The AIF-C01 study guide: what to learn, in what order

A payments company has three years of transactions labeled as fraudulent or legitimate. Its data science team wants to build and deploy its own fraud detection model with full control over the algorithm and features. Which approach BEST meets these requirements?

• AUse Amazon Comprehend custom classification on the transaction records.
  Comprehend offers custom classification, so the word classification is tempting. It classifies natural language text, not tabular transaction features, and it gives no control over the algorithm.
• BUse Amazon Personalize to rank transactions by customer preference.
  Personalize builds recommendations from user preferences. It is not designed to classify transactions as fraudulent.
• CPrompt a foundation model in Amazon Bedrock to label each transaction.
  Bedrock is the headline GenAI service, so this is tempting. Prompting an FM does not give the team control over the algorithm and features, and it is a poor fit for high-volume tabular fraud scoring.
• DTrain a classification model in Amazon SageMaker AI on the labeled data.
  Correct. Fraud detection is a supervised classification problem. SageMaker AI lets the team choose algorithms and features, then train, evaluate, and deploy the model.

A support assistant includes an entire 300-page product manual in every prompt. Responses are slow, costs are high, and answers sometimes miss the relevant section. Which change BEST applies context engineering?

• AFine-tune the model on the full manual every time a new version of it is published
  Frequent retraining is costly and slow compared with retrieving current passages at query time.
• BUse prompt caching so the repeated 300-page manual costs less per request
  Prompt caching does cut cost and latency for repeated prefixes, so it is tempting. The model still receives the whole manual, so answers can still miss the relevant section; retrieving only relevant parts fixes that.
• CSwitch to a model with a larger context window and keep sending the full manual
  A bigger window allows more text but still pays for and dilutes attention across irrelevant pages.
• DRetrieve only the manual sections relevant to each question and add them to the prompt
  Correct. Targeted retrieval keeps the context small and on topic, cutting tokens and improving answer quality.

An HR department wants a chatbot that answers employee questions using company policies. The policy library runs to thousands of pages and changes every week, and answers must reflect the latest version. Which approach meets these requirements with the LEAST ongoing effort?

• AFine-tune a foundation model on the policies every week
  Weekly fine-tuning adds recurring training cost and still lags behind the latest changes.
• BUse RAG with Amazon Bedrock Knowledge Bases and sync the policies
  Correct. A knowledge base can be re-synced when policies change, so answers use current content without retraining any model.
• CPut the policy library in a cached prompt prefix for every request
  Caching lowers the cost of a repeated prefix, but thousands of pages will not fit in a context window, and weekly updates would break the cache anyway.
• DPaste the complete, current policy library into every user prompt as context
  This quickly exceeds context limits, raises token costs, and is hard to maintain.

A lender's credit model passed a bias review at launch. Over the following year the applicant population changed, and the compliance team wants to be alerted automatically if bias in the live model's predictions starts to exceed agreed limits. Which approach BEST meets this requirement?

• ARerun the original SageMaker Clarify pre-training bias report each year
  A pre-training report analyzes the original training dataset. It says nothing about how the live model behaves on today's applicants.
• BRecord the launch results in a SageMaker Model Card
  Model Cards document a model's purpose and evaluation results. They are a governance record, not a continuous monitoring and alerting mechanism.
• CRoute every live prediction to Amazon A2I human reviewers
  A2I adds human review of individual predictions, but it does not compute bias metrics against a baseline or raise automatic alerts when bias drifts past agreed limits.
• DUse SageMaker Model Monitor to track bias drift on the endpoint
  Correct. Model Monitor integrates with Clarify to monitor deployed models for bias drift on a schedule and can raise alerts through Amazon CloudWatch when metrics cross thresholds.

An e-commerce company deploys a customer-support agent that can call a refund tool. Leadership requires that the agent can never issue a refund above USD 500, and that this limit is enforced deterministically even if the model is manipulated. Which approach BEST meets this requirement?

• AAdd the limit as an instruction in the agent's system prompt
  Prompt instructions guide the model but are not deterministic. A manipulated or confused model can still ignore them.
• BAdd refunds above USD 500 as a denied topic in Amazon Bedrock Guardrails
  Denied topics filter conversation content. They do not reliably inspect and block the parameters of an agent's tool calls.
• CLog all refund tool calls with AWS CloudTrail and review them weekly
  Weekly log review is a detective control. Refunds above the limit would already have been issued before anyone noticed.
• DEnforce the limit with Policy in Amazon Bedrock AgentCore
  Correct. Policy in AgentCore evaluates tool calls against defined rules outside the model's reasoning, so a request that breaks the limit is blocked regardless of the prompt.