Overview
Fundamentals of GenAI has three task statements. Task 2.1 covers the building blocks: tokens, chunking, embeddings, vectors, transformer-based LLMs, multi-modal and diffusion models, the foundation model lifecycle, token-based pricing, context engineering and agentic AI concepts such as multi-agent patterns, the Model Context Protocol, memory, tool use and orchestration. Task 2.2 weighs GenAI advantages against hallucinations, nondeterminism and limited interpretability, and covers model selection factors and business metrics.
Task 2.3 moves to AWS: Amazon Bedrock, SageMaker AI, SageMaker JumpStart, Amazon Quick, Kiro, Strands Agents and Amazon Bedrock AgentCore, plus why managed GenAI services lower the barrier to entry and how on-demand, batch and provisioned throughput pricing trade cost against performance. Much of the agentic content entered the guide during the 2026 revisions, so it deserves extra attention.
How it's tested
Domain 2 questions describe a GenAI goal or problem and ask which concept explains it, which limitation applies, which AWS service fits, or which pricing option suits the workload. Expect plain definitions mixed with business framing, such as choosing a model on cost and latency rather than on raw capability.
-
Explain the mechanism
A behavior such as a long prompt costing more or a model inventing facts is described, and you name the concept behind it, such as tokens or hallucination.
-
Pick the AWS GenAI service
A team's skills and goal are described, and you choose among Amazon Bedrock, SageMaker JumpStart, Amazon Quick, Kiro, Strands Agents or AgentCore.
-
Weigh cost and capacity
Traffic patterns and budget are given, and you choose on-demand, batch or provisioned throughput, or a smaller model, to balance cost with performance.
A company wants a customer service agent that looks up orders in an internal system and issues refunds. A plain chat model can only generate text, so it cannot act. The design needs an agent that can call tools, keep session memory and authenticate to back-end systems. Amazon Bedrock AgentCore provides managed runtime, memory, gateway and identity for this, and MCP offers a standard way to connect the agent to the order system.
Task statements 3
Official task statements from the Amazon Web Services outline, in outline order.
2.1 Explain the basic concepts of generative AI (GenAI)Tokens, embeddings, model types, FM lifecycle, pricing, context engineering and agentic AI.
- What it means
- This task statement covers the core ideas behind GenAI: tokens, chunking, embeddings, vectors, prompt engineering, transformer-based LLMs, foundation models, multi-modal and diffusion models. It also includes common GenAI use cases, the foundation model lifecycle from data selection to feedback, token-based pricing, context engineering, and agentic AI concepts including multi-agent patterns, the Model Context Protocol, memory management, tool use and workflow orchestration.
- What to understand
- A token is a chunk of text the model reads and writes, and pricing is usually per input and output token. An embedding is a numeric vector that captures meaning, so similar text sits close together in vector space. Transformers use attention to model relationships across a sequence; diffusion models create images by gradually removing noise. Context engineering means deciding what instructions, history, retrieved data and tool results enter the context window. Agents combine a model with tools, memory and orchestration, and MCP is an open protocol that standardizes how agents connect to external tools and data.
- How it shows up
- A GenAI behavior, cost or architecture is described, and you name the concept, model type or agentic pattern it reflects.
- Common confusion
- Confusing tokens with embeddings; tokens are units of text, while embeddings are vectors that represent meaning.
2.2 Understand the capabilities and limitations of GenAI for solving business problemsAdvantages, disadvantages, model selection factors and business value metrics.
- What it means
- This task statement covers what GenAI does well, such as adaptability, responsiveness, conversation and content generation, and where it falls short, including hallucinations, poor interpretability, inaccuracy and nondeterminism. It also covers factors for selecting a GenAI model (type, performance, capabilities, constraints, compliance, cost, latency, complexity) and metrics that show business value, such as ROI, efficiency, conversion rate, average revenue per user and customer lifetime value.
- What to understand
- Hallucination means confident output that is false, which happens because the model predicts likely text rather than checking facts. Nondeterminism means the same prompt can return different answers, which matters for audits and testing. Model choice is a tradeoff: a larger model may be more capable but slower and more expensive. Business value should be measured with outcomes such as conversion or handling time, not by how impressive the output looks.
- How it shows up
- A business use is proposed, and you identify the main limitation or the right selection factor or value metric.
- Common confusion
- Assuming a bigger model is always the best choice; cost, latency and compliance often favor a smaller one.
2.3 Describe AWS infrastructure and technologies for building GenAI applicationsAWS GenAI services, their advantages, infrastructure benefits and cost tradeoffs.
- What it means
- This task statement covers the AWS services for building GenAI applications, including Amazon Bedrock, Amazon SageMaker AI, SageMaker JumpStart, Amazon Quick, Kiro, Strands Agents and Amazon Bedrock AgentCore. It also covers why managed services help (lower barrier to entry, speed to market, cost-effectiveness), the security and compliance benefits of AWS infrastructure, and cost tradeoffs such as Regional coverage, token-based pricing, provisioned throughput and custom models.
