How the exam works
This is a foundational exam. You need to recognize which AWS service or concept fits a business problem. You will not write code, tune models, or do math.
Question types
| Type | What you do |
|---|---|
| Multiple choice | Pick 1 correct answer out of 4. |
| Multiple response | Pick 2 or more correct answers out of 5 or more. All must be right to get credit. |
| Ordering | Put 3 to 5 steps in the correct order. |
| Matching | Match 3 to 7 prompts to the right answers. All pairs must be right. |
Scoring is compensatory: you only need to pass the exam overall, not each domain.
Where the points are
Exam guide versions
| Version | Published | What changed |
|---|---|---|
| 1.0 | March 26, 2026 | First version in AWS's change history. |
| 1.1 | April 30, 2026 | Big content update: agentic AI and new services (details below). |
| 1.2 (current) | October 29, 2026 | AWS's change log lists one change: six services removed from the out-of-scope list (Elastic Transcoder, Elemental MediaStore, IoT 1-Click, IoT Analytics, IoT Events, Nimble Studio). The five domains and their weightings are unchanged. |
In scope and out of scope services (v1.2)
AWS publishes a list of services that can appear as the right answer, and a list that should not. Both lists are "non-exhaustive", so features of an in-scope service (like SageMaker Clarify or Bedrock Guardrails) count too. In the Services A–Z tab, each card shows a green In scope or red Out of scope tag. A grey Not listed tag means AWS puts the service on neither list (for example Amazon Kendra, AWS Audit Manager, Amazon A2I and AWS HealthScribe). Such services can still show up in questions, but they are less likely to be the right answer than in-scope ones. If two options look equally good, lean toward the in-scope service.
| Category | In scope |
|---|---|
| Machine learning | Amazon Bedrock, Bedrock AgentCore, Comprehend, Lex, Nova, Personalize, Polly, Rekognition, SageMaker AI, SageMaker JumpStart, Textract, Transcribe, Translate, AWS Transform |
| Developer tools | Kiro, Strands Agents |
| Analytics | AWS Data Exchange, Amazon EMR, AWS Glue, AWS Glue DataBrew, AWS Lake Formation, Amazon OpenSearch Service, Amazon Quick, Amazon Redshift |
| Database | Amazon Aurora, Amazon DocumentDB, Amazon DynamoDB, Amazon ElastiCache, Amazon Neptune, Amazon RDS |
| Compute & containers | Amazon EC2, AWS Lambda, Amazon ECS, Amazon EKS |
| Security, identity & compliance | AWS Artifact, IAM, Amazon Inspector, AWS KMS, Amazon Macie, AWS Secrets Manager |
| Management & governance | AWS CloudTrail, Amazon CloudWatch, AWS Config, AWS Trusted Advisor, AWS Well-Architected Tool |
| Networking | Amazon CloudFront, Amazon VPC |
| Storage | Amazon S3, Amazon S3 Glacier |
| Cost management | AWS Budgets, AWS Cost Explorer |
Out of scope, but you may see them as wrong answers
Amazon GuardDuty, AWS Security Hub, AWS Shield, AWS WAF, AWS Certificate Manager, Amazon Cognito, Amazon Detective, Amazon SES, AWS DMS, AWS Backup, AWS X-Ray, AWS Control Tower, Amazon Route 53, AWS HealthImaging and AWS HealthOmics, among others.
What changed in v1.1
AWS updated the exam guide in April 2026. If your study materials are older, they probably miss these newer topics, all covered in this guide:
- Agentic AI: AI agents, multi-agent systems, Model Context Protocol (MCP), memory, tool use.
- New AWS services: Amazon Quick, Kiro, Strands Agents, Amazon Bedrock AgentCore (including AgentCore Identity and Policy).
- Context engineering, prompt caching, model distillation, Bedrock Prompt Management.
- Evaluating AI apps (RAG, agents, workflows) and LLM-as-a-judge.
- Security for AI: prompt injection, data leakage, output filtering, hallucination detection, audit logging.
Out of scope (don't over-study)
The guide says you are not expected to: code models, do feature engineering, tune hyperparameters, build pipelines or infrastructure, do math or statistics, or implement security and governance frameworks yourself. Know what things are and when to use them.
Fundamentals of AI and ML
1.1How AI, ML, deep learning, GenAI and agentic AI fit together
Key terms in one line each
| Term | Plain meaning |
|---|---|
| Model | The trained "brain" that makes predictions. |
| Algorithm | The recipe used to train the model. |
| Training | Teaching the model with data. |
| Inference | Using the trained model to get answers. |
| Neural network | Layers of connected nodes, loosely inspired by the brain. |
| Computer vision | AI that understands images and video. |
| NLP | Natural language processing: AI that understands human language. |
| LLM | Large language model: a huge model trained on text that predicts the next word. |
| Bias | Systematic unfairness or error, often from skewed data. |
| Fairness | Results that treat different groups equitably. |
| Fit | How well the model learned: underfit, good fit, or overfit. |
| Feature | One input variable the model learns from, like a column in a spreadsheet (age, income, city). |
| Feature engineering | Creating new or better features from existing data, such as "days since last purchase" from a purchase date. |
| Weights | The strength of each connection between nodes in a neural network. Learning = adjusting weights. |
How a neural network learns
The network makes a guess, compares it with the correct label, and measures the error. It then nudges its weights (connection strengths) to shrink that error. This step is called backpropagation. Repeating it across thousands of examples is how it learns patterns. It does not memorize examples in a database or follow rules a developer wrote.
1.1Types of inference
| Type | Analogy | Use when… |
|---|---|---|
| Real-time | Asking a cashier and getting an instant answer | You need a low-latency answer right now (fraud check at checkout). Always-on endpoint. |
| Serverless | A taxi: only there when you call it | Traffic is spiky or idle a lot. No servers to manage; may have a short "cold start" delay. |
| Asynchronous | Dropping off dry cleaning and getting a text when it's ready | Large inputs or long processing. Requests wait in a queue. |
| Batch | A photo lab developing a whole roll of film | Scoring a big dataset at once, offline (nightly reports). No endpoint needed. |
Fastest to slowest response: real-time → serverless (can add a short cold start) → asynchronous → batch.
Pick the inference type
1.1Types of data
Labeled vs unlabeled
Labeled data has the answer attached ("this email is spam"). Unlabeled data has no answers.
Structured vs unstructured
Structured fits in rows and columns (spreadsheets, databases). Unstructured doesn't (emails, images, audio).
Tabular
Rows and columns, like a spreadsheet of customers.
Time-series
Values recorded over time, like daily sales or sensor readings.
Image
Photos, scans, video frames. Used for computer vision.
Text
Documents, reviews, chat logs. Used for NLP and LLMs.
1.1Types of learning
1.2Picking the right technique
| Technique | Question it answers | Example |
|---|---|---|
| Classification | Which category? | Spam or not spam; approve or deny a loan |
| Regression | How much / how many? | House price; next month's sales |
| Clustering | Which items are similar? | Customer segments for marketing |
| Anomaly detection | Which items are unusual? | Rare claims that look very different from normal ones |
Binary classification
Exactly two possible answers: fraud or not fraud, spam or not spam.
Multiclass classification
One of three or more answers: route a call to billing, tech support or sales; approve, investigate or deny a claim.
Where AI/ML adds value
Three reasons to use AI: it can assist human decision-making (flag risky loans for a person to review), give scalability (handle millions of requests no team could), and enable automation (process forms with no manual typing). Common real-world uses: computer vision, NLP, speech recognition, recommendation systems, fraud detection, forecasting (predicting future values such as next month's demand), knowledge bases, and agentic AI.
When AI/ML is not the right choice
- You need an exact, guaranteed outcome (a tax calculation). Use normal rules-based code. ML gives predictions, not certainty.
- The cost is more than the benefit (a cost-benefit analysis says no).
- You don't have enough good data.
- A simple, fixed rule already solves it ("every invoice over $10,000 goes to a manager").
