cert-bank

Guides

이 글을 한국어로 읽기

GCP-GAIL Study Guide: Place the Product on the Right Layer, Then Match the Problem to Its Fix

The Generative AI Leader exam asks two kinds of question again and again. The first is "which offering fits this business?", where several Google products sound plausible and the answer depends on how much the company wants to build. The second is "the output is wrong in this particular way; what fixes it?", where the remedies sound alike and fix different things. This guide gives you a map for each: the five layers of the gen AI landscape for the first, and a problem-to-fix table for the second.

Study guide · Google Cloud Certified Generative AI Leader (GCP-GAIL)

GCP-GAIL Study Guide: Place the Product on the Right Layer, Then Match the Problem to Its Fix

The exam

50 to 60 multiple choice questions in 90 minutes. Google does not publish a pass mark, so check the official exam page for current details. Four sections, with the exam guide's approximate weights: Fundamentals of gen AI (~30%), Google Cloud's gen AI offerings (~35%), Techniques to improve gen AI model output (~20%) and Business strategies for a successful gen AI solution (~15%). It is a business-level exam: no code and no console steps.

Two figures from the 274 GCP-GAIL questions we wrote for this question bank: grounding, RAG or retrieval appears in the stem or options of 51 of them (19%), across all four sections, and fine-tuning in 15. In this bank, grounding is the usual fix for fact problems and tuning a narrow one.

Map 1: the five layers

The exam guide names five layers of the gen AI landscape; read them as a stack, each building on the one below. The higher the layer, the faster the time to value and the fewer ways to change behaviour; the lower, the more control and the more skills you need.

LayerGoogle Cloud examplesPick it when the stem says
ApplicationsGemini for Google Workspace; Gemini Enterprise"Employees need help now", "no data science team", "summaries and emails"
AgentsPrebuilt and custom agents; Conversational Agents; Agent Assist"Complete a task", "book, change or look up", "act through tools"
PlatformGemini Enterprise Agent Platform (formerly Vertex AI): Model Garden, Agent Studio, RAG Engine, Agent Runtime"Our developers will build", "governed", "choose and tune models"
ModelsGemini (multimodal), Gemma (open), Imagen (images), Veo (video)"Which model", "open weights", "generate video"
InfrastructureGPUs, TPUs, AI Hypercomputer"Train or serve at very large scale"

The offerings section (35%) is largely about placing a need on this table. A safe default in answers is to start high: managed services and pre-trained models first, look in Model Garden for a model that already does the job, and build lower down only when evaluation shows a gap.

The five layers of Google's gen AI landscape
Applications
Gemini for Google Workspace · Gemini Enterprise
Agents
Prebuilt and custom agents · Conversational Agents · Agent Assist
Platform
Gemini Enterprise Agent Platform · Model Garden, Agent Studio · RAG Engine, Agent Runtime
Models
Gemini · Gemma · Imagen · Veo
Infrastