AWS Certified AI Practitioner (AIF-C01) Study Notes

I sat the AWS Certified AI Practitioner (AIF-C01) and passed. This is the retrospective I wish I had read before starting - what the exam actually tests, the services and concepts to know well, and which prep resources earned their place.

AIF-C01 covers more ground than the name suggests: classical ML, generative AI, foundation models, and responsible AI. Domains 2 and 3 carry the most weight and the vocabulary is denser than it looks. Your starting point will shape where you need to spend time, so calibrate accordingly.

For more content on other relevant certifications, check Certifications.

The exam at a glance

Confirm these against the live AWS certification page before your exam - costs and question counts occasionally change.

The five domains

AIF-C01 has five domains. The percentages are from the official exam guide - weight your study time accordingly.

Domain 1 - Fundamentals of AI and ML (20%)

Focus areas:

• Types of ML: supervised, unsupervised, reinforcement learning - and when each applies
• Core concepts: training vs inference, features, labels, overfitting, underfitting
• Model evaluation metrics: accuracy, precision, recall, F1, AUC-ROC
• The ML pipeline: data preparation, training, evaluation, deployment, monitoring
• AWS ML services overview: SageMaker, Rekognition, Comprehend, Polly, Transcribe, Translate

Domain 2 - Fundamentals of Generative AI (24%)

Focus areas:

• Foundation models vs traditional ML models - pre-trained, general-purpose, fine-tunable
• Large language models: tokenization, embeddings, context windows, attention mechanisms (conceptual)
• Generative AI output types: text, image, code, audio
• Key inference parameters: temperature, top-p, max tokens - what they control
• Amazon Bedrock: managed access to foundation models; providers and model IDs
• Prompt engineering basics: zero-shot, few-shot, chain-of-thought

Domain 3 - Applications of Foundation Models (28%)

Focus areas:

• Retrieval-Augmented Generation (RAG): how it works, why it reduces hallucinations, when to prefer it over fine-tuning
• Fine-tuning vs RAG vs prompt engineering - the trade-off matrix: cost, data needs, latency, freshness
• Amazon Bedrock Knowledge Bases - managed RAG on AWS
• Amazon Bedrock Agents - multi-step task orchestration with foundation models
• Vector databases and embeddings: how semantic search enables RAG
• Evaluating generative AI outputs: human review, automated metrics, hallucination detection

Domain 4 - Guidelines for Responsible AI (14%)

Focus areas:

• AWS responsible AI pillars: fairness, explainability, transparency, privacy, robustness, safety, controllability
• Bias in AI: types of bias, where it enters the pipeline, how to detect and mitigate it
• Explainability: what it means, why it matters for regulated use cases
• Amazon SageMaker Clarify - bias detection and explainability tooling
• Human oversight: when to require human review of AI outputs

Domain 5 - Security, Compliance, and Governance for AI Solutions (14%)

Focus areas:

• Shared responsibility for managed AI services vs self-hosted models
• Amazon Bedrock Guardrails - content filtering, denied topics, PII redaction
• Data privacy for training and inference: where data goes, what AWS controls
• IAM for Bedrock and SageMaker - least privilege for ML workloads
• Compliance considerations: data residency, model provenance, audit logging

Services to know well

The exam tests AWS services at a conceptual level, not deep implementation detail. Know what each does, which use case it fits, and what it is commonly confused with:

Easy things to mix up

• Bedrock vs SageMaker - Bedrock is for consuming pre-built foundation models via API; SageMaker is for building, training, and deploying your own models. The exam constructs scenarios specifically to distinguish them.
• RAG vs fine-tuning - RAG retrieves relevant context at inference time (no retraining, data stays fresh); fine-tuning bakes new knowledge into model weights (requires retraining, more expensive, knowledge becomes stale). Fine-tuning suits tone and style; RAG suits factual, updatable knowledge.
• Bedrock Knowledge Bases vs Kendra - both do search, but Bedrock Knowledge Bases is semantic vector search to feed a generative model; Kendra is keyword and ML-powered enterprise search without a GenAI layer.
• Temperature vs top-p - temperature scales the probability distribution across all tokens (higher = more random, more creative output; lower = more deterministic and conservative); top-p limits the token pool to the smallest set whose cumulative probability meets a threshold. Both control randomness but from different angles.
• Supervised vs unsupervised vs reinforcement learning - supervised learns from labelled examples; unsupervised finds structure in unlabelled data; reinforcement learns from reward signals. Clustering = unsupervised, classification = supervised.
• Overfitting vs underfitting - overfitting is when a model memorises training data and fails to generalise (high variance: performs well on training data, poorly on new data); underfitting is when the model is too simple to capture the pattern (high bias: performs poorly on both). The bias-variance trade-off is the tension between them - more model complexity reduces bias but increases variance.
• Precision vs recall - precision is “of what I predicted positive, how many were actually positive” (minimise false positives); recall is “of all actual positives, how many did I catch” (minimise false negatives). The trade-off matters in content moderation and medical diagnosis scenarios.
• Hallucination - model produces confident but factually wrong output. RAG is the primary mitigation the exam points to; scenario questions will describe it and ask how to address it.
• Foundation model vs LLM - foundation model is the broader term for large pre-trained general-purpose models (includes image and multimodal); LLM is specifically language-based. All LLMs are foundation models; not all foundation models are LLMs.

Resources

More so than most AWS certs, the official Skill Builder material is worth starting with here - it’s closely aligned to what the exam actually tests.

Essential

• Official Exam Guide (AIF-C01) - read this first. The five domains, their weights, and the in-scope services list are the syllabus. Available on the AWS certification page.
• Exam Prep Standard Course: AWS Certified AI Practitioner (AIF-C01) - the AWS-authored prep course. More directly exam-relevant here than third-party alternatives.
• Standard Exam Prep Plan: AWS Certified AI Practitioner (AIF-C01) - a structured plan that sequences all the prep material.
• Official Practice Question Set - AWS-authored practice questions. The style matches the real exam closely.
• Official Practice Exam - the full-length timed version; available via AWS Skill Builder. Do one complete timed run before sitting the real exam.

Useful

• Tutorials Dojo practice exams (Jon Bonso) - if available for AIF-C01, worth it for the explanations. Check current availability.
• Amazon Bedrock documentation - the Bedrock service is central to Domains 2 and 3. The concepts overview and Guardrails docs are high-yield reading.
• AWS Responsible AI resources - AWS publishes a responsible AI FAQ and whitepaper; Domain 4 pulls from this framing directly.

Skip if you are tight on time

• General ML textbooks and courses - this exam tests conceptual awareness, not implementation depth. Andrew Ng’s ML course is excellent but goes far deeper than required. Save it for after.
• SageMaker deep-dives - SageMaker appears in the exam but at a high level. A full SageMaker course is overkill for AIF-C01 specifically.

Notes