Complete guide
AIF-C01 AWS Certified AI Practitioner
Tap any service to see more details and official docs.
Exam breakdown
40–60
Questions
~45
Min read
700
Passing score
AI and ML Fundamentals
20% of the examTypes of machine learning
| Type | How it learns | Example |
|---|---|---|
| Supervised | With labeled data (known input → known output) | Classifying emails as spam/not spam |
| Unsupervised | Finds patterns without labels | Grouping customers by behavior (clustering) |
| Reinforcement | Trial and error, receives rewards or penalties | An agent that learns to play by optimizing its score |
Lifecycle of an ML project
Data collection
Gather relevant and sufficient data for the problem.
Data preparation
Clean, transform, and label (if applicable).
Training
The model learns patterns from the data.
Evaluation
Measure accuracy with data the model hasn't seen.
Deployment
Put the model into production (endpoint).
Monitoring
Detect model degradation or drift over time.
Service for building your own models
Click any service to see more details
Overfitting = the model memorizes the training data but fails on new data (too complex). Underfitting = the model is too simple and doesn't even learn the training data well.
Fundamentals of Generative AI
24% of the examWhat a foundation model is
A foundation model is a model trained on massive amounts of data, capable of adapting to many different tasks (text, image, code) without being trained from scratch for each one — unlike traditional ML, where each model solves one specific problem.
Key generative AI concepts
| Concept | What it means |
|---|---|
| Token | The smallest unit of text the model processes (a word or part of a word) |
| Embedding | A numerical representation (vector) of the meaning of a text, used to search for semantic similarity |
| Context window | The maximum number of tokens the model can "remember" within a single conversation or prompt |
| Temperature | Controls randomness: high = more creative/varied, low = more deterministic/consistent |
| Prompt engineering | Designing the input text to get better responses, without modifying the model |
Amazon Bedrock — access to foundation models
Click any service to see more details
Foundation models can "hallucinate": generate information that sounds convincing but is false. RAG reduces this risk by grounding responses in real documents, but doesn't eliminate it entirely.
Applications of Foundation Models
28% of the examPrompt engineering vs RAG vs Fine-tuning
This is the single heaviest-weighted topic on the exam: identifying which technique to use for a given scenario.
| Technique | When to use it | Effort |
|---|---|---|
| Prompt engineering | Adjusting how you ask for something, without additional data or retraining | Low — immediate |
| RAG (Bedrock Knowledge Bases) | The model needs to answer using your own, up-to-date document data | Medium — no model retraining |
| Fine-tuning | You need the model to permanently and consistently change its behavior or tone | High — requires training data and compute |
Pre-trained AI services by use case
Click any service to see more details
D3 carries the most weight on the exam (28%). Memorize the right service for each use case: images→Rekognition, documents→Textract, speech→text→Transcribe, text→speech→Polly, text analysis→Comprehend, chatbots→Lex, internal search→Kendra. And for generative AI: prompt engineering vs RAG vs fine-tuning depending on the scenario.
Guidelines for Responsible AI
14% of the examDimensions of responsible AI
⚖️ Fairness
The model does not discriminate against or harm specific groups
🔍 Explainability
You can understand why the model made a decision
🛡️ Robustness
The model performs consistently against unexpected inputs
👁️ Transparency
It's clear what data and limitations the model has
🔒 Privacy
Personal data used is adequately protected
🚨 Safety
The system avoids generating harmful or dangerous content
Responsible AI tools on AWS
| Tool | Function |
|---|---|
| SageMaker Clarify | Detects bias in training data and explains model predictions |
| SageMaker Model Monitor | Detects drift or degradation of a model already in production |
| Bedrock Guardrails | Filters harmful content, prohibited topics, and sensitive data in responses |
| Model Cards | Standard documentation about a model's purpose, limitations, and performance |
Bias in a model almost always comes from the training data, not the algorithm itself — if the historical data reflects an inequality, the model learns it and reproduces it.
Security, Compliance, and Governance
14% of the examShared responsibility applied to AI
🏢 AWS
- • Compute infrastructure for AI
- • Security of the base foundation model
- • Availability of the AI services
👤 Customer
- • Your data and prompts
- • Access control (IAM) to the AI services
- • Regulatory compliance for your use case
Relevant security and governance services
| Service | Function in an AI context |
|---|---|
| AWS IAM | Controls who can invoke models, create endpoints, or access training data |
| AWS KMS | Encrypts training data, model artifacts, and data in transit |
| AWS CloudTrail | Audits which calls were made to the AI services and who made them |
| Amazon Macie | Discovers and classifies sensitive data (PII) before it's used to train or provide context to a model |
The exam does not evaluate how to configure each security service in detail — it evaluates whether you understand which part of AI security is yours (data, prompts, access) and which part is AWS's (infrastructure, the base model).
Tips for exam day
Supervised learning uses labeled data (input → known output). Unsupervised learning finds patterns without labels (clustering). Reinforcement learning learns by trial and error with rewards.
Bedrock = use already-trained foundation models via API, with no infrastructure of your own. SageMaker = build and train your own models from scratch. The exam constantly compares the two.
Prompt engineering (fast, cheap, no data) vs RAG (connects your data without retraining) vs Fine-tuning (slow, expensive, changes the model's behavior). In that order of increasing effort.
RAG = Retrieval-Augmented Generation: it searches your documents for relevant information and gives it to the model as context before it answers. It does not modify the model's weights.
High temperature = more creative/varied responses. Low temperature = more deterministic/consistent responses. Top-p and top-k control the pool of words the model chooses from.
Responsible AI on the exam: fairness, explainability, robustness, transparency, privacy, and safety are the dimensions AWS evaluates.
SageMaker Clarify detects bias in training data and explains why a model made a decision (feature importance). It shows up in almost every D4 question.
Bedrock Guardrails filters harmful content, prohibited topics, and PII in responses — it's a safety layer, it does not improve model accuracy.
Pre-trained AI services by use case: Rekognition = images/video, Textract = structured documents, Transcribe = speech→text, Polly = text→speech, Comprehend = NLP/sentiment, Lex = chatbots, Kendra = semantic search.
Overfitting = the model memorizes the training data and fails on new data. Underfitting = the model is too simple and doesn't even learn the training data well.
The shared responsibility model also applies to AI: AWS secures the infrastructure and the base model's weights; you're responsible for your prompts, your data, and access control to your application.
The exam doesn't expect you to memorize exact prices, but it does expect you to understand the relative order of cost/effort: using a pre-trained model < prompt engineering < RAG < fine-tuning < training from scratch.
Want to go deeper?
This guide is a quick summary. Visit the Deep Dive section for an exhaustive walkthrough with real cases, cost analysis and architecture decisions.
Practice AIF-C01 with interactive simulations
Click-through labs: navigate a step-by-step portal simulation, with clear objectives and verification on every action — no real costs.
Ready to practice?
Put what you learned to the test with original AIF-C01 exam questions.
Start AIF-C01 mock examUnlimited mock exams with the Pro plan
Was this guide helpful?
Not ready to start yet?
Leave your email and create your free account whenever you're ready — start with 15 AIF-C01 questions, no credit card required.
Already have an account? Go straight to the mock exam →