Traditional ML or a foundation model?
Choose a traditional ML model when regulatory concerns require strong explainability (regulated industries like lending), operational constraints are tight, a narrow, well-defined prediction, or low cost and latency. Choose a foundation model for open-ended language, content generation, or many tasks at once.
1.2AWS managed AI services (no ML expertise needed)
| Service | What it does | Hook |
|---|---|---|
| Amazon Comprehend | Finds meaning in text: sentiment, key phrases, entities, PII (personally identifiable information, such as names, addresses and account numbers) | Comprehend = understand text |
| Amazon Transcribe | Speech to text. For specialist words (medical, legal, product names) use a custom vocabulary or a custom language model to raise accuracy. Language identification only detects which language is spoken. | A transcript of what was said |
| Amazon Polly | Text to speech (lifelike voices) | Polly the parrot talks |
| Amazon Translate | Translates text between languages | Says it in the name |
| Amazon Lex | Builds chatbots and voice bots (the tech behind Alexa) | Lex = lexicon, conversation |
| Amazon Rekognition | Analyzes images and video: objects, faces, unsafe content | Recognize what's in the picture |
| Amazon Textract | Extracts text, forms and tables from scanned documents | Extract text |
| Amazon Kendra | Intelligent enterprise search across documents | Search your company's knowledge |
| Amazon Personalize | Real-time product and content recommendations | "Customers like you also bought…" |
| Amazon SageMaker AI | Build, train and deploy your own ML models | The full ML workshop |
| Rekognition Custom Labels | Teach Rekognition your own objects or defects from a small labeled set | Your own labels |
| AWS HealthScribe | Turns patient–clinician conversations into transcripts and draft clinical notes. HIPAA-eligible. | A medical scribe |
| Transcribe Medical | Speech to text for medical terms (transcript only) | Transcribe for doctors |
| Comprehend Medical | Pulls conditions, medications and dosages out of clinical text | Comprehend for doctors |
| Amazon Fraud Detector | Managed fraud scoring. No longer open to new customers (Nov 2025), but it can still appear in questions. | Says it in the name |
The first five services and SageMaker AI are named in the exam guide. The others are long-standing AWS AI services that commonly appear in scenarios. Search any service name at the top of the page for more detail.
1.3The ML development lifecycle
| Stage | Helpful AWS service or feature |
|---|---|
| Label data | SageMaker Ground Truth (humans label data) |
| Prepare data | SageMaker Data Wrangler; SageMaker Feature Store (stores reusable features) |
| No-code ML | SageMaker Canvas (point-and-click models for business users) |
| Use a pre-trained model | SageMaker JumpStart; Amazon Bedrock |
| Check bias / explain | SageMaker Clarify |
| Work environment | SageMaker Studio (one IDE, or integrated development environment, for all ML work) |
| Train with your own script | Prebuilt framework containers (PyTorch, TensorFlow) in "script mode": least overhead |
| Train huge models | SageMaker HyperPod (large distributed clusters) |
| Track & compare training runs | SageMaker Experiments (managed MLflow): parameters, datasets, metrics |
| Trace data → model → endpoint | SageMaker ML Lineage Tracking |
| Automate & track | SageMaker Pipelines; Model Registry |
| Watch in production | SageMaker Model Monitor |
| AI for business users | Amazon Quick (chat agents, research, BI and automation) |
| AI for developers | Kiro (agentic coding IDE) |
| Get outside data | AWS Data Exchange (find and subscribe to third-party datasets) |
| Prepare data at scale | AWS Glue (serverless ETL and Data Catalog); AWS Glue DataBrew (no-code data cleaning); Amazon EMR (big-data processing with Spark) |
Need more variables (features) to improve a model? They come from two steps: data collection (bring in new data sources) and feature engineering (create new features from the data you have). Hyperparameter tuning changes model settings, not the data.
Ways to get and run a model
- Sources: use an open-source pre-trained model, or train your own custom model (expensive).
- Run it: as a managed API service (API: application programming interface, the way one program asks another to do something) (Bedrock: AWS runs everything) or self-hosted (you deploy it, for example on a SageMaker endpoint, and manage it).
MLOps in plain terms
MLOps (machine learning operations) is DevOps (development plus IT operations) for ML: making ML repeatable, scalable and reliable. Key ideas from the exam guide: experimentation (track and compare runs), repeatable processes (automated pipelines), scalable systems, managing technical debt (shortcuts that cost more later), production readiness (tested, secure, monitored before launch), model monitoring, and model re-training when they drift.
1.3SageMaker Experiments vs Lineage vs Registry
These three sound alike. Each one answers a different question.
| Feature | Question it answers | Pick it when the question says… |
|---|---|---|
| SageMaker Experiments | "Which training run did best, and with what settings?" | Track, organize or compare training runs, hyperparameters, datasets and metrics; reproducibility |
| ML Lineage Tracking | "Where did this model's data come from, and what touched it?" | Trace the full journey from data source through transformations to the model and endpoint, for auditors |
| Model Registry | "Which model version is approved for production?" | Version, approve and promote models |
In the current SageMaker Studio, experiment tracking is done with fully managed MLflow. Exam questions may still say "SageMaker Experiments".
1.3Training, validation and test data
- Training set: what the model learns from.
- Validation set: used again and again while you tune settings and pick the best model.
- Test set: held back until the end for one final, unbiased score before production.
1.3Model performance metrics
Move the decision threshold
Each dot is a transaction the model scored from 0 (surely fine) to 1 (surely fraud). Anything at or right of the line is flagged as fraud.
| Metric | Simple meaning | Use when… |
|---|---|---|
| Accuracy | % of all predictions that were right | Classes are balanced. Misleading if not: always saying "not fraud" can be 99% accurate. |
| Precision | Of the ones flagged, how many were real? | False alarms are costly (blocking good customers). |
| Recall | Of the real ones, how many did we catch? | Misses are costly (disease, fraud). |
| F1 score | Balance of precision and recall (their harmonic mean). "F1" is a name, not an acronym. | You need both, or data is imbalanced. |
| AUC | Area under the ROC (receiver operating characteristic) curve. How well the model separates the classes (0.5 = guessing, 1 = perfect) | Comparing classification models overall. |
| RMSE / MAE | Root mean squared error / mean absolute error: the average size of the error | Regression (predicting numbers). Lower is better. |
Business metrics
A technically great model still has to pay off. Know: cost per user, development costs, customer feedback, and ROI (return on investment).
Model fit: underfitting vs overfitting
| Problem | Training score | Test score | Also called | Fix |
|---|---|---|---|---|
| Underfitting | Poor | Poor | High bias | More features, train longer, more complex model, less regularization |
| Overfitting | Great | Poor | High variance | More data, regularization, simpler model, fewer features (PCA, principal component analysis), early stopping |
Reading the numbers
| Training error | Validation error | Diagnosis |
|---|---|---|
| 2% | 35% | Overfitting: big gap |
| 30% | 31% | Underfitting: high error on both |
| 6% | 8% | Good fit: low on both, small gap |
Turn up model complexity
Fundamentals of generative AI
2.1Core GenAI vocabulary
| Term | Plain meaning |
|---|---|
| Token | A chunk of text the model reads, roughly part of a word. You pay per token. |
| Chunking | Splitting long documents into smaller pieces so they can be searched and fed to a model. |
| Embedding | Turning text (or images) into a list of numbers that captures its meaning. |
| Vector | That list of numbers. Similar meanings produce nearby vectors. |
| Foundation model (FM) | A huge model pre-trained on broad data that can do many tasks. |
| Transformer LLM | The architecture behind modern LLMs. It uses self-attention to weigh how every token in a sequence relates to every other token. |
| Context window | The maximum number of tokens a model can consider in one request, including input and output. Its "short-term memory". |
| Multi-modal model | Handles more than one type of data, like text and images together. |
| Diffusion model | Generates images by starting with random noise and removing it step by step. |
| Prompt engineering | Writing inputs that get better outputs. |
| Context engineering | Choosing everything the model sees: instructions, retrieved documents, memory and tool results, to fit the context window. |
Chunk a document
2.1Watch a diffusion model work
Diffusion starts from random static and removes a little noise at each step, guided by the prompt. New static gives a new picture for the same prompt.
