Prepare and test your skills
Prepare and test your skills
Assessing build versus buy strategies is the decision between adopting a ready-made pretrained service and building a custom model for a specific business 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). The right choice depends on required customization, speed to market, and how much operational work the team can manage.
A decision tree that branches on whether generalized categories fit to choose pretrained Buy services, or custom labels and training data require a Vertex AI Build solution.
Buying means using pretrained models that work through an API without training data or model management. These services cover common tasks with generalized capabilities:
TEXT_DETECTION or DOCUMENT_TEXT_DETECTION enums.AI.IF, AI.SCORE, AI.CLASSIFY, and AI.AGG use streamlined syntax and are optimized for cost and quality.Because these services arrive already trained, traffic flows from the application to the Google-hosted model and back, and model upkeep stays with Google, which shortens time to market.
Building a custom solution means training or tuning a model so it matches company labels, entities, or behaviors that pretrained models do not cover. Vertex AI brings AutoML and AI Platform together into a unified API, client library, and platform for this custom work. Within that platform, AutoML Natural Language classification creates a model that classifies content with custom category labels, and AutoML entity extraction builds a model that identifies entities specific to a company or use case. Vertex AI also supports tuning and custom training, where a pretrained embedding model can be fine-tuned on a specific dataset, supervised tuning jobs can be started, retrieved, or canceled, and custom training pipelines can be created for custom jobs.
The trade-off between buying and building centers on speed, cost, customization, and accuracy risk. Off-the-shelf services are fast because they require no training step, while custom work is slower because it requires data collection, training or tuning jobs, and evaluation. Managed functions shift upkeep to Google and are optimized for cost and quality, while a build team pays for training, evaluation, and ongoing management in exchange for full control over model, prompt, parameters, and custom labels.
| 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 |
| Accuracy risk | Strong for general tasks but can underperform on niche data | Can match niche needs only if training data is suitable |
A custom model is not automatically better. AutoML Vision classification is generally not recommended for document classification because its accuracy tends to be less than text-based methods, though exceptions exist, so the model type must fit the data type.
Operational requirements for a build approach include project setup, API enablement, and Identity and Access Management (IAM) before any model runs. The team must create a project with the Project Creator role (roles/resourcemanager.projectCreator), enable billing, and enable required APIs such as the Vertex AI API, Cloud SQL (opens in a new tab) Admin API, and Compute Engine API. Access between services is granted through IAM, for example by creating a custom role with the aiplatform.endpoints.predict permission and granting it to a service account so one service can call a Vertex AI endpoint. These policy changes generally take effect within 2 minutes but can take up to 7 minutes to propagate fully.
Exam tip: Choose a pretrained or managed service when generalized categories or standard processors 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.
Aligning solutions with enterprise security and compliance means choosing a generative AI solution that meets the business need while respecting policies for privacy, residency, and ethical use. This balance keeps AI work close to governed data and uses the right level of control over models and infrastructure. BigQuery (opens in a new tab) supports this balance by letting teams run machine learning and generative AI directly where the data already lives.
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 through the console, bq command-line tool, REST API, or Colab Enterprise notebooks. Google Cloud also provides a secure data environments architecture that secures sensitive datasets against accidental exposure and malicious exfiltration. A borderless open data lakehouse pattern establishes a highly governed, secure pipeline that flows from raw silos to AI-driven actions, so governance applies from storage through analysis.
Choosing the right BigQuery AI function means matching control, cost, and transparency to the task and to governance rules. BigQuery AI functions accept natural language inputs and use pretrained Gemini Enterprise Agent Platform models, partner large language model (LLM) models, Cloud AI APIs, or built-in BigQuery models. General-purpose functions give full control and transparency over model, prompt, and parameters, while managed functions use streamlined syntax and let BigQuery choose a model to optimize for cost and quality.
| Function category | Level of control | Choose when |
|---|---|---|
| General-purpose AI functions | Full control over model, prompt, and parameters | Content generation, analysis, summarization, extraction, classification, or embeddings need transparency and flexibility |
| Managed AI functions | Streamlined syntax with model chosen for cost and quality | Routine filtering, rating, classification, or aggregation must run at large scale |
| Task-specific functions | Uses Cloud AI APIs for a defined task | The need is natural language processing, translation, document processing, audio transcription, or computer vision |
General-purpose functions include AI.GENERATE for flexible inference on structured or unstructured data, AI.GENERATE_TEXT as a table-valued version that also supports partner and open models, and structured-output options such as AI.GENERATE_TABLE, AI.GENERATE_BOOL, AI.GENERATE_DOUBLE, and AI.GENERATE_INT. Managed functions include AI.IF for filtering with natural language conditions, AI.SCORE for rating by quality or sentiment, AI.CLASSIFY for classifying into user-defined categories, and AI.AGG for summarizing or analyzing data. With optimized mode in Preview, managed functions scale to millions or billions of rows.
