Hard — Gen AI Leader Practice Quiz

Question 1

A retail company ran a successful gen AI pilot for product description generation using a premium model. Finance approved a budget based on the pilot's monthly costs. Six months after full rollout, actual costs are three times the projected budget, even though usage grew only modestly. The leadership team wants to understand the most likely root cause of the cost overrun so they can plan future initiatives more accurately. What is the MOST likely explanation?

Accepted Answer

The pilot budget only accounted for per-request inference costs and overlooked ongoing costs such as monitoring, human review, data pipeline maintenance, and prompt/output token growth at scale.

Answer

Premium foundation models automatically increase their per-token price once an organization exceeds a fixed number of monthly requests.

Answer

Gen AI models require full retraining every month, and retraining costs were excluded from the pilot budget.

Answer

The company should have used a predictive ML model instead, since generative models cannot be budgeted for at all.

Question 2

A retail company launched six gen AI initiatives 18 months ago. During a portfolio review, an AI-powered product-description generator shows high monthly inference costs but a marginal, hard-to-attribute lift in conversion, while consuming significant engineering support time. Three other initiatives show strong, measurable ROI. As the gen AI program leader, what is the most appropriate governance action for the product-description generator?

Accepted Answer

Decommission or pause it after documenting the decision, reallocating budget and engineering capacity to the higher-ROI initiatives.

Answer

Immediately scale it to all product categories, since sunk investment justifies broader deployment.

Answer

Keep it running indefinitely without review, because removing a live AI feature would signal program failure to executives.

Answer

Switch it to a larger, more expensive foundation model to try to improve the conversion lift before evaluating cost.

Question 3

A retail company rolled out a gen AI product-description generator across its e-commerce team six months ago. Leadership now claims the tool 'drove a 12% increase in online sales' and wants to expand the investment. The analytics lead is skeptical because online sales also benefited from a new seasonal ad campaign and a website redesign during the same period. What should the leader recommend before attributing the sales lift to the gen AI tool and approving expansion?

Accepted Answer

Isolate the tool's impact using a controlled comparison (e.g., an A/B test or holdout group) so the sales lift can be causally attributed rather than correlated with concurrent initiatives

Answer

Approve the expansion immediately, since a 12% sales increase clearly exceeds any reasonable ROI threshold for the tool

Answer

Measure the number of product descriptions generated per week and report that as the primary business impact metric

Answer

Survey the e-commerce team about whether they feel the tool is helpful and use their sentiment as the ROI justification

Question 4

A retail company ran three successful gen AI pilots in isolated teams, each showing strong ROI. However, when leadership tried to scale these solutions enterprise-wide, adoption stalled: budgets were fragmented, teams duplicated effort on the same tooling, and no one owned enterprise-wide decisions on vendor selection or Responsible AI standards. As the gen AI program leader, what is the most effective organizational action to unblock scaling?

Accepted Answer

Establish executive sponsorship and a centralized operating model (such as a Center of Excellence) to coordinate funding, standards, and shared tooling across teams

Answer

Ask each team to independently expand its own pilot and negotiate separate vendor contracts to preserve their autonomy and speed

Answer

Pause all gen AI work until every employee completes prompt-engineering training to close the skills gap

Answer

Select the single highest-ROI pilot and cancel the other two to focus all resources on one initiative

Question 5

A retail company's newly formed AI steering committee has approved four gen AI initiatives but has limited engineering capacity to launch them one at a time over the next year. The CFO wants to build organizational momentum and secure continued funding, while the CEO wants to ensure the program eventually delivers transformational value. Which sequencing approach best balances these leadership priorities when deciding what to build first?

Accepted Answer

Start with a low-effort, high-visibility use case that delivers a measurable early win, then reinvest the credibility and learnings into the more complex, higher-value initiatives

Answer

Launch the most technically ambitious and highest-value initiative first, since it will justify the entire program's budget in a single deliverable

Answer

Sequence the initiatives strictly by which teams volunteered first, to reward enthusiasm and maximize internal buy-in

Answer

Delay all launches until every initiative has a fully validated ROI model, so the committee can rank them purely by projected financial return

Question 6

A large insurer is launching several gen AI initiatives at once: an internal tool that summarizes meeting notes, a customer-facing chatbot that answers policy coverage questions, and a marketing team's brainstorming assistant. The Responsible AI governance board has limited review capacity and must decide where to focus its most rigorous oversight first. Which approach best reflects a risk-based governance strategy?

