Google Cloud PCA - Section 7
AI, ML, and GenAI
Goal: recognize which Google Cloud AI service or architecture pattern fits the business requirement without going deep into data science implementation details.
PCA framing: This section is mostly about choosing the right AI pattern. Use managed Gemini/Agent Platform when the org wants enterprise GenAI or agents, pre-trained APIs when the org does not need custom model training, Vertex AI/Agent Platform ML lifecycle tools when the org needs custom training or governed ML operations, and AI Hypercomputer/accelerators when the requirement is large-scale training or serving infrastructure. |
Fast Decision Matrix
Requirement | Choose | PCA Reasoning |
|---|---|---|
Build enterprise AI agents with governance, tools, and production controls | Gemini Enterprise Agent Platform / Agent Builder | Use when the org needs agent development, deployment, governance, observability, and access to models/tools at enterprise scale. |
Give business users an enterprise AI assistant over company content | Gemini Enterprise app | Use for knowledge work, enterprise search, productivity, and agent access without every team building custom ML apps. |
Use a frontier Google model for text, image, code, or multimodal prompts | Gemini models | Use when the business need is GenAI inference rather than training a model from scratch. |
Evaluate or deploy Google, partner, or open models | Model Garden | Use when the team needs a catalog of available models and deployment/customization options. |
Build a low-code/no-code conversational or search agent | Agent Builder / Vertex AI Search / Conversation | Use when the org needs an AI app over websites, documents, or enterprise data with less custom code. |
Automate ML training, evaluation, and deployment workflows | Vertex AI / Agent Platform Pipelines | Use for repeatable ML workflows, governed MLOps, retraining, and batch pipeline orchestration. |
Run very large-scale AI training/tuning/serving | AI Hypercomputer | Use for optimized GPU/TPU clusters, high-performance networking/storage, and flexible consumption for AI workloads. |
Need acceleration for training or inference | GPUs / TPUs | Use accelerators when CPUs cannot meet model training or inference latency/cost needs. |
Extract text, labels, or objects from images | Vision AI / Cloud Vision API | Use pre-trained vision instead of building an image recognition model. |
Extract entities, sentiment, or meaning from text | Natural Language AI | Use pre-trained NLP for classification, sentiment, and entity analysis. |
Transcribe audio into text | Speech-to-Text | Use for call center recordings, captions, voice input, and transcription pipelines. |
Turn text into spoken audio | Text-to-Speech | Use for accessibility, IVR, voice assistants, and generated speech. |
Translate content across languages | Translation AI | Use for multilingual apps, localization, and document/text translation. |
Extract structured data from documents | Document AI | Use for invoices, forms, contracts, IDs, claims, and document processing. |
Protect prompts/responses from unsafe content or prompt injection | Model Armor | Use as an AI security guardrail around Gemini/Vertex/Gemini Enterprise workloads. |
Find, classify, redact, or de-identify sensitive data used in AI | Sensitive Data Protection | Use before training/tuning/inference workflows when PII or regulated data may be present. |
Service Deep Dives
Service / Topic | Why an org would use it | PCA Exam Trigger | Common trap / wrong answer |
|---|---|---|---|
Gemini Enterprise Agent Platform | An org uses this to build, deploy, govern, and optimize production AI agents and GenAI applications on Google Cloud. | Questions mention enterprise agents, governed GenAI apps, model/tool access, production deployment, observability, or agent lifecycle. | Do not pick raw APIs or self-managed LLM hosting when the question wants enterprise agent governance and managed platform capabilities. |
Vertex AI | An org uses Vertex AI/Agent Platform capabilities when it needs a managed ML/AI platform for custom models, GenAI, deployment, training, tuning, and operations. | Questions mention managed ML lifecycle, custom training, endpoints, batch prediction, tuning, model monitoring, or MLOps. | Do not confuse with BigQuery ML for SQL-based analytics modeling or pre-trained APIs for simple packaged use cases. |
Gemini Enterprise app | An org uses this when business users need a governed AI workspace or assistant connected to enterprise knowledge, agents, and tools. | Questions mention enterprise productivity, company knowledge search, user-facing AI assistant, or broad workforce AI enablement. | Not the same as building custom model infrastructure. It is closer to a managed enterprise AI experience. |
