AIF-C01 Exam Cheat Sheet: Complete Quick Reference for AWS AI Practitioner (2026)
The definitive AIF-C01 cheat sheet for 2026: every AWS AI service mapped to its use case, Bedrock component breakdown (Knowledge Bases, Agents, Guardrails, Flows), evaluation metrics, responsible AI vocabulary, and the 10 most common exam traps — all in one scannable reference. Review this the night before your exam.
This is the final-review cheat sheet for the AWS Certified AI Practitioner (AIF-C01). It is designed to be read the evening before your exam — a dense, scannable reference that hits every domain. It assumes you have already studied the material in depth. If something here feels unfamiliar, go back to the relevant deep dive. If everything feels familiar, you are ready. The exam is 65 questions in 90 minutes, passing score 700/1000. Let's make sure every point counts.
- Domain Weights at a Glance
- AWS AI Services: The Full Map
- Amazon Bedrock: Every Component Explained
- RAG vs. Fine-Tuning vs. Prompt Engineering
- Core ML Concepts Quick Reference
- Evaluation Metrics Card
- Responsible AI Vocabulary
- Security and Governance Reference
- Amazon Q: Developer vs. Business
- 10 Most Common Exam Traps
Domain Weights at a Glance
| Domain | Name | Weight | ~Qs |
|---|---|---|---|
| D1 | Fundamentals of AI and ML | 20% | ~13 |
| D2 | Fundamentals of Generative AI | 24% | ~16 |
| D3 | Applications of Foundation Models | 28% | ~18 |
| D4 | Guidelines for Responsible AI | 14% | ~9 |
| D5 | Security, Compliance, and Governance | 14% | ~9 |
Prioritization: D3 + D2 = 52% of the exam. Master Bedrock. Then D1 and D4/D5 become your point multipliers.
AWS AI Services: The Full Map
| Use Case | AWS Service | Key Distinguisher |
|---|---|---|
| Foundation models (LLMs, diffusion) | Amazon Bedrock | Managed API access to Claude, Titan, Llama, Mistral, Stability AI |
| Custom ML model training and deployment | Amazon SageMaker | Full MLOps platform; code-required |
| Image and video analysis | Amazon Rekognition | Object detection, faces, content moderation, PPE, celebrity |
| NLP: sentiment, entities, PII | Amazon Comprehend | Comprehend Medical for clinical notes |
| Speech to text | Amazon Transcribe | Speaker diarization; Transcribe Medical for HIPAA |
| Text to speech | Amazon Polly | Neural vs. standard voices; SSML support |
| Language translation | Amazon Translate | 75+ language pairs; custom terminology |
| Structured conversational bots | Amazon Lex | Intents + slots; rule-based dialogue flows |
| Document / form extraction | Amazon Textract | Key-value pairs, tables, layout from scanned docs |
| Personalized recommendations | Amazon Personalize | Behavioral data → real-time recs; no ML experience needed |
| Time-series forecasting | Amazon Forecast | DeepAR, ARIMA, ETS auto-selection; related time series |
| Online fraud detection | Amazon Fraud Detector | Payment fraud, account takeover; uses your fraud history |
| Enterprise document search | Amazon Kendra | Intelligent search across internal docs; often used as data source for Bedrock Agents RAG |
| Bias detection + model explainability | SageMaker Clarify | SHAP values; pre/post-training bias metrics |
| Human-in-the-loop review | Amazon Augmented AI (A2I) | Route low-confidence predictions to human workers |
| Code generation + completion | Amazon Q Developer | IDE plugin for code suggestions; replaces CodeWhisperer |
| Enterprise AI assistant | Amazon Q Business | Q&A over internal company data; connects to SharePoint, Slack, S3 |
Amazon Bedrock: Every Component Explained
Domain 3 (28%) is almost entirely about Bedrock. Know every component and when to use it:
Connect your data (S3, Confluence, SharePoint) to a foundation model via RAG. Bedrock automatically embeds, chunks, indexes, and retrieves relevant documents at query time. Exam: the go-to answer when a scenario says "answer questions based on our internal documentation without fine-tuning."
Orchestrate multi-step tasks by calling APIs (via Lambda action groups) and querying Knowledge Bases. The agent decides which tools to call and in what order to complete a goal. Multi-agent orchestration: a supervisor agent delegates subtasks to specialized sub-agents. Exam: answer for "take actions on behalf of the user" or "automate multi-step business workflows."
Add safety controls to any Bedrock model: content filtering (hate speech, violence), topic restrictions (block off-topic requests), PII detection and masking, grounding check (reduce hallucinations by comparing output to source documents), word filters. Exam: answer for "prevent the model from answering questions about competitors" or "ensure the model does not reveal customer PII."
