AWS Certified AI Practitioner (AIF-C01) Exam Guide
AWS Certified AI Practitioner (AIF-C01) Exam Guide
The AWS Certified AI Practitioner (AIF-C01, Exam Guide v1.1) is a foundational AWS certification validating comprehensive knowledge of artificial intelligence, machine learning concepts, generative AI technologies, and associated AWS services.
Interactive Practice Available: Drill the 200 curated exam questions (100 Core + 100 Scenario questions with Easy/Medium/Hard ratings) or take a full-length 90-minute mock exam with instant scoring in our Exam Crack Buddy Console.
📋 Exam Blueprint & Facts
| Attribute | Specification |
|---|---|
| Exam Code | AIF-C01 (Exam Guide version 1.1) |
| Questions | 65 questions (50 scored + 15 unscored pretest items) |
| Time Limit | 90 minutes (approx. 80 seconds per question) |
| Passing Score | 700 on a 100–1,000 scaled score |
| Question Types | Multiple choice, multiple response, ordering, matching |
| Scoring Model | Compensatory: pass overall, no negative marking for incorrect guesses |
Domain Weights
┌─────────────────┬─────────────────┬──────────────────────┬─────────────────┬─────────────────┐
│ Domain 1: AI/ML │ Domain 2: GenAI │ Domain 3: FM Apps │ Domain 4: Resp. │ Domain 5: Sec. │
│ 20% │ 24% │ 28% │ 14% │ 14% │
└─────────────────┴─────────────────┴──────────────────────┴─────────────────┴─────────────────┘
- Domain 1: Fundamentals of AI and ML — 20% (~13 questions)
- Domain 2: Fundamentals of Generative AI — 24% (~16 questions)
- Domain 3: Applications of Foundation Models — 28% (~18 questions)
- Domain 4: Guidelines for Responsible AI — 14% (~9 questions)
- Domain 5: Security, Compliance, and Governance for AI Solutions — 14% (~9 questions)
🗓️ 10-Day Sprint Plan
- Day 1 — Domain 1 (Part A): AI/ML/DL/GenAI/agentic hierarchy, supervised vs unsupervised vs reinforcement learning, inference types, model fit, evaluation metrics. Practice D1 Q1–10.
- Day 2 — Domain 1 (Part B): The machine learning lifecycle, SageMaker AI capabilities, AWS managed AI services map, traditional ML vs foundation models. Practice D1 Q11–20.
- Day 3 — Domain 2 (Part A): Tokenization, embeddings, vector spaces, transformers, diffusion models, hallucinations, token pricing models, context engineering. Practice D2 Q1–12.
- Day 4 — Domain 2 (Part B): Agentic workflows (Model Context Protocol, multi-agent systems, memory), Amazon Bedrock architectures, SageMaker JumpStart, Amazon Q business & developer tools. Practice D2 Q13–24.
- Day 5 — Domain 3 (Part A): Model inference parameters (temperature, top-p, top-k, stop sequences), RAG concepts, Amazon Bedrock Knowledge Bases, vector database options, customization cost ladder. Practice D3 Q1–14.
- Day 6 — Domain 3 (Part B): Advanced prompting strategies (Zero-shot, Few-shot, Chain-of-Thought), prompt security vulnerabilities, fine-tuning requirements, RLHF, FM evaluation metrics. Practice D3 Q15–28.
- Day 7 — Domain 4: The dimensions of responsible AI, bias mitigation, explainability, SageMaker Clarify, Model Cards, Guardrails for Amazon Bedrock, Amazon Augmented AI (A2I), AWS AI Service Cards. Practice all Domain 4 questions.
- Day 8 — Domain 5: The AWS Shared Responsibility Model for AI, IAM least-privilege policies, AWS KMS encryption, Amazon Macie, PrivateLink VPC endpoints, CloudTrail, AWS Config, AWS Artifact, AWS Audit Manager. Practice all Domain 5 questions.
- Day 9 — Mock Exam 1: Sit for a full 65-question timed mock test. Review all incorrect and flagged items against memory hooks.
