AWS AIF-C01 cheat sheet
AWS (Amazon Web Services) · AI Practitioner (AIF-C01)
QH Verified
By SpyderMan
This is not authorized to be used elsewhere but I hope this helps others practice for this exam.
Cheat Sheet
AWS AI Practitioner Exam Mastery
AIF-C01Please use my guide to finalize your training for the AWS Certified AI Practitioner exam. Reproduction of this is not allowed.
Created by:Mr. Sparkles
The Golden Rules
- Domain 3 is 28%. If you only drilled AI vocabulary, you are underprepared.
- Know the distinctions cold: training vs inference, RAG vs fine-tuning, transparency vs explainability, grounding vs guardrails.
- Service job first, buzzword second. Bedrock = FM access. SageMaker AI = build/train/deploy. Clarify = bias/explainability.
- Need current facts? Prefer RAG. Need behavior/style change? Consider fine-tuning. Do not swap those.
- Compensatory scoring: 700 / 1000 pass. You do not need every domain individually.
- An agent can plan and call tools. It does not replace IAM or business authorization.
1
Blueprint Weights AIF-C01
| Domain | Weight | Focus |
|---|---|---|
| 1 · Fundamentals of AI & ML | 20% | Terms, learning types, inference, data, lifecycle, metrics, MLOps |
| 2 · Fundamentals of GenAI | 24% | Tokens, embeddings, FMs, limits, context, agents, AWS GenAI services |
| 3 · Applications of Foundation Models | 28% | FM selection, RAG, prompting, customization, evaluation, cost |
| 4 · Responsible AI | 14% | Bias, fairness, safety, veracity, transparency, explainability |
| 5 · Security, Compliance & Governance | 14% | IAM, encryption, AI threats, grounding, logging, compliance services |
Scoring: Scaled 100–1,000; 700 to pass. Compensatory scoring — weak domains can be offset. Confirm the live exam guide before you book.
2
Study priorities
| Bucket | Drill |
|---|---|
| Distinctions | Training vs inference · classification vs regression · RAG vs fine-tuning · transparency vs explainability |
| Service jobs | Bedrock, SageMaker AI, Lex, Polly, Transcribe, Translate, Comprehend, Rekognition, Textract |
| GenAI controls | Temperature, token limits, context, grounding, Guardrails, Prompt Management |
| Security | IAM, KMS, Macie, PrivateLink, CloudTrail, Config, Audit Manager, Artifact |
Exam angle: Domain 3 alone is bigger than Domains 4 + 5 combined. Spend most of your last week there.
3
AI / ML hierarchy Domain 1 · 20%
AI → ML → Deep Learning → Neural Nets
- AI: systems performing tasks associated with intelligence
- ML: learns patterns from data instead of hard-coded rules
- Deep learning: ML with multi-layer neural networks
- GenAI: generates new content; often built on foundation models
- Agentic AI: pursues goals via reasoning, tools, memory/context, and actions
Learning types
| Type | Cue | Typical use |
|---|---|---|
| Supervised | Labeled examples | Classification, regression |
| Unsupervised | No target labels | Clustering, pattern discovery |
| Reinforcement | Rewards / penalties | Learn actions through interaction |
Task → technique
| Need | Technique |
|---|---|
| Predict a category | Classification |
| Predict a number | Regression |
| Find natural groups | Clustering |
4
Inference, data & lifecycle
| Inference mode | Think |
|---|---|
| Batch | Many records together; latency less critical |
| Real-time | Immediate response for interactive requests |
| Asynchronous | Completes later; caller need not wait |
| Serverless | Provider manages infra; app focuses on requests |
Data shapes
- Structured: schema, tables
- Unstructured: text, images, audio, video
- Labeled / unlabeled: supervised vs not
- Time series / tabular: time-linked vs rows/cols
ML lifecycle
- Collect → Explore → Prepare
- Features → Train → Tune
- Evaluate → Deploy → Monitor
MLOps: repeatable experiments, scalable production, monitoring, retraining, tech-debt control, production readiness.