- What to understand
- Amazon Bedrock gives serverless API access to foundation models from several providers. SageMaker JumpStart offers pre-trained models and solutions you deploy and can fine-tune in SageMaker AI, with more control and more operational work. Strands Agents is an open-source SDK for building agents in code, and AgentCore supplies managed runtime, memory, gateway, identity and observability for running agents securely. Kiro is an agentic development environment, and Amazon Quick serves business users. On-demand pricing suits variable use, while Provisioned Throughput reserves capacity for steady, high-volume or custom-model workloads.
- How it shows up
- A team's needs and constraints are described, and you choose the AWS GenAI service or pricing option that fits.
- Common confusion
- Mixing up Bedrock and JumpStart; Bedrock is a managed API, while JumpStart deploys models into SageMaker AI infrastructure you manage.
High-yield concepts
The terms, rules and formulas this domain turns into questions most often.
-
Token
A unit of text that a model reads or generates.
- Remember
- Cost and context limits are counted in tokens.
- Tested as
- Pricing and limit questions.
- Trap
- Equating one token with one word.
-
Embedding
A vector of numbers representing the meaning of text, images or other data.
- Remember
- Similar meanings have nearby vectors.
- Tested as
- Search and RAG questions.
- Trap
- Calling embeddings compressed tokens.
-
Transformer vs. diffusion model
Transformers predict sequences with attention; diffusion models generate images by denoising.
- Remember
- LLMs are transformers.
- Tested as
- Model type questions.
- Trap
- Assuming diffusion models generate text.
-
Hallucination
Fluent output that is factually wrong or unsupported.
- Remember
- It comes from prediction, not lookup.
- Tested as
- Limitation questions.
- Trap
- Thinking a bigger model removes it.
-
Nondeterminism
The same input can produce different outputs.
- Remember
- Matters for testing and audit.
- Tested as
- Limitation questions.
- Trap
- Expecting identical answers every run.
-
Context engineering
Choosing what instructions, history, retrieved data and tool results fill the context window.
- Remember
- Wider than prompt wording alone.
- Tested as
- New 2026 concept.
- Trap
- Treating it as a synonym for fine-tuning.
-
Model Context Protocol (MCP)
An open standard for connecting agents and models to external tools and data sources.
- Remember
- It is a connection protocol, not a model.
- Tested as
- Agentic AI questions.
- Trap
- Confusing it with a vector database.
-
Amazon Bedrock AgentCore
Managed services to run agents securely: runtime, memory, gateway, identity, policy and observability.
- Remember
- Operates agents at scale.
- Tested as
- Service selection questions.
- Trap
- Thinking it is a foundation model.
-
Strands Agents
An open-source SDK for building AI agents in code.
- Remember
- Build with Strands, run on AgentCore.
- Tested as
- Service selection questions.
- Trap
- Confusing it with a managed hosting service.
-
Bedrock vs. SageMaker JumpStart
Serverless model API versus deployable model hub in SageMaker AI.
- Remember
- Bedrock means no infrastructure to manage.
- Tested as
- Service selection questions.
- Trap
- Picking JumpStart when no-ops is required.
-
On-demand vs. Provisioned Throughput
Pay per token versus reserved model capacity for a term.
- Remember
- Provisioned suits steady high volume and many custom models.
- Tested as
- Cost tradeoff questions.
- Trap
- Buying provisioned capacity for sporadic use.
Side by side
Concepts candidates mix up, compared on the point that decides the answer.
GenAI building blocks
| Concept | Meaning | Key difference | Common trap |
|---|---|---|---|
| Token | Piece of text processed by the model | Unit of cost and length | Treating it as a word |
| Embedding | Numeric vector of meaning | Enables similarity search | Confusing it with a token |
| Chunk | Segment of a document prepared for retrieval | Unit of indexed content | Making chunks too large to be precise |
| Context window | Maximum tokens a model handles at once | Caps prompt plus output | Ignoring output tokens in the limit |
AWS options for building GenAI
| Concept | Meaning | Key difference | Common trap |
|---|---|---|---|
| Amazon Bedrock | Managed API to many FMs | No infrastructure to run | Thinking you must host the model |
| SageMaker JumpStart | Hub of models deployed in SageMaker AI | More control, more operations | Choosing it for a no-ops requirement |
| Strands Agents and AgentCore | Agent SDK plus managed agent runtime services | Build versus run agents | Treating them as models |
| Kiro and Amazon Quick | Agentic IDE for builders; agentic workspace for business users | Different audiences | Swapping their audiences |
GenAI pricing choices
| Concept | Meaning | Key difference | Common trap |
|---|---|---|---|
| On-demand | Pay per input and output token | No commitment | Ignoring output token cost |
| Batch inference | Submit many prompts for offline processing | Lower cost, not immediate | Using it for live chat |
| Provisioned Throughput | Reserved model units for a term | Dedicated capacity | Using it for light, spiky traffic |
Exam traps
Each trap is an answer that looks right. Learn the clue that should change your mind.
-
Model looks facts up
- What candidates pick
- Assume an LLM retrieves facts from a database when it answers.
- Why it looks right
- Answers sound authoritative.
- Why it's wrong
- An LLM predicts likely tokens from training patterns; without retrieval it can state false facts confidently.
