Back to learning centerHome

Section 7

AI, ML, and GenAI

Review Vertex AI, Gemini, Agent Builder, Model Garden, AI Hypercomputer, pre-trained APIs, Document AI, Model Armor, and Sensitive Data Protection.

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.