5
Metrics picker
| Metric | Meaning | Watch for |
|---|---|---|
| Accuracy | Correct ÷ all predictions | Misleading with severe class imbalance |
| Precision | TP ÷ predicted positives | False positives are costly |
| Recall | TP ÷ actual positives | Missing positives is costly |
| F1 | Harmonic mean of precision & recall | Need balance of both |
| AUC | Ranking / discrimination across thresholds | Higher ≈ better separation |
| ROI | Business return vs investment | Business metric, not model quality |
| Cost / user or interaction | Economic cost of the AI capability | Production viability |
| Customer feedback | User-reported value / quality | Links model quality to outcomes |
Trap: High accuracy on an imbalanced dataset can still be useless. Ask what error type hurts the business.
6
GenAI vocabulary Domain 2 · 24%
| Term | Quick recognition |
|---|---|
| Foundation model | Broadly trained; adaptable to many tasks |
| LLM | FM specialized in language |
| Transformer | Architecture built on attention |
| Token | Unit of text the model processes |
| Embedding | Numeric representation of meaning |
| Vector | Ordered numbers for similarity math |
| Chunking | Split content into retrievable pieces |
| Context engineering | Deliberately build the context fed to an FM |
| Multimodal | Text + images / audio / video |
| Diffusion | Common generative path for image/audio |
FM lifecycle
Data selection Model selection Pre-training Fine-tuning Evaluation Deployment Feedback
7
GenAI reality & tokens
| Strength | Tradeoff |
|---|---|
| Adaptability | Inconsistent across domains |
| Responsiveness | Latency and token limits still matter |
| Conversation | Context can be incomplete or wrong |
| Content generation | Hallucination risk remains |
Model selection factors
Capability Modality Performance Cost Latency Complexity Compliance Availability Regional coverage Context / I/O length
- Token pricing: input and/or output volume drives cost
- Longer context = more info, more compute and money
- Output limits cap length, latency, and cost
- Prompt caching can cut repeated-input work where supported
Agentic AI: goal → plan → authorized tools → state/memory → observe → continue/stop. Still needs IAM.
8
FM applications Domain 3 · 28%
| Requirement | Compare |
|---|---|
| Interactive UX | Latency, throughput, response length, availability |
| Global audience | Languages, modality, regional availability |
| Cost-sensitive | Token pricing, model size, caching, routing |
| Specialized domain | Prompting / RAG first; fine-tune or continued pre-training when warranted |
| High-risk use | Compliance, explainability, human oversight, grounding, validation |
Prompt engineering
- Instruction: what to do
- Context: info needed for the task
- Negative prompt: what to avoid
- Prompt template: reusable structure + variables
- Zero-shot: no examples
- Single-shot: one example
- Few-shot: multiple examples
- Chain-of-thought: structured reasoning guidance — do not force private reasoning as required output
9
RAG vs customization
| Approach | Best recognition | Tradeoff |
|---|---|---|
| In-context learning | Examples/context in the request | Fast to change; eats context |
| RAG | Retrieve current knowledge at inference | Needs retrieval pipeline; grounds without weight changes |
| Fine-tuning | Adapt behavior with training data | Prep/cost; not first answer for changing facts |
| Continuous pre-training | Further pre-train on domain data | Expensive and operationally heavy |
| Distillation | Transfer capability into a smaller model | Lower cost/latency; may lose capability |
RAG pipeline
Source Chunk Embed Vector store Retrieve Augment prompt Generate Validate
Vector-store examples in guide: OpenSearch Service, Aurora, Neptune, RDS for PostgreSQL.
10
Inference controls & evaluation
| Control | Effect |
|---|---|
| Temperature | Higher → more variability; lower → more predictable |
| Input length | More context, more tokens / cost / latency |
| Output length | Caps response size, cost, latency |
| Prompt caching | Cuts repeated processing where supported |
| Prompt Management | Bedrock reusable prompts / versions |
Prompt hygiene: specific + concise, clear output contract, relevant context, examples when helpful, experiment systematically, version meaningful changes.
