Prepare and test your skills
Prepare and test your skills
Recognizing the different types of generative AI solutions means matching the kind of output a business needs to the right Google Cloud model and pattern. Text, image, code, video, music, and personalized conversational experiences are the main solution types. Personalized solutions often combine several types by adding enterprise data to a foundation model. Foundation models (opens in a new tab) are large models trained on vast amounts of text, images, video, and audio that predict likely responses and generate new content. Google offers foundation models through a managed API, and the Vertex AI model garden provides access to available models.
Personalized virtual agents are conversational systems that synthesize information from many enterprise sources into specific, reasoned, and actionable responses. A virtual agent can draw on historical support cases, websites, product documentation, inventory, emails, and old chat conversations that include resolutions. Virtual agents understand conversational queries, summarize complex information, and base answers on provided data in line with business expectations. Generative AI chatbots differ from traditional AI chatbots in understanding, inputs, and resolution. Traditional chatbots focus on specific keywords and text input, while generative AI chatbots interpret complex language and full context and accept text, voice, and image input. For example, a customer can share a photo of a damaged product, and the chatbot combines the description with the photo to improve troubleshooting accuracy.
| Chatbot type | How it understands a request | Input methods | Effect on issue resolution |
|---|---|---|---|
| Traditional AI chatbot | Focuses on specific keywords | Text-based keyword input | Limited to keyword matching |
| Generative AI chatbot | Interprets complex language, nuances, and full context | Text, voice, and image | Combines all input to improve diagnostic accuracy |
Exam tip: A traditional AI chatbot matches keywords, while a generative AI chatbot interprets full context and can combine text, voice, and images.
Intelligent search for personalized needs is a meaning-based retrieval system that uses embeddings and retrieval-augmented generation (RAG) (opens in a new tab) to find information by meaning rather than exact words. An embedding is a numerical representation of data such as text, images, or video that captures relationships between inputs. Embeddings are created during training by converting data into arrays of floating point numbers called vectors, where closer proximity in vector space means similar meanings. The RAG retrieval flow runs in order: first the user query is converted into an embedding, then the query embedding is compared with document embeddings in a knowledge base, then a similarity search retrieves the most semantically relevant documents for the response. This flow lets the system pull from enterprise knowledge instead of relying only on the model’s internal knowledge. Function calling then connects large language models (LLMs) to external tools such as APIs, databases, customer relationship management systems, and document repositories by letting the model choose a tool and format input for precise interaction.
A four-step process flow showing the RAG retrieval flow: the user query is converted into an embedding, compared against document embeddings in the enterprise knowledge base, the most semantically relevant documents are retrieved by similarity search, and the model generates a response grounded in that retrieved enterprise data.
A large context window lets a personalized solution handle large amounts of user or enterprise context. A context window of 1 million tokens can process 50,000 lines of code, all text messages sent in the last 5 years, 8 average-length English novels, transcripts of over 200 average-length podcast episodes, 1 hour of video without audio, approximately 45 minutes of video with audio, or 9.5 hours of audio.
Matching modalities to models means selecting the Google Cloud model or function based on the output type the task produces. Gemini models handle text and multimodal work, Imagen generates and edits images, Veo generates high-resolution video, Lyria generates music with instrumentals and vocals, and embedding models create vectors for search and comparison. A single task can combine modalities when a text prompt is embedded and compared against image embeddings for multimodal search. Text generation in BigQuery uses AI.GENERATE as a flexible scalar function that returns a single value with no remote model, while AI.GENERATE_TEXT is a table-valued version that takes a table, returns a table, and requires a remote model to be created first. Both process between 1 million and 10 million rows per 6-hour query job. Imagen edits image content using a mask, adds content with mask-based inpainting, adds content with automatic mask detection, and produces short-form captions, while Vision AI and Cloud Vision API analyze images that already exist. Code generation covers batch code prediction, embeddings for code retrieval, and prompts that generate unit tests with a publisher chat model.
| Modality | Google Cloud model or function | Output produced |
|---|---|---|
| Text | Gemini, AI.GENERATE, AI.GENERATE_TEXT | Written content, answers, summaries, structured data |
| Image | Imagen | Generated or edited images, image captions |
| Code | Pre-trained code generation models | Code predictions, unit tests, code retrieval embeddings |
| Video and music | Veo, Lyria | Videos, music with instrumentals and vocals |
| Embeddings | AI.EMBED, AI.GENERATE_EMBEDDING | Numerical vectors for search, comparison, and clustering |
Identifying the key factors that influence generative AI needs means working backward from business goals and then checking technical, regulatory, and ethical limits. Business requirements decide if an initiative is worth doing, while technical constraints decide which model and deployment path can deliver it. Regulatory and security rules shape project scope from the start by limiting data, access, and release steps.