Prompt: "a red sun over green hills"
Step 0 is pure noise. Each step is a small, easy cleanup. Try "New random noise": the sun and hills land somewhere different, which is why the same prompt gives a new image each time.
2.1GenAI use cases
Image, video and audio generation; summarization; AI assistants; translation; code generation; customer service agents; search; recommendation engines.
2.1Amazon's own foundation models: Nova
Practice questions often ask you to pick the right Amazon Nova model. All are used through Amazon Bedrock.
| Model | Input → output | Pick it for… |
|---|---|---|
| Nova Micro | Text → text | Cheapest and fastest. High-volume text tasks: classification, routing, extraction. |
| Nova Lite | Text, image, video → text | Low-cost multimodal tasks and fast agents. |
| Nova Pro | Text, image, video → text | Complex, multi-step reasoning over long or mixed documents. |
| Nova Premier | Text, image, video → text | The hardest tasks. Also a "teacher" for distillation. |
| Nova Canvas | Text, image → image | Marketing images, editing (inpainting, outpainting), brand style and colors. |
| Nova Reel | Text, image → video | Short videos with camera motion control and built-in safety controls. |
| Nova Sonic | Speech → speech | Real-time voice conversations. |
Other image models in Bedrock: Stability AI's Stable Diffusion. Amazon Titan Text Embeddings is a common embeddings model for RAG. A newer Nova 2 generation also exists. Don't confuse Nova Canvas with SageMaker Canvas (no-code ML).
2.1The foundation model lifecycle
2.1Token-based pricing
- You pay for input tokens (your prompt) plus output tokens (the answer). Output usually costs more per token.
- Longer prompts and longer answers cost more and take longer (higher latency).
- Ways to save: shorter prompts, limit max output tokens, prompt caching for repeated prompt text, a smaller model, or batch inference for non-urgent jobs.
See what drives the bill
Example prices only, to show proportions: $3 per million input tokens, $15 per million output tokens; cached tokens at 10% of the input price; batch at 50% off; smaller model at one fifth of the price. Real prices vary by model and change over time.
2.1Agentic AI concepts New in v1.1
| Concept | Plain meaning |
|---|---|
| AI agent | An LLM that can plan steps and take actions with tools to reach a goal. |
| Tool usage | The agent calls APIs, searches, runs code or queries databases. |
| Model Context Protocol (MCP) | An open standard for connecting agents to external tools and data in a consistent way. |
| Memory management | Short-term memory (this conversation) and long-term memory (facts and preferences kept across sessions). |
| Multi-agent systems | Several specialized agents working together, often with a supervisor agent handing out sub-tasks. |
| Workflow orchestration | Coordinating the order of steps, agents and tools. |
2.1MCP's three building blocks
Tools
Functions the agent can call to do something: create a ticket, query a database.
Resources
Data or context the agent can read: files, records, documents.
Prompts
Reusable, pre-approved prompt templates stored on the server, so interactions stay standard and can be updated in one place.
A typical MCP tool call, in order
- The model reasons about the goal and decides it needs a tool.
- The agent calls the tool the MCP server exposes, passing the details.
- The MCP server runs the tool against the real system and returns the result.
- The model observes the result and answers the user.
Step through an agent using an MCP tool
2.1Perceive, reason, act
AWS describes every agent as a loop of three modules.
2.1Kinds of agent designs
| Design | How it works | Use when… |
|---|---|---|
| Workflow agent | Follows a fixed, pre-defined sequence of steps | The steps are known and the same every time |
| Autonomous agent | Plans its own steps (self-orchestration), uses tools and memory, adapts to results | The steps can't be known in advance and vary per request |
| Multi-agent (supervisor) | A supervisor splits work among specialist agents and combines results | Large tasks needing several specialties; more components to manage |
Enterprise adoption phases
Experiment / proof of concept → production → scale across the company. The hardest jump is into production, where agents meet real enterprise systems. That brings integration complexity and the need to observe and govern what autonomous agents do.
2.2Strengths and weaknesses of GenAI
- Adaptability: one model handles many tasks
- Responsive and fast
- Natural conversation
- Generates new content
- Hallucinations: confident but false answers
- Interpretability: hard to explain why
- Inaccuracy
- Nondeterminism: same prompt, different answers
Choosing a GenAI model, consider:
Model type, performance needs, capabilities, constraints, compliance, cost, latency, and model complexity.
Business value metrics
| Metric | Question it answers |
|---|---|
| Conversion rate | Did more visitors buy (or sign up) because of the AI feature? |
| ARPU | Average revenue per user: revenue ÷ active users. Compare across products. |
| CLV | Customer lifetime value: total revenue from a customer over the whole relationship. |
| Efficiency / cost per interaction | Did the AI lower the cost of handling each request? |
| ROI | Did the benefit outweigh what we spent? |
| Cross-domain performance | Does the model work well across different tasks or business areas? |
2.3AWS services for building GenAI
| Service | What it is | Pick it when… |
|---|---|---|
| Amazon Bedrock | Fully managed, serverless API access to many foundation models from Amazon and other providers. Includes Knowledge Bases (RAG), Agents, Guardrails, Model Evaluation and Prompt Management. | You want to use FMs without managing servers. |
| SageMaker AI | Build, train, fine-tune and host your own models with full control. | You need deep customization or your own ML. |
| SageMaker JumpStart | A hub of pre-trained models and solutions you can deploy in a few clicks. | You want a pre-trained model to deploy or fine-tune in SageMaker. |
| Amazon Quick New | An agentic AI workspace for business users: chat agents, research, BI dashboards (Quick Sight) and automation (Quick Flows / Automate). | Business users want answers and automation without coding. |
| Kiro New | AWS's agentic IDE. Turns specs into code, docs and tests (spec-driven development). | Developers want AI-assisted coding. |
| Strands Agents New | An open-source SDK (software development kit, a ready-made code library) for building AI agents in a few lines of code. The model does the planning. | Developers want to build an agent. |
| Bedrock AgentCore New | Services to deploy and run agents securely at scale: Runtime, Memory, Gateway (tools/MCP), Identity, Policy, Observability, plus built-in tools. | You need to run agents in production, securely. |
| AWS Transform | Agentic AI that migrates and modernizes legacy workloads: mainframe, VMware, Windows/.NET, and code upgrades. | The company wants AI agents to speed up migration or modernization of old applications. |
AgentCore's parts
| Part | What it does |
|---|---|
| Runtime | Serverless, isolated hosting for agents |
| Memory | Short-term (within a session) and long-term (across sessions) memory; shareable across agents; customer-managed KMS keys supported |
| Gateway | Turns APIs and Lambda functions into MCP tools agents can call |
| Identity | Agent and user identities and credentials (such as OAuth tokens) |
| Policy | Rules (Cedar) for which tools and actions an agent may use |
| Observability | Traces and metrics of what agents do |
2.3Kiro in more detail
Kiro is AWS's agentic IDE. Practice questions test two features: specs and steering files.
EARS (Easy Approach to Requirements Syntax) is a simple template for writing testable requirements, for example: "WHEN a customer requests a refund, THE system SHALL create a refund ticket."
Steering files
Persistent project context Kiro always follows: product.md (product purpose), structure.md (folder layout), tech.md (approved frameworks and tools).
Hooks
Automations that run on events, like updating tests or docs when a file is saved.