Prompt risks
Prompt injection Prompt hijacking Prompt poisoning Jailbreaking Sensitive-data exposure
Evaluation toolkit
| Method | Use |
|---|---|
| Human evaluation | Usefulness, quality, safety, task fit |
| Benchmark dataset | Repeatable comparison |
| ROUGE | Overlap-oriented; often summarization |
| BLEU | Overlap-oriented; often translation |
| BERTScore | Semantic similarity via contextual reps |
| LLM-as-a-judge | Another model scores against criteria |
| Task completion / satisfaction / cost | Business outcome metrics |
11
AWS GenAI capability map
| Service / capability | Fast recognition | Exam association |
|---|---|---|
| Amazon Bedrock | Managed FM access + GenAI app capabilities | FM access, inference, app building |
| Bedrock Knowledge Bases | Managed retrieval / grounding for RAG | External knowledge + RAG |
| Bedrock Agents | Multi-step orchestration with tools/actions | Agents and workflows |
| Bedrock Guardrails | Configurable input/output safeguards | Safety, topics, PII / harmful content |
| Bedrock Prompt Management | Templates and versions | Prompt reuse |
| Bedrock Model Evaluation | Evaluate FM performance | Model comparison |
| Amazon SageMaker AI | Build / train / deploy / operate ML | Traditional ML lifecycle |
| SageMaker JumpStart | Pre-trained models + templates | Accelerate selection/dev |
| Amazon Q | GenAI assistants for work / AWS tasks | Business / developer assistance |
| Strands Agents | Agent development framework | Agentic app development |
| Bedrock AgentCore | Deploy / operate agents | Current agentic-AI scope |
| Amazon Nova | AWS foundation-model family | FM selection |
| Kiro | AWS AI-powered development environment | In-scope developer tooling |
12
Responsible AI Domain 4 · 14%
| Feature | Meaning |
|---|---|
| Fairness | Avoid unjustified outcome disparities |
| Inclusivity | Design for diverse users / populations |
| Robustness | Reliable under varied conditions |
| Safety | Reduce harmful / unintended outcomes |
| Veracity | Truthfulness / factual reliability |
| Transparency | Purpose, behavior, limits understandable |
| Explainability | Understandable reasons for outputs |
| Interpretability | How understandable model behavior is |
- Diverse, curated, balanced datasets; check label quality
- Subgroup analysis, human audits, continuous monitoring
- Overfitting and underfitting both hurt generalization
Trap: Transparency ≠ explainability. Transparency = system purpose/limits. Explainability = why this output.
13
Responsible-AI tools & legal risk
| Tool | Recognition |
|---|---|
| SageMaker Clarify | Bias detection + explainability support |
| SageMaker Model Monitor | Watch deployed behavior / data drift |
| Amazon A2I | Human review / human-in-the-loop |
| SageMaker Model Cards | Document purpose, limits, evaluations |
| Bedrock Guardrails | GenAI input/output safeguards |
Legal / trust risks
- IP infringement claims
- Biased / discriminatory outputs
- Hallucinations and bad advice
- End-user harm and lost trust
- Licensing, provenance, data/model-use constraints
Human-centered explainability: meaningful transparency, feedback paths, and a way to challenge high-impact decisions.
14
Security fundamentals & AI threats Domain 5 · 14%
| Concept | Recognition |
|---|---|
| Authentication | Who are you? |
| Authorization | What may you do? |
| Least privilege | Only required permissions |
| Encryption at rest / in transit | Stored vs moving data |
| Data leakage prevention | Stop sensitive data escaping |
| Audit trail | Investigate / demonstrate later |
| Data lineage | Where data came from and how it changed |
| Threat | Recognition |
|---|---|
| Prompt injection | Malicious instructions manipulate behavior |
| Jailbreaking | Bypass safety restrictions |
| Data poisoning | Bad data influences training / pipelines |
| Adversarial input | Crafted input to force wrong behavior |
| Model theft | Unauthorized acquisition of model capability |
| Data leakage | Secrets in prompts, logs, retrieval, or outputs |
15
Security & governance services
| Service | Remember it as… |
|---|---|
| IAM | Identity, roles, policies, permissions |
| KMS | Keys and encryption control |
| Macie | Discover / protect sensitive S3 data |
| PrivateLink | Private connectivity to endpoints |
| CloudTrail | API activity / audit history |
| Config | Resource config + compliance assessment |
| Audit Manager | Collect evidence for audits |
| Artifact | AWS compliance reports / agreements |
| Inspector | Vulnerability assessment |
| Trusted Advisor | Cost, performance, security, limits advice |
| CloudWatch | Metrics, logs, observability |
| Bedrock Guardrails | GenAI safety controls |
Shared responsibility: AWS secures the cloud; you secure how you configure services and protect your data.