Analyzing business objectives means stating a measurable goal and the value to be achieved before selecting technology. Common goals include increasing efficiency and productivity, reducing costs while maintaining quality, enhancing customer experience and satisfaction, and creating competitive advantage. Business leaders approve the selection when it aligns with business priorities, and end users must understand the output and how to use it for informed decisions. Return on Investment (ROI) is a key measure used to judge AI project success and set its priority. ROI for generative AI is tracked through direct financial gains such as increased revenue or reduced costs, operational efficiency such as faster time-to-market or quicker issue resolution, and customer experience such as increased satisfaction scores or improved retention. When expected ROI is strong and linked to a priority goal, the initiative has strategic necessity and ranks higher for funding.
Exam tip: Work backward from the desired business outcome to the technology choice, and use ROI to prove the link.
Technical constraints and infrastructure readiness are the limits and available resources that decide which AI model can be chosen and how it can be deployed. The three main technical constraints are data quality and readiness, latency requirements, and computational costs. These constraints determine whether a team uses a pre-trained traditional model, a custom-trained traditional model, or a generative AI model, and whether the solution runs on Vertex AI (opens in a new tab), Google Kubernetes Engine (GKE), or Compute Engine. Model selection starts by asking if a pre-trained traditional AI model such as Document AI, Vision AI, Natural Language API, or Video Intelligence API can already meet the requirements. If the pre-trained model cannot meet requirements, the next question is whether enough training data exists to custom train a model. Traditional predictive AI works particularly well with structured data for prediction and classification, while generative AI is used for summarization, content generation, advanced transcription, and work that combines text, images, videos, or audio. Generative AI workloads can use Graphics Processing Units (GPUs) on Compute Engine, GKE provides AI orchestration with high performance and flexible infrastructure, and Vertex AI reduces operational work for testing, tuning, and deploying foundation models including Gemini.
| Option | When to choose | Data requirement | Tradeoff |
|---|---|---|---|
| Pre-trained traditional AI model | Classification or detection use case and the pre-trained model meets requirements | No custom training data needed | Fastest path, but less control |
| Custom-trained traditional AI model | Pre-trained model does not meet requirements and high control is priority | Enough training data to custom train | More control over algorithm and hyperparameters; slower go-to-market |
| Generative AI with prompt engineering | No training dataset exists, datasets are too small, or faster go-to-market is priority | Can start with prompts; can be tuned with examples | Faster go-to-market and multimodal support; less control than custom training |
The generative AI development flow has six stages in order: model selection, prompt engineering, tuning, optimization, deployment through Model Registry, online prediction, or batch prediction, and monitoring. Model selection, prompt engineering, tuning, and optimization form a looping evaluation subcycle where results are tested and refined before deployment. Combining approaches is also common, where output from a traditional AI model becomes part of the prompt for a generative AI model, such as churn forecasts explored through a conversational chatbot.
Regulatory, security, and ethical considerations are the business constraints that shape what a project is allowed to do and how its outputs are governed. Regulatory requirements include industry compliance rules and country-specific laws that set boundaries for which data sources can be ingested, where data can be stored, and which users can access it. These constraints are identified during the business goal and success criteria step because they determine scope before technology is chosen. Security protocols protect data as it moves from source systems into storage and processing services. For RAG applications, private connectivity keeps network traffic within trusted boundaries by preventing data from traveling over the public internet. Security log analytics (opens in a new tab) collects, exports, and analyzes logs from Google Cloud to audit usage and detect threats. Responsible AI adds human decision points, such as a human review step before generated marketing content is uploaded to Cloud Storage, plus feedback loops so the model learns from the impact of its outputs over time.