2.3Other Bedrock and assistant features you may see
| Feature | What it does |
|---|---|
| Bedrock playgrounds | Try and compare models in the console with no code (text, chat, image). |
| Bedrock model access | Grants your account permission to use specific models. It doesn't test them. |
| Amazon Q Business | Managed enterprise assistant over company data. Connectors index data and keep existing permissions; plugins (ServiceNow, Jira) take actions. Closed to new customers; AWS points to Amazon Quick. |
| Bedrock data privacy | Bedrock does not share your prompts or outputs with model providers and does not use them to train models. Nothing to switch on. |
2.3Why AWS, and GenAI cost tradeoffs
Why build GenAI on AWS
Accessibility, lower barrier to entry, efficiency, cost-effectiveness, speed to market, and meeting business goals. AWS infrastructure adds security, compliance, shared responsibility and safety.
Cost tradeoffs
| Option | How you pay | Best for |
|---|---|---|
| On-demand | Pay per token, no commitment | Variable or unpredictable usage |
| Provisioned throughput | Reserve model capacity for a time period | Steady, high-volume, guaranteed performance; often needed for custom models |
| Batch | Lower price, results later | Large, non-urgent jobs |
| Custom models | Training plus hosting costs | Only when prompting and RAG aren't enough |
| Prompt caching | Repeated prompt prefix is cached and billed at a lower rate | Every request repeats the same long context (manual, policy) |
Tracking spend: AWS Cost Explorer shows where the money went (for example, Bedrock spend by month). AWS Budgets alerts you before spend passes a limit you set.
Also weigh responsiveness, availability, redundancy, performance and regional coverage (not every model is in every Region).
Applications of foundation models
3.1Choosing a foundation model
Selection criteria: cost, modality (text, image, audio), latency, multi-lingual support, model size, complexity, customization options, input/output length (context window), and prompt caching.
3.1Inference parameters (try it)
At each step the model ranks possible next words. AWS's own example: after "I hear the hoof beats of…" the candidates are horses, zebras and unicorns. Move the sliders to see what each setting does.
Crossed-out words are removed from the pool. Bars show the chance each remaining word is picked.
| Setting | What it does | Hook |
|---|---|---|
| Temperature | Creativity dial. Low = safe, predictable. High = varied, creative. | Hot = wild, cold = calm |
| Top K | Keep only the top K candidate words. | K = Kount |
| Top P | Keep the top words until their probabilities add up to P. | P = Percent |
| Max length / max tokens | Caps how long the answer can be. Saves cost. | A word limit |
| Stop sequences | Text that tells the model to stop generating. | A stop sign |
3.1Retrieval Augmented Generation (RAG)
- Why use RAG: answers use your latest, private data; fewer hallucinations; you can cite sources; cheaper than fine-tuning.
- AWS service: Amazon Bedrock Knowledge Bases is the managed way to build RAG.
- Business uses: customer support bots, internal policy Q&A, document search assistants.
Vector databases on AWS (store embeddings)
Amazon OpenSearch Service
Search and analytics with vector search. A common default for Knowledge Bases.
Amazon Aurora (PostgreSQL)
Relational database with vector support through the pgvector extension.
Amazon RDS for PostgreSQL
Also supports pgvector for vector search.
Amazon DocumentDB
JSON document database with built-in vector search. Fits when the data already lives in DocumentDB.
Amazon ElastiCache
In-memory cache for very fast lookups. Often used to cache frequent model answers to cut cost and latency.
Amazon Neptune
Graph database. Neptune Analytics stores vectors and powers GraphRAG when relationships between facts matter.
RAG steps, in order
- Split documents into chunks.
- Turn chunks into embeddings with an embeddings model.
- Store the vectors in a vector store.
- At question time, retrieve the most similar chunks.
- Send the question plus chunks to the FM, which answers with citations.
3.1Customization approaches and cost
The trade-offs of customizing your own FM (instead of using a ready, pre-trained one) are higher cost and more implementation complexity: preparing and labeling data, running training jobs, evaluating, and hosting the custom model.
Which approach fits?
Think of your answer, then tap a card to check it.
3.1AI agents in business
Agents complete multi-step tasks on their own: handling a customer return end-to-end, processing invoices, triaging IT tickets, or researching and drafting reports. On AWS: Amazon Bedrock Agents, Strands Agents, and Bedrock AgentCore.
3.2Prompt engineering
Parts of a good prompt
Instruction
What to do. "Summarize this email in 3 bullets."
Context
Background info. "You are a support agent for a bank."
Input data
The content to work on: the email, the document, the question.
Output indicator
The format of the answer: "Return a table with 3 columns" or "bullet points only".
Negative prompt
What not to do. "Don't give diagnoses or medication advice."
Examples
Optional samples of good answers (one-shot or few-shot).
Prompting image and video models
For Amazon Nova Reel, put camera directions at the start or end of the prompt and use recognized terms such as "dolly in" or "pan left". Describe the subject and scene separately.
Techniques
| Technique | What it means | Hook |
|---|---|---|
| Zero-shot | No examples, just the task | 0 examples |
| Single-shot (one-shot) | One example of the desired output | 1 example |
| Few-shot | A few examples to show the pattern | 2+ examples |
| Chain-of-thought | Ask the model to reason step by step | "Show your work" |
| Prompt templates | Reusable prompts with fill-in blanks | A form letter |
Best practices
Best practices: specificity and concision (clear, short instructions), experimentation (try and compare versions), guardrails, and using multiple comments or steps to break complex tasks apart. Benefits: better response quality and discovery of what the model can do.
Prompt risks
| Risk | Plain meaning |
|---|---|
| Exposure / prompt leaking | The model reveals sensitive data or its confidential system prompt. |
| Poisoning | Bad data is slipped into training or retrieval data to corrupt outputs. Biased or toxic output that comes from the training data points here; the fix is retraining on curated, clean data. |
| Hijacking / prompt injection | Hidden instructions take over the model ("ignore your rules and…"). |
| Jailbreaking | Tricking the model into bypassing its safety rules. |
Amazon Bedrock Prompt Management lets you save, version, test and reuse prompts, like version control for prompts.
3.3Training and fine-tuning
| Method | Data used | Plain meaning |
|---|---|---|
| Pre-training | Massive unlabeled data | Builds the base model's general knowledge. |
| Continued pre-training | Unlabeled domain data | Teaches a domain (medical, legal) to an existing model. |
| Instruction tuning | Labeled prompt → response pairs | Teaches the model to follow instructions a certain way. |
| Domain adaptation | Domain-specific examples | Makes the model better at one industry's tasks. |
| Transfer learning | Small task data | Reuse what a model already learned for a new, related task. |
| Distillation | Teacher model outputs | A small model learns to copy a big model. |
| RLHF | Human ratings of answers | Reinforcement learning from human feedback aligns answers with what people prefer. |
Three kinds of data in an FM's life
| Data | When it's used | Example |
|---|---|---|
| Pre-training data | Building the base model | Huge public text and code collections |
| Fine-tuning data | Adapting a pre-trained model; changes its weights | De-identified clinical records to adapt a model for medicine |
| Inference data | Sent at run time; doesn't change the model | A doctor's note submitted for a summary today |
Preparing fine-tuning data
Use data curation to pick high-quality data, apply governance, use the right size, label accurately, and make it representative of real users. Check it for bias and rebalance it before training.
Fine-tuning in Bedrock, in order
- Curate and label prompt–completion examples; check quality, bias and representativeness.
- Run a model customization (fine-tuning) job with the data in S3.
- Evaluate the custom model against the base model on held-out data.
- Buy Provisioned Throughput and use the custom model in production.
3.4Evaluating foundation models
Human evaluation
People rate answers. Best for nuance like tone and helpfulness. Slow and costly.
Benchmark datasets
Standard test sets that compare models on the same questions.
