AWS Certified AI Practitioner (AIF-C01)

Separate deterministic rules from prediction and content generation. A fixed policy calculation may need ordinary software, while a fraud-risk estimate may need classification and a first-draft response may suit generative AI. Define who will use the result, the cost of an error, the baseline process and a measurable improvement before choosing a tool.

Your practice outputCreate a one-page use-case brief with an input, expected output, human owner, failure cost and success metric. Reject one idea for which an AI model adds no clear value.

Compare labeled-data classification and regression with unsupervised clustering. Trace an example from data preparation and training through evaluation, inference and monitoring. Distinguish accuracy from precision and recall when false positives and false negatives have different consequences. Recognize overfitting, data leakage and a changing real-world data distribution.

Your practice outputFor a fictional support-ticket routing model, identify the label, a possible leakage feature and why accuracy alone could conceal poor performance on a rare urgent class.

Describe foundation models, tokens, context windows, embeddings and multimodal input without treating a fluent answer as a verified fact. Compare a summarization, translation, retrieval and generation request. Evaluate latency, input/output length and cost per interaction alongside quality. Know why a model may hallucinate, repeat bias or expose information it should not have received.

Your practice outputWrite a risk register for a proposed document assistant: include stale facts, private data, ungrounded claims, accessibility and the person who must review a consequential answer.

Compare managed AI services for speech, language and document extraction with Amazon Bedrock for applications using foundation models and Amazon SageMaker AI for a broader custom ML lifecycle. Select a model by modality, language, latency, context length, governance needs and cost, rather than by brand alone. Understand that the AIF-C01 candidate is expected to recognize these choices, not build production pipelines or tune model hyperparameters.

Your practice outputFor three short requests—transcribe a call, extract fields from a form and draft a grounded answer—state the capability needed, a plausible AWS service and one validation question.

Write explicit instructions and examples, then test how missing context changes the answer. Compare retrieval-augmented generation (RAG), in-context examples and fine-tuning: frequently changing source documents usually call for retrieval, not retraining the model whenever a policy changes. Explain that citations help a reader verify a source but do not prove every generated statement. Evaluate groundedness, task completion, latency and cost with representative test cases and human review.

Your practice outputDraft a prompt and an evaluation table for a fictional policy assistant. Include a question whose answer is absent from the approved documents and require the assistant to say so.

A model can perform well overall while failing a subgroup. Review source-data balance, harmful output, explainability and feedback routes before launch. Separate guardrail configuration from broader product governance: a content filter alone cannot validate factual accuracy, settle intellectual-property questions or replace a human decision on a high-impact case.

Your practice outputDescribe a review plan for an AI-assisted support triage pilot: subgroup checks, a safe escalation path, an appeal or correction channel and a reason to stop the pilot.

Apply least-privilege IAM access, data classification, encryption, logging and retention to the full application, not only to the model API. Identify where user input, retrieved documents and generated output can expose sensitive information. Distinguish AWS service controls from the customer's responsibility to approve data sources, set access rules, configure the application and investigate incidents.

Your practice outputDraw a data-flow sketch for a fictional internal knowledge assistant. Mark data owners, retrieval permissions, logging, retention, output checks and who can revoke access.

Work through the official domain guide first, then use an original practice set as a diagnostic. For each missed item, explain why the chosen option fails, identify the governing concept and revisit the AWS source. Repeat with a different set after studying, paying particular attention to the high-weight foundation-model domain. Time management matters, but a raw mock percentage is not AWS's scaled score.

Your practice outputKeep an error log grouped by the five official domains. For each weak domain, write a corrected decision rule and a practical example before attempting another set.