6.3. Practice Questions

Question 1. A regional retailer receives thousands of complaints each week about damaged deliveries. Resolving each case requires checking the order status, confirming inventory, choosing between a replacement and a refund according to policy, executing the choice in the order system, and escalating unusual cases to a supervisor. Which solution is most appropriate?

• A. A conversational chatbot trained to explain the returns policy to customers
• B. An AI agent integrated with order, inventory, and payment systems, with spending limits and escalation rules
• C. A retrieval-augmented assistant that answers questions using the latest policy documents
• D. A rules-based workflow with a fixed decision tree covering each known complaint type

Answer: B. The task requires multi-step actions across business systems and judgment among next steps, which is what agents do; limits and escalation rules provide oversight. A and C only produce text and cannot act. D breaks on the variety of real cases and cannot adapt to results.

Question 2. A company's legal team wants an assistant that answers employee questions about internal policies. Policies are updated several times a month, and employees must be able to see which document each answer came from. What should the team use?

• A. Fine-tune a foundation model on the current policy documents each quarter
• B. Train a custom model on the policy library using Amazon SageMaker AI
• C. Use retrieval augmented generation over the policy library with citations
• D. Write a longer system prompt that includes a summary of all policies

Answer: C. RAG retrieves current content at question time and supports citations, so updates take effect immediately. A and B freeze knowledge at training time, cost more, and cannot cite sources. D consumes tokens, goes stale, and cannot hold detailed, changing policies reliably.

Question 3. A consulting firm discovers that many employees use a free public generative AI tool to draft client deliverables, sometimes pasting confidential client data. Which TWO actions should the firm take? (Select TWO.)

• A. Block the unapproved tool and publish a register classifying AI tools as approved, under evaluation, or blocked
• B. Ask employees to use their judgment about which client data is safe to paste
• C. Provide an approved enterprise generative AI tool and a fast, organization-wide process for requesting new tools
• D. Monitor network traffic to the public tool and review the logs at year end
• E. Allow each department head to approve tools for their own team without central review

Answer: A and C. Together they combine a preventive control with a governed path and a sanctioned alternative, which contains shadow AI without driving it underground. B relies on goodwill. D detects too late and prevents nothing. E fragments governance and bypasses organization-wide review.

Question 4. A program manager is preparing to launch an AI assistant for 800 customer-service agents next quarter. The CFO will want proof of impact in six months. What should the program manager do first?

• A. Capture current handle time, resolution rate, and satisfaction for comparable agent groups before launch
• B. Select the most accurate foundation model available to maximize measurable improvement
• C. Launch to all agents at once so results accumulate as quickly as possible
• D. Survey agents after launch to estimate how much time the assistant saves them

Answer: A. Without a pre-implementation baseline and a comparison group, impact cannot be attributed or defended. B is a technical choice that does not enable measurement. C removes the possibility of a comparison group. D produces after-the-fact estimates rather than measured change.

Question 5. An insurer's pilot that summarizes claim documents met every technical success criterion. However, measured time savings are small, and the projected monthly usage cost at full rollout exceeds the projected savings. What is the most appropriate decision?

• A. Pause or terminate the initiative and capture reusable components and lessons
• B. Scale the pilot immediately because it met its technical success criteria
• C. Replace the foundation model with a larger one to improve summary quality
• D. Expand the pilot to more teams to increase total time savings

Answer: A. Prioritization weighs value and sustainability, not only technical success; when cost exceeds value, pausing or stopping is mature governance. B confuses technical success with business value. C adds cost without addressing the value gap. D scales an uneconomic solution.

Question 6. A bank uses a model to recommend credit limits. Regulators require that customers receive reasons when they are offered a lower limit. A more complex model is slightly more accurate but much harder to explain. How should the bank's leadership proceed?

• A. Choose the more accurate model, because accuracy is the most important responsible AI dimension
• B. Make the trade-off explicit, involve legal and risk owners, document the decision, and ensure required explanations
• C. Let the data science team decide, because model selection is a technical choice
• D. Delay any decision until a model is available that is both the most accurate and fully explainable

Answer: B. Trade-offs should be surfaced, decided by accountable owners with legal input, documented, and kept within regulatory floors such as explanation requirements. A ignores a legal obligation. C leaves a business and legal risk decision to a technical team. D avoids the decision indefinitely.