Bedrock Model Evaluation
Managed evaluation using automatic metrics, human reviewers, or LLM-as-a-judge.
| Metric | Simple meaning | Best for | Hook |
|---|---|---|---|
| ROUGE | Did it cover the reference's key words? | Summarization | R = Recall = Reduce |
| BLEU | Are its words precise vs the reference? | Translation | B = Bilingual |
| BERTScore | Does it mean the same thing, even with different words? | Meaning similarity | BERT = Brains |
| LLM-as-a-judge | Another LLM grades the answer | Scaling quality reviews, open-ended answers with no reference | An AI teacher |
The names: ROUGE = Recall-Oriented Understudy for Gisting Evaluation. BLEU = Bilingual Evaluation Understudy. BERTScore uses BERT (Bidirectional Encoder Representations from Transformers), a language model that understands meaning. A bit more detail: BLEU scores n-gram (word sequence) precision against reference translations and adds a brevity penalty so short answers can't cheat. ROUGE scores how much of the reference's content was recalled.
Evaluating whole AI applications
- RAG: Did it retrieve the right chunks? Is the answer faithful to them (grounded)?
- Agents and workflows: Did it complete the task? Did it use the right tools in the right order?
Does it meet business goals?
Look at productivity, user engagement, task completion rate, user satisfaction, and cost per interaction.
Guidelines for responsible AI
4.1Features of responsible AI
Fairness & bias
Works equally well for all groups of people.
Inclusivity
Designed for diverse users and needs.
Robustness
Reliable even with unexpected or tricky inputs.
Safety
Avoids harmful outputs.
Veracity
Truthful and accurate.
Transparency
People know how and when AI is used.
Match the control to the dimension
| Control | Dimension |
|---|---|
| Compare approval rates across age and gender groups | Fairness |
| Show customers the top factors behind their decision | Explainability |
| Mask account numbers before prompts are logged | Privacy and security |
| Check answers against sources to catch made-up facts | Veracity and robustness |
| Block responses that encourage self-harm | Safety |
4.1Responsible AI across the lifecycle
- Responsible AI applies to every phase (design, development, deployment, monitoring, evaluation) for both traditional ML and GenAI.
- GenAI adds its own risks, such as hallucinations and veracity problems, so it needs extra controls: grounding (RAG), output validation and guardrails.
- Define each use case narrowly. A catalog for all shoppers should favor neutral, complete information. A campaign aimed at one group focuses on what matters most to that group. Different goals mean different tuning and different risks.
4.1Amazon Bedrock Guardrails
Which Guardrails feature handles it?
- Content filters: block hate, insults, sexual content, violence, misconduct, and prompt attacks. Each category has its own strength setting, and you can create separate guardrails with stricter settings for different teams.
- Denied topics: block subjects you choose (for example, "no investment advice").
- Word filters: block specific words or phrases.
- Sensitive information filters: detect and block or mask PII.
- Contextual grounding checks: catch hallucinations that aren't supported by the source.
- Automated Reasoning checks: verify answers against your rules using logic.
4.1Responsible model selection and legal risks
Responsible selection: consider environmental impact and sustainability. A smaller, efficient model uses less energy.
| Legal risk | Example |
|---|---|
| Intellectual property | Generated content copies copyrighted work. |
| Biased outputs | A hiring model favors one group. |
| Loss of customer trust | A chatbot gives offensive or wrong answers publicly. |
| End-user risk | Users act on harmful advice. |
| Hallucinations | A bot invents a refund policy the company must honor. |
4.1Good datasets, bias and variance
Good datasets show inclusivity and diversity : they are inclusive, diverse, curated from trusted sources, and balanced across groups.
Effects of bias and variance
- Effects on demographic groups: a model trained mostly on one group performs worse on others.
- Inaccuracy: wrong predictions.
- Underfitting (high bias) and overfitting (high variance). See Domain 1 for the diagram.
| Fix | Simple idea | Helps with |
|---|---|---|
| Cross-validation | Test on several different splits of the data | Detects overfitting |
| More data | More varied examples, harder to memorize | Overfitting |
| Regularization | Penalize overly complex models | Overfitting |
| Simpler model | Fewer parameters | Overfitting |
| Dimension reduction (PCA: principal component analysis) | Combine many features into fewer | Overfitting |
| Early stopping | Stop training before it memorizes | Overfitting |
Tools to detect and monitor bias
| Tool or method | What it does |
|---|---|
| SageMaker Clarify | Detects bias in data and models, and explains which features drove a prediction (SHAP). No longer open to new customers, but it can still appear on the exam. |
| SageMaker Model Monitor | Watches models in production for drift in quality and bias. |
| Amazon Augmented AI (A2I) | Sends low-confidence predictions to human reviewers. |
| Label quality analysis | Checks that training labels are correct. |
| Human audits | People review outputs for problems. |
| Subgroup analysis | Compares performance across groups (age, region, gender). |
4.1A responsible bias workflow on SageMaker
- Underrepresented group? Create more examples for it: data augmentation (rotate or flip images, for example) or oversampling, in the Data Wrangler workflow you already use. This is the cheapest fix.
- Explain individual predictions: Clarify gives SHAP feature attributions for each prediction (it can return them with the endpoint's response).
- Governance loop: Model Monitor runs scheduled checks, Clarify writes bias reports, reports are stored in S3, and remediation is tracked (for example in Systems Manager OpsCenter).
4.2What is SHAP?
SHAP (SHapley Additive exPlanations) is a way to explain a single prediction by giving each input feature a score for how much it pushed the result up or down. It comes from Shapley values in game theory, which split a team's prize fairly by each player's contribution. Here the "game" is one prediction and the "players" are the features.
Start from the model's average prediction (the baseline). Each feature then adds or subtracts its share, and the shares add up exactly to the final prediction.
Why was this loan approved or denied?
| Term | Plain meaning |
|---|---|
| Feature attribution | How much each input contributed to a prediction. SHAP is one way to calculate it. |
| Local explanation | Why this one prediction came out the way it did (one applicant). |
| Global explanation | Which features matter most to the model overall, across all predictions. |
| Post-hoc | An explanation calculated after the model makes its prediction, by testing how the output changes. SHAP and LIME (Local Interpretable Model-agnostic Explanations) are post-hoc. They estimate the model's reasoning rather than read it directly. |
4.2Transparent and explainable models
- Transparent / explainable: you can see how it reached a decision (needed for loans, healthcare, legal).
- Not transparent ("black box"): accurate but hard to explain, like most deep learning and LLMs.
Tools for transparency
- SageMaker Model Cards: a "nutrition label" documenting a model's purpose, training data, metrics and risks.
- Amazon Bedrock Model Evaluations: compare models with documented results.
- Open-source models, open data and clear licensing add transparency.
