Task 2.2: Understand the capabilities and limitations of GenAI for solving business problems.

• ADeterminism, because a foundation model returns the same output for the same input

  Incorrect

  Foundation models sample each next token from a probability distribution, so the same prompt can produce different responses.

• BExact recall, because a foundation model repeats facts from its training data without error

  Incorrect

  Foundation models can produce incorrect or misleading output, known as hallucinations; factual errors are their most common form.

• CAdaptability, because one pre-trained model can perform many different tasks

  Correct

  Foundation models are trained on broad data and can perform a wide variety of general tasks, such as answering questions, writing text and summarizing.

• DZero inference cost, because a foundation model is already trained

  Incorrect

  Using a foundation model is still billed; on-demand inference in Amazon Bedrock, for example, is priced per input and output token.

One pre-trained model, many tasks: that adaptability is the main reason foundation models lower the barrier to building AI features.

• AAmazon Bedrock retrains the base model on each prompt that the company sends

  Incorrect

  Amazon Bedrock does not use your content to improve the base models, so the model does not change between the ten requests.

• BThe temperature is set to its lowest value, which forces a random choice on every run

  Incorrect

  It works the other way: a lower temperature makes the model favor higher-probability tokens and leads to more deterministic responses.

• CPrompt caching is turned on, so each request returns a different cached answer

  Incorrect

  Prompt caching reuses repeated prompt prefixes to cut latency and input token cost. It does not store or vary the model's answers.

• DThe model samples each next token from a probability distribution

  Correct

  To generate each token, the model samples from a probability distribution of possible next tokens, so repeated runs can differ. Inference parameters such as temperature change how much they vary.

Nondeterminism is built into how generative models produce text. Lowering temperature, Top K or Top P narrows the choices and makes output more consistent.

• AModel X, because a model with more parameters is always the better production choice

  Incorrect

  Larger models can perform better, but they are more expensive to operate, and here Model X misses the two-second latency requirement.

• BModel Y, because it meets the accuracy and latency needs at a lower cost

  Correct

  AWS recommends choosing the model that gives the best balance of performance, cost and risk for the use case. Model Y meets every stated requirement and costs less.

• CModel X, because public leaderboard rankings outweigh the company's own evaluation

  Incorrect

  Public benchmarks give only a general view of performance. Evaluation on your own data is what shows fit for your use case, and the scenario already has those results.

• DNeither, because a new model must be pre-trained to meet a latency target

  Incorrect

  Building a foundation model from scratch is expensive and can take months. Model Y already meets the target.

Model selection balances quality against latency, cost and risk. The best model is the one that meets the requirements, not the one with the top score.

• AThe change in conversion rate among shoppers who use the assistant

  Correct

  Conversion rate is one of the business value metrics the exam guide lists for generative AI applications, alongside average revenue per user and customer lifetime value.

• BThe BLEU score of the assistant's answers against human-written reference answers

  Incorrect

  BLEU is a metric for assessing a foundation model's output quality. It does not show whether the business earns more.

• CThe number of parameters in the model behind the assistant

  Incorrect

  The number of parameters describes model size, a selection factor that affects capability and cost, not business results.

• DThe total number of tokens the assistant processes each day

  Incorrect

  Token volume drives the inference bill. It measures usage and cost, not the value the assistant creates.

Separate model metrics (BLEU, ROUGE), usage and cost drivers (tokens) and business metrics (conversion rate, revenue per user). Leadership asked for the last kind.

• AThe model can only process structured, tabular data

  Incorrect

  Generative AI creates content such as conversations, stories, images, videos and music; it is not limited to tabular data.

• BThe model can produce confident answers that are false

  Correct

  An AI hallucination is incorrect or misleading output from a foundation model; factual errors are the most common kind, and they reduce user trust.

• CThe model needs a separate training run for every new question

  Incorrect

  Foundation models are pre-trained and can answer a wide variety of questions without new training for each one.

• DThe model cannot answer in more than one language

  Incorrect

  Foundation models can work across languages; AWS lists translation among the tasks they handle.

• EIt is hard to explain how the model reached a particular answer

  Correct

  ML models are similar to black boxes whose inner workings are hidden. Explaining an individual decision needs extra explainability methods, which matters in regulated areas such as finance.

