AIAI Agent Dev Labaiagentdev-lab.com
Sign in to save progressSign in
Menu

STEP 1 · Study

GAIL StudyGenerative AI Leader

Read the key points for each exam domain, then use “Practice this domain” to answer questions from that domain only. Domains with a larger weight appear more often on the exam.

01

Section 1: Fundamentals of generative AI

約30%

Learning goals

  • Explain the relationship among AI, ML, generative AI, foundation models and LLMs
  • Explain the stages of the ML lifecycle and the criteria for choosing a foundation model that fits the business
  • Explain data types and quality viewpoints, the five layers of generative AI, and the differences among Gemini, Gemma, Imagen and Veo

Key points

  • AI ⊃ machine learning ⊃ deep learning ⊃ foundation models (LLMs etc.). Generative AI is AI that creates new text, images, code, audio and video.
  • Foundation models are large models pretrained on vast data that can adapt to diverse tasks. Multimodal means handling several data types such as text, images, audio and video.
  • Diffusion models generate images and video by learning a noise-removal process.
  • Prompt engineering is crafting the input instructions; prompt tuning adapts a model by training a small number of parameters (soft prompts).
  • ML lifecycle: data ingestion → preparation → training → deployment → management (monitoring, retraining). On Google Cloud you use BigQuery, Cloud Storage, Agent Platform (formerly Vertex AI) and more.
  • Criteria for choosing a foundation model: modality, context window (how much input it handles at once), security, availability and reliability, cost, performance, ease of customization.
  • Structured data is tabular (sales, customer DBs); unstructured data is documents, images, audio and video. Labeled data has answers; unlabeled data doesn't.
  • Layers of generative AI: infrastructure (TPU, GPU) → models (Gemini etc.) → platform (Agent Platform) → agents → applications (Gemini app, Workspace etc.).
  • Gemini is the flagship multimodal model, Gemma a lightweight open model, Imagen image generation, and Veo video generation.

Workplace uses

  • SalesMap where generative AI fits by splitting work into proposal writing (create), meeting-note summaries (summarize), internal document search (discover) and auto-drafting quotes (automate).

Exam objectives (from the exam guide)

  • 生成 AI の主要な概念（AI・自然言語処理・機械学習・生成 AI・基盤モデル・マルチモーダル・拡散モデル・LLM・プロンプトエンジニアリング・プロンプトチューニング）
  • 機械学習の手法（教師あり・教師なし・強化学習）と ML のライフサイクル（データの取り込み・準備・学習・デプロイ・管理）と各段階の Google Cloud のツール
  • 業務に合う基盤モデルの選び方（モダリティ、コンテキストウィンドウ、セキュリティ、可用性、コスト、性能、ファインチューニング）と、作成・要約・発見・自動化のユースケース
  • データの種類（構造化・非構造化、ラベルあり・なし）とデータの品質・アクセス性（完全性・一貫性・関連性・可用性・コスト・形式）
  • 生成 AI の層（インフラ・モデル・プラットフォーム・エージェント・アプリケーション）
  • Google の基盤モデル（Gemini、Gemma、Imagen、Veo）の用途と強み

Learning resources

Learn in detail (6 topics with mini quizzes) →Practice this domain (90 questions) →Next domain ↓
02

Section 2: Google Cloud's generative AI offerings

約35%

Learning goals

  • Explain Google Cloud's generative AI strengths and infrastructure elements
  • Choose the Google offering that fits a business problem (Gemini app, Gemini Enterprise, Workspace, Customer Engagement Suite, Agent Platform)
  • Choose agent tool types and pretrained APIs that fit the goal

Key points

  • Enterprise-ready: customer data isn't used to train models; security, privacy and governance; reliability and scalability.
  • AI Hypercomputer is AI infrastructure integrating TPUs, GPUs, storage, networking and software. TPUs are Google-designed processors for AI.
  • Open approach: Model Garden lets you pick Google, partner and open models (Gemma, Llama etc.), avoiding lock-in.
  • Gemini app is conversational AI for personal productivity. Gems are custom Gemini instances with saved instructions for specific uses.
  • Gemini Enterprise is a business platform that searches and summarizes across internal data and builds agents to automate work.
  • Gemini for Google Workspace is generative AI built into Gmail, Docs, Sheets, Slides, Meet and more.
  • Customer Engagement Suite: Conversational Agents (chat and voice agents for customers), Agent Assist (real-time support for operators), Conversational Insights (conversation analytics), CCaaS (cloud contact center).
  • Agent Platform (Gemini Enterprise Agent Platform, formerly Vertex AI) is a unified developer platform for choosing, customizing, evaluating and deploying models and building agents. Agent Search (formerly Vertex AI Search) handles enterprise data search and RAG.
  • Google AI Studio is a developer tool for quickly trying the Gemini API; Agent Studio (formerly Vertex AI Studio) takes you from prototype to production under enterprise management and security.
  • Pretrained APIs: Speech-to-Text, Text-to-Speech, Translation, Document AI (extracting information from documents), Vision, Video Intelligence, Natural Language.

Workplace uses

  • Customer supportHand first-line inquiries to Conversational Agents, pass hard ones to operators, and show answer suggestions with Agent Assist.