| Need | Tool |
|---|---|
| WHY was this specific prediction made? | SageMaker Clarify (per-prediction feature attributions) |
| HOW was the model built, and what is it for? | SageMaker Model Cards |
4.2Human-centered design for explainable AI
Build in user-feedback mechanisms (thumbs up/down) and make AI decisions transparent (AI decision transparency) to users (tell them when AI is involved and why it decided something). AWS training groups the principles into three areas:
| Area | Principles |
|---|---|
| Design for amplified decision-making | Clarity (present information so it isn't misunderstood), simplicity, usability, reflexivity (prompt people to reflect on their own assumptions), accountability (who owns the decision) |
| Design for unbiased decision-making | Spot and reduce bias in the data, model and process |
| Design for human and AI learning | Cognitive apprenticeship (AI and people learn from each other), personalization, user-centered design |
Security, compliance and governance for AI
5.1The shared responsibility model
5.1AWS services and features to secure AI
| Service / feature | What it does | Hook |
|---|---|---|
| IAM roles, policies, permissions | Control who can do what (least privilege) | The keycard system |
| Encryption (AWS KMS) | Protect data at rest and in transit | A locked safe |
| Amazon Macie | Finds sensitive data (PII) in Amazon S3 | Macie finds secrets |
| AWS PrivateLink | Private connection to AWS services without the public internet | A private tunnel |
| Bedrock AgentCore Identity New | Manages identity and credentials so agents access tools and data securely on behalf of users | An agent's ID badge |
| Policy in AgentCore New | Sets rules for what actions and tools an agent is allowed to use | The agent's rulebook |
| Amazon Bedrock Guardrails | Filters harmful input and output, masks PII | The bouncer |
| AWS Secrets Manager | Stores and rotates secrets such as API keys, database passwords and tokens that apps and agents use | The locked key cabinet |
| Amazon VPC | Your private network in AWS; with PrivateLink, AI traffic stays off the public internet | Your own fenced yard |
| Amazon GuardDuty Out of scope | Detects threats and anomalous access (including to S3 data) | The guard on duty |
| IAM Identity Center | Single sign-on with corporate identities; lets apps respect existing user permissions | One badge for everything |
| IAM Access Analyzer | Finds resources shared outside the account; tightens policies | The leak finder |
| AWS Security Hub Out of scope | One dashboard of security findings and posture checks | The control room |
5.1Common access and data-protection patterns
| Requirement | Pattern |
|---|---|
| Each data scientist may only reach their own project's S3 data, with no workflow changes | Attribute-based access control (ABAC): tag each role with its project, allow access only where aws:PrincipalTag matches, and cap roles with a permissions boundary. |
| Traffic to Bedrock must not cross the public internet | VPC interface endpoint (AWS PrivateLink; a VPC, or virtual private cloud, is your private network in AWS). TLS (Transport Layer Security, the "S" in HTTPS) already protects data in transit. |
| Company must control the encryption keys | Customer-managed AWS KMS keys (Bedrock, SageMaker and AgentCore Memory accept them). |
| Provider must not train on our prompts | Nothing to configure: Bedrock never shares prompts or outputs with model providers or uses them for training. |
| EU data must never leave the EU | Store, train and run in an EU Region; bucket policy denies requests from outside that Region. (Encryption alone does not keep data in place.) |
| PII in training data and in outputs | Macie finds PII in S3 (you then remove or mask it); Guardrails masks PII in model outputs. |
| Agents need safe credentials and tool limits | AgentCore Identity (credentials) + Policy in AgentCore (allowed tools) + Guardrails prompt-attack filter (injection). |
5.1Security services people mix up
| Service | Finds or does… | Not for… |
|---|---|---|
| Amazon Macie | Sensitive data (PII) sitting in S3 | Anonymizing data, or spotting suspicious access |
| Amazon GuardDuty Out of scope | Suspicious or anomalous activity and access | Software vulnerabilities |
| Amazon Inspector | Software vulnerabilities in EC2, containers, Lambda | Active threats or PII |
| Bedrock Guardrails | Harmful content and PII in prompts and responses | PII stored in S3 |
Which service?
Think of your answer, then tap a card to check it.
5.1Documenting where data comes from
- Source citation: show which documents an answer came from (RAG makes this possible).
- Data lineage: track where data came from and how it changed.
- Data cataloging: an inventory of your datasets (for example, AWS Glue Data Catalog).
- SageMaker Model Cards: document the model and its training data.
- SageMaker ML Lineage Tracking: automatically records the chain from data source through processing and training jobs to the model and endpoint.
- SageMaker Experiments: records each training run's parameters, datasets and metrics, for reproducibility.
Secure data engineering best practices
Assess data quality, use privacy-enhancing technologies (masking, anonymization), control data access, and protect data integrity.
5.1Security and privacy risks for AI
| Risk / control | Plain meaning |
|---|---|
| Prompt injection | Malicious instructions hidden in input take over the model. |
| Data leakage prevention | Stop the model from revealing private data. |
| Output filtering & validation | Check answers before users see them. |
| Encryption at rest & in transit | Protect stored and moving data. |
| Audit trails & logging | Record AI interactions for review (who asked what, and what came back). |
| Toxicity | Harmful or offensive output; filter it. |
| Threat detection & vulnerability management | Find attacks and weaknesses early. |
| Infrastructure protection & application security | Secure the networks and apps around the model. |
Detecting and reducing hallucinations
RAG grounding
Base answers on retrieved trusted documents.
Output validation
Check answers against rules or sources (e.g., Guardrails contextual grounding checks).
Confidence scoring
Flag low-confidence answers for human review.
5.2Governance and compliance services
| Service | What it does | Hook |
|---|---|---|
| AWS CloudTrail | Logs every API call: who did what, and when | The security camera |
| AWS Config | Tracks resource settings and checks them against rules | The configuration inspector |
| Amazon Inspector | Scans workloads for software vulnerabilities | The vulnerability scanner |
| AWS Artifact | Download AWS compliance reports and agreements (SOC, ISO, HIPAA BAA) | The compliance paperwork folder |
| AWS Trusted Advisor | Best-practice recommendations for cost, security, performance | The advisor's checklist |
| AWS Audit Manager | Continuously collects evidence from your accounts and maps it to frameworks such as SOC 2, GDPR, HIPAA | The audit binder that fills itself |
| Amazon S3 Object Lock | Write-once-read-many (WORM): logs can't be changed or deleted for a set period | A sealed evidence bag |
| S3 Versioning & lifecycle | Keep every version of a dataset; keep or expire data on a schedule (e.g., 7 years) | The filing cabinet with dates |
| Bedrock invocation logging | Logs prompts and responses to CloudWatch Logs or S3 | The conversation recorder |
| AWS Well-Architected Tool | Reviews a workload against AWS best practices. Includes lenses for Generative AI and Responsible AI. | The architecture checkup |
Compliance names you'll see
| Name | Stands for | What it is |
|---|---|---|
| SOC 2 | System and Organization Controls | An audit report on how a company protects customer data |
| ISO 27001 | International Organization for Standardization | A global security-management standard |
| HIPAA | Health Insurance Portability and Accountability Act | US law protecting patient health information |
| BAA | Business Associate Addendum (Agreement) | The HIPAA contract you accept with AWS in AWS Artifact |
| GDPR | General Data Protection Regulation | EU (European Union) privacy law; drives data residency questions |
| PCI DSS | Payment Card Industry Data Security Standard | Rules for handling card numbers |
5.2Data governance and governance processes
- Data governance strategies: data lifecycles, logging, residency (where data is stored), monitoring, observation, and retention (how long data is kept).
- Governance processes: written policies, regular review cadence and strategies, transparency standards, and team training requirements.
Generative AI Security Scoping Matrix
Service families
Big services like SageMaker AI and Bedrock are really families of features, and exam questions usually name the feature, not the family. Each table lists the members, what each one does, and the clue that points to it. Click any name to open its full card.
