Reference

Domain 3: Innovating with Google Cloud Artificial Intelligence (~16%)

Domain 3 of the Google Cloud Digital Leader exam covers artificial intelligence and machine learning concepts, Google Cloud's AI/ML product portfolio, and how to select the right solution for a given business problem. This domain accounts for approximately 8-10 questions on the exam.

The exam tests three areas within this domain:

  1. AI and ML fundamentals -- definitions, problem types, data quality, responsible AI
  2. Google Cloud's AI and ML solutions -- the product spectrum from pre-trained APIs to custom models
  3. Building and using AI/ML solutions -- choosing the right approach and understanding the ML workflow

This is a business-oriented exam. Questions emphasize when and why to use each product, not how to implement them. Expect scenario questions that describe a business problem and ask you to pick the correct Google Cloud service.

1. AI and ML Fundamentals

AI vs. ML vs. Deep Learning

These three terms form a nested hierarchy. The exam expects you to distinguish them clearly.

Term Definition Relationship
Artificial Intelligence (AI) Broad field of computer science focused on creating systems that can perform tasks that normally require human intelligence Umbrella term -- ML and deep learning are subsets
Machine Learning (ML) Subset of AI where systems learn patterns from data without being explicitly programmed for each task Uses algorithms that improve through experience (training data)
Deep Learning Subset of ML that uses artificial neural networks with multiple layers to learn complex patterns Requires large datasets and significant compute; powers image recognition, natural language processing, speech recognition

Exam trap: AI is not synonymous with ML. AI includes rule-based systems, expert systems, and robotics -- not all AI involves learning from data. When the exam says "ML," it specifically means systems that learn from data.

AI vs. Data Analytics vs. Business Intelligence

The exam tests whether you can distinguish AI/ML capabilities from traditional analytics.

Capability What It Does Example
Business Intelligence (BI) Reports on what happened using historical data; dashboards, visualizations, KPIs "Sales dropped 12% last quarter in the Northeast region"
Data Analytics Explores why something happened; statistical analysis, trend identification "Sales dropped because competitor launched a similar product at a lower price point"
Machine Learning Predicts what will happen or automates decisions; learns patterns from data to make predictions on new data "Based on current trends, sales will drop another 8% next quarter unless pricing changes"

Key distinction: BI and analytics look backward (descriptive/diagnostic). ML looks forward (predictive/prescriptive). The exam frequently presents a business scenario and asks whether it needs BI, analytics, or ML.

Types of ML Problems

Know these four problem types and be able to match them to business scenarios.

Problem Type What It Does Business Example
Classification Assigns input to a discrete category Spam detection (spam/not spam), image labeling (cat/dog), fraud detection (fraudulent/legitimate)
Regression Predicts a continuous numerical value Forecasting revenue, predicting housing prices, estimating delivery times
Clustering Groups similar data points together without predefined labels Customer segmentation, anomaly detection, document organization
Recommendation Suggests items based on user behavior or item similarity Product recommendations, content suggestions, ad targeting

Exam trap: Classification predicts categories, regression predicts numbers. If the question asks about predicting a customer's lifetime value (a dollar amount), that is regression. If it asks about predicting whether a customer will churn (yes/no), that is classification.

Supervised vs. Unsupervised vs. Reinforcement Learning

Learning Type Data Requirement How It Works Example Problem Types
Supervised Learning Labeled data (input-output pairs) Model learns the mapping from inputs to known correct outputs Classification, regression
Unsupervised Learning Unlabeled data Model discovers hidden patterns or groupings in data without guidance Clustering, dimensionality reduction, anomaly detection
Reinforcement Learning Environment with reward signals Agent learns by taking actions and receiving rewards or penalties Game playing, robotics, resource optimization

Exam trap: Supervised learning requires labeled training data -- someone must tag examples as "spam" or "not spam" before the model can learn. If the question says "the company has millions of customer records but no labels," the answer involves unsupervised learning (clustering), not supervised learning.

Data Quality for ML

The exam emphasizes that ML models are only as good as their training data. Key points:

  • Garbage in, garbage out: Inaccurate, incomplete, or biased data produces unreliable models
  • Volume: ML generally requires large datasets to learn meaningful patterns
  • Labeling accuracy: For supervised learning, mislabeled data directly degrades model performance
  • Bias in data: If training data reflects historical biases (gender, racial, geographic), the model will reproduce and potentially amplify those biases
  • Data freshness: Models trained on stale data may not reflect current conditions

Exam perspective: Questions often present a scenario where an ML model is performing poorly and ask for the most likely cause. Data quality is almost always the answer over model complexity or compute resources.