Visual workflow builder (formerly "Prompt Flows") for chaining prompts, conditions, and data transformations into automated pipelines — without writing Lambda code. Think of it as low-code orchestration for GenAI workflows. Exam: answer when a scenario emphasizes "no-code" or "business analysts building workflows."
Version-controlled storage for prompt templates used across your Bedrock applications. Allows teams to manage, version, share, and audit prompts centrally. Exam: answer for "ensure prompt consistency across teams" or "audit which prompt version was used in production."
Run automatic or human-based evaluation jobs to compare model responses across multiple foundation models. Generates reports with quality metrics. Exam: answer for "compare which Bedrock model best fits our use case before choosing one."
- Anthropic Claude — conversational, long context, safety-focused; default choice for complex NLU
- Amazon Titan — AWS-built; Titan Text (generation), Titan Embeddings (RAG), Titan Image Generator
- Meta Llama — open-weight; good for fine-tuning on custom tasks
- Mistral AI — efficient, strong instruction following; cost-effective option
- Stability AI — Stable Diffusion; image generation and editing
- Cohere — embeddings and retrieval; often used for Bedrock Knowledge Bases embedding model
RAG vs. Fine-Tuning vs. Prompt Engineering
| Approach | When to Use | Cost | Data Freshness |
|---|---|---|---|
| Prompt Engineering | Simple tasks, well-scoped queries, few-shot examples fit in context | Lowest | N/A |
| RAG (Knowledge Bases) | Need up-to-date or proprietary knowledge; don't want to update the model | Medium (retrieval + inference) | High (docs updated anytime) |
| Fine-Tuning | Consistent output format, domain-specific style, specialized vocabulary the base model lacks | Highest (training + storage) | Low (requires retraining) |
Core ML Concepts Quick Reference
| Supervised Learning | Labeled training data. Examples: classification, regression. |
| Unsupervised Learning | Unlabeled data, find patterns. Examples: clustering, dimensionality reduction. |
| Reinforcement Learning | Agent + environment + rewards. Examples: game AI, robotics, ad bidding. |
| Self-supervised Learning | Model creates its own labels (next-token prediction). How LLMs are pretrained. |
| Overfitting | High train accuracy, low test accuracy. Fix: regularization, more data, simpler model. |
| Underfitting | Low accuracy on both train and test. Fix: more complex model, more features. |
| Tokenization | Breaking text into tokens (subwords). ~1 token ≈ 0.75 English words. |
| Embedding | Dense numerical vector representing semantic meaning of text. Similar meanings → similar vectors. |
| Context Window | Maximum tokens the model can process in one call (input + output). |
| Hallucination | Model generates plausible but factually incorrect outputs. Mitigate with RAG + Guardrails grounding check. |
| Temperature | Controls randomness. Low (0.0–0.3) → deterministic, factual. High (0.7–1.0) → creative, varied. |
| Top-P / Top-K | Alternative sampling controls. Top-P: nucleus sampling by cumulative probability. Top-K: sample from top K tokens. |
| Zero-shot | No examples in prompt; rely on model pre-training. Works when task is well-understood by the model. |
| Few-shot | Provide 2–5 examples in prompt. Improves accuracy for specific formats or edge cases without fine-tuning. |
| Chain-of-Thought | Prompt the model to reason step-by-step before answering. Improves complex reasoning accuracy. |
Evaluation Metrics Card
| Metric | Optimizes For | Use When |
|---|---|---|
| Accuracy | Overall correctness | Balanced classes only |
| Precision | Minimize false positives | Spam, ads, irrelevant alerts |
| Recall | Minimize false negatives | Medical diagnosis, fraud, security threats |
| F1 Score | Balance precision + recall | Imbalanced classes with both types of errors |
| AUC-ROC | Threshold-independent discrimination | Comparing classifiers regardless of threshold |
| RMSE / MAE | Regression error magnitude | Price prediction, forecasting |
| BLEU | N-gram overlap with reference | Translation quality evaluation |
| ROUGE | Recall-oriented n-gram overlap | Summarization evaluation |
| BERTScore | Semantic similarity (not just surface matching) | GenAI output quality; more nuanced than BLEU/ROUGE |
| Perplexity | Language model confidence | LLM fluency — lower perplexity = more confident/coherent |
Responsible AI Vocabulary
| Bias (pre-training) | Skew in training data that leads to unfair predictions (e.g., underrepresentation of demographic groups). Detected with SageMaker Clarify. |
| Bias (post-training) | Disparate outcomes for different groups in deployed model. Also measured with Clarify after model evaluation. |
| Explainability | Ability to understand why a model made a prediction. SageMaker Clarify uses SHAP values to assign feature importance. |