Describing how to choose the right generative AI solution means starting from the business objective and narrowing through governance, build-versus-buy, and model selection. Governance decides where data can live and how much control is required, build-versus-buy decides whether pretrained services are enough, and Model Garden selection tunes performance, latency, cost, and infrastructure. Together these decisions balance innovation, cost-efficiency, and operational feasibility.
Assessing build versus buy is the decision between adopting a ready-made pretrained service and building a custom model for a specific need. The buy path calls models that Google already trained through an API, while the build path trains or tunes a model on company data with Vertex AI (opens in a new tab). Vertex AI brings AutoML and AI Platform together into a unified API, client library, and platform, where AutoML Natural Language classification creates custom category labels and AutoML entity extraction identifies company-specific entities.
Buying covers common tasks with generalized capabilities such as Cloud Natural Language API for classification, Cloud Translation API for translation between thousands of language pairs, Cloud Vision API for text detection, Document AI for pretrained processors, and BigQuery managed functions for routine work. Because these services arrive already trained, traffic flows from the application to the Google-hosted model and back, and model upkeep stays with Google. Building requires project setup, API enablement, and Identity and Access Management (IAM), including the Project Creator role, billing enablement, Vertex AI API enablement, and permissions such as aiplatform.endpoints.predict for a service account to call an endpoint.
| Dimension | Off-the-shelf (buy) | Custom on Vertex AI (build) |
|---|---|---|
| Speed to market | Fast, called through an API with no training | Slower, requires data, training, and evaluation |
| Cost profile | Optimized for cost and quality with no model upkeep | Pays for training, tuning, and ongoing management |
| Customization | Limited to generalized categories or pretrained processors | Full control over model, prompt, parameters, and labels |
Exam tip: Choose a pretrained or managed service when generalized categories fit the need, and choose AutoML or Vertex AI tuning only when the business requires custom labels, custom entities, or control over the model and prompt.
Meeting data governance by bringing AI to the data reduces unnecessary data movement and keeps sensitive data under existing controls. BigQuery ML lets SQL practitioners train, evaluate, and run inference with GoogleSQL queries, so forecasting, anomaly detection, classification, regression, clustering, and recommendations can run inside BigQuery (opens in a new tab). A secure data environments architecture secures sensitive datasets against accidental exposure and exfiltration, while a borderless open data lakehouse establishes a governed pipeline from raw silos to AI-driven actions.
Choosing the right BigQuery AI function means matching control, cost, and transparency to the task. General-purpose functions give full control over model, prompt, and parameters, managed functions use streamlined syntax and let BigQuery choose a model to optimize cost and quality, and task-specific functions use Cloud AI APIs for a defined task. General-purpose functions include AI.GENERATE, AI.GENERATE_TEXT, AI.GENERATE_TABLE, AI.GENERATE_BOOL, AI.GENERATE_DOUBLE, and AI.GENERATE_INT, while managed functions include AI.IF for filtering, AI.SCORE for rating, AI.CLASSIFY for classifying, and AI.AGG for summarizing. The AI.COUNT_TOKENS utility estimates tokens before a query runs to forecast cost and stay within limits, and IAM policy changes for access generally take effect within 2 minutes but can take up to 7 minutes to propagate fully.
Selecting models via Vertex AI Model Garden means matching a foundation model to the business need by comparing performance, latency, and cost. Performance is how well the model handles the task, latency is how fast it responds, and cost is what it takes to run it. The performance, cost, and scalability of an application depend directly on both the model and the underlying infrastructure of compute, storage, and networking. Gemini models are pretrained, general-purpose models that accept natural language and can generate content, answer questions, summarize, and analyze text or unstructured data such as images. Partner models and open models give alternatives when a specific task, license, or cost profile fits better. Cost and latency both grow with how much text the model processes, so teams track token usage, track costs, and set up monitoring and error management. The right model balances innovation against cost-efficiency: a model that is powerful but too slow or too expensive fails just as surely as one that is cheap but inaccurate.
Identifying the steps to integrate generative AI means combining trusted data, governed infrastructure, high-value use cases, and change management into one ordered framework. Data readiness provides reliable inputs, infrastructure provides scalable compute and models, use-case selection provides business focus, and governance provides safe adoption. Each part depends on the others before AI outputs reach daily workflows.