- Day 10 — Mock Exam 2 & Final Polish: Target weak domains, retake missed questions, take Mock Exam 2. Schedule your exam when scoring 85%+ consistently.
🧩 Domain 1: Fundamentals of AI & ML (20%)
1. Fundamental Terminology
- Machine Learning (ML): Developing algorithms that generalize patterns from empirical training data rather than relying on explicit procedural code.
- Deep Learning (DL): Complex artificial neural networks with multiple stacked hidden layers capable of learning non-linear feature representations.
- Agentic AI: Autonomous multi-step systems leveraging foundation models with tool-use (function calling), scratchpad memory, and iterative reasoning loops to accomplish open-ended goals.
Memory Hook:
- Classification: Answers “Which bucket?” (discrete classes).
- Regression: Answers “How much / what number?” (continuous numerical values).
- Clustering: Answers “Who is similar?” (unsupervised grouping without labels).
2. Amazon SageMaker AI Inference Types
| Inference Type | Latency | Traffic Pattern | Payload Size | Typical Use Case |
|---|---|---|---|---|
| Real-time | Milliseconds | Steady, consistent | Up to 6 MB | Interactive customer applications, live web fraud scoring |
| Serverless | Milliseconds (cold-start tolerant) | Intermittent, spiky, idle periods | Up to 4 MB | Internal apps with sporadic usage, cost minimization |
| Asynchronous | Minutes (queued) | Large payloads, long inference runs | Up to 1 GB (via S3) | Complex image rendering, batch document OCR, video analysis |
| Batch Transform | Hours | Scheduled bulk processing | Entire S3 datasets | Weekly customer churn scoring, overnight reporting |
Memory Hook: Now (Real-time), Sometimes (Serverless), Later (Asynchronous), All at once (Batch Transform).
3. Model Fit & Classical ML Metrics
- Overfitting (High Variance): Model memorizes the training dataset noise and performs poorly on unseen validation data.
- Remedies: Collect more/diverse training data, apply regularization ($L_1 / L_2$), apply dropout, implement early stopping, reduce model complexity.
- Underfitting (High Bias): Model is too simple to capture the underlying pattern in data.
- Remedies: Add more features, engineer new domain features, increase model capacity/parameters.
- Classification Metrics:
- Precision: $\frac{TP}{TP + FP}$ — Of all positive predictions, how many were actually positive? Critical when false alarms are costly (e.g., spam filtering).
- Recall (Sensitivity): $\frac{TP}{TP + FN}$ — Of all actual positive cases, how many did we catch? Critical when missing a positive is catastrophic (e.g., cancer detection, fraud detection).
- F1 Score: Harmonic mean of precision and recall ($2 \times \frac{P \times R}{P + R}$). Best metric for imbalanced classes.
- ROC-AUC: Area under the Receiver Operating Characteristic curve. Measures trade-off between True Positive Rate and False Positive Rate across all classification thresholds.
⚡ Domain 2: Fundamentals of Generative AI (24%)
1. Mechanics of Generative Models
- Tokens: The fundamental discrete atomic chunks (words, characters, or subwords) processed by language models. In English, $1,000 \text{ tokens} \approx 750 \text{ words}$.
- Embeddings: High-dimensional vector representations that capture semantic meaning. Concepts with similar meanings map to proximal vectors in geometric embedding space (e.g., via cosine similarity).
- Transformer Architecture:
- Self-Attention: Enables the model to calculate contextual relationships between all words in a sequence simultaneously, regardless of their distance.
- Feed-Forward Layers: Project attention outputs into dense semantic feature representations.
- Diffusion Models: Image generation architectures that iteratively reverse Gaussian noise added to an image during training to synthesize photo-realistic visuals from text prompts (e.g., Amazon Titan Image Generator).
2. Hallucinations & Context Engineering
- Hallucination: An LLM generating factually incorrect, fabricated, or ungrounded assertions with high statistical confidence.
- Mitigation Strategies:
- Use Retrieval Augmented Generation (RAG) to anchor responses in verified enterprise knowledge bases.
- Set lower temperature parameters ($0.0 - 0.2$).
- Prompt instructions requiring the model to state “I don’t know based on the provided context” when information is missing.
