Fundamentals of GenAI
AWS Certified AI Practitioner exam topic, about 24% of the exam. Updated .
The Support Bot: play by chapter
Chapter 1: What is it actually doing?
2 questions
Chapter 2: Where it goes wrong
2 questions
Chapter 3: Building it on AWS
2 questions
Storyline: The Support Bot
An online retailer builds its first AI support assistant and learns what generative AI costs, where it fails, and which AWS service to build it on.
The company is Shelfline, a 300-person online homewares retailer. You meet Leo (head of customer support, owns the assistant), Mei (engineering lead, builds it), Sam (finance analyst, watches the bill) and Nora (senior support agent, spots its mistakes). Across three chapters, you are their advisor.
Chapter 1: What is it actually doing?
Generative AI models read and write tokens, small chunks of text. You pay for both: the input tokens in your prompt and the output tokens in the reply. Long prompts cost money on every request. On their own, models only produce text; to act on other systems, agents connect to tools and data. The Model Context Protocol (MCP) is an open standard for those connections, so each tool is exposed once and any compatible agent can use it.
Chapter 2: Where it goes wrong
Generative AI is adaptable and conversational, but it has limits. It can hallucinate: produce fluent, confident answers that aren't true. It is nondeterministic: the same question can get different answers. Choosing a model is a trade-off between capability, latency, cost, language support and compliance. The biggest model is rarely the right default.
Chapter 3: Building it on AWS
Amazon Bedrock gives you foundation models from several providers through one API, fully managed. Amazon SageMaker AI is for building, training and hosting your own models, and SageMaker JumpStart deploys pre-trained models to endpoints you run. On Bedrock, on-demand pricing charges per token and suits variable traffic; Provisioned Throughput reserves capacity for steady, high-volume traffic.