AWS AI Practitioner Study Guide & Cheat Sheet

AWS Certified AI Practitioner (AIF-C01) Exam Guide

Exam domains

Who it’s for: Business and technical professionals - analysts, product and project managers, sales and marketing staff, and developers - who need foundational fluency in AI/ML concepts and the AWS AI service portfolio. No coding or hands-on model-building experience is required; AWS suggests up to ~6 months of general exposure to AI/ML technologies on AWS.

Study & test-day tips

• Spend the most prep time on generative AI: Domains 2 and 3 together are 52% of the exam, so master foundation models, prompt engineering, RAG, fine-tuning, and Amazon Bedrock before anything else.
• Learn the 'when to use what' decision tree for managed AI services - Comprehend (text/NLP), Rekognition (images/video), Textract (document data extraction), Transcribe (speech-to-text), Polly (text-to-speech), Translate, Lex (chatbots), Kendra (enterprise search), and Personalize (recommendations).
• Memorize the RAG vs. fine-tuning vs. continuous pre-training vs. in-context learning distinction: RAG adds external knowledge at inference without retraining, fine-tuning adapts weights to a task, continuous pre-training extends domain knowledge, and in-context learning teaches via the prompt only.
• Know inference parameters cold: lower temperature and lower top-p make output more deterministic and focused; higher values increase randomness and creativity. Expect a scenario question asking which to adjust.
• Map each SageMaker capability to its job: Clarify (bias detection + explainability), Model Monitor (drift), Model Cards (documentation), Data Wrangler (data prep), Feature Store (feature reuse), Autopilot (AutoML), JumpStart (prebuilt/foundation models).
• Pin responsible-AI tooling to scenarios: Bedrock Guardrails (filter harmful content/PII in GenAI apps), SageMaker Clarify (fairness/bias), and Amazon A2I (human review of low-confidence predictions).
• For security questions, default to AWS-native answers: IAM for access, KMS for encryption at rest, Macie for discovering sensitive data in S3, and PrivateLink for private connectivity - and remember the shared responsibility model (AWS secures the cloud, you secure what's in it).
• Know your evaluation metrics by output type: accuracy/precision/recall/F1/AUC for classification; ROUGE and BLEU for text summarization/translation; BERTScore for semantic similarity of generated text.
• Treat it as a vocabulary exam - many questions test whether you recognize a term, so drill the glossary (tokens, embeddings, hallucination, overfitting, RLHF, vector store) until definitions are automatic.
• Budget your time and answer everything: ~90 minutes for 65 questions is generous, there's no penalty for guessing, 15 questions are unscored, and you only need 700/1000 overall - flag hard items, eliminate wrong options, and never leave a blank.

Cheat sheet

Core AI/ML concepts

• Supervised learning = labeled data (classification, regression); unsupervised = unlabeled (clustering); reinforcement = reward-driven trial and error.
• Overfitting = memorizes training data, fails on new data (high variance); underfitting = too simple, misses patterns (high bias).
• Classification metrics: accuracy (overall correct), precision (of predicted positives, how many right), recall (of actual positives, how many found), F1 (precision/recall balance), AUC (ranking quality).
• ML lifecycle: collect/prepare data, train, evaluate, deploy, then monitor for drift and retrain.
• AI > ML > deep learning are nested; not every problem needs ML - rules-based logic can be simpler and cheaper.

Generative AI & foundation models

• Foundation models (FMs) are large models pre-trained on broad data and adapted to many tasks; LLMs are FMs for text.
• Tokens are the chunks of text a model processes; embeddings are numeric vectors capturing meaning; similar meanings sit close in vector space.
• Prompt engineering styles: zero-shot (no examples), one-shot (one example), few-shot (several), chain-of-thought (show reasoning steps).
• Temperature and top-p control randomness: low = focused/deterministic, high = diverse/creative.
• Hallucination = confident but false output; mitigate with RAG, grounding, and guardrails.
• RLHF (reinforcement learning from human feedback) aligns model behavior with human preferences.

Amazon Bedrock & the AWS GenAI stack

• Amazon Bedrock = fully managed, serverless access to FMs from multiple providers via one API; no infrastructure to manage.
• Bedrock Knowledge Bases = managed RAG: connect your data so the model answers from it without retraining.
• Bedrock Agents = orchestrate multi-step tasks and call APIs/tools to take action.
• Bedrock Guardrails = configurable content filters, denied topics, and PII redaction for safer GenAI apps.
• Amazon Q = generative AI assistant for work (business) and for developers; SageMaker JumpStart = prebuilt models and solutions; PartyRock = no-code app playground.
• Customization spectrum: prompt engineering -> RAG -> fine-tuning -> continuous pre-training, in increasing cost/effort.

AWS AI services (when to use what)

• Comprehend = NLP (sentiment, entities, key phrases, PII detection); Translate = language translation.
• Rekognition = image/video analysis (objects, faces, moderation); Textract = extract text/forms/tables from documents.
• Transcribe = speech-to-text; Polly = text-to-speech; Lex = conversational bots (powers Alexa-style chat).
• Kendra = ML-powered enterprise search; Personalize = real-time recommendations.
• Amazon SageMaker = build/train/deploy custom ML; Clarify, Model Monitor, Model Cards, Data Wrangler, Feature Store, Autopilot are its key features.
• Amazon A2I = adds human review for low-confidence predictions.