Data Readiness and Infrastructure Foundation is the combination of trusted data and scalable cloud services that allows generative AI to create reliable business value. Data readiness means organizational data is accurate, complete, relevant, secure, and accessible to approved teams. Infrastructure foundation means compute, storage, and pre-trained models are ready so teams can test and deploy without building everything from scratch. Google Cloud manages data through its lifecycle without requiring all datasets to be consolidated in one place. Knowledge Catalog organizes distributed data into logical data lakes and zones, controls access with Google Groups and catalog roles, and streamlines quality checks with auto data quality. BigQuery sharing exchanges curated assets through data exchanges and data clean rooms, while Agent Platform Feature Store serves reusable tabular features for online and offline use, Cloud Storage holds unstructured data governed through object tables and signed URLs, and Vector Search manages vector embeddings for retrieval. Reuse avoids redundant engineering, which speeds model development and lowers infrastructure costs.
| Approach | Best for | Data requirement |
|---|---|---|
| Pre-trained traditional AI model such as Document AI or Vision AI | Classification or detection when an existing API meets requirements | No custom training data needed |
| Traditional predictive AI | Forecasting and classifying from historical patterns, especially structured tabular data | Requires sufficient historical training data |
| Generative AI model such as Gemini | Summarization, content generation, advanced transcription, and multimodal inputs | Works with prompt engineering without large dataset, with optional tuning |
Exam tip: When no training dataset is available or datasets are too small for custom training, use generative AI models with prompt engineering rather than custom training a traditional model.
Responsible AI governance and change management is the organizational foundation for integrating generative AI into workflows. A borderless open data lakehouse establishes a highly governed, secure pipeline from raw silos to AI and agentic driven actions, while secure data environments place protection on each side of the trust boundary between sensitive data and users, applications, and agents. Data moves forward only when exposure and exfiltration controls are in place. Machine learning operations (MLOps) connects model development to deployment and operation, so the same practices cover classification models, generative AI, and agentic AI systems. Google Cloud organizes planning knowledge into design guides (opens in a new tab) for prescriptive cross-product planning, reference architectures for detailed workload examples including RAG and single-agent and multi-agent systems, use cases for high-level business problems, and deployment guides for step-by-step instructions. Change follows an ordered migration path where plan comes first, then design, then implementation.
A Google Cloud landing zone guides identity onboarding, resource hierarchy, network design, and security, while Cloud Identity and Google Workspace provide managed accounts for workforce users. This order matters because design depends on the plan and implementation depends on the design.
Strategic use case identification is the practice of finding generative AI opportunities that serve measurable business goals rather than standalone experiments. The business value-driven decision process works backward from the outcome to the technology choice in five steps: define the business goal and success criteria with ROI and compliance constraints, decide the type of AI or whether AI is needed, define user experience expectations, connect the solution to business requirements and user expectations, and identify required business process changes. Traditional AI learns from historical data to classify or predict, while generative AI creates summaries, uncovers correlations, and generates new content. For classification or detection, first check whether a pre-trained traditional model meets the requirement; if not and enough data exists, choose custom training for high control or generative AI for faster go-to-market. A common combined pattern feeds traditional AI output into a generative AI prompt, such as forecasting churn probability and letting a chatbot explain it, identifying customer segments and generating personalized marketing content, or detecting sign language with computer vision and generating voice output. Example Google Cloud flows include ingesting marketing data into BigQuery, deriving insights with Dataflow, triggering Cloud Run through Eventarc, and prompting Gemini to generate assets stored in Cloud Storage, plus RAG support flows and vector similarity recommenders with human review and feedback loops where brand safety matters.
Continue with this study guide: Identifying the steps to integrate gen AI into an organization
Identifying techniques to measure the impact of generative AI initiatives means separating how well the system runs from what it delivers to the organization. Technical metrics guide engineers tuning prompts and models, while business metrics guide leaders deciding whether to continue, expand, or stop the work. Long-term return on investment (ROI) requires both sides together, tracked over time with continuous monitoring (opens in a new tab).
