Generative AI Leader Advanced Practice Exam: Hard Questions 2025

A financial services company is deploying an internal generative AI assistant for relationship managers. Users complain that the model’s answers are fluent but sometimes wrong, and the mistakes are hard to detect. The company wants to reduce hallucinations while ensuring responses cite authoritative internal policy documents. Latency must remain under a few seconds, and the solution must be maintainable by a small team. Which approach best addresses the requirements?

A support organization built a RAG-based assistant. In offline evaluation, retrieval quality is high, but in production the assistant frequently answers with irrelevant details from older documents. The team discovers that their document store contains multiple versions of the same SOP with inconsistent metadata, and users often ask time-sensitive questions ("What is the current escalation path?"). They want to minimize wrong-version answers without significantly increasing operational complexity. What is the best next step?

A media company wants to generate localized marketing copy in 25 languages using a single prompt template. They notice that for some languages the tone becomes overly formal and the brand voice drifts. They must preserve brand voice, avoid prohibited claims, and keep costs controlled. Which solution best balances quality, governance, and maintainability?

A retailer is building a multi-turn shopping assistant on Google Cloud that integrates with inventory and order APIs. The assistant must: (1) call tools only when necessary, (2) avoid exposing sensitive user data in logs, (3) provide consistent behavior across environments, and (4) support rapid iteration by multiple teams. Which architecture is most appropriate?

A company wants to let employees query internal documents (Drive exports, PDFs, and Confluence pages) through a generative AI interface. Requirements: enforce per-user access controls (employees should only retrieve documents they are authorized to see), minimize custom security code, and support citations. Which approach best meets these requirements on Google Cloud?

A startup is prototyping a customer-facing agent that uses tool calling to issue refunds. During testing, the model sometimes triggers the refund tool when a customer merely asks about refund policy. The team wants to reduce accidental tool execution without significantly degrading user experience. What is the best mitigation strategy?

A global manufacturer wants to deploy a generative AI assistant to help technicians troubleshoot equipment. Field connectivity is inconsistent, and technicians need answers even when offline. However, the company requires that: the assistant uses the latest maintenance bulletins when online, does not leak proprietary manuals, and supports auditability of answers. Which design best fits these constraints?

A healthcare provider wants to summarize clinician notes into discharge instructions. They must reduce clinician time while ensuring patient safety and compliance. The model output must never introduce new medical facts, and any uncertain or missing information should be flagged for clinician review. Which workflow is most appropriate?

A bank is evaluating generative AI for customer email responses. Legal requires that the system: (1) avoids giving regulated financial advice, (2) provides consistent responses across branches, and (3) can be audited to explain why a particular response was produced. Which combination of controls best satisfies these requirements?

A company deploys a generative AI summarization feature for customer support tickets. After launch, they discover that summaries occasionally include personal data that is not necessary for the summary (e.g., full addresses) and sometimes reflect bias in describing customer tone. They need to reduce these risks and demonstrate governance to auditors without halting the product. What should they do first?