Exam tip: Choose general-purpose AI functions when governance requires full control over model, prompt, and parameters, and choose managed AI functions when BigQuery can select the model to optimize cost and quality.
Using architecture guidance and operational controls helps teams select models and infrastructure that remain feasible and compliant in production. Google Cloud organizes guidance into four types that support different stages of selection and implementation:
Operational review also covers locations, limitations, pricing, token usage, cost tracking, monitoring, and error management. The AI.COUNT_TOKENS utility function estimates tokens in an input prompt before running a query, which helps forecast cost and stay within limits. Checking edition availability, location support, and monitoring controls before implementation ensures the solution respects residency needs and can be operated reliably.
Selecting models via Vertex AI Model Garden (opens in a new tab) 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.
Foundation model types are grouped by the kind of work they do and the shape of output they produce. 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. Beyond Gemini, the platform also supports partner models and open models, which give alternatives when a specific task, license, or cost profile fits better. Models are also chosen by output shape:
The data flows from the prompt to the selected model type, and the model returns output in the shape the application expects.
Matching control, cost, and latency means choosing how much tuning power the team needs versus how much simplicity and scale it wants. General-purpose AI functions give full control and transparency over model, prompt, and parameters for teams that need to tune behavior. Managed AI functions use streamlined syntax and are optimized for cost and quality for teams that want good results without tuning every setting. Cost and latency both grow with how much text the model processes, so the AI.COUNT_TOKENS utility estimates tokens before the query runs to predict cost and response time. Teams then track token usage, track costs, and set up monitoring and error management so model choice stays aligned with budget and reliability targets.
Comparing model options requires looking at five considerations together, because a model that is cheap but inaccurate or powerful but too slow still fails the business need. Performance decides whether the model handles the specific task well enough to be worth its price. Latency decides whether interactive use gets a fast response or batch work can tolerate slower processing. Cost reflects token usage and price per model, where larger or more capable models usually cost more per request. Control reflects the choice between general-purpose functions for tuning and managed functions for cost and quality out of the box. Infrastructure reflects the compute, storage, and networking behind the model, which directly shape performance, cost, and scalability.
Exam tip: General-purpose AI functions give full control over model, prompt, and parameters, while managed AI functions are optimized for cost and quality with simpler syntax.
Why model selection is business-driven is that the right model balances innovation against cost-efficiency and operational feasibility. The choice should start from the business objective and work backward to model type, level of control, and supporting infrastructure. A model that is powerful but too slow or too expensive for the workload fails just as surely as one that is cheap but inaccurate, so technical comparison must serve the business outcome.
Choosing the right generative AI solution starts from the business objective, then narrows through governance, build-versus-buy, and model selection. Governance decides where data can live and how much model control is required, build-versus-buy decides whether pretrained services are enough or custom Vertex AI work is needed, and Model Garden selection tunes performance, latency, cost, and infrastructure to the workload. Together these decisions keep innovation, cost-efficiency, and operational feasibility in balance.
AI.COUNT_TOKENS utility function estimates tokens in an input prompt before running a query to help forecast cost and response time.Buying means using pretrained models called through an API without training data or model management, while building means training or tuning a model on company data with Vertex AI for custom labels, entities, or behaviors. Off-the-shelf services are fast with model upkeep staying with Google, while custom work is slower because it requires data collection, training or tuning jobs, and evaluation in exchange for full control over model, prompt, parameters, and labels.
Use general-purpose BigQuery AI functions when content generation, analysis, summarization, extraction, classification, or embeddings need transparency and flexibility with full control over model, prompt, and parameters. Use managed BigQuery AI functions when routine filtering, rating, classification, or aggregation must run at large scale with streamlined syntax where BigQuery chooses a model to optimize for cost and quality.
BigQuery ML lets SQL practitioners train, evaluate, and run inference with GoogleSQL queries inside BigQuery so forecasting, classification, and recommendations can run where the data already lives without requiring data movement. This approach reduces unnecessary data movement and keeps sensitive data under existing controls.
An enterprise is deciding between purchasing an off-the-shelf Software-as-a-Service (SaaS) application and building a custom generative AI solution using Google Cloud Vertex AI to handle specialized customer workflows.
Which statement accurately describes a key strategic trade-off when choosing to build a custom solution on Vertex AI instead of buying a SaaS product?
A custom build on Vertex AI enables deep customization, model fine-tuning, and direct integration with proprietary systems, but requires higher engineering effort and initial investment.
A custom build on Vertex AI provides the fastest time-to-market with zero software development overhead or maintenance responsibility.
A custom build on Vertex AI locks the organization into fixed vendor feature roadmaps with no ability to modify model parameters or search behavior.
A custom build on Vertex AI eliminates the need for Identity and Access Management (IAM) permissions and internal security monitoring.
Last updated