Hallucinations and limited interpretability are the two limitations that matter most when answers carry financial or legal weight.

• ADrafting first versions of product descriptions for editors to review

  Incorrect

  Generating content for human review plays to generative AI's strength and keeps a person checking the output.

• BCalculating each employee's exact payroll tax under fixed published rules

  Correct

  Generative models sample their output and can produce confident errors (hallucinations), so an exact, rule-defined calculation that must be right every time is better done in ordinary code.

• CSummarizing long meeting transcripts into action items

  Incorrect

  Summarization is a common use of foundation models.

• DAnswering customer questions in natural language from the company's help center articles

  Incorrect

  Conversational question answering is a core generative AI use case, especially when grounded in the company's own content.

Use generative AI where variety and language matter; use deterministic code where one exact answer is required.

• ASupport for image generation

  Incorrect

  The input and output here are text; image generation is not needed.

• BThe lowest possible price per input and output token on the market

  Incorrect

  Cost matters, but a cheap model that cannot take in the whole contract fails the requirement.

• CA context window large enough for the whole contract

  Correct

  The context window is the number of tokens a model can take into account at once; larger windows are preferred for tasks such as summarizing long documents.

• DA high default temperature

  Incorrect

  Temperature controls randomness, not how much input the model can read.

Input length is a hard limit: if the document does not fit in the context window, the model cannot consider all of it in one request.

• AIt always costs more to run than any proprietary model

  Incorrect

  Costs vary by model and workload; AWS notes that some open-source models have lower computational costs than usage-based proprietary models.

• BIt cannot be customized with the company's own data

  Incorrect

  Open-source models can be adapted; for example, SageMaker JumpStart lets you fine-tune pretrained open-source models.

• CIt is automatically covered by the model provider's intellectual property indemnity

  Incorrect

  Indemnification for training data and outputs is described as a typical feature of proprietary models, not open-source ones.

• DIt may bring security and legal risks, so licensing terms need vetting

  Correct

  Open-source models may introduce security and legal risks; the organization must enforce its own compliance measures and thoroughly vet the licensing terms.

Open-source models offer control and flexibility but shift compliance and licensing work onto the user; proprietary models often include more built-in assurances.

• AIts training data has a cut-off date; add retrieval of current data (RAG)

  Correct

  LLM training data is static with a cut-off date, so recent changes are unknown to the model; RAG supplies current information from an authoritative source without retraining.

• BThe temperature is set too low; raise it so the model will consider newer facts

  Incorrect

  Temperature only changes how randomly the model picks among tokens. It cannot add knowledge the model never saw.

• CThe model is too small; switch to a model with more parameters

  Incorrect

  Any model trained before last week is missing the change; a bigger model has the same cut-off problem.

• DThe response length is too short; raise the maximum token count

  Incorrect

  Response length limits how long the answer is, not whether it is current.

Out-of-date answers are a knowledge problem, not a style problem. Ground the model in current data rather than tuning its parameters.

• AThe ROUGE score of the drafts against the agents' final replies

  Incorrect

  ROUGE measures how closely generated text overlaps with reference text. It is a model quality metric, neither a cost nor a business value.

• BThe size of the context window of the model behind the assistant

  Incorrect

  The context window is the number of tokens a model can keep in context, a criterion for choosing a model, not a figure in an ROI calculation.

• CThe average latency of the assistant's responses

  Incorrect

  Latency is a technical metric that engineering teams track; on its own it is neither a cost nor a business value figure.

• DThe agent hours the assistant saves, expressed as a money value

  Correct

  AWS guidance says an ROI dashboard should integrate business value metrics such as hours saved; this is the return side of the calculation.

• EThe total spend on the assistant, including inference and infrastructure costs

  Correct

  The same guidance names financial metrics such as total infrastructure spend and cost per interaction; token consumption drives the cost of operating the application.

ROI weighs what an application returns against what it costs, so the dashboard needs both sides: value delivered (hours saved, revenue lift, satisfaction) and money spent (inference, infrastructure, maintenance). Technical metrics matter to engineers but only feed ROI once they are traced to a business outcome.