Business Value of ML

The exam focuses on three key value propositions, as stated in the official exam guide:

Value Proposition What It Means
Work with large datasets ML can process volumes of data that humans cannot manually analyze -- extracting patterns across millions of records
Scale business decisions Automate repetitive decision-making (loan approvals, content moderation, routing) so humans focus on exceptions
Unlock unstructured data ML can extract meaning from images, audio, video, and free-form text that traditional analytics cannot process

Responsible AI and Google's AI Principles

Google published AI Principles that govern how they develop and deploy AI. The exam tests awareness of responsible AI concepts.

Current framework (revised February 2025): Google updated its AI Principles to three core tenets:

  1. Bold Innovation -- push the boundaries of AI capability while maintaining safety
  2. Responsible Development and Deployment -- substantial benefits must outweigh foreseeable risks; risk-benefit assessment replaces categorical bans
  3. Collaborative Progress -- work with governments, academia, and civil society on AI governance

Historical context (2018-2025): Google's original framework listed seven objectives (socially beneficial, avoid bias, safety, accountability, privacy, scientific excellence, principled use) and four areas they would not pursue (overall harm, weapons, surveillance violating norms, violations of international law). The 2025 revision replaced categorical bans with a risk-benefit assessment model. The exam may reference either the original or updated principles.

Exam trap: The shift from "we will not pursue" to "benefits must outweigh risks" is significant. The 2025 framework is more permissive but adds structured risk assessment. Know that Google's approach evolved from rules-based to risk-based.

Key concepts for the exam:

Concept Definition
Explainable AI (XAI) The ability to understand and interpret how an ML model makes decisions; makes the "black box" transparent
Fairness Ensuring ML models do not discriminate against particular groups
Accountability Humans remain responsible for AI system outcomes; there must be clear ownership
Transparency Organizations should be open about how AI systems work and what data they use

2. Google Cloud's AI and ML Solutions

The AI/ML Product Spectrum

This is the single most important concept in Domain 3. Google Cloud offers a spectrum of AI/ML solutions that trade off between ease of use and customization.

Solution ML Expertise Required Customization Time to Deploy Cost When to Use
Pre-trained APIs None None (use as-is) Hours Lowest Standard tasks (vision, language, speech, translation) where Google's general models are sufficient
AutoML (via Vertex AI) Minimal Moderate (custom data, auto-tuned) Days to weeks Medium Custom classification or detection tasks with your own labeled data, but without ML expertise
BigQuery ML SQL knowledge Moderate (SQL-based training) Hours to days Medium ML on data already in BigQuery; data analysts who know SQL but not Python
Custom Models (via Vertex AI) High (data scientists) Full control Weeks to months Highest Unique business problems requiring specialized architectures, novel research, or maximum differentiation

Exam trap: The exam loves asking "which approach should the company use?" The decision factors from the official exam guide are: speed (time to deploy), effort (resources required), differentiation (competitive advantage), and required expertise (team skill level). If the scenario mentions "no ML expertise on staff," the answer is pre-trained APIs or AutoML, never custom models.

Pre-trained APIs

These are ready-to-use ML models accessed via REST API calls. No training required -- you send data in, get predictions back. The exam expects you to match each API to its use case.

API What It Does Use Cases
Natural Language API Analyzes text for sentiment, entities, syntax, and content classification Customer feedback analysis, content categorization, entity extraction from documents
Vision API Analyzes images for labels, text (OCR), faces, landmarks, logos, objects, and explicit content Product image cataloging, document scanning, content moderation, accessibility (image descriptions)
Cloud Translation API Translates text between 100+ languages Multilingual customer support, website localization, document translation
Speech-to-Text API Converts audio to text transcription Call center transcription, meeting notes, voice commands, accessibility
Text-to-Speech API Converts text to natural-sounding audio Voice assistants, audiobook generation, accessibility (screen readers)

Natural Language API features: Sentiment analysis (positive/negative score per sentence and document), entity analysis (named entities like people, places, organizations), entity sentiment analysis (sentiment per entity), syntax analysis (part-of-speech and dependency trees), and content classification (hierarchical topic categories).

Vision API features: Label detection (objects, scenes, activities), text detection/OCR (printed and handwritten text), face detection (locations and expressions), landmark detection (famous locations with coordinates), logo detection, object localization (multiple objects with bounding boxes), and SafeSearch detection (content moderation for explicit material).

AutoML (via Vertex AI)

AutoML automates the ML model-building process -- feature engineering, architecture search, hyperparameter tuning -- so users with minimal ML expertise can train custom models on their own data.