Responsible AI

• Core pillars: fairness, explainability, transparency, robustness, privacy/security, safety, and governance.
• Bias = systematic unfairness in outcomes; SageMaker Clarify detects bias and explains feature importance.
• SageMaker Model Cards document a model's purpose, data, and limits; Model Monitor watches for drift in production.
• Bedrock Guardrails enforce content safety and block harmful or off-topic responses.
• Trade-offs are real: maximizing accuracy can conflict with fairness, transparency, or cost - design for the use case.
• Keep humans in the loop (Amazon A2I) for high-stakes or low-confidence decisions.

Security, governance & compliance

• Shared responsibility model: AWS secures the cloud infrastructure; you secure your data, access, and configuration.
• IAM = identity and least-privilege access; KMS = encryption keys for data at rest.
• Amazon Macie = discovers and protects sensitive data (e.g., PII) in Amazon S3; PrivateLink = private, in-VPC connectivity to services.
• CloudTrail = API activity/audit logs; AWS Config = resource configuration and compliance tracking.
• AWS Audit Manager = automates evidence collection for audits; AWS Artifact = on-demand compliance reports and agreements.
• Protect training and inference data with encryption, access controls, and data-residency awareness.

Glossary

Artificial Intelligence (AI)

The broad field of building systems that perform tasks normally requiring human intelligence, such as understanding language or recognizing images.

Machine Learning (ML)

A subset of AI where models learn patterns from data rather than being explicitly programmed with rules.

Deep Learning

ML using multi-layered neural networks; powers modern image, speech, and language models.

Supervised Learning

Training on labeled examples so the model learns to predict the correct output (classification or regression).

Unsupervised Learning

Finding structure in unlabeled data, such as grouping similar items via clustering.

Reinforcement Learning

Learning by trial and error, where an agent maximizes a reward signal from its environment.

Foundation Model (FM)

A large model pre-trained on broad data that can be adapted to many downstream tasks.

Large Language Model (LLM)

A foundation model specialized for understanding and generating human language.

Token

A unit of text (word or word-piece) that a model reads or generates; usage and cost are often measured in tokens.

Embedding

A numeric vector representation of text or data that captures meaning so similar items are close together.

Vector / Vector Store

A vector encodes meaning numerically; a vector store (e.g., OpenSearch, Aurora pgvector, RDS PostgreSQL, Neptune, DocumentDB) indexes them for similarity search in RAG.

Prompt Engineering

Crafting model inputs to get better outputs, using techniques like zero-, one-, and few-shot prompting.

Zero/One/Few-Shot

Providing zero, one, or several examples in the prompt to guide the model's response.

Chain-of-Thought

A prompting technique that asks the model to show step-by-step reasoning to improve complex answers.

Temperature

An inference setting controlling randomness; lower is more focused and deterministic, higher is more creative.

Top-p (Nucleus Sampling)

An inference setting that limits choices to the most probable tokens whose cumulative probability reaches p, balancing diversity and coherence.

RAG (Retrieval-Augmented Generation)

Supplying a model with relevant external data at inference time so answers are grounded without retraining.

Fine-Tuning

Further training a foundation model on task- or domain-specific labeled data to adapt its behavior.

Continuous Pre-Training

Extending a model's general knowledge by training further on large amounts of unlabeled domain data.

In-Context Learning

Teaching a model a task purely through examples and instructions in the prompt, without changing its weights.

RLHF

Reinforcement Learning from Human Feedback - using human preference signals to align model outputs with what people want.

Hallucination

When a model produces fluent but factually incorrect or fabricated output.

Bias / Variance

Bias is error from oversimplified assumptions (underfitting); variance is sensitivity to training data (overfitting); good models balance the two.

Overfitting / Underfitting

Overfitting memorizes training data and generalizes poorly; underfitting is too simple to capture the underlying pattern.

Precision / Recall / F1

Precision is the share of predicted positives that are correct; recall is the share of actual positives found; F1 is their harmonic mean.

AUC

Area Under the ROC Curve - measures how well a classifier ranks positives above negatives across thresholds.

ROUGE / BLEU / BERTScore

Text-generation metrics: ROUGE for summarization overlap, BLEU for translation overlap, BERTScore for semantic similarity.

Amazon Bedrock

A fully managed service offering serverless API access to foundation models, with Knowledge Bases, Agents, and Guardrails.

Bedrock Guardrails

Configurable safety controls that filter harmful content, block denied topics, and redact PII in generative AI apps.

Amazon Q

AWS's generative AI assistant for work, available in business and developer variants to answer questions and help with tasks.

Amazon SageMaker

AWS's platform to build, train, tune, deploy, and monitor custom ML models, with features like Clarify, Model Monitor, and JumpStart.

SageMaker Clarify

A SageMaker feature that detects bias in data and models and explains feature importance for transparency.

Amazon A2I

Amazon Augmented AI - adds human review workflows for low-confidence or high-stakes ML predictions.

Shared Responsibility Model

AWS secures the underlying cloud infrastructure, while the customer secures their data, access, and configurations in the cloud.