Gemini models / Gemini LLMs | An org uses Gemini models for text generation, summarization, coding, multimodal reasoning, document understanding, and image/audio/video-aware experiences. | Questions mention GenAI, multimodal prompts, summarization, natural language answers, or generating content. | Do not train a custom model unless the question requires domain-specific model training beyond prompt/RAG/tuning patterns. |
Agent Builder / AI Agents | An org uses Agent Builder when it wants to create conversational agents, search agents, or task agents with lower engineering overhead. | Questions mention building agents over websites, docs, enterprise sources, workflows, or user support experiences. | Do not pick Cloud Functions/Workflows alone when the main requirement is an AI agent experience. |
Model Garden | An org uses Model Garden to discover, test, customize, and deploy Google, partner, and open models from a central catalog. | Questions mention comparing available models, using prebuilt/foundation models, or deploying a model without starting from scratch. | Do not confuse model catalog selection with model training pipelines or serving infrastructure. |
Vertex AI / Agent Platform Pipelines | An org uses pipelines to automate repeatable ML workflows such as preprocessing, training, evaluation, deployment, and retraining. | Questions mention MLOps, repeatability, scheduled ML workflows, Kubeflow Pipelines, TensorFlow Extended, lineage, or governed training flows. | Do not use Cloud Composer by default if the workflow is specifically an ML pipeline managed in the AI platform. |
AI Hypercomputer | An org uses AI Hypercomputer for large AI workloads needing optimized accelerators, networking, storage, cluster scheduling, and consumption choices. | Questions mention very large model training/tuning/serving, GPU/TPU cluster efficiency, goodput, or specialized AI infrastructure. | Overkill for ordinary inference, basic AutoML, or small app-level AI features. |
GPUs / TPUs for ML | An org uses accelerators when training or inference requires parallel compute for deep learning or large model serving. | Questions mention model training time, inference latency, accelerator capacity, GPU/TPU choice, or cost/performance for ML workloads. | Do not attach accelerators to every workload. They add cost and operational complexity. |
Pre-trained ML APIs | An org uses pre-trained APIs when it needs AI capability quickly and does not need to build or maintain a custom model. | Questions mention vision, speech, translation, natural language, or document extraction as packaged capabilities. | Do not build a custom model when the question says standard recognition, translation, sentiment, or transcription is enough. |
Vision AI / Cloud Vision API | An org uses Vision AI to detect labels, objects, faces, logos, landmarks, OCR text, or unsafe content in images. | Questions mention image analysis, OCR from images, object detection, or image content classification. | Do not confuse with Document AI when the input is structured documents like invoices and forms. |
Natural Language AI | An org uses Natural Language AI to extract entities, classify text, and analyze sentiment from unstructured text. | Questions mention sentiment analysis, entity extraction, content classification, or text insights. | Do not pick Gemini by default when deterministic packaged NLP is sufficient and cheaper/simpler. |
Speech-to-Text | An org uses Speech-to-Text to transcribe audio into text for captions, call analytics, voice input, and recordings. | Questions mention transcription, voice recognition, captions, or converting audio recordings to searchable text. | Do not confuse with Text-to-Speech, which goes the opposite direction. |
Text-to-Speech | An org uses Text-to-Speech to generate natural-sounding audio from text. | Questions mention voice output, accessibility, IVR, narration, or spoken responses. | Do not confuse with Speech-to-Text. |
Translation AI | An org uses Translation AI to translate app content, documents, support messages, or multilingual text. | Questions mention localization, multilingual apps, or real-time translation. | Do not train custom NLP unless the translation requirement is highly domain-specific and cannot be handled with glossary/customization options. |