16
Governance checklist & grounding
- Approved use cases + owners
- Classify data; residency / retention
- Least-privilege access
- Encrypt at rest and in transit
- Document sources and lineage
- Log AI interactions + admin actions
- Validate before high-impact actions
- Monitor drift, bias, toxicity, harmful outputs
- Review cadence + escalation paths
- Train teams on approved AI practices
Hallucination decision chain
Trusted sources Retrieve Ground prompt Generate Validate Confidence / human review
| Technique | Purpose |
|---|---|
| RAG grounding | Supply retrieved facts to improve grounding |
| Output validation | Check structure, rules, sources, constraints |
| Confidence scoring | Flag outputs that need more review |
| Human review | Escalate high-impact / ambiguous decisions |
17
High-value distinctions
| Don't confuse | Distinction |
|---|---|
| Training vs inference | Training changes parameters; inference produces outputs |
| Classification vs regression | Category vs continuous number |
| Supervised vs unsupervised | Labeled targets vs structure without labels |
| RAG vs fine-tuning | External knowledge at inference vs trained behavior change |
| Prompting vs fine-tuning | Instructions/context vs learned weights |
| Grounding vs guardrails | Factual basis vs unsafe/unwanted behavior controls |
| Transparency vs explainability | System purpose/limits vs reasons for an output |
| Accuracy vs precision vs recall | Overall correct vs positive correctness vs positive coverage |
| CloudTrail vs CloudWatch | API history vs metrics/logs/ops |
| Config vs Audit Manager | Resource config compliance vs audit evidence collection |
| Artifact vs Audit Manager | AWS compliance docs vs your audit evidence |
| AuthN vs AuthZ | Identity vs allowed actions |
| Model vs business evaluation | Model behavior metrics vs value metrics |
18
Classic AI/ML service recognition
| Service | Fast answer |
|---|---|
| Transcribe | Speech → text |
| Translate | Language translation |
| Polly | Text → speech |
| Lex | Conversational interfaces (voice/text) |
| Comprehend | NLP: sentiment, entities, topics |
| Rekognition | Image / video analysis |
| Textract | Extract text + structure from documents |
| Personalize | Personalized recommendations |
| SageMaker AI | Build, train, customize, deploy, operate ML |
| Kendra | Enterprise / intelligent search |
| A2I | Human review workflows |
19
2026 guide scope watch
Guide v1.1 (Apr 30, 2026): expanded GenAI / agentic terms, async/serverless inference, traditional-ML vs FM selection, token pricing, context engineering, MCP / multi-agent, prompt management, LLM-as-a-judge, business metrics, hallucination/grounding, AgentCore security concepts.
- Emphasized services: Bedrock AgentCore, Kiro, Strands Agents, Amazon Q, JumpStart, AWS Transform, Aurora in vector scope
- Agentic vocab: MCP, multi-agent, memory, tool use, orchestration, AgentCore Identity / policy
- Eval: LLM-as-a-judge + task completion, satisfaction, cost per interaction
- Security: leakage prevention, output filtering, AI audit logging, toxicity, grounding / hallucination detection
✓
Last-minute recall
AI = umbrella ML = learns patterns DL = multilayer nets GenAI = creates content Agent = reasons + acts RAG = retrieve + ground Fine-tune = adapt behavior Temperature = variability IAM = permissions KMS = keys CloudTrail = API history Config = configuration Artifact = AWS compliance docs Audit Manager = audit evidence Macie = sensitive S3 data Guardrails = GenAI safeguards
- Domain 3 (28%) first: RAG, prompting, FM selection, evaluation
- Service jobs cold for Bedrock family + classic AI services
- Distinctions table until you cannot miss them
- Security: AuthN/AuthZ, least privilege, shared responsibility
- Responsible AI: fairness stack + Clarify / Guardrails / A2I