Distinguishing between technical performance and business value metrics means separating how well a system runs from what it delivers. Technical performance metrics are model-centric measurements such as accuracy and latency, while business value metrics are Key Performance Indicators (KPIs) such as cost savings, revenue generation, and time-to-market acceleration. BigQuery generative AI functions accept natural language inputs and run on pre-trained models, so the choice of model, prompt, and parameters directly shapes accuracy, latency, and resource use. BigQuery exposes signals that support both views: AI.GENERATE gives full control over model, prompt, and parameters, managed functions such as AI.IF, AI.SCORE, AI.CLASSIFY, and AI.AGG are optimized for cost and quality, AI.COUNT_TOKENS estimates tokens before execution, and monitoring and error management track behavior over time. Token usage is a technical measurement, but once multiplied by pricing it becomes a tracked business cost. Latency and accuracy affect adoption, and adoption produces revenue or time savings.
| Dimension | Technical performance metrics | Business value metrics |
|---|---|---|
| What it measures | Model behavior, such as accuracy and latency | Organizational outcomes, such as cost savings, revenue, and time-to-market |
| Example signal | Token count from AI.COUNT_TOKENS | Tracked costs and pricing for the same usage |
| Primary audience | Engineers and data teams tuning prompts and models | Business leaders funding and steering the initiative |
| Decision it informs | Which model, prompt, or parameters to use | Whether to continue, expand, or stop the initiative |
Exam tip: A token count is a technical metric; the same token count multiplied by pricing becomes a cost metric.
Qualitative impact measurement captures non-numeric signals such as sentiment, satisfaction, creativity, and experience from surveys, support transcripts, and human review. Managed AI functions are the most direct fit for scoring and sorting this feedback at scale, and with optimized mode they scale to millions or billions of rows. AI.IF filters data using natural language conditions, AI.CLASSIFY sorts input into user-defined categories, AI.SCORE rates input by quality or sentiment, and AI.AGG aggregates input to summarize or analyze data. These four functions move a team from raw text to structured signals in order: AI.IF narrows the dataset, AI.CLASSIFY sorts responses, AI.SCORE attaches a rating, and AI.AGG summarizes the result. Human-in-the-loop evaluation lets people set the rules while the model applies them, because AI.CLASSIFY uses human-defined categories and AI.IF uses human-written conditions. General-purpose functions such as AI.GENERATE offer full control and transparency over model, prompt, and parameters, while managed functions trade some control for streamlined syntax and cost-and-quality optimization.
Tracking long-term ROI means comparing cumulative business value against cumulative AI cost over time. Vertex AI evaluation tools measure model quality, BigQuery tracks token usage and query costs where the data lives, and dashboards such as Looker visualize both sides together. Running measurement inside BigQuery avoids moving data, so impact analysis can scale to large tables for sentiment detection, summarization, extraction, classification, embeddings, and rating. Vertex AI evaluation (opens in a new tab) follows a fixed order: define the evaluation specification with model, prompts, and metrics, evaluate the model against test data or reference responses, and retrieve metrics that show accuracy or summarization quality. The ROUGE metric compares a model summary against a ground-truth reference, and an evaluation slice shows metrics for a specific segment of results. After queries run, token usage tracking shows consumption, cost tracking shows spend, and monitoring reveals trends such as rising usage, falling cost per successful answer, or quality drops after changes.
Continue with this study guide: Identifying techniques to measure the impact of gen AI initiatives
A traditional AI chatbot focuses on specific keywords with text-based input and is limited to keyword matching, while a generative AI chatbot interprets complex language, nuances and full context. Generative AI chatbots accept text, voice and image input and combine all inputs to improve diagnostic accuracy, such as combining a product description with a photo of damage to improve troubleshooting.
Use generative AI with prompt engineering when no training dataset exists, datasets are too small, or faster go-to-market is a priority. If a pre-trained traditional AI model such as Document AI, Vision AI, Natural Language API or Video Intelligence API already meets requirements, use it with no custom training data, and choose custom training only when the pre-trained model does not meet requirements and high control is a priority.
General-purpose BigQuery AI functions give full control over model, prompt and parameters, while managed BigQuery AI functions use streamlined syntax and let BigQuery choose a model to optimize cost and quality. General-purpose functions include AI.GENERATE, AI.GENERATE_TEXT, AI.GENERATE_TABLE, AI.GENERATE_BOOL, AI.GENERATE_DOUBLE and AI.GENERATE_INT, while managed functions include AI.IF for filtering, AI.SCORE for rating, AI.CLASSIFY for classifying and AI.AGG for summarizing.
Last updated