Document AI | An org uses Document AI to parse, classify, and extract structured fields from documents. | Questions mention invoices, forms, contracts, IDs, claims, receipts, or document workflow automation. | Do not pick Vision OCR alone when the task requires structured document extraction and field parsing. |
Video AI / video analysis | An org uses video intelligence capabilities to analyze video content for labels, shots, explicit content, objects, or searchable metadata. | Questions mention extracting insight from video libraries, media moderation, or video indexing. | Do not use standard Vision image APIs for full video analysis unless the question is about individual frames. |
Imagen / image generation APIs | An org uses image generation when it needs to create or edit images from prompts under an enterprise AI platform. | Questions mention generating product imagery, marketing images, or visual content using AI. | Do not confuse image generation with Vision AI, which analyzes existing images. |
NotebookLM Enterprise | An org uses NotebookLM Enterprise when users need to summarize, query, and synthesize dense sources in a compliant enterprise product. | Questions mention AI-powered research, source-grounded summaries, notebooks over PDFs/Docs/Slides, or enterprise-ready NotebookLM. | Do not treat personal NotebookLM/NotebookLM Plus as the same thing as the Google Cloud enterprise product. |
Model Armor | An org uses Model Armor to inspect or block unsafe prompts and responses, including prompt-injection and safety-related risks. | Questions mention AI security guardrails, prompt/response filtering, unsafe content controls, or protecting Gemini workloads. | Do not solve prompt injection only with IAM or network controls. Those do not inspect prompt content. |
Sensitive Data Protection for AI | An org uses Sensitive Data Protection to discover, classify, redact, and de-identify sensitive data before or during AI workflows. | Questions mention PII, regulated data, data profiling, de-identification, or safe use of training/tuning datasets. | Do not rely only on encryption if the real requirement is identifying or removing sensitive content from data. |
Secure model deployment | An org uses secure deployment patterns to control model access, isolate data, enforce IAM, log usage, monitor models, and apply guardrails. | Questions mention deploying AI safely in production, private endpoints, auditability, governance, or compliance. | Do not focus only on model accuracy. PCA often tests governance, access, monitoring, and data protection. |
Data integration for AI/ML | An org uses data integration patterns to connect Cloud Storage, BigQuery, Pub/Sub, Dataflow, Dataplex, and databases into AI workflows. | Questions mention feeding models, RAG over enterprise data, streaming inputs, feature data, or data governance for ML. | Do not choose an AI service without thinking about where the data lives, how it is secured, and how it updates. |
Consumption models for AI | An org chooses consumption models based on workload duration, capacity needs, cost predictability, and performance requirements. | Questions mention reserved capacity, bursty inference, long training jobs, cost control, or predictable accelerator needs. | Do not pick the biggest accelerator option without considering cost, availability, duration, and utilization. |
Common PCA Comparison Patterns
Comparison | How to decide |
|---|---|
Pre-trained API vs custom model | Use pre-trained APIs when the use case is common and the org wants speed. Use custom model training/tuning when the requirement is domain-specific or the packaged API does not meet accuracy/business needs. |
Gemini prompt/RAG vs training a model | Use prompt engineering or retrieval-augmented generation first when the model can answer from enterprise data. Train/tune only when behavior or domain adaptation requires it. |
Gemini Enterprise app vs Agent Platform | Use Gemini Enterprise app for end-user enterprise AI experiences. Use Agent Platform when builders need to develop, deploy, and govern production agents. |
Agent Builder vs Workflows/Cloud Tasks | Use Agent Builder for AI-driven conversational/search/task agents. Use Workflows/Tasks for deterministic orchestration without AI reasoning. |
Document AI vs Vision OCR | Use Document AI for structured document extraction. Use Vision OCR when the need is simple text extraction from images. |
BigQuery ML vs Vertex AI/Agent Platform | Use BigQuery ML when analysts want ML directly in BigQuery with SQL. Use Vertex AI/Agent Platform for broader custom ML lifecycle, endpoints, and MLOps. |
Cloud Composer vs Vertex AI Pipelines | Use Composer for general DAG orchestration. Use Vertex AI/Agent Platform Pipelines for ML pipeline workflows and lifecycle management. |
