Introduction
It’s tempting to walk into the AWS Certified AI Practitioner (AIF-C01) exam thinking it’s all about generative AI — Bedrock, prompts, foundation models. Those topics matter, but the exam’s foundation is Domain 1: Fundamentals of AI and ML, which carries roughly 20% of your score. This domain establishes the vocabulary and mental models that every other question builds on. Get it solid and the rest of the exam — generative AI, applications of foundation models, responsible AI, security — becomes far easier to reason about.
The good news is that this domain is conceptual and approachable even if you’ve never trained a model. AWS isn’t asking you to do the math behind gradient descent. It’s asking whether you can tell supervised learning from unsupervised, recognize where a problem fits in the machine learning lifecycle, read an evaluation metric, and map a use case onto the right layer of the AWS ML stack. Those are learnable patterns, not feats of memorization.
This guide walks Domain 1 the way the exam frames it: defining AI, ML, and deep learning; the types of machine learning; the end-to-end ML lifecycle; how models are evaluated; and the AWS services that implement each layer. If you want the full exam picture first, start with the AWS AI Practitioner Exam Guide 2026 and the domains breakdown, then come back here to go deep on the fundamentals.
AI vs Machine Learning vs Deep Learning
The single most common foundational question type asks you to place these three terms in their proper nesting. They are concentric, not interchangeable:
| Term | Definition | Relationship |
|---|---|---|
| Artificial Intelligence (AI) | The broad field of building systems that perform tasks normally requiring human intelligence | The widest circle |
| Machine Learning (ML) | A subset of AI where systems learn patterns from data rather than being explicitly programmed | Inside AI |
| Deep Learning (DL) | A subset of ML using multi-layered neural networks to learn complex patterns | Inside ML |
| Generative AI | Applications of deep learning (foundation models) that generate new content | A specialization of DL |
The exam cue: if a question describes a system that learns from data and improves with more examples, that’s machine learning. If it emphasizes neural networks with many layers handling images, audio, or language, that’s deep learning. If the system creates new text, images, or code, that’s generative AI. And the broadest umbrella covering all of them is AI.
A useful contrast the exam likes: traditional programming takes rules + data and produces answers, while machine learning takes data + answers (examples) and produces the rules (a model). If a scenario says “the rules are too complex or change too often to hand-code,” that’s a signal for an ML approach.
The Types of Machine Learning
Machine learning splits into three primary paradigms. Knowing which one fits a described problem is one of the highest-yield skills for this domain.
Supervised Learning
The model learns from labeled data — inputs paired with known correct outputs — and then predicts labels for new inputs. Two sub-types:
- Classification — predicts a discrete category (spam vs not-spam, fraud vs legitimate, which of five product categories).
- Regression — predicts a continuous number (house price, expected demand, temperature).
Exam cue: the presence of historical labeled examples and a goal of predicting a known target signals supervised learning. “Predict whether a transaction is fraudulent” (classification) and “forecast next month’s sales” (regression) are both supervised.
Unsupervised Learning
The model works with unlabeled data and finds structure on its own. Common tasks:
- Clustering — grouping similar items (customer segmentation).
- Dimensionality reduction — compressing many features into fewer while keeping signal.
- Anomaly detection — flagging unusual points without labeled examples of “unusual.”
Exam cue: no labels and a goal of discovering hidden groupings or patterns signals unsupervised learning. “Segment customers into groups we didn’t predefine” is the classic example.
Reinforcement Learning
An agent learns by interacting with an environment, taking actions, and receiving rewards or penalties, optimizing its behavior over time. Think robotics, game-playing, and dynamic resource allocation.
Exam cue: language about an agent, actions, rewards, and trial and error points to reinforcement learning. AWS even has a learning-focused service, AWS DeepRacer, built around it.
| Paradigm | Data | Goal | Example |
|---|---|---|---|
| Supervised | Labeled | Predict a known target | Fraud detection, sales forecasting |
| Unsupervised | Unlabeled | Find structure | Customer segmentation, anomaly detection |
| Reinforcement | Reward signal | Learn optimal actions | Robotics, game AI, DeepRacer |
A bonus concept worth knowing: self-supervised learning, where the model generates its own labels from raw data, is how large foundation models are pre-trained. It bridges this domain to the generative-AI material covered in the Amazon Bedrock guide.