- Utilize Guardrails for Amazon Bedrock to verify factual consistency and grounding.
3. Managed Generative AI on AWS
- Amazon Bedrock: Fully managed service offering choice of high-performing Foundation Models (FMs) from leading AI companies (Anthropic Claude, Meta Llama, Mistral, AI21, Cohere, Amazon Titan) via a single unified serverless API, with built-in security, privacy, and responsible AI guardrails.
- Amazon SageMaker JumpStart: Machine learning hub offering pretrained open-source and proprietary foundation models with complete infrastructure customization, private VPC deployment, and full hyperparameter control.
- Amazon Q Business: Generative AI-powered conversational assistant connected to enterprise data stores (Wikis, Jira, Salesforce, S3) with role-based access control and source citations.
- Amazon Q Developer: Coding companion that provides real-time code recommendations, vulnerability scanning, and automated application upgrades (Java, .NET) in the IDE.
🚀 Domain 3: Applications of Foundation Models (28%)
1. Inference Parameters
| Parameter | Behavior | Low Value Effect | High Value Effect |
|---|---|---|---|
| Temperature | Controls randomness of token selection | Deterministic, focused, analytical (ideal for coding/math: $0.0 - 0.2$) | Creative, varied, surprising (ideal for marketing/creative writing: $0.7 - 1.0$) |
| Top-P (Nucleus Sampling) | Considers smallest set of tokens whose cumulative probability exceeds $P$ | Strictly high-probability words | Considers broader tail of vocabulary words |
| Top-K | Limits choice to only the $K$ most probable next tokens | Narrow, predictable choices | Broader selection pool |
| Stop Sequences | Specific character strings that halt generation | N/A | Ensures clean termination (e.g., \n\nUser:) |
Best Practice: Adjust either Temperature or Top-P when tuning generation, not both simultaneously.
2. Retrieval-Augmented Generation (RAG) Architecture
User Query ──► Embedding Model ──► Vector Search (Knowledge Base)
│
▼ Retrieved Relevant Chunks
User Query + Contextual Chunks ────────► Foundation Model ──► Grounded Response + Citations
- Amazon Bedrock Knowledge Bases: Managed RAG pipeline that handles document ingestion from S3, automated text chunking, embedding generation via Titan Multimodal, and storage in a vector store.
- Supported Vector Stores:
- Amazon OpenSearch Serverless (Vector Engine)
- Amazon Aurora PostgreSQL with
pgvector - Amazon Neptune Analytics
- Pinecone / MongoDB Atlas
3. The Customization Cost & Complexity Ladder
Low Cost / Fast ────────────────────────────────────────────────────────► High Cost / Complex
1. Prompt Engineering 2. RAG (Knowledge Bases) 3. Model Fine-Tuning 4. Pre-training from Scratch
- Prompt Engineering: Optimize instructions, system prompts, and formatting. Zero infrastructure cost.
- RAG: Dynamically supply external, real-time, private enterprise data into the prompt context window without retraining the model.
- Fine-Tuning: Update internal model weights using a curated dataset of prompt-completion pairs to adapt tone, format, or specialized domain vocabulary. Requires labeled data and GPU compute hours.
- Pre-training from Scratch: Train a new model from random initialization on billions of tokens. Multi-million dollar endeavor reserved for sovereign or highly proprietary domains.
4. Advanced Prompting Techniques
- Zero-Shot: Providing a prompt with no prior examples.
- Few-Shot: Providing 2–5 structured input-output exemplars in the prompt context to demonstrate the desired format or reasoning pattern.
- Chain-of-Thought (CoT): Guiding the model to output intermediate reasoning steps before providing the final answer (“Think step by step”). Dramatically improves math and multi-step logic.
- Prompt Injection: An attack where user input manipulates the model into ignoring its system safety instructions and executing unauthorized actions.
🛡️ Domain 4: Guidelines for Responsible AI (14%)
1. The Core Pillars of Responsible AI
- Fairness: Ensuring model predictions do not exhibit algorithmic bias or discriminate against protected groups.
- Explainability & Interpretability: Providing transparent insights into why a model reached a specific decision.