Amazon SageMaker AI: build, train and run your own models
| Feature | What it does | Clue in the question |
|---|---|---|
| Web IDE for all ML work | "one development environment for data scientists" | |
| Point-and-click ML, no code | "business analysts with no ML or coding skills" | |
| Pre-trained and open-source models to deploy or fine-tune | "deploy a pre-trained model on SageMaker" | |
| Visual data prep, transforms, augmentation | "prepare data with little code", "balance an underrepresented group" | |
| Humans label training data | "data has no labels yet" | |
| Shared store of ML features | "teams keep rebuilding the same features" | |
| Built-in algorithms, prebuilt framework containers (script mode), or your own container | "existing PyTorch script, least overhead" → prebuilt framework container | |
| Large, resilient clusters for training foundation models | "train or pre-train a very large model across many GPUs" | |
| Tracks runs: parameters, datasets, metrics (managed MLflow) | "track and compare training runs", "reproducibility" | |
| Bias reports before and after training; SHAP explanations | "bias across groups", "explain why this prediction" | |
| Versions and approvals for models | "which version is approved for production" | |
| Automated, repeatable ML workflows | "repeatable retraining", "MLOps automation" | |
| Real-time, Serverless, Asynchronous, Batch Transform | latency, payload size and traffic pattern | |
| Watches live models for drift and bias drift | "accuracy dropped as customer behavior changed" | |
| Documents purpose, training, risks | "document how the model works for auditors" | |
| One view of all models and their monitoring status | "see every model's health in one place" | |
| Quickly creates least-privilege IAM roles for ML users | "set up permissions for data scientists" | |
| Records the chain from data to endpoint | "trace data through every transformation to the model" | |
| Human review of low-confidence or high-risk predictions | "human oversight before the decision is final" |
Amazon Bedrock: use and customize foundation models
| Feature | What it does | Clue in the question |
|---|---|---|
| Try and compare models in the console | "test models without writing code" | |
| Grants permission to use a model | "account can't use the model yet" | |
| Managed RAG with citations | "answer from company documents", "current data without retraining" | |
| FM plans steps and calls APIs | "take actions, multi-step tasks" | |
| Filters harmful content, PII, prompt attacks; grounding checks | "block topics", "mask PII", "stop jailbreaks", "catch hallucinations" | |
| Compare models on your data (metrics, LLM judge, humans) | "is the smaller or fine-tuned model good enough" | |
| Save, version and reuse prompts | "version prompts across teams" | |
| Make a private custom model | "change tone or format" / "learn domain language" | |
| Big teacher trains small student | "cheaper, faster, similar quality" | |
| Reserved capacity | "steady high volume", "custom model in production" | |
| Pay per token / cheaper delayed jobs | "unpredictable usage" / "overnight, not urgent" | |
| Reuse a repeated prompt prefix cheaply | "same long context in every request" | |
| Logs prompts and responses | "audit what users asked the AI" |
Agents and AI assistants
| Service | Who it's for | Clue in the question |
|---|---|---|
| Developers writing agent code | "open-source SDK, a few lines of Python" | |
| Builders who want a managed agent inside Bedrock | "managed agent with action groups" | |
| Teams running agents in production | "deploy securely at scale", any framework | |
| Developers writing applications | "agentic IDE", "specs", "steering files" | |
| Business users | "no-code chat agents, research, dashboards, automation" | |
| Business users (older, closed to new customers) | "permission-aware assistant", connectors and plugins |
AgentCore's parts
| Part | Job | Clue |
|---|---|---|
| Hosts agents, serverless | "run the agent without servers" | |
| Short-term and long-term memory | "remember across sessions", "share memory between agents" | |
| APIs become MCP tools | "give agents access to our APIs as tools" | |
| Who the agent is; its credentials | "OAuth tokens", "act on behalf of a user" | |
| What the agent may do (Cedar rules) | "limit which tools each agent can call" | |
| Traces and metrics | "see what agents did, debug them" |
Ready-made AI services, by type of data
: meaning, sentiment, PII
: medical facts
: other languages
: speech → text
: clinical speech
: visit → clinical notes
: text → speech
: photos, video
: your own objects
: forms, tables
: chatbots with intents
: live voice (FM)
: enterprise search
: recommendations
: fraud scoring (closed to new customers)
Data, databases and cost
| Job | Services |
|---|---|
| Get and prepare data | (buy third-party data) · (ETL, Data Catalog) · (no-code cleaning) · (big data) · (ML data prep) |
| Store and analyze | · · · (BI) |
| Vector search for RAG | · · · |
| Other databases | (key-value) · (in-memory cache) |
| Run your code | · · · |
| Watch the cost | (where money went) · (alerts before overspend) |
Security and governance, by job
| Job | Services |
|---|---|
| Control who gets in | (roles, ABAC, permissions boundaries) · · · · |
| Protect the data | (encryption keys) · (API keys, passwords) · (private network) · · |
| Find problems | (PII in S3) · (vulnerabilities) · out of scope: (threats), (all findings) |
| Keep AI output safe | · |
| Prove compliance | (who called what) · (settings over time) · (your evidence) · (AWS's reports) · (advice) · (best-practice review) |
| Keep records | · · · · |
Services A–Z
Every AWS service and feature you are likely to see on the exam, including the ones used as wrong answers. Each card says what it is, when to pick it on the exam, and what it gets confused with. Supporting services are usually distractors or glue. Tags follow AWS's exam guide v1.2 lists: In scope on the in-scope list (or a feature of a service on it), Out of scope on the out-of-scope list, Not listed on neither list.
Service flashcards
Read the situation, name the service in your head, then reveal.
Quick review: "If you see this, pick that"
Scan this the night before. Each row maps a common exam clue to the answer it usually points to.
AWS services
| If the question says… | Think… |
|---|---|
| Use FMs via API, no servers | Amazon Bedrock |
| Answer questions from company documents | RAG with Bedrock Knowledge Bases |
| Block harmful content or mask PII in GenAI | Bedrock Guardrails |
| Compare or evaluate FMs | Bedrock Model Evaluation |
| Version and reuse prompts | Bedrock Prompt Management |
| Build an agent in code (open source) | Strands Agents |
| Run agents securely in production | Bedrock AgentCore |
| AI assistant for business users, no code | Amazon Quick |
| AI-powered coding IDE | Kiro |
| Build and train your own ML model | SageMaker AI |
| Deploy a pre-trained model in SageMaker | SageMaker JumpStart |
| No-code ML for analysts | SageMaker Canvas |
| Detect bias / explain predictions | SageMaker Clarify |
| Model drift in production | SageMaker Model Monitor |
| Document a model's purpose and risks | SageMaker Model Cards |
| Humans label training data | SageMaker Ground Truth |
| Humans review low-confidence predictions | Amazon Augmented AI (A2I) |
| Sentiment, entities, PII in text | Amazon Comprehend |
| Speech → text | Amazon Transcribe |
| Text → speech | Amazon Polly |
| Translate languages | Amazon Translate |
| Chatbot or voice bot | Amazon Lex |
| Analyze images and video | Amazon Rekognition |
| Extract text and tables from forms | Amazon Textract |
| Store embeddings / vector search | OpenSearch Service, Aurora / RDS PostgreSQL (pgvector), Neptune |
| Find PII in S3 | Amazon Macie |
| Private access, no internet | AWS PrivateLink |
| Audit who called which API | AWS CloudTrail |
| Track configuration compliance | AWS Config |
| Scan for vulnerabilities | Amazon Inspector |
| Download compliance reports | AWS Artifact |
| Best-practice recommendations | AWS Trusted Advisor |
| Automate audit evidence for SOC 2 / HIPAA | AWS Audit Manager |
| Tamper-proof audit logs | CloudTrail → S3 with Object Lock |
| Detect anomalous access to data | Amazon GuardDuty (out of scope; expect it only as a distractor) |
| Store API keys and passwords safely | AWS Secrets Manager |
| Review a GenAI workload against best practices | AWS Well-Architected Tool (Generative AI Lens, Responsible AI Lens) |
| AI agents to modernize legacy apps | AWS Transform |
| Alert before spend passes a limit | AWS Budgets |
| Track training runs, hyperparameters, metrics | SageMaker Experiments (managed MLflow) |
| Trace data → transformations → model | SageMaker ML Lineage Tracking |
| Run my existing PyTorch script with least effort | SageMaker prebuilt framework container (script mode) |
| Train a huge model on a cluster | SageMaker HyperPod |
| Clinical notes from doctor–patient audio | AWS HealthScribe |
| Recommendations with no ML code | Amazon Personalize |
| Test models in the console, no code | Bedrock playgrounds |
| Cheapest, fastest Amazon text model | Amazon Nova Micro |
| Generate images / video with Nova | Nova Canvas / Nova Reel |
| Agent remembers across sessions | AgentCore Memory |
| Agent credentials / allowed tools | AgentCore Identity / Policy in AgentCore |
| Project context for Kiro | Steering files: product.md, structure.md, tech.md |
| Align requirements before coding | Kiro specs: requirements.md → design.md → tasks.md |
| Restrict S3 access by project tag | ABAC with aws:PrincipalTag + permissions boundary |
Concepts
| If the question says… | Think… |
|---|---|
| Labeled data, predict a category | Supervised · classification |
| Labeled data, predict a number | Supervised · regression |
| No labels, find groups | Unsupervised · clustering |
| Learn by reward and penalty | Reinforcement learning |
| Instant prediction | Real-time inference |
| Large dataset, no rush | Batch inference |
| Great on training, bad on new data | Overfitting (high variance) |
| Bad on everything | Underfitting (high bias) |
| Missing positives is costly | Recall |
| False alarms are costly | Precision |
| Balance precision and recall | F1 score |
| Evaluate a summary | ROUGE |
| Evaluate a translation | BLEU |
| Same meaning, different words | BERTScore |
| More consistent / factual answers | Lower temperature |
| Use latest private data without retraining | RAG |
| Change model tone or task behavior | Fine-tuning |
| Smaller, cheaper copy of a big model | Distillation |
| Show a few examples in the prompt | Few-shot prompting |
| Ask the model to reason step by step | Chain-of-thought |
| "Ignore your instructions and…" | Prompt injection / hijacking |
| Standard way to connect agents to tools | Model Context Protocol (MCP) |
| Confident but false answer | Hallucination → RAG grounding, validation |
| Same prompt, different answers | Nondeterminism |
| Need a guaranteed exact answer | Don't use ML; use rules |
| Need to explain decisions (regulated) | Simpler, interpretable model |
| Which features drove this prediction? | SHAP feature attributions (SageMaker Clarify) |
| Improve transcription of specialist terms | Transcribe custom language model (or custom vocabulary) |
| Add more variables to the training data | Data collection + feature engineering |
| Tune settings while iterating | Validation set (test set is the final check) |
| Pick one of 3+ categories | Multiclass classification |
| Wrong output format | Output indicator / strong format directives |
| Model revealed its system prompt | Prompt leaking (exposure) |
| Standard pre-approved prompts on an MCP server | MCP prompts |
| Steps unknown in advance, agent must adapt | Autonomous agent (self-orchestration) |
| Agent must read voice, forms and sensors | Perception layer |
| Agent moves physical hardware | Actuators channel |
| Users misread AI output, decide inconsistently | Clarity (human-centered design) |
| Underrepresented group in images | Data augmentation (Data Wrangler) |
Acronyms A–Z
Every acronym used in this guide. Type any of them in the search box to jump here.