| Fairness | Consistent outcomes across demographic groups (gender, race, age). Evaluated with equalized odds, demographic parity metrics. |
| AI Service Card | AWS document describing the intended use, limitations, and responsible AI considerations of a specific AWS AI service (e.g., Rekognition, Transcribe). Published by AWS. |
| Model Card | Documentation created by a model developer (you or a third party) describing training data, evaluation results, limitations, and intended use cases of a specific ML model. |
| Amazon Augmented AI (A2I) | Route low-confidence AI predictions to human reviewers. Supports built-in workflows for Textract and Rekognition, or custom workflows. |
| Model Drift | Degradation in model performance over time as real-world data distribution changes. Detected with SageMaker Model Monitor. |
| Transparency | Users knowing when they are interacting with AI and understanding how decisions are made. |
Security and Governance Reference
| Shared Responsibility (AI) | AWS secures the infrastructure and managed services. Customer secures data, access controls, prompt content, and output handling. |
| IAM for Bedrock | Control which users/roles can invoke which Bedrock models or create Knowledge Bases. Use least-privilege IAM policies. |
| VPC Endpoints for Bedrock | Keep Bedrock API calls within the AWS network — no data traverses the public internet. |
| Data Privacy in Training | AWS does not use your data sent to Bedrock to train its models (opt-in model customization is separate). Your prompts are not shared with model providers. |
| AWS CloudTrail for AI | Log all API calls to Bedrock, SageMaker, Rekognition for audit trail and compliance. |
| AWS Config | Track configuration changes to AI resources for governance and compliance. |
| Bedrock Guardrails (security angle) | Prevent prompt injection attacks; block sensitive topic disclosure; mask PII in model outputs. |
Amazon Q: Developer vs. Business
Amazon Q appears increasingly as both answer option and distractor. Know the distinction cold:
- AI-powered coding assistant
- IDE plugin (VS Code, JetBrains)
- Code completion, generation, security scanning
- Explain and transform legacy code
- Successor to Amazon CodeWhisperer
- Target user: individual developers
- Enterprise AI assistant / chatbot
- Connects to internal data (S3, SharePoint, Confluence, Slack, Salesforce)
- Employees ask questions about company policies, procedures, data
- Access controls: users only see data they're authorized to access
- Target user: non-technical business users
10 Most Common Exam Traps
-
Rekognition vs. Textract for "text in images"
Text in a natural scene (sign, label) → Rekognition. Structured text in a document/form → Textract. -
RAG vs. Fine-Tuning: "proprietary knowledge" vs. "consistent style"
Access to internal/up-to-date documents → RAG (Bedrock Knowledge Bases). Consistent output format or specialized vocabulary baked into the model → fine-tuning. -
AI Service Card vs. Model Card
AI Service Card = published by AWS for their pre-built services. Model Card = published by the model developer for a specific trained model. -
SageMaker Clarify vs. Bedrock Guardrails for "bias"
Clarify detects bias in custom ML models during training and evaluation. Guardrails prevents harmful or off-topic model outputs at inference time. -
Amazon A2I vs. Bedrock Agents for "human review"
A2I = route low-confidence ML predictions to a human workforce for review. Agents = autonomous multi-step task execution (no human in the loop by default). -
Amazon Lex vs. Amazon Bedrock for chatbots
Lex = structured intent-based dialogue (book a flight, check account balance). Bedrock = open-ended generative AI conversations (any topic, long-form answers). -
Q Developer vs. Q Business
Q Developer = developer tool for code generation. Q Business = enterprise assistant connecting to internal company data for all employees. -
Precision vs. Recall: read the cost of errors
The scenario will tell you which error is more costly. "Missing a positive" (e.g., disease, fraud, security threat) → maximize Recall. "False alarm costs" (spam, unnecessary procedures) → maximize Precision. -
Bedrock Flows vs. Bedrock Agents
Flows = visual low-code workflow builder for chaining prompts. Agents = autonomous LLM that calls tools (APIs) and reasons about multi-step tasks. -
Accuracy on imbalanced datasets
A model with 99% accuracy on a 1% fraud dataset is detecting nothing. Always flag accuracy as misleading for imbalanced classes — use F1, Precision, Recall, or AUC-ROC instead.
Run one full 65-question practice exam on CertLand. Review every incorrect answer — not to memorize the answer, but to understand the reasoning. If you're scoring above 75%, you are in good shape for exam day. Good luck!
Comments
No comments yet. Be the first!
Comments are reviewed before publication.