Key characteristics:

  • Trains models on your labeled data (unlike pre-trained APIs which use Google's data)
  • Automates model selection and tuning (no need to choose algorithms manually)
  • Supports image classification, object detection, text classification, tabular data, and video analysis
  • Accessed through the Vertex AI console -- AutoML is now part of Vertex AI, not a separate product
  • Produces models that can be deployed as endpoints for online predictions

When to choose AutoML over pre-trained APIs: Your data is domain-specific and Google's general models do not recognize your categories. Example: a retail company wants to classify product defects specific to their manufacturing process -- the Vision API cannot do this, but AutoML trained on their labeled defect images can.

Vertex AI

Vertex AI is Google Cloud's unified ML platform. It brings together AutoML and custom model training under one interface.

Key capabilities for the exam:

Feature What It Does
AutoML Training Code-free model training with automated architecture search
Custom Training Full-control training using TensorFlow, PyTorch, scikit-learn, or XGBoost in managed environments
Model Registry Central repository for trained models with versioning
Prediction Endpoints Deploy models for online (real-time) or batch predictions
Pipelines Orchestrate and automate ML workflows end-to-end
Feature Store Centralized repository for ML features, ensuring consistency between training and serving
Model Monitoring Detect data drift and model performance degradation in production
Vertex AI Studio Interface for experimenting with generative AI models (Gemini)

Exam trap: Vertex AI is the platform, not a model. It hosts both AutoML and custom models. If a question asks "where do you build custom ML models on Google Cloud?" the answer is Vertex AI. If it asks "how do you build ML models without coding?" the answer is AutoML on Vertex AI.

Gemini

Gemini is Google's flagship multimodal large language model (LLM) family, replacing the earlier Bard branding. Gemini is central to Google Cloud's generative AI strategy and is available through Vertex AI.

Tier Capability Use Case
Gemini Ultra Most capable model; complex reasoning, multimodal (text, images, audio, video, code) Enterprise AI applications requiring advanced reasoning
Gemini Pro Balanced performance and cost; strong general-purpose capabilities Most production workloads, chatbots, content generation
Gemini Nano Lightweight on-device model Mobile apps, edge devices, offline scenarios

Key exam points:

  • Gemini is accessed via Vertex AI Studio (experimentation) and Vertex AI API (production)
  • Gemini is multimodal -- it processes text, images, audio, video, and code in a single model
  • Gemini powers features across Google Workspace (Docs, Sheets, Gmail) and Google Cloud Console
  • For the exam, Gemini fits at the "pre-trained model" end of the AI spectrum -- no ML expertise required to use it

Exam trap: Gemini is a model family, not a platform. Vertex AI is the platform; Gemini runs on Vertex AI. If a question asks about Google's generative AI model, the answer is Gemini. If it asks where you access and manage it, the answer is Vertex AI.

BigQuery ML

BigQuery ML lets you create and run ML models using standard SQL queries directly inside BigQuery. Models are stored alongside your data in BigQuery datasets.

Supported model types:

Category Models
Regression & Classification Linear regression, logistic regression, boosted trees, random forest, deep neural networks
Clustering K-means
Recommendation Matrix factorization
Time Series ARIMA_PLUS for forecasting
Dimensionality Reduction Principal Component Analysis (PCA)

Key advantages:

  • SQL-based: Data analysts who know SQL can build ML models without learning Python or R
  • No data movement: Train on data already in BigQuery -- no ETL to a separate ML platform
  • Democratizes ML: Empowers data analysts (not just data scientists) to build predictive models

Exam trap: BigQuery ML is the answer when the scenario says "the team consists of SQL-proficient data analysts" or "the data is already in BigQuery." It is not the answer when the scenario requires complex custom architectures or computer vision.

TensorFlow

TensorFlow is Google's open-source end-to-end platform for building and training ML models. Key exam facts:

  • Open-source: Free to use, not a paid Google Cloud service
  • Framework: Used for building custom models, particularly deep learning
  • Relationship to Google Cloud: TensorFlow models can be trained on Google Cloud (Vertex AI, Compute Engine with GPUs/TPUs) and deployed via Vertex AI endpoints
  • Keras: High-level API within TensorFlow that simplifies model building
  • TensorFlow Lite: Optimized version for mobile and edge devices

Exam context: TensorFlow appears when the scenario involves data scientists building custom deep learning models. It represents the highest-expertise, highest-customization end of the spectrum.

Cloud TPU

Cloud TPUs (Tensor Processing Units) are Google's custom-designed ASICs (application-specific integrated circuits) optimized specifically for ML workloads.

Aspect Detail
What they are Custom hardware accelerators designed by Google, optimized for matrix operations
Best for Large-scale model training, models dominated by matrix computations, training jobs lasting weeks or months
Not suitable for Workloads with frequent branching, high-precision arithmetic, or many custom operations
vs. GPUs TPUs excel at pure matrix computation; GPUs are more flexible and better for models with many custom operations
Access methods Cloud TPU VMs, Google Kubernetes Engine (GKE), Vertex AI
Architecture 128x128 systolic array -- optimal performance when tensor dimensions are multiples of 128

Exam trap: TPUs are specialized hardware, not general-purpose. The exam may ask when to use TPUs vs. GPUs. Answer: TPUs for large-scale matrix-heavy training (especially NLP and recommendation models); GPUs for more diverse workloads or models with custom operations.

