Prepare and test your skills
Prepare and test your skills
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 on Google Cloud. 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.
Data quality and training data constraints describe whether useful data exists for training or tuning and whether that data fits the use case. Model selection starts by asking if a pre-trained traditional AI model can already meet the requirements before considering custom training. If the pre-trained model cannot meet requirements, the next question is whether enough training data exists to custom train a model. When sufficient data exists, the team chooses between high control from custom training and faster go-to-market (GTM) from generative AI with prompt engineering.
Traditional predictive AI works particularly well with structured data for prediction and classification. Generative AI is used for summarization, content generation, advanced transcription, and work that combines text, images, videos, or audio. Pre-trained traditional options include Document AI (opens in a new tab), Vision AI, Natural Language API, and Video Intelligence API. If no training dataset exists or available datasets are too small to custom train, generative AI with prompt engineering is used and can later be tuned with data examples.
Latency and serving requirements define how quickly an AI system must respond, especially for real-time use cases. Expected serving latency is one factor in a deeper model selection analysis that also considers the data, anticipated outcomes, specific model metrics, and other influential factors. The generative AI development flow supports deployment through online prediction for real-time requests and batch prediction for non-urgent work. When a use case requires an immediate answer, latency becomes a primary constraint on both model choice and deployment architecture.
Computational costs and deployment architecture describe the price of training and serving models and where those workloads run. Generative AI workloads can use Graphics Processing Units (GPUs) on Compute Engine to accelerate processing on virtual machine (VM) instances. Google Kubernetes Engine (GKE) (opens in a new tab) provides AI and machine learning (ML) orchestration with high performance, cost-effective serving and training, large scale, and flexible infrastructure options. Vertex AI provides foundation models, including Gemini (opens in a new tab), while third-party models can be deployed to Vertex AI Model Garden or self-hosted on GKE or Compute Engine.
These deployment paths let architects trade control, scale, and cost. Vertex AI reduces operational work for testing, tuning, and deploying foundation models. GKE and Compute Engine give more control over infrastructure but require more management of scaling and resources. The right choice depends on how much control the team needs and how much compute the model consumes.
The model selection decision flow applies data, control, and speed constraints in order, starting from the use case. The flow first checks pre-trained models, then checks data for custom training, then considers generative AI for speed and flexibility:
A decision tree starting from the use case that checks whether a pre-trained traditional AI model meets requirements, whether enough training data exists for custom training, and whether control or faster go-to-market is the priority, ending in a pre-trained model, custom-trained traditional model, or generative AI.
| 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 over model behavior |
| Custom-trained traditional AI model | Pre-trained model does not meet requirements and high control is the priority | Enough training data to custom train | More control over algorithm, loss functions, explainability, layers, learning rate, and hyperparameters; slower GTM |
| Generative AI with prompt engineering | No training dataset exists, datasets are too small to custom train, or faster GTM is the priority | Can start with prompts; can be tuned with data examples | Faster GTM and multimodal support; less control than custom training |
| Traditional predictive AI | Predictive use case with structured data | Particularly effective with structured data | Best fit for prediction with structured data |
| Generative AI | Summarization, content generation, advanced transcription, or multimodal input | Varies by use case | Handles text, images, videos, and audio; choice depends on latency and cost |
Exam tip: If a pre-trained traditional AI model meets the requirements, use it; custom training is for when the pre-trained model cannot meet requirements and sufficient training data exists.
Infrastructure readiness on Google Cloud means having the services and lifecycle tools needed to select, tune, deploy, and monitor the chosen model. Vertex AI provides access to Google foundation models, including Gemini, for testing, tuning, and deploying AI-powered applications. Vertex AI Model Garden lets teams discover, test, customize, and deploy select open-source models and assets. Third-party models can be deployed to Vertex AI Model Garden or self-hosted on GKE or Compute Engine with GPUs for acceleration.
The generative AI development flow has six stages in order:
Model selection, prompt engineering, tuning, and optimization form a looping evaluation subcycle where results are tested and refined before deployment. This lifecycle shows that readiness includes compute plus the tools for evaluation, deployment, and monitoring.
Combining traditional AI and generative AI means using both approaches together when neither approach alone satisfies the technical constraints. Output from a traditional AI model can become part of the prompt for a generative AI model. For example, traditional predictive AI can analyze historical data to forecast customer churn probability, and then a generative AI-powered chatbot can let a sales team explore those predictions through natural language conversation.
Another pattern uses traditional predictive AI to identify customer segments and generative AI to create personalized marketing content for each segment. In this combined architecture, traditional AI handles prediction or classification while generative AI handles generation, explanation, or conversation. Infrastructure readiness therefore means matching data readiness, latency needs, and computational cost to the right combination of models and deployment paths.
Analyzing business objectives and strategic alignment means working backward from measurable business outcomes to decide if a generative AI initiative is necessary and how urgent it is. Generative AI (opens in a new tab) and traditional AI should support business goals and should not exist in isolation. This business-value approach connects organizational goals, target use cases, and expected return to executive priorities and user needs before any technology is chosen.
Defining measurable business goals first means stating the specific goal and the value to be achieved before selecting technology. The team then checks that the proposed solution directly contributes to that goal and fits within business constraints. Common business goals that drive AI investment include:
Business constraints also shape which goals are realistic. Security and privacy requirements must meet specific industry compliances or country regulatory requirements. Those constraints can raise the priority of compliant solutions or rule out other use cases entirely.