- Robustness: Ensuring model stability and predictable behavior when encountering out-of-distribution or noisy inputs.
- Privacy & Security: Preventing leakage of personally identifiable information (PII) or proprietary training data.
- Transparency: Disclosing when a user is interacting with an AI system and publishing clear system documentation.
2. AWS Responsible AI Toolset
- Amazon SageMaker Clarify:
- Evaluates pre-training dataset bias and post-training model prediction bias across demographic groups.
- Generates feature attribution reports using SHAP (SHapley Additive exPlanations) values to explain which features influenced predictions most.
- SageMaker Model Cards: Standardized digital fact sheets documenting model architecture, intended use cases, training datasets, performance benchmarks, and ethical considerations.
- Guardrails for Amazon Bedrock:
- Enforces customized content filtering across inputs and model responses.
- Redacts or masks Sensitive PII (Social Security numbers, credit cards, emails).
- Filters harmful categories (Hate speech, violence, sexual content).
- Detects and prevents prompt injection attacks and validates factual grounding to prevent hallucinations.
- Amazon Augmented AI (A2I): Coordinates human-in-the-loop (HITL) review workflows for low-confidence model predictions or sensitive business decisions.
- AWS AI Service Cards: Public documentation provided by AWS outlining the responsible use, limitations, and operational best practices for AWS pre-built AI services (Rekognition, Textract, Transcribe).
🔒 Domain 5: Security, Compliance, & Governance (14%)
1. Shared Responsibility Model for Generative AI
- AWS Responsibility: Physical data center security, host infrastructure, hardware virtualization, foundational model security (for AWS-provided models), and compliance certifications (SOC, HIPAA, ISO).
- Customer Responsibility: Prompt inputs, training data security, access control policies (IAM), encryption key management (KMS), output validation, data classification, and regulatory compliance.
Bedrock Privacy Guarantee: AWS does NOT use customer prompts, responses, or fine-tuning data to train or improve any base AWS foundation models, nor is customer data shared with third-party model providers.
2. Security Services Architecture
- AWS Identity and Access Management (IAM): Enforces least-privilege access to Bedrock model invocation APIs (
bedrock:InvokeModel). - AWS Key Management Service (KMS): Encrypts model weights, fine-tuning training datasets in S3, Bedrock Knowledge Base vector embeddings, and transient caches using customer-managed KMS keys (CMKs).
- AWS PrivateLink (VPC Endpoints): Establishes private network connectivity between VPC subnets and Amazon Bedrock APIs without sending traffic over the public internet.
- Amazon Macie: Employs pattern matching and machine learning to automatically discover, classify, and protect sensitive PII in S3 buckets used for model training or RAG ingestion.
- AWS CloudTrail: Logs every API call made across all AI services (who invoked what model, from which IP, and when) for forensic auditing.
- AWS Artifact: Enterprise portal providing on-demand access to AWS compliance reports, certifications, and Business Associate Agreements (BAAs) for HIPAA compliance.
- AWS Audit Manager: Automates continuous evidence collection to demonstrate compliance with industry frameworks and regulations (GDPR, EU AI Act, NIST AI RMF).
🎯 Test Day Traps & Exam Tips
- Unscored Questions: 15 of the 65 questions are pretest items that do not count toward your score. Never stress over an unusually difficult question; maintain pace and answer every item.
- No Negative Marking: Unanswered questions receive zero marks. Never leave any question blank. If uncertain, eliminate obviously flawed options and select your best guess.
- Qualifiers: Always look for qualifiers: LEAST operational overhead usually points to serverless managed solutions (Amazon Bedrock, pre-built AI services) rather than configuring EC2 or self-managed clusters.
- Clarify vs Guardrails: SageMaker Clarify is for bias analysis and feature explainability. Guardrails for Bedrock is for real-time content filtering, PII redaction, and prompt injection defense.
- Human Review: If a question asks how to involve human reviewers when model confidence drops below a threshold, the correct answer is Amazon Augmented AI (A2I).
🚀 Take a Timed Practice Exam
Master the concepts above by practicing with the official question pool and testing under full exam timer conditions:
How would you rate this article?
Was this article helpful?