Domain Weights
| Domain | % |
| 1 · Fundamentals of AI and ML | 20 |
| 2 · Fundamentals of Generative AI | 24 |
| 3 · Applications of Foundation Models | 28 |
| 4 · Guidelines for Responsible AI | 14 |
| 5 · Security, Compliance, and Governance for AI Solutions | 14 |
Compensatory scoring — pass the whole exam, not each domain. No penalty for guessing. Question types: multiple choice, multiple response (all-or-nothing), ordering, matching.
AI vs ML vs DL vs GenAI
- AI = broadest field, machines performing tasks that normally need human intelligence.
- ML = subset of AI, systems learn patterns from data instead of explicit rules.
- Deep Learning = subset of ML, multi-layer neural networks, needs lots of data + compute.
- GenAI = subset of DL, models (often foundation models) that generate new content: text, image, code, audio.
- Foundation Model = huge model pretrained on broad data, adaptable to many downstream tasks.
Learning Types
| Type | What / example |
| Supervised | labeled data → classification (spam/not-spam) or regression (predict price) |
| Unsupervised | no labels → clustering (customer segmentation), dim reduction |
| Reinforcement | agent + reward signal → game playing, RLHF for FM alignment |
Classification = discrete label. Regression = continuous number. Clustering = group similar unlabeled points.
Training / Fit / Bias-Variance
- Training = model learns weights from data. Inference = trained model produces predictions on new input.
- Overfitting = memorizes training data, fails on new data. Fix: more data, regularization, simpler model, dropout, early stopping.
- Underfitting = too simple, misses patterns even in training data. Fix: more features, bigger/complex model, train longer.
- High bias = systematic error, underfits. High variance = overreacts to training data noise, overfits. Fix bias → more complex model. Fix variance → more data / regularization.
FM Vocabulary
- Tokens = chunks of text (word/sub-word) the model processes; billing + limits are per-token.
- Embeddings = numeric vector representation capturing semantic meaning, used for similarity search.
- Context window = max tokens (input + output) the model can consider at once.
- Temperature = randomness knob; low = deterministic/focused, high = creative/diverse.
- Top-k = sample from the k most likely next tokens. Top-p (nucleus) = sample from smallest set whose cumulative probability ≥ p.
- Hallucination = model states confident but false/unfounded output; mitigate with RAG, grounding, lower temperature, Guardrails.
Customization Ladder (cheap → expensive)
| Approach | When |
| Prompt engineering | fastest, no training, reshape instructions/examples |
| RAG | need current/proprietary facts, cite sources, reduce hallucination |
| Fine-tuning | need consistent style/format/domain behavior, have labeled examples |
| Continued pre-training | large unlabeled domain corpus, teach new domain knowledge broadly |
| Train from scratch | rare, huge cost, only when nothing existing fits (highly specialized/proprietary) |
Highest-yield table on the exam. Read scenario for "needs facts to stay current" (RAG) vs "needs a different tone/format every time" (fine-tune).
Prompt Engineering
- Zero-shot = ask directly, no examples. Few-shot = give a handful of example input/output pairs.
- Chain-of-thought = ask model to reason step by step before answering; improves multi-step/logic tasks.
- Prompt templates = reusable structured prompt with variable slots for consistency at scale.
- Negative prompting = tell the model what to avoid / not include in the output.
Amazon Bedrock
- Fully managed, serverless access to multiple foundation models from AWS + 3rd parties (Anthropic Claude, Meta Llama, Amazon Nova/Titan, etc.) via one API.
- Model access = request/enable specific FMs per account before use.
- Agents = orchestrate multi-step tasks, call APIs/Lambda, use a reasoning loop.
- Knowledge Bases = managed RAG: ingest docs, chunk, embed, store in a vector DB, retrieve at query time.
- Guardrails = policy layer to filter harmful content, PII, topics, and block prompt attacks.
- Model Evaluation = automatic or human-based comparison of FM outputs on your data/metrics.
- On-demand = pay per token, no commitment. Provisioned Throughput = reserved capacity, $/hr commitment, needed for high sustained volume or custom/fine-tuned models.
Amazon SageMaker AI
- Studio = IDE for the full ML lifecycle (build/train/deploy).
- JumpStart = hub of pretrained models + prebuilt solutions, one-click deploy/fine-tune.
- Canvas = no-code visual ML for business analysts.
- Data Wrangler = visual data prep/cleaning/feature transforms.
- Ground Truth = data labeling (human + automated) to build training sets.
- Clarify = detect bias in data/models and explain predictions (feature attribution).
- Model Monitor = detect data/model drift and quality decay in production.
- Feature Store = central repo of reusable ML features for training + inference consistency.
- Model Cards = document a model's intended use, training data, performance, and limitations.
Managed AI Services (by input type)
| Input | Service · does |
| Text | Comprehend — NLP: entities, sentiment, PII, key phrases |
| Documents | Textract — extract text/forms/tables from scans |
| Images/video | Rekognition — object/face/moderation detection |
| Speech → text | Transcribe |
| Text → speech | Polly |
| Language | Translate |
| Conversational bot | Lex — voice/text chatbots |
| Recommendations | Personalize |
| AWS-built FM family | Amazon Nova — text/image/video FMs in Bedrock |
| Agent runtime at scale | Bedrock AgentCore — deploy/operate agents |
| Build agents in code | Strands Agents (SDK) · Kiro (agentic IDE) |
| Modernize/migrate workloads | AWS Transform |
| BI & dashboards over data | Amazon Quick |
| Assistant / business Q&A / coding | Amazon Q (Business / Developer) |
Not on the AIF-C01 in-scope list: Kendra, Forecast, Fraud Detector, Augmented AI (A2I). They appear in older AIF study material; the current exam guide dropped them.