AI Hypercomputer vs ordinary GPU VM | Use AI Hypercomputer for large-scale AI infrastructure optimization. Use ordinary GPU/TPU VMs for smaller or more direct accelerator workloads. |
Model Armor vs Sensitive Data Protection | Model Armor screens prompts/responses for AI safety and prompt risks. Sensitive Data Protection discovers/classifies/redacts sensitive data like PII. |
IAM/network security vs AI guardrails | IAM and private networking control who can call services and from where. Model Armor and data protection address what content flows through the AI system. |
Architecture Patterns PCA Likes
Pattern | Typical architecture logic |
|---|---|
RAG over enterprise data | Use Gemini/Agent Platform with search or retrieval over approved enterprise sources. Secure the data sources, use IAM, audit logs, and data protection. |
Event-driven AI processing | Use Pub/Sub/Eventarc/Dataflow/Cloud Run with AI APIs to process images, audio, documents, or text as new files/events arrive. |
Governed ML lifecycle | Use Vertex AI/Agent Platform tools for training, pipelines, registry/deployment, monitoring, and repeatable operations. |
AI document automation | Use Cloud Storage ingestion, Document AI extraction, Dataflow/Cloud Run processing, and BigQuery/Cloud SQL output depending on analytics vs operational needs. |
Enterprise AI security | Use IAM, service accounts, VPC-SC where applicable, audit logs, Model Armor, Sensitive Data Protection, and least privilege. |
Large-model infrastructure | Use AI Hypercomputer, GPUs/TPUs, optimized storage, and capacity planning when the workload is large enough to justify specialized infrastructure. |
High-Value Exam Traps
- If the question says the org does not have ML expertise and needs common AI capability quickly, pick a pre-trained API or managed GenAI product, not custom training.
- If the question says enterprise users need to search and summarize internal knowledge, think Gemini Enterprise / Agent Builder / RAG before building a custom model from scratch.
- If the question says invoices, forms, contracts, or structured fields, think Document AI before generic OCR.
- If the question says prompt injection, unsafe prompts, or response filtering, think Model Armor. IAM alone does not inspect prompts.
- If the question says PII or regulated data inside training/tuning data, think Sensitive Data Protection plus governance before model training.
- If the question says huge AI training or serving with accelerator efficiency, think AI Hypercomputer, GPUs, TPUs, storage, and capacity planning.
- If the question says repeatable ML workflow, retraining, or Kubeflow/TensorFlow Extended, think Vertex AI/Agent Platform Pipelines.
- If the question says translate, transcribe, synthesize speech, classify text, or analyze images, do not over-engineer with a custom model unless there is a stated accuracy/domain reason.
Cram Summary
Service / Topic | One-line memory hook |
|---|---|
Gemini Enterprise Agent Platform | Build and govern enterprise agents and GenAI apps. |
Gemini Enterprise app | Enterprise AI assistant/workspace for business users. |
Gemini models | Frontier multimodal GenAI models for text, code, image, audio/video-aware use cases. |
Model Garden | Find, compare, customize, and deploy models. |
Agent Builder | Low-code/no-code AI agents and enterprise search/conversation experiences. |
Vertex AI / Agent Platform Pipelines | Automated ML workflows and MLOps. |
AI Hypercomputer | Optimized AI infrastructure for large training/tuning/serving. |
GPUs / TPUs | Accelerators for ML training and inference. |
Vision AI | Analyze images and OCR image text. |
Natural Language AI | Entity, sentiment, and text classification. |
Speech-to-Text | Audio to text. |
Text-to-Speech | Text to audio. |
Translation AI | Translate text/documents across languages. |
Document AI | Extract structured data from documents. |
NotebookLM Enterprise | Enterprise source-grounded research and summaries. |
Model Armor | Prompt/response AI security guardrails. |
Sensitive Data Protection | Discover, classify, redact, and de-identify sensitive data. |
Official docs checked
This section was built against Google Cloud PCA guidance and current Google Cloud docs for Gemini Enterprise Agent Platform, Vertex AI/Agent Platform, Model Garden, Agent Builder, Vertex AI/Agent Platform Pipelines, AI Hypercomputer, Model Armor, Sensitive Data Protection, NotebookLM Enterprise, and Google AI APIs including Vision, Natural Language, Speech-to-Text, Text-to-Speech, Translation, and Document AI.