| Acronym | Stands for | In plain words |
|---|---|---|
| A2I | Amazon Augmented AI | Human review of ML predictions |
| ABAC | Attribute-based access control | Access decided by tags, e.g., project = blue |
| ACM | AWS Certificate Manager | Manages TLS certificates |
| AI | Artificial intelligence | Machines doing tasks that need human-like thinking |
| AIF-C01 | AWS Certified AI Practitioner exam code | This exam |
| API | Application programming interface | How one program asks another to do something |
| ARPU | Average revenue per user | Revenue ÷ active users |
| AUC | Area under the (ROC) curve | How well a classifier separates two classes; 0.5 = guessing, 1 = perfect |
| AWS | Amazon Web Services | Amazon's cloud platform |
| BAA | Business Associate Addendum | HIPAA agreement accepted in AWS Artifact |
| BERT | Bidirectional Encoder Representations from Transformers | A language model; BERTScore uses it to compare meaning |
| BI | Business intelligence | Dashboards and reports (Quick Sight) |
| BLEU | Bilingual Evaluation Understudy | Translation metric: word-sequence precision |
| CDN | Content delivery network | Caches content close to users (CloudFront) |
| CI/CD | Continuous integration / continuous delivery | Automatically testing and shipping changes |
| CLV | Customer lifetime value | Total revenue from a customer over the relationship |
| CMK | Customer-managed key | A KMS key you control |
| DDoS | Distributed denial of service | Flooding a site with traffic to knock it offline |
| DMS | Database Migration Service | Moves databases to AWS |
| EARS | Easy Approach to Requirements Syntax | Template for testable requirements, used in Kiro specs |
| EC2 | Elastic Compute Cloud | Virtual servers |
| ECR | Elastic Container Registry | Stores container images |
| ECS | Elastic Container Service | Runs containers |
| EKS | Elastic Kubernetes Service | Runs Kubernetes containers |
| EMR | Elastic MapReduce (Amazon EMR) | Big-data processing with Spark and Hadoop |
| ETL | Extract, transform, load | Moving and reshaping data between systems |
| EU | European Union | Region with strict data residency rules (GDPR) |
| F1 | Not an acronym | Harmonic mean of precision and recall |
| FM | Foundation model | A large pre-trained model that can do many tasks |
| FN / FP | False negative / false positive | A missed case / a false alarm |
| GDPR | General Data Protection Regulation | EU privacy law |
| GenAI | Generative AI | AI that creates new content |
| GPU | Graphics processing unit | The chip that trains and runs most models |
| HIPAA | Health Insurance Portability and Accountability Act | US health data privacy law |
| HTTPS | Hypertext Transfer Protocol Secure | Web traffic encrypted with TLS |
| IAM | Identity and Access Management | Who can do what in AWS |
| IDE | Integrated development environment | A code editor with built-in tools (Kiro, SageMaker Studio) |
| IoT | Internet of Things | Connected devices and sensors |
| ISO | International Organization for Standardization | Publishes standards such as ISO 27001 |
| JSON | JavaScript Object Notation | A common text format for structured data |
| k-NN | k-nearest neighbors | Finding the closest vectors; how vector search works |
| KMS | Key Management Service | Creates and controls encryption keys |
| LIME | Local Interpretable Model-agnostic Explanations | A post-hoc explanation method, like SHAP |
| LLM | Large language model | A foundation model for text |
| MAE | Mean absolute error | Average size of a regression model's error |
| MCP | Model Context Protocol | Open standard connecting agents to tools and data |
| ML | Machine learning | Learning patterns from data |
| MLflow | MLflow (a name) | Open-source tool for tracking ML experiments; SageMaker hosts it for you |
| MLOps | Machine learning operations | DevOps practices applied to ML |
| NLP | Natural language processing | AI that understands human language |
| NoSQL | "Not only SQL" | Databases that don't use tables and joins, such as DynamoDB |
| OAuth | Open Authorization | Lets an app act for a user without sharing a password |
| OCR | Optical character recognition | Reading text from images or scans (Textract) |
| PCA | Principal component analysis | Combines many features into fewer |
| PCI DSS | Payment Card Industry Data Security Standard | Rules for card data |
| PII | Personally identifiable information | Names, addresses, account numbers, and so on |
| PT | Provisioned Throughput | Reserved model capacity in Bedrock |
| PyTorch | PyTorch / TensorFlow (names) | Popular open-source frameworks for building ML models |
| RAG | Retrieval Augmented Generation | Look up documents, then answer from them |
| RDS | Relational Database Service | Managed SQL databases |
| REST | Representational State Transfer | The common style for web APIs |
| RLHF | Reinforcement learning from human feedback | Tuning a model with human ratings |
| RMSE | Root mean squared error | Regression error that punishes big misses |
| ROC | Receiver operating characteristic | The curve behind AUC |
| ROI | Return on investment | Benefit compared with cost |
| ROUGE | Recall-Oriented Understudy for Gisting Evaluation | Summarization metric: how much content was recalled |
| S3 | Simple Storage Service | Object storage |
| SaaS | Software as a service | Software you use online, such as Salesforce |
| SDK | Software development kit | A ready-made code library |
| SDXL | Stable Diffusion XL | A Stability AI image model |
| SES / SNS / SQS | Simple Email / Notification / Queue Service | Email, alerts, message queues |
| SHAP | SHapley Additive exPlanations | Scores each feature's push on one prediction |
| SOC | System and Organization Controls | Audit reports such as SOC 2 |
| SQL | Structured Query Language | The language for querying databases |
| SSO | Single sign-on | One login for many apps (IAM Identity Center) |
| TLS | Transport Layer Security | Encryption for data in transit |
| TP / TN | True positive / true negative | A correct catch / a correct ignore |
| VPC | Virtual private cloud | Your private network in AWS |
| WAF | Web application firewall | Filters malicious web traffic |
| WORM | Write once, read many | Storage that can't be changed (S3 Object Lock) |
| XGBoost | Extreme Gradient Boosting | A popular, accurate tree-ensemble algorithm; harder to explain than one tree |