Question 7. A company's AI governance committee consists of the CIO and three data scientists. An AI tool that screens job applicants is awaiting approval. What is the most important improvement to the governance structure?

• A. Add more data scientists so the committee can review the model's architecture in depth
• B. Transfer approval authority to the vendor that supplied the screening tool
• C. Add business, legal, compliance, HR, risk, and privacy representation, and name one accountable owner
• D. Replace the committee with an automated checklist that every project completes

Answer: C. Governance needs cross-functional representation and clear accountability, especially for a high-risk employment use case. A adds more of the same perspective. B outsources accountability that remains with the company. D removes human judgment from a high-risk decision.

Question 8. A global manufacturer has run 25 AI pilots over two years. Only one has reached production. Pilots use different tools and approval processes, and plant data is held in disconnected systems with no clear owners. What best describes the situation and the next step?

• A. The company is highly mature and should launch more pilots in new areas
• B. The company should hire more data scientists so that pilots finish faster
• C. The company lacks advanced models and should adopt the newest foundation models
• D. The company is in pilot purgatory and should build governance, data ownership, and a production path

Answer: D. Many pilots with little production signals low maturity caused by missing foundations. A treats activity as maturity. B and C respond to an organizational problem with technical or staffing fixes that leave data ownership and governance gaps in place.

Question 9. Six months after launching AI tools, adoption among frontline staff is low. Surveys show that employees fear job losses and that a manager was reassigned after a failed experiment. Which TWO leadership interventions are most appropriate? (Select TWO.)

• A. Communicate specifically about timelines, expected outcomes, and how roles will change, through managers and two-way forums
• B. Make AI tool usage mandatory and track individual usage in performance reviews
• C. Postpone all further communication until the workforce impact is fully known
• D. Recognize well-run experiments that fail and share their lessons openly
• E. Restrict AI use to the central AI team until employees are more comfortable

Answer: A and D. A addresses fear of replacement with transparency; D addresses fear of failure with psychological safety. B creates compliance without trust. C lets rumors grow. E stalls adoption and does nothing about the underlying barriers.

Question 10. A generative AI assistant used by 30 analysts in a pilot will be rolled out to 4,000 employees. Which approach best prepares it for production?

• A. Increase the token budget and roll the pilot configuration out to all employees on the same day
• B. Keep the pilot team as the only support channel to preserve continuity
• C. Remove human review steps to reduce the operating cost at scale
• D. Complete risk classification and sign-off, and add monitoring, a support model, service levels, and user training

Answer: D. Production requires both governance and operational requirements that pilots usually lack. A copies a fragile pilot and creates a big-bang risk. B overloads builders who are not a support organization. C removes oversight exactly when exposure increases.

Question 11. A company plans to estimate the cost of a new generative AI application before approving its business case, and then track actual spending against that estimate after launch. Which pairing of AWS tools fits these two needs?

• A. AWS Cost Explorer to estimate, then AWS Pricing Calculator to track
• B. AWS Marketplace to estimate, then AWS Cost Explorer to track
• C. AWS Pricing Calculator to estimate, then AWS Cost Explorer to track
• D. AWS Pricing Calculator to estimate, then AWS Marketplace to track

Answer: C. AWS Pricing Calculator estimates costs before deployment; AWS Cost Explorer analyzes actual spend afterward. A reverses the tools. B and D misuse AWS Marketplace, which is for buying third-party products, not estimating or tracking spend.

Question 12. A city permit office has a 12-week backlog. Analysis shows that most of the delay comes from three sequential sign-offs that could run in parallel and a form that asks for the same information four times. An executive proposes an AI assistant to answer applicants' status questions. What should the strategist recommend first?

• A. Deploy the AI assistant to reduce the number of status calls from applicants
• B. Fine-tune a model on past permit applications to speed up reviews
• C. Build a predictive model to forecast which permits will be delayed
• D. Redesign the approval process and the form before considering AI

Answer: D. The root cause is process design, and AI would only automate the waiting. A treats a symptom. C predicts the delay instead of removing it. B applies an expensive technique to a problem that is not about model behavior.