Which Service for Which Job
- "extract text from scanned invoice/form" → Textract
- "understand sentiment / entities in reviews" → Comprehend
- "search across internal enterprise documents" → Amazon Q Business
- "build my own RAG app with any FM" → Bedrock Knowledge Bases
- "recommend products to a shopper" → Personalize
- "deploy and operate an AI agent at scale" → Bedrock AgentCore
- "assistant answering questions about my company data" → Amazon Q Business
- "AI pair programmer in my IDE" → Amazon Q Developer
- "no-code ML for a business analyst" → SageMaker Canvas
- "detect bias / explain a model's prediction" → SageMaker Clarify
RAG Pipeline & Vector Stores
Pipeline: ingest docs → chunk → embed (embedding model) → store vectors → user query embedded → similarity search retrieves relevant chunks → chunks + query sent to FM as context → grounded answer.
| Vector store | Note |
| OpenSearch Serverless | common Bedrock KB default, managed vector search |
| Aurora (pgvector) | vectors alongside relational data |
| Neptune Analytics | graph + vector similarity |
| DocumentDB | vector search on document/JSON data |
| OpenSearch Service | the provisioned sibling; you manage the domain |
| ElastiCache | in-memory, low-latency vector search |
Evaluation Metrics
- Classical ML: Accuracy (overall correct), Precision (of predicted positives, how many right — false-positive cost), Recall (of actual positives, how many caught — false-negative cost), F1 (harmonic mean of P/R), AUC-ROC (ranking quality across thresholds).
- GenAI text: BLEU (n-gram overlap, translation), ROUGE (recall-oriented overlap, summarization), BERTScore (semantic similarity via embeddings), Perplexity (how "surprised" the model is — lower is more confident/fluent).
- Business metrics: user satisfaction, task completion rate, cost per inference, latency, adoption/conversion — tie model choice back to business value, not just benchmark scores.
Responsible AI
- Bias = systematic skew in data/model favoring/disfavoring a group. Fairness = mitigating that skew across outcomes.
- Explainability = why did the model make THIS decision (feature attribution, SHAP-style). Interpretability = how the model works overall (simpler models are inherently more interpretable).
- AI Service Cards = AWS-published transparency docs for its own AI services (intended use, limitations). Model Cards = you document your own model similarly.
- SageMaker Clarify = detects bias pre/post-training and explains predictions.
- Bedrock Guardrails filters: denied topics, content filters (hate/violence/insults/sexual), word filters, PII redaction, contextual grounding checks (reduce hallucination).
- Human review = keep a human in the loop on low-confidence or sensitive predictions; SageMaker Ground Truth is the in-scope labeling/review path.
Security & Governance
- IAM = least-privilege access to Bedrock/SageMaker APIs and models.
- KMS = encrypt training data, model artifacts, endpoints at rest.
- PrivateLink / VPC endpoints = keep Bedrock/SageMaker traffic off the public internet.
- CloudTrail = audit who invoked which model/API when.
- Macie = discover PII in S3 data used for training/RAG.
- Data residency = choose Region to keep data/processing within required geography.
- Prompt injection / jailbreak = malicious input tries to override system instructions or extract hidden data — mitigate with Guardrails, input validation, least-privilege tool access for Agents.
- Model poisoning = attacker corrupts training data to bias/backdoor the model — mitigate with data validation, provenance, access control on training pipelines.
Compliance Frameworks
- ISO 42001 = AI management system standard.
- NIST AI RMF = US risk management framework for trustworthy AI.
- EU AI Act = risk-tiered EU regulation on AI systems.
- GDPR = EU data protection/privacy, relevant when AI processes personal data.
- AWS Config = track resource configuration/compliance over time.
- Audit Manager = continuously collect evidence for audits.
- Artifact = on-demand access to AWS's own compliance reports/agreements.
Cost Levers
- Model choice = smaller/cheaper model when task allows; don't over-provision capability.
- Token count = shorter prompts/outputs = lower cost; trim unnecessary context.
- Provisioned vs on-demand = provisioned for steady high volume, on-demand for spiky/light use.
- Batch inference = cheaper for large non-real-time workloads vs real-time endpoints.
- Caching = reuse repeated prompt/response pairs to skip re-inference.
- Small vs large model = route simple tasks to a small/fast model, reserve large FM for complex reasoning.
Top Traps
RAG (fresh/proprietary facts, no retrain) vs fine-tuning (new style/behavior, needs labeled data) · Bedrock (managed FM API, no infra) vs SageMaker (full custom ML build/train/deploy control) · Amazon Q Business (org knowledge assistant) vs Q Developer (code assistant) · Amazon Q Business (managed assistant over your data, minimal setup) vs Bedrock Knowledge Bases (you choose FM, chunking and vector store) · Comprehend (analyze existing text) vs Textract (extract text from documents/images) · Personalize (recommendations) vs Bedrock AgentCore (running agents, not predicting) · explainability (why this output) vs interpretability (how the model works overall)
Read the last sentence first: "least operational overhead," "most cost-effective," "without retraining" — that qualifier picks the winner between two right-looking answers.