Exam objectives (from the exam guide)

  • Google Cloud の強み（AI ファーストの取り組み、エンタープライズ対応、エコシステム、オープンなアプローチ、AI Hypercomputer・TPU・GPU、データの管理、ローコード・ノーコードによる AI の民主化）
  • すぐに使える生成 AI（Gemini app と Gems、Gemini Enterprise、Gemini for Google Workspace）
  • 顧客体験の向上（Agent Search・Google 検索、Customer Engagement Suite：Conversational Agents・Agent Assist・Conversational Insights・Contact Center as a Service）
  • 開発者向け（Agent Platform：Model Garden・Agent Search・AutoML、RAG のサービス、カスタムエージェントの構築）
  • エージェントのツール（拡張機能・関数・データストア・プラグイン、Cloud Storage・Cloud Run・各種 AI API）と Agent Studio・Google AI Studio の使い分け

Learning resources

Learn in detail (5 topics with mini quizzes) →Practice this domain (105 questions) →Next domain ↓
03

Section 3: Techniques to improve generative AI model output

約20%

Learning goals

  • Choose countermeasures that fit each limitation of foundation models
  • Use prompt techniques according to the goal
  • Explain grounding methods and the effects of sampling parameters

Key points

  • Knowledge cutoff: the model doesn't know information newer than its training data. Counter it with grounding via Google Search or RAG.
  • Hallucination: generating plausible errors. Countermeasures are grounding, RAG, citing sources and human review.
  • Fine-tuning further trains a model on your data to fit specific formats or specialist phrasing. RAG suits adding up-to-date facts.
  • Human in the Loop (HITL) is a mechanism where humans check and correct important outputs.
  • Zero-shot gives no examples, one-shot one example, few-shot several. Role prompts assign a role; prompt chaining links several prompts in sequence.
  • Chain-of-Thought has the model think step by step to improve reasoning; ReAct alternates reasoning and actions (tool calls).
  • Grounding ties outputs to verifiable sources: your own data (internal documents), third-party data, and world data (Google Search).
  • Low temperature makes outputs stable and deterministic; high makes them varied and creative. top-p picks from candidates until cumulative probability reaches p. Maximum output tokens limit length.

Workplace uses

  • HRIn a bot answering work-rule questions, ground answers in internal regulations (your own data) and lower temperature to reduce variation between answers.

Exam objectives (from the exam guide)

  • 基盤モデルの限界（データへの依存、知識のカットオフ、バイアス、公平性、ハルシネーション、エッジケース）と対策（グラウンディング、RAG、プロンプトエンジニアリング、ファインチューニング、Human in the Loop）
  • 継続的な監視と評価（モデルの自動アップグレード、KPI、セキュリティパッチ、バージョン管理、性能の追跡、ドリフトの監視、Feature Store）
  • プロンプトの技法（Zero-shot・One-shot・Few-shot、ロールプロンプト、プロンプトチェーン、Chain-of-Thought、ReAct）
  • グラウンディング（自社データ・第三者データ・世界のデータ、RAG、Agent Search、RAG API、Google 検索によるグラウンディング）
  • サンプリングのパラメータ（トークン数、temperature、top-p、安全性の設定、出力の長さ）

Learning resources

Learn in detail (5 topics with mini quizzes) →Practice this domain (60 questions) →Next domain ↓
04

Section 4: Business strategies for a successful generative AI solution

約15%

Learning goals

  • Choose a generative AI solution from business requirements and technical constraints, and explain the steps of adoption and measuring results
  • Explain the purpose of SAIF and the roles of Google Cloud security tools
  • Apply responsible AI viewpoints (transparency, privacy, fairness, accountability) to cases

Key points

  • For adoption, set goals and KPIs, start small with high-impact, feasible use cases, measure results and expand.
  • When choosing a solution, consider business requirements (accuracy, response speed, target users) and technical constraints (data, budget, existing systems, skills). Consider off-the-shelf products, then customization, then custom development.
  • Measuring results: time saved, quality, customer satisfaction, adoption (usage) rate and return on investment.
  • SAIF (Secure AI Framework) is Google's conceptual framework for securing AI systems. It extends existing security foundations to AI and includes detection and response and automated defenses.
  • IAM manages who can access what with least privilege; Security Command Center gives a single view of risks and threats.
  • Responsible AI: transparency (show AI use and limits), privacy (anonymization, pseudonymization, data minimization), fairness (checking data bias), accountability (owners and records).

Workplace uses

  • Corporate planningFor internal generative AI adoption, set KPIs for a 3-month trial in advance (50% less time writing meeting notes, 60% usage), then decide from the results whether to roll out company-wide.

Exam objectives (from the exam guide)

  • 生成 AI の導入の進め方（ソリューションの種類、ニーズに影響する要因、適切なソリューションの選択、組織への統合の手順、効果の測定）
  • セキュアな AI（ML のライフサイクル全体のセキュリティ、SAIF、セキュア・バイ・デザインのインフラ、IAM、Security Command Center、ワークロードの監視）
  • 責任ある AI（透明性、プライバシー、匿名化・仮名化、データの品質・バイアス・公平性、説明責任・説明可能性）

Learning resources

Learn in detail (3 topics with mini quizzes) →Practice this domain (42 questions) →Try the mock exam →

When you finish reading

Practice all domains (297 questions)Mock exam (60 questions · 90 min)GAIL home
Notes (unofficial, original questions, no pass guarantee)