The Machine Learning Lifecycle
AIF-C01 expects you to recognize the end-to-end stages of building and operating an ML solution. You won’t execute these, but you must know what happens at each stage and which AWS service supports it.
- Business problem framing — define the goal and success criteria. Not every problem needs ML.
- Data collection — gather raw data (often into Amazon S3).
- Data preparation & feature engineering — clean, transform, and select features. Tools: SageMaker Data Wrangler, AWS Glue, SageMaker Feature Store.
- Model training — fit an algorithm to the prepared data. Tool: Amazon SageMaker training.
- Model evaluation — measure quality against held-out data using metrics (below).
- Model deployment — serve the model for real-time (endpoints) or batch inference.
- Monitoring & maintenance — watch for model drift and retrain as needed. Tool: SageMaker Model Monitor.
The single most testable idea here: ML is iterative, not linear. Poor evaluation results loop you back to data prep or training. Another favorite: the distinction between training (learning from data, compute-heavy, periodic) and inference (using the trained model to make predictions, latency-sensitive, continuous). And remember that model drift — performance decaying as real-world data shifts away from the training data — is why monitoring and retraining exist.
Frame problem → Collect data → Prepare data → Train → Evaluate → Deploy → Monitor
↑__________________ iterate as needed ______________________|How Models Are Evaluated
Evaluation metrics appear throughout the exam, so it’s worth a clear pass. The metric depends on the problem type.
Classification metrics (built from the confusion matrix of true/false positives and negatives):
| Metric | What it measures | When it matters |
|---|---|---|
| Accuracy | Overall % correct | Balanced classes; misleading when classes are skewed |
| Precision | Of predicted positives, how many were right | When false positives are costly (e.g., flagging good email as spam) |
| Recall | Of actual positives, how many you caught | When false negatives are costly (e.g., missing fraud or disease) |
| F1 score | Harmonic mean of precision and recall | When you need a balance and classes are imbalanced |
| AUC-ROC | Ability to separate classes across thresholds | Comparing classifiers overall |
Regression metrics: MAE, MSE, and RMSE — all measure how far predictions are from actual values, with lower being better.
Exam cue: a scenario about rare events (fraud, disease, defects) where missing a positive is dangerous emphasizes recall; a scenario where false alarms are expensive emphasizes precision. Knowing accuracy is misleading on imbalanced datasets is a classic AIF-C01 insight.
You’ll also meet two failure modes:
- Overfitting — the model memorizes training data and fails on new data (high training accuracy, low test accuracy).
- Underfitting — the model is too simple to capture the pattern (poor on both).
The fix space — more data, regularization, simpler/more complex models — is fair game at a conceptual level.
The AWS ML Stack
AWS frames its AI/ML services as a three-layer stack, and mapping a use case to the right layer is a recurring question pattern.
| Layer | What it is | Who it’s for | Examples |
|---|---|---|---|
| AI Services | Pre-trained, fully managed APIs — no ML expertise needed | Developers solving common tasks | Rekognition, Comprehend, Textract, Transcribe, Polly, Translate, Lex |
| ML Services / Platform | Build, train, and deploy custom models | Data scientists & ML engineers | Amazon SageMaker (and its suite) |
| Frameworks & Infrastructure | Low-level compute and frameworks | ML researchers, advanced users | EC2 (GPU/Trainium/Inferentia), TensorFlow, PyTorch |
Plus the generative AI layer that sits alongside: Amazon Bedrock (managed access to foundation models) and Amazon Q (AI assistants).
The exam cue is about effort vs control:
- Need a common capability (detect objects, extract text, transcribe audio) with no ML work? Pick the relevant AI service. Our AWS AI Services guide covers exactly which service maps to which task.
- Need a custom model trained on your own data? That’s SageMaker.
- Need to build on foundation models for generation? That’s Bedrock.
Memorize the AI-service-to-task mapping (Rekognition → images/video, Comprehend → NLP/sentiment, Textract → document text/forms, Transcribe → speech-to-text, Polly → text-to-speech, Translate → translation, Lex → chatbots). These are some of the easiest points on the entire exam.