Accepted Answer

Prioritize the deepest review on the customer-facing policy chatbot, because incorrect coverage answers carry the highest regulatory and financial harm

Answer

Apply the same intensive review process equally to all three initiatives to ensure consistency

Answer

Prioritize the internal meeting-notes tool first, since it is deployed soonest and touches the most employees

Answer

Prioritize the marketing brainstorming assistant, because creative outputs are hardest to predict

Question 7

A retail company is choosing a gen AI platform to power several customer-facing and internal solutions over the next three years. During evaluation, an executive raises concern that committing deeply to one provider's proprietary APIs, prompt formats, and fine-tuned model artifacts could make it very costly to switch providers later if pricing or performance changes. From an organizational cost and risk governance perspective, which action best addresses this concern?

Accepted Answer

Assess vendor lock-in risk and build portability into the strategy by favoring standard interfaces, abstraction layers, and exportable data/prompts where practical

Answer

Select the single lowest-cost provider immediately to minimize spend and revisit the decision only if a problem occurs

Answer

Avoid all managed platforms and build every model in-house to guarantee full control

Answer

Sign the longest available contract term to lock in current pricing before it rises

Question 8

A company deploys a RAG-based internal support assistant that retrieves from a document store before answering. Employees complain that answers are still frequently off-topic or incomplete, even though the source documents clearly contain the correct information. The generation model itself performs well when given the right context manually. What is the MOST likely area to investigate first to improve output quality?

Accepted Answer

The quality of the retrieval step, including how documents are chunked and how relevant passages are matched to the query

Answer

Fine-tuning the base model on a larger general-purpose dataset to increase its overall knowledge

Answer

Lowering the model's temperature setting to make responses more deterministic

Answer

Increasing the model's maximum output token limit so it can produce longer answers

Question 9

A financial analytics team uses an LLM to solve multi-step budget calculations. When they use chain-of-thought prompting, the model shows its reasoning, but occasionally a single arithmetic slip early in the chain leads to a wrong final answer. They want to increase the reliability of the final answers without switching models or fine-tuning. Which technique should they apply?

Accepted Answer

Sample multiple independent reasoning chains for the same prompt and select the answer that appears most consistently across them

Answer

Set the temperature to 0 so the model always produces the same single reasoning chain

Answer

Switch to zero-shot prompting to remove the intermediate reasoning steps that introduce errors

Answer

Increase the maximum output token limit so the model has more room to reason

Question 10

A logistics company is building an AI assistant that handles complex customer requests. A single request may require checking shipment status, calculating refund eligibility, and drafting a customer email. The team wants to break these responsibilities across several specialized agents that coordinate with one another, rather than cramming all logic into one monolithic agent. Which Vertex AI Agent Builder capability best supports this design?

Accepted Answer

A multi-agent system where a primary orchestrator agent delegates subtasks to specialized sub-agents

Answer

A single conversational agent with a larger context window to hold all instructions

Answer

A data store grounding connection to a single knowledge base document

Answer

A batch prediction pipeline that processes all requests overnight

Question 11

A retail company runs a large, highly capable Gemini model to power its product recommendation explanations, but inference costs are becoming unsustainable at their transaction volume. They want a smaller model that mimics the behavior and quality of the large model for this specific task, so they can deploy it at lower cost and latency. Which Vertex AI capability best addresses this need?

Accepted Answer

Model distillation, which trains a smaller student model to replicate the outputs of the larger teacher model

Answer

Grounding with Google Search to reduce the model's reliance on its parameters

Answer

Deploying additional TPU pods to horizontally scale the existing large model

Answer

Retrieval-augmented generation (RAG) to fetch recommendations from a vector database