cert-bank

Guides

이 글을 한국어로 읽기

AIF-C01 Study Guide: The AWS AI Practitioner Exam Is About Choosing, Not Building

The AWS Certified AI Practitioner exam (AIF-C01) is a foundational exam, but people with a machine learning background often find it harder than expected, and people without one often find it easier. The reason is simple: it never asks you to train a model or write code. It asks you to choose: which approach, which AWS service, which safeguard, for a business need described in a few sentences. This guide is organised around those choices.

Study guide · AWS Certified AI Practitioner (AIF-C01)

AIF-C01 Study Guide: The AWS AI Practitioner Exam Is About Choosing, Not Building

The exam in one paragraph

65 questions (50 scored and 15 unscored that are not marked), 90 minutes, a passing score of 700 on a scale of 100–1,000. Besides single and multiple choice you may see ordering, matching and short case-study questions. 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%). More than half of the exam is about generative AI and foundation models.

Seven decisions the exam keeps asking you to make

We wrote 356 original AIF-C01 questions for this site. Amazon Bedrock appears in more of them than any other service, by a wide margin; SageMaker AI, Bedrock Knowledge Bases and Bedrock Guardrails follow. Almost every question comes down to one of these seven decisions.

1. Is this even a machine learning problem?

If the rules are known and stable ("approve refunds under $50"), plain code is the answer. ML fits when the pattern has to be learned from data and some error is acceptable. The exam rewards recognising when AI is not needed.

2. Which kind of learning?

Labelled outcomes to predict → supervised (classification for categories, regression for numbers). No labels, find structure → unsupervised (clustering, anomaly detection). An agent learning from rewards → reinforcement learning. Know the vocabulary: features, labels, training versus inference, overfitting (great on training data, poor on new data) and underfitting.

3. A ready-made AI service, or a foundation model?

For a narrow, common task, the purpose-built service is cheaper and simpler: Amazon Comprehend for sentiment and entities, Transcribe for speech to text, Polly for text to speech, Translate, Rekognition for images and video, Textract for documents, Lex for chatbots with defined intents. Reach for Amazon Bedrock when the task needs open-ended generation, summarising, or reasoning over varied text.

4. Prompting, RAG, fine-tuning, or continued pre-training?

This is the single most tested decision. Work through it in order of cost and effort:

  • Prompt engineering first: clear instructions, examples (few-shot), a defined output format.
  • Retrieval Augmented Generation (RAG), for example with Bedrock Knowledge Bases, when answers must come from your own documents or from information that changes. It needs no training and the source data stays up to date.
  • Fine-tuning with labelled examples when the model must learn a style, format or specialised task that prompting cannot reach.
  • Continued pre-training on large amounts of unlabelled domain text when the model lacks a whole domain's vocabulary.

A frequent wrong answer is fine-tuning to "teach the model last week's prices". Data that changes belongs in RAG.

Going deeper: why RAG reduces hallucinations but does not remove them

A foundation model generates the most likely next tokens; it has no built-in notion of whether a sentence is true. RAG changes what the model conditions on: the question is turned into an embedding, the closest chunks of your documents are retrieved by vector similarity, and they are placed in the prompt. The answer improves only as far as retrieval does. If the chunks are cut in the wrong places, if the right passage is not among the top results, or if the prompt does not tell the model to answer only from the context, it can still invent details.

That is why the exam pairs RAG with other controls: chunking and retrieval settings in Bedrock Knowledge Bases, contextual grounding checks in Bedrock Guardrails, and evaluation of answers against sources. When a question asks how to make answers traceable, "return the source documents with the answer" is part of the solution.

Prompting, RAG, fine-tuning or continued pre-training: cheapest first
Can clear instructions, few-shot examples and a defined output format get there?
YES →
Prompt engineering
↓ no
Must answers come from your own documents or from information that changes?
YES →
RAG, for example Bedrock Knowledge Bases
↓ no
Must the model learn a style, format or specialised task that prompting cannot reach?
YES →
Fine-tuning with labelled examples
↓ no
OTHERWISE →
Continued pre-training on large amounts of unlabelled domain text
Work down in order of cost and effort; the first yes is the answer, and data that changes belongs in RAG.

5. Which inference setting?

Temperature (and top-p) control randomness: low for consistent, factual answers, higher for creative variety. Maximum tokens limits the response length and cost. Tokens are also how you are billed with on-demand pricing; provisioned throughput buys reserved capacity for steady, high volume.

Beginner tip: open the Amazon Bedrock playground (text or chat), pick one model and send the same prompt three times at temperature 0, then three times at the highest setting. The first three answers barely change and the last three differ every time. After that, questions about temperature, top-p and consistent output need no memorising.

6. How do we keep it responsible?

Bias and explainability in ML models → Amazon SageMaker Clarify. Blocking harmful content, denied topics or personal data in a generative application → Amazon Bedrock Guardrails. Documenting what a model is for and how it was evaluated → model cards. Human review of low-confidence predictions → Amazon Augmented AI (A2I). Hallucinations are reduced by grounding (RAG) and checked by evaluation, not removed by a higher temperature.

7. Who secures what?

The shared responsibility model still applies: AWS secures the service, you control access (IAM), encryption keys (KMS), network paths (VPC endpoints through PrivateLink) and logging (CloudTrail). With Amazon Bedrock, your prompts and outputs are not used to train the base models, which is often the deciding fact in a data-privacy question.

How to measure a model, briefly

TaskMetrics to recognise
ClassificationAccuracy, precision (few false alarms), recall (few misses), F1, AUC
RegressionMAE, RMSE, R²
Text generationROUGE (summaries), BLEU (translation), BERTScore (meaning), plus human evaluation
Business valueCost per task, time saved, conversion, customer satisfaction

Read what the scenario fears: "we must not miss fraud" means recall; "we must not flag good customers" means precision.

A three-week plan

  • Week 1: AI and ML fundamentals and the AWS AI services (decisions 1–3). Practise the Fundamentals of AI and ML domain.
  • Week 2: Generative AI and Amazon Bedrock in depth (decisions 4–5): models, prompts, Knowledge Bases, Agents, evaluation. This is half the exam.
  • Week 3: Responsible AI and security (decisions 6–7), then mixed practice sets under time pressure.

Practise by domain

Practise in the question bank: every question explains each option, right and wrong. Start with the domain you just studied, then take a timed mock exam.

Exam facts as published in the AWS exam guide for AIF-C01; AWS revises the guide and the services change quickly, so check the current version before you book.