple percentage of correct answers. Use practice results mainly to find concepts you cannot explain and scenario patterns where you select a plausible but unsupported option.
AWS lists the standard exam cost as USD 100 and directs candidates to its exam-pricing information for additional pricing and foreign-exchange details. Confirm the current registration information through AWS before making a purchase, because operational details can change.
AWS states that the exam can be taken at a Pearson VUE testing center or as an online-proctored exam. It lists Arabic, English, French, German, Italian, Japanese, Korean, Portuguese (Brazil), Spanish (Latin America and Spain), Simplified Chinese, and Traditional Chinese as offered languages. Check the current scheduling page for availability in your chosen location, language, and delivery format.
Follow a practical preparation roadmap
Work from concepts to decisions to exam-style reasoning: establish vocabulary, apply it to use cases, connect it to AWS services and controls, then use timed review to expose weak explanations.
First, read the current AIF-C01 exam guide and turn the five domains into a personal checklist. Do not begin with random service videos or broad AI news. The checklist should include every published task statement: AI/ML fundamentals and lifecycle; GenAI fundamentals and AWS technologies; foundation-model design, prompting, customization, and evaluation; responsible AI and explainability; and security, compliance, and governance.
Next, establish the conceptual foundation. Study Domain 1 before attempting detailed foundation-model architecture choices. Learn the contrasts among AI, ML, GenAI, foundation models, and agentic AI. Add common data types, learning methods, inference patterns, and metrics. For every concept, create one use case and one example where it would be inappropriate. This follows the outline’s emphasis on deciding whether AI/ML is appropriate.
Then build Domain 2 and Domain 3 together. Pair each GenAI term with a business implication: tokens with inference cost and performance, context engineering with quality and relevance, embeddings with retrieval, model selection with latency or modality, and RAG with grounded contextual responses. Follow with prompt construction, prompt risks, customization tradeoffs, and evaluation methods.
After that, study Domains 4 and 5 as cross-cutting decision filters. Take the same scenario used for your GenAI practice and ask different questions: Is the dataset inclusive and appropriately governed? Could a biased or inaccurate output cause harm? Does the user need transparency? Who should access the data? How are interactions logged? What prevents prompt injection or data leakage? This method makes responsible AI, security, and governance memorable because they are applied to a real decision.
Use retrieval practice rather than rereading. Close your notes and explain why RAG might be chosen over another customization approach, why a business metric differs from a model metric, or why a particular control addresses a named threat. When you cannot give a concise explanation, return to the official objective and correct the gap.
Use practice questions carefully. Their best use is diagnosing whether you misread requirements, confused concepts, or lacked service knowledge. After every incorrect answer, write one sentence identifying the missed clue and one sentence explaining why the selected option was less appropriate. Do not simply record the answer letter; that method does not build transferable judgment.
In the final review stage, perform mixed-domain practice. A strong scenario may combine business value, model choice, latency, prompt risk, output accuracy, human oversight, and access control. Train yourself to identify the central requirement first, then filter choices through the constraints. Revisit the official content outline immediately before scheduling or sitting the exam so your study plan remains aligned with AWS’s published scope.
Avoid preparation mistakes that waste effort
The most avoidable AIF-C01 mistakes are overstudying implementation, memorizing service names without use cases, and separating model quality from business, security, and human-impact requirements.
Do not spend most of your preparation time learning to code models, tune hyperparameters, engineer features, or build pipelines. Those tasks are explicitly outside the target candidate scope. Technical curiosity is useful, but it should not displace study of the published decisions the exam actually covers.
Do not equate GenAI with a universal solution. The outline repeatedly emphasizes capabilities and limitations, cost and latency considerations, hallucinations, nondeterminism, model constraints, and situations where AI/ML is not appropriate. In a scenario, a deterministic requirement or weak cost-benefit case should change your answer.
Do not memorize a service inventory without associating services with roles. Service names are easier to retain when attached to a need such as foundation-model access, managed ML, embedding storage, access management, encryption, auditing, or compliance support. A service name alone rarely demonstrates the judgment measured by the exam.
Do not treat the highest-weight domain as the only domain. Content Domain 3, Applications of Foundation Models, represents 28% of scored content, but a candidate who ignores Content Domain 4, Guidelines for Responsible AI, or Content Domain 5, Security, Compliance, and Governance for AI Solutions, each representing 14% of scored content, leaves important published objectives unprepared.
Finally, do not confuse an output-quality metric with success for the organization. AWS includes technical measures, business measures, and alignment metrics. A useful answer must match the metric to the purpose: model performance, user experience, task completion, cost, productivity, or another stated objective.
Conclusion
AIF-C01 preparation is most effective when you can explain a business requirement, select an appropriate AI or GenAI approach, connect it to relevant AWS capabilities, and identify the associated quality, cost, responsible-AI, security, or governance concern. Use the official outline as your checklist, give Domain 3 repeated practice, and keep the other domains active through mixed scenarios. When you can justify choices in plain language rather than recite terms, you are ready to make a well-informed scheduling decision.
