Business Strategies for a Successful Gen AI Solution
Sample questions with answers
8 of the 15 questions in this set, with the correct answer marked and every option explained.
1. A business leader is trying to explain to their team how generative AI differs from the traditional machine learning models the company has used for years, such as a model that predicts which customers are likely to churn. What is the key distinguishing characteristic of generative AI?
✓It creates new content -- such as text, images, or code -- rather than only predicting a label or number from existing data
Traditional ML models like a churn predictor classify or predict a value from input data; generative AI's defining trait is producing new content -- text, images, audio, code -- which is a fundamentally different kind of output.
✗It always requires more training data than traditional machine learning
Data requirements vary by approach on both sides -- this isn't the defining conceptual difference between generative and traditional ML, and foundation models are often adapted with comparatively little task-specific data.
✗It can only be used for customer service chatbots
Generative AI spans text, image, code, and video generation across many business functions, not just chatbots -- this understates its scope rather than defining it.
✗It runs exclusively on a company's own private infrastructure
Where a model runs (cloud, on-prem, managed service) is a deployment choice, not the conceptual distinction between generative and traditional machine learning.
2. A product team is evaluating foundation models for three different projects: one needs to generate marketing copy, one needs to generate product photos from text descriptions, and one needs to analyze a mix of uploaded images and text together. Which model type distinction is most relevant for choosing the right model for the third project?
✓Multimodal models, which can process and reason across more than one type of input (such as text and images) at once
The third project needs to handle images and text together in a single request -- that's exactly what a multimodal model is built for, as distinct from a model that only handles one input type.
✗Diffusion models, which are used specifically for generating images from text prompts
Diffusion models are a strong fit for the second project (generating images from text), but the third project needs to jointly reason over existing images and text, which is a multimodal requirement, not a pure image-generation one.
✗Large language models with no image capability at all
A text-only LLM couldn't process the uploaded images in the third project's requirement -- it lacks the visual input capability entirely.
✗Open-source lightweight models, chosen only for their smaller size
Model size (lightweight vs. large) is a separate dimension from modality support -- a small model isn't inherently able to jointly process images and text just by virtue of being lightweight.
3. A data team has a large archive of customer support call transcripts that have never been categorized by topic or outcome. Before this data can be used to fine-tune a model for categorizing new calls, what is the most accurate way to describe its current state?
✓Unstructured and unlabeled data
Free-form call transcripts don't fit a predefined schema (unstructured), and they have no assigned category or outcome tags yet (unlabeled) -- both descriptions apply and matter for what has to happen before this data is trainable for a categorization task.
✗Structured and labeled data
Structured data fits a predefined schema like a database table, and labeled data already has target tags assigned -- raw call transcripts with no category tags are neither.
✗Structured and unlabeled data
Free-form transcripts are not structured -- they don't conform to a fixed schema like rows and columns -- so this only gets the labeling half right.
✗It doesn't matter whether the data is labeled, since generative models never need labels
This use case specifically wants to fine-tune a model to categorize calls -- that supervised task needs labeled examples, so labeling status directly matters here.
4. A company is mapping out its plan to build a generative AI-powered internal search tool, from collecting source documents through to keeping the deployed tool accurate over time. Which sequence best reflects the standard machine learning lifecycle stages this plan should follow?
✓Data ingestion, data preparation, model training, model deployment, model management
This is the standard ML lifecycle: data is first ingested, then prepared for use, a model is trained (or a foundation model is adapted), it's deployed into production, and then ongoing model management keeps it accurate over time -- matching the 'collect through to keep accurate over time' scope described.
✗Model deployment, then data ingestion, then data preparation
Deploying a model before ingesting or preparing any data is out of order -- there would be nothing trained yet to deploy.
✗Model training only, with no data preparation or ongoing management step
Skipping data preparation risks training on messy data, and skipping model management means there's no plan to keep the tool accurate over time, which the scenario explicitly asks for.
✗Model management first, followed by data ingestion
Model management (monitoring, retraining, versioning) applies to a model that's already deployed -- it can't logically come before any data has even been ingested.
5. An executive asks why a company's internal gen AI assistant sometimes confidently states outdated pricing information for a product that was repriced last month, even though the assistant was never told to lie. Which model limitation best explains this behavior?
✓Knowledge cutoff -- the model's training data has a fixed end date and doesn't include events or changes after that point
A model trained before the repricing simply has no knowledge of it -- it isn't fabricating the old price maliciously, it's reflecting the most recent information available at training time, which is exactly what 'knowledge cutoff' describes.
✗The model's temperature setting was set too low
Temperature controls output randomness/creativity; it doesn't determine whether the model has current pricing information -- lowering or raising it wouldn't fix outdated knowledge.
✗The model was never given a system prompt
A missing system prompt affects instruction-following and tone, not whether the model's underlying training data includes recent events -- it doesn't explain outdated factual knowledge.
✗The model has too large a context window
Context window size affects how much text fits in a single request; it has no bearing on whether the model's training data includes last month's price change.
6. A team wants to quickly try out several different pre-trained foundation models for a summarization task, without training any model from scratch. Which Vertex AI capability is built specifically for browsing and trying pre-trained models?
✓Model Garden
Model Garden is Vertex AI's centralized repository of pre-trained, ready-to-use models, letting a team browse and experiment with multiple model options for a task like summarization without training anything from scratch.
✗BigQuery
BigQuery is a data warehouse for storing and querying structured data; it does not provide pre-trained generative models to browse or try out.
✗Cloud Run
Cloud Run is a serverless compute platform for running containerized applications; it manages infrastructure, not a catalog of pre-trained AI models.
✗Cloud Storage
Cloud Storage is an object storage service for files and data; it has no role in browsing or selecting pre-trained foundation models.
7. A company wants employees to be able to ask natural-language questions and get answers grounded in the company's own internal documents, going beyond what simple keyword search over a document repository could provide. Which Google Cloud offering is purpose-built for this enterprise search scenario?
✓Vertex AI Search
Vertex AI Search is built for exactly this: enterprise search enhanced with gen AI capabilities, so employees get natural-language, grounded answers from internal documents rather than a plain keyword-matched list of results.
✗Imagen
Imagen is a text-to-image generation model; it has no role in searching or answering questions from a document repository.
✗Veo
Veo generates video content from text prompts; it's unrelated to enterprise document search.
✗Cloud Functions
Cloud Functions is a serverless compute service for running small pieces of code in response to events; it isn't a search product and provides no built-in grounding or retrieval capability.
8. A large enterprise wants a business-grade AI assistant with enterprise security controls, multimodal search across internal content, and the ability to build custom agents -- distinct from what an individual consumer would get from a personal AI assistant app. Which Google offering matches this description?
✓Gemini Enterprise
Gemini Enterprise is specifically the business-grade tier: enterprise security and customization, multimodal search, and custom agent capabilities, which is exactly the enterprise-vs-consumer distinction the scenario is asking about.
✗The consumer Gemini App
The consumer Gemini App targets individual, personal use (like Gems as custom chatbots for personal productivity); it isn't built around enterprise security controls or custom agent development for a business.
✗Gemini for Google Workspace only
Gemini for Google Workspace focuses specifically on AI assistance inside Gmail, Docs, Sheets, Slides, and Meet -- it's a real enterprise product, but it's narrower than the broader search-and-custom-agent platform described.
✗Google AI Studio
Google AI Studio is a lightweight environment for rapid prototyping and experimentation, not a business-grade assistant platform with enterprise search and agent-building for production use.
7 more questions in the app
Practise the full 15-question set with a timer, scoring and progress tracking.