Where Domain 1 Connects to the Rest of the Exam
Domain 1 is the trunk; the other domains are branches:
- Generative AI fundamentals (Domain 2) builds directly on “deep learning” and “foundation models” — see the prompt engineering guide.
- Applications of foundation models (Domain 3) assumes you understand the ML lifecycle (now with prompts, RAG, and fine-tuning).
- Responsible AI (Domain 4) extends evaluation into bias and fairness — covered in the Responsible AI guide.
- Security & governance (Domain 5) wraps the lifecycle in IAM, encryption, and compliance.
Treat Domain 1 as the investment that pays off everywhere else. Candidates who rush past it tend to second-guess questions in every other domain.
Study Strategy for Domain 1
This domain rewards crisp definitions and fast pattern-matching. An efficient approach:
- Drill the three ML types until you can classify any described problem in seconds.
- Memorize the lifecycle stages and the AWS service attached to each.
- Internalize precision vs recall with the “costly false positive vs costly false negative” framing.
- Lock in the three-layer stack and the AI-service-to-task table.
Reading builds the model, but the AIF-C01 is a timed, scenario-based exam — 65 questions in 90 minutes — and the only reliable way to confirm you can apply these fundamentals is to practice under exam conditions. Warm up with our free AWS AI Practitioner practice questions, and when you want full-length, timed mocks spanning every domain with detailed explanations, the AWS Certified AI Practitioner Mock Exam Bundle gives you eight full exams so you can find and close gaps before exam day.
For a structured week-by-week plan that sequences this domain with the rest, see the AWS AI Practitioner Study Plan, and if you’re mapping your broader AWS AI journey, the AWS AI certification path guide shows where AIF-C01 fits.
Conclusion
Domain 1 of the AIF-C01 isn’t the flashiest part of the exam, but it’s the most leveraged. Understand the nesting of AI, ML, and deep learning; classify problems into supervised, unsupervised, and reinforcement learning; trace the iterative ML lifecycle; read evaluation metrics with an eye for precision-versus-recall trade-offs; and map use cases to the right layer of the AWS ML stack. Those skills don’t just answer Domain 1 questions — they make every other domain easier to reason about.
Pair this conceptual foundation with timed, scenario-based practice, and the fundamentals become reliable points you can bank quickly, freeing time for the trickier generative-AI and governance scenarios later in the exam.
FAQ
How much of the AIF-C01 exam is the Fundamentals of AI and ML domain?
Domain 1 (Fundamentals of AI and ML) accounts for roughly 20% of the AWS Certified AI Practitioner exam. With 65 scored-and-unscored questions, expect a meaningful cluster of items testing core ML concepts and vocabulary.
Do I need a data science background to pass Domain 1?
No. AIF-C01 is a foundational, conceptual exam. You need to understand what the concepts mean and when to apply them — not the underlying mathematics. No coding or model training is required.
What’s the difference between supervised and unsupervised learning?
Supervised learning trains on labeled data to predict a known target (classification or regression). Unsupervised learning works with unlabeled data to discover structure, such as clustering or anomaly detection. The presence or absence of labels is the deciding factor.
When should I care about precision versus recall?
Emphasize recall when missing a positive is costly — fraud, disease, or defect detection, where false negatives are dangerous. Emphasize precision when false positives are costly — like flagging legitimate emails as spam. The F1 score balances both and is preferred on imbalanced datasets, where plain accuracy is misleading.
What is the AWS ML stack?
AWS organizes its AI/ML offerings into three layers: AI Services (pre-trained APIs like Rekognition and Comprehend that need no ML expertise), ML Services (Amazon SageMaker for building custom models), and Frameworks & Infrastructure (low-level compute and frameworks). Generative AI services like Amazon Bedrock sit alongside this stack.
What is model drift and why does it matter?
Model drift is the gradual decline in a model’s performance as real-world data shifts away from the data it was trained on. It’s the reason the ML lifecycle includes ongoing monitoring and retraining — a deployed model isn’t “done,” it needs maintenance to stay accurate.