Return on Investment (ROI) is one of the key measures used to judge AI project success and set its priority. ROI clarifies how the business plans to measure success for each identified goal and objective. ROI for generative AI can be measured through three related categories:
When expected ROI is strong and clearly linked to a priority goal, the initiative has strategic necessity and ranks higher for funding. When ROI is weak, unclear, or disconnected from executive priorities, the use case should be deprioritized even if the technology is interesting.
Aligning use cases with executive priorities and user value means connecting the best-fit technology use case to measurable business requirements, organizational priorities, and user expectations. Business leaders approve the selection when it aligns with business priorities. End users must understand the output and how to use it for more informed decision-making. The team also identifies the end users, how they will interact with the AI-powered application (opens in a new tab), and what changes to existing processes or workflows are needed for adoption.
Strategic value comes from three related paths. The business can achieve more outcomes at a faster pace with less operational complexity and effort, which often brings cost savings. The business can enhance customer or product experience through better service, personalization, and quality. The business can also create innovative value by improving existing offerings, enhancing creativity, and exploring new possibilities that set it apart from competitors.
Exam tip: Work backward from the desired business outcome to the technology choice, and use ROI to prove the link.
Regulatory, security, and ethical considerations are the business constraints that shape what a generative AI project is allowed to do, how its data is protected, and how its outputs are governed. These factors are identified early during the business goal and success criteria step because they determine the project's scope, the data it can use, and the review steps it must include. A project that addresses them early is more likely to stay within legal boundaries and produce outputs the organization can stand behind.
Regulatory and compliance requirements are the industry compliance rules and country-specific laws that a generative AI solution must satisfy. When a team defines measurable business goals, it must also identify business constraints such as ensuring that security and privacy aspects meet those specific requirements. These constraints set the boundaries for which data sources can be ingested, where that data can be stored, and which users can access it. A project that ignores these constraints risks failing compliance audits or violating data privacy laws, which can halt deployment entirely.
Security protocols are the technical controls that protect data as it moves through a generative AI system. For retrieval-augmented generation (RAG) (opens in a new tab) applications, private connectivity can keep network traffic within trusted boundaries by preventing data from traveling over the public internet. For auditing and threat detection, security log analytics (opens in a new tab) collects, exports, and analyzes logs from Google Cloud to help audit usage and detect threats to data and workloads. These controls apply to the data pipeline itself, where user data flows from source systems into storage and processing services and each hop must be protected by network and access controls.
Responsible AI and ethical principles are the guidelines that ensure generated content is safe, fair, and aligned with organizational standards. In a generative AI marketing architecture, for example, a human review step can be added before generated content is uploaded to Cloud Storage. That review verifies that the content aligns with safety guidelines and brand standards before public release. A feedback loop can also be built so the model learns from the impact of its outputs over time. These measures place a human decision point between the model's output and its release, which reduces the risk of harmful, biased, or off-brand content reaching end users.
Governance and stakeholder roles define who has authority over model behavior, who approves its use, and who is accountable for its outputs. Although data scientists and machine learning engineers commonly lead model selection, other stakeholders have defined roles:
Each role sits at a different point in the governance chain. Business leaders set the scope, product owners shape behavior, domain experts refine quality, and end users consume the results for decisions.
How these considerations shape project scope describes how legal, technical, and ethical limits combine to define what a project can and cannot do. Regulatory requirements limit which data can be used and where that data can flow. Security protocols protect that data at each stage of the pipeline from source systems to storage to processing. Responsible AI principles add review steps and feedback loops that gate the release of generated content. Governance roles assign accountability across all of these areas from scope-setting to daily use.
Exam tip: Regulatory constraints are identified during the business goal step, not after deployment — they shape the project's scope from the start.
Use a pre-trained traditional AI model when it meets the requirements for a classification or detection use case because it needs no custom training data and is the fastest path. Use a custom-trained traditional AI model when the pre-trained model does not meet requirements, high control is the priority, and enough training data exists to custom train. Use generative AI with prompt engineering when no training dataset exists, datasets are too small to custom train, or faster go-to-market is the priority, since it can start with prompts and later be tuned with data examples.
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, Google Kubernetes Engine (GKE), or Compute Engine.
Return on Investment (ROI) for generative AI is measured through direct financial gains, operational efficiency, and customer experience. Direct financial gains include increased revenue or reduced costs, operational efficiency includes faster time-to-market or quicker issue resolution, and customer experience includes increased satisfaction scores or improved retention. When expected ROI is strong and clearly linked to a priority goal, the initiative has strategic necessity and ranks higher for funding.
Private connectivity keeps retrieval-augmented generation (RAG) network traffic within trusted boundaries by preventing data from traveling over the public internet, while security log analytics collects, exports, and analyzes logs from Google Cloud to audit usage and detect threats. Regulatory requirements set boundaries for which data sources can be ingested, where data can be stored, and which users can access it, and they are identified during the business goal step. A human review step before generated content is uploaded to Cloud Storage verifies alignment with safety guidelines and brand standards before public release.
A financial services firm is preparing to deploy a generative AI assistant on Google Cloud to support representatives handling live customer inquiries. During the infrastructure readiness assessment, the architecture team must determine which technical constraints will dictate their model selection, resource provisioning, and deployment topology.
The deployment must satisfy three core conditions:
Which set of technical constraints and architectural factors directly governs these decisions on Google Cloud?
Last updated