Generative AI is 52% of the AWS Certified AI Practitioner (AIF-C01) exam: Fundamentals of Generative AI (24%) and Applications of Foundation Models (28%) together cover roughly 26 of the 50 scored questions, and Amazon Bedrock is the AWS service that delivers nearly all of it, from model choice to prompting, retrieval-augmented generation, and Guardrails.
Last updated August 2026.
Domain 2 (Fundamentals of Generative AI, 24%) tests the vocabulary: what foundation models are, how they consume and produce text, and what they can and cannot do. Domain 3 (Applications of Foundation Models, 28%) tests the decisions: how to adapt a model to your use case and how to judge its output. Together they are the majority of the exam, and Amazon Bedrock recurs across both as the service that makes each concept concrete.
The depth bar matters as much as the weight. AIF-C01 is a foundational exam: you will not write code against the Bedrock API or tune hyperparameters. Questions ask which capability or approach fits a described need, so the preparation is knowing what each thing is for.
Domain weights from the official AWS Certified AI Practitioner (AIF-C01) exam guide.
Amazon Bedrock is AWS's managed service for building with foundation models. The three properties the exam cares about: it offers models from multiple providers through a single interface, it is fully managed (no infrastructure to provision to run inference), and it wraps the surrounding needs of a generative application: grounding answers in your data, taking multi-step actions, and filtering unsafe content.
That last part is where the named features attach: Knowledge Bases for retrieval-augmented generation, Agents for multi-step tasks, and Guardrails for content safety. A question that describes one of those needs is usually asking you to name the matching capability.
The single most-tested decision on AIF-C01: how much adaptation does a use case need? The options form a ladder of rising cost and effort, and the right answer is the lowest rung that meets the need.
| Approach | What it changes | Reach for it when |
|---|---|---|
| Prompt engineering | Only the input | Instructions and examples get the output you need. Always try this first |
| Retrieval-augmented generation (RAG) | The context, at query time | Answers must be grounded in your documents, and that knowledge changes often |
| Fine-tuning | The model's weights | The model's behavior, style, or domain fluency itself must change |
| Training from scratch | Everything | Almost never on this exam: the most expensive option, for needs no existing FM can meet |
The RAG-versus-fine-tuning distractor pair is worth rehearsing: changing knowledge points at RAG, changing behavior points at fine-tuning. Updating a fine-tuned model means retraining, while a RAG system picks up new documents as soon as they are indexed, which is why frequently changing information favors RAG. On Bedrock, Knowledge Bases is the managed RAG capability and model customization covers fine-tuning.
Two more Bedrock capabilities close out the exam's generative AI story. Agents handle multi-step tasks: breaking a request into steps and calling APIs and data sources to complete them. Guardrails for Amazon Bedrock applies safety filters to prompts and responses, blocking configured topics and content.
Guardrails also bridges into Domain 4 (Guidelines for Responsible AI, 14%), where it is the named answer to output-safety concerns, alongside SageMaker Clarify for bias and explainability. Learning Guardrails once therefore earns marks in two domains, which is good value on a 90-minute paper.
Bedrock is the central service across the two generative AI domains, which together are 52% of scored content, roughly 26 of the 50 scored questions. Not every one of those questions names Bedrock, but the concepts they test (foundation models, prompting, RAG, fine-tuning, content safety) are delivered through it on AWS.
No, the exam is conceptual and never asks you to code against the API. That said, an hour running prompts in Bedrock makes tokens, inference parameters, and model choice concrete far faster than reading, so light hands-on time is the highest-value preparation per minute spent.
RAG retrieves your documents at query time and grounds the model's answer in them, so knowledge stays current without touching the model. Fine-tuning changes the model's weights to alter its behavior, style, or domain fluency. Scenarios about frequently changing information point at RAG; scenarios about changing how the model itself responds point at fine-tuning.
Guardrails apply configurable safety filters to a generative application's prompts and responses, blocking defined topics and harmful content. On the exam they are the named answer to output-safety concerns, appearing in both the foundation-model domains and the responsible AI domain.
The AIF-C01 exam guide lists the task statements and in-scope services for every domain, including exactly which Bedrock capabilities can appear. Questions are written from those statements.
Open the AWS exam guide →