AWS Certified AI Practitioner (AIF-C01) study guide

How to use this guide

Know these cold:

• Core terminology — model, inference, training, embedding, token — used precisely
• Supervised, unsupervised and reinforcement learning, and what each is for
• When traditional ML beats generative AI, and when neither is the right tool
• The AWS ML stack: SageMaker, and the managed AI services above it

Know these cold:

• Foundation models, and how pre-training differs from fine-tuning
• Tokens, context windows, embeddings and vector representations
• Amazon Bedrock and Amazon Q — what each is for and how they differ
• Where generative AI genuinely fits a business problem, and where it does not

Know these cold:

• Prompt engineering: zero-shot, few-shot, chain-of-thought, and prompt injection risk
• Retrieval-augmented generation — vector stores, chunking, and why retrieval quality dominates output quality
• Choosing between prompting, RAG, fine-tuning and continued pre-training on cost and effort
• Evaluating outputs: relevance, groundedness, hallucination, and human review

Know these cold:

• Bias and fairness — where it enters, and how it is measured
• Transparency and explainability, including model cards
• Guardrails for Amazon Bedrock and content filtering
• Human-in-the-loop review, and when it is not optional

Know these cold:

• Protecting training and inference data — encryption, IAM, and least privilege for model access
• Data residency, retention, and what leaves your account when you call a model
• Auditability and monitoring of AI workloads
• Regulatory and IP considerations around model output

Common mistakes

• ✕Treating generative AI as the answer to every scenario. A good share of questions are testing whether you recognise the case where classic ML, or no AI at all, is correct.
• ✕Confusing fine-tuning with RAG. Fine-tuning changes model behaviour; RAG changes what the model can see. Questions turn on which problem you actually have.
• ✕Underestimating responsible AI. Bias, transparency and guardrails are 14% on their own, and they are tested as judgment rather than definitions.
• ✕Mixing up Amazon Bedrock and Amazon Q — one is model access for building, the other is an assistant product.
• ✕Forgetting that data governance still applies. Questions frequently hinge on what happens to prompts and outputs, not on the model itself.
• ✕Ignoring the non-multiple-choice formats and meeting ordering or matching items cold.

Exam-day tips

• ✓Scoring is compensatory — a weak domain cannot fail you alone. Answer everything.
• ✓There is no penalty for guessing. Never leave an item blank.
• ✓Ordering questions are graded on the whole sequence — read all the steps before placing any of them.
• ✓When an option promises to eliminate bias or guarantee accuracy, it is almost always wrong. Responsible-AI answers are hedged for a reason.
• ✓Applications of Foundation Models is 28% — if you are short on study time, spend it on prompting and RAG.
• ✓At 65 questions in 90 minutes you have under 90 seconds each. Flag and move.

How to know you’re ready

One good practice score can be luck. The signal that holds up: consistently at or above the real pass mark (700 on a 100–1000 scaled score) across multiple full-length sets, with no single domain dragging far below the rest. Kwizza tracks your per-domain readiness automatically as you practice.