3. Building and Using AI/ML Solutions

Choosing the Right Approach

This decision framework is critical for the exam. Given a business scenario, you must select the right product.

Scenario Best Solution Why
"Detect sentiment in customer emails" Natural Language API Standard NLP task; pre-trained API handles it out of the box
"Classify product defects unique to our factory" AutoML (Vertex AI) Domain-specific categories require custom training; AutoML needs minimal ML expertise
"Predict quarterly sales from data in our data warehouse" BigQuery ML Data already in BigQuery; SQL-based forecasting with ARIMA_PLUS
"Build a recommendation engine with a novel architecture" Custom model on Vertex AI Requires full control over model architecture; data science team needed
"Transcribe customer support phone calls" Speech-to-Text API Standard transcription task; pre-trained API
"Translate product listings into 50 languages" Cloud Translation API Standard translation task; pre-trained API
"Identify objects in satellite imagery for a defense contractor" Custom model on Vertex AI Specialized domain; pre-trained Vision API will not recognize military objects
"An SQL analyst wants to segment customers" BigQuery ML (K-means) SQL-based clustering; no Python needed; data in BigQuery

The ML Workflow

The exam tests awareness of the end-to-end ML lifecycle, not implementation details.

Phase What Happens Google Cloud Tools
Data Preparation Collect, clean, label, and split data into training/validation/test sets BigQuery, Dataflow, Dataprep, Cloud Storage
Feature Engineering Transform raw data into meaningful inputs (features) for the model BigQuery, Vertex AI Feature Store
Model Training Algorithm learns patterns from training data Vertex AI (AutoML or custom), BigQuery ML
Model Evaluation Assess model performance on held-out test data using metrics (accuracy, precision, recall, F1) Vertex AI evaluation tools
Model Deployment Make the trained model available for predictions Vertex AI endpoints (online or batch)
Model Monitoring Track prediction quality over time; detect data drift or performance degradation Vertex AI Model Monitoring

MLOps Concepts

MLOps applies DevOps principles to ML systems. The exam tests basic awareness.

Concept What It Means
Continuous Training Automatically retrain models when new data arrives or performance degrades
Model Versioning Track different versions of a model; roll back if a new version underperforms
Data Drift Input data distribution changes over time, causing model accuracy to degrade
Model Monitoring Ongoing measurement of prediction quality, latency, and resource usage in production
Feature Store Centralized, reusable repository of ML features ensuring consistency between training and serving
Pipeline Orchestration Automate the end-to-end ML workflow (data prep through deployment)

Cost and Complexity Tradeoffs

Factor Pre-trained APIs AutoML BigQuery ML Custom Models
Cost Pay per API call Training + prediction costs BigQuery compute costs Highest (compute, storage, expertise)
Time to value Hours Days to weeks Hours to days Weeks to months
Team expertise Developers Domain experts SQL analysts Data scientists / ML engineers
Maintenance Google maintains the model Retraining on new data SQL-based retraining Full lifecycle management
Differentiation None (same model for all) Moderate (your data) Moderate (your data) Maximum (your architecture + data)

Exam Tips for Domain 3

  1. Match the scenario to the product: If the scenario describes a standard task (sentiment, OCR, translation, transcription), the answer is a pre-trained API. If it requires custom categories on the company's own data, AutoML. If the data is in BigQuery and the team knows SQL, BigQuery ML. If the team has data scientists and needs full control, custom training on Vertex AI.

  2. "No ML expertise" in the scenario always eliminates custom models. The answer is pre-trained APIs (for standard tasks) or AutoML (for custom tasks).

  3. Data quality is always the first thing to fix. If a model is underperforming, the exam expects you to look at the data before looking at the algorithm or compute.

  4. Responsible AI questions focus on bias in training data, explainability of model decisions, and Google's commitment to not causing harm. Know that Google has published AI Principles and that Explainable AI tools exist.

  5. Vertex AI is the platform, not a model. It hosts AutoML, custom training, model deployment, and monitoring. Do not confuse it with a specific model type.

  6. BigQuery ML democratizes ML for SQL users. It does not replace Vertex AI for complex use cases -- it supplements it for data warehouse-centric analytics.

  7. TPUs are for large-scale training, not general compute. If the question mentions training a large model for weeks, TPUs are relevant. For standard workloads, GPUs or managed services (AutoML) are more appropriate.

References