Clue to watch for The scenario reports confident but wrong answers.
-
Bigger is better
- What candidates pick
- Choose the most capable, largest model for a simple classification task.
- Why it looks right
- Larger models score higher on benchmarks.
- Why it's wrong
- Larger models cost more per token and add latency; a smaller model often meets the requirement more cheaply.
Clue to watch for The question stresses cost or response time.
-
Provisioned for spiky traffic
- What candidates pick
- Buy Provisioned Throughput for a pilot with occasional use.
- Why it looks right
- Reserved capacity sounds reliable.
- Why it's wrong
- Provisioned Throughput charges for reserved capacity whether used or not; on-demand fits low or unpredictable volume.
Clue to watch for Usage is light, irregular or experimental.
-
MCP is a model
- What candidates pick
- Treat the Model Context Protocol as a model or a vector store.
- Why it looks right
- The name contains the word model.
- Why it's wrong
- MCP is an open protocol that standardizes how agents reach tools and data; it does not generate content or store embeddings.
Clue to watch for The need is connecting an agent to external systems.
-
Chatbot equals agent
- What candidates pick
- Call a model that only answers questions an agent.
- Why it looks right
- Both are conversational.
- Why it's wrong
- An agent plans, uses tools, keeps memory and takes actions; a chat model only generates responses.
Clue to watch for The task requires actions in other systems.
-
JumpStart for zero operations
- What candidates pick
- Choose SageMaker JumpStart when the team wants no infrastructure.
- Why it looks right
- JumpStart offers ready-made models.
- Why it's wrong
- JumpStart models run on SageMaker AI endpoints the team manages; Amazon Bedrock is the serverless API option.
Clue to watch for The requirement says no infrastructure management.
Answer framework
Work through Domain 2 questions about GenAI concepts, limitations and AWS service choices with these steps.
- Name the underlying concept first, such as tokens, embeddings, context window, hallucination or nondeterminism, before reading the answer options.
- Decide whether the task needs content generation only or actions in other systems, since actions point to agents, tools, memory and MCP.
- Choose the service by required control: Amazon Bedrock for managed model access, SageMaker AI or JumpStart for more control, AgentCore for running agents.
- Apply selection factors in order of the stated constraint, such as compliance, latency, cost, modality or model size, rather than raw capability.
- Match pricing to usage: on-demand for variable traffic, batch for offline volume, Provisioned Throughput for steady heavy load or custom models.
Practice question
Want more like this? The AIF-C01 practice test has questions for every domain.
Checklist
0 of 13
Tick each one you can do without notes. Your ticks stay in this browser.
Study strategy
Begin with the mechanics. Read a short explanation of how an LLM turns text into tokens, maps them to embeddings and predicts the next token, then use the Amazon Bedrock console or playground to watch token counts change as you lengthen a prompt. Seeing cost and output length move together makes pricing questions easy.
Second, write a limitations list with a business example for each: hallucination in a policy answer, nondeterminism in a compliance report, weak interpretability in a lending tool. Pair each with the selection factor or control that reduces it.
Third, give the agentic material its own session, because it is the newest content in the guide. Learn what an agent adds to a model, what MCP standardizes, and the separate roles of Strands Agents, AgentCore, Kiro and Amazon Quick. A simple table of service, audience and purpose works well.
Finally, compare Bedrock with SageMaker JumpStart and the three pricing options, then drill practice questions until you can name the concept behind each scenario in a few seconds.
FAQ
How much of AIF-C01 is Domain 2?
Fundamentals of GenAI is 24% of the scored content, roughly twelve of the fifty scored questions, across three task statements.
What is the difference between a token and an embedding?
A token is a piece of text the model processes. An embedding is a vector of numbers that represents meaning and supports similarity search.
Why do GenAI models hallucinate?
They generate statistically likely text rather than verifying facts, so they can produce fluent answers that are wrong. Grounding with retrieval reduces this.
What is the Model Context Protocol?
An open protocol that gives agents and models a standard way to connect to external tools, APIs and data sources.
How are Strands Agents and Amazon Bedrock AgentCore related?
Strands Agents is an open-source SDK for writing agents. AgentCore provides managed services to deploy and operate agents, including runtime, memory, gateway and identity.
When should I use Provisioned Throughput?
For steady, high-volume workloads that need dedicated capacity, and for many customized models. Light or unpredictable use is cheaper on demand.
Does this domain test prompt engineering techniques?
It defines prompt engineering as a concept. Specific techniques such as few-shot and chain-of-thought are tested in Domain 3.
Other Artificial Intelligence & Machine Learning certifications
Where the same subject shows up in other exams. Cards marked "Same subject" open the matching domain.
Test your Domain 2 knowledge
You have covered GenAI building blocks, agentic AI and MCP, GenAI limits and the AWS services and pricing used to build GenAI apps. See how well you apply them to scenarios.
Keep going
- Training Training Built on your Weak areas
- Practice test Questions by domain, with explanations
- Mock exam Full length, timed like the real exam
- Readiness test Find your weakest domain in minutes
- Cheat sheet Key terms and formulas on one page
- Ask the AIF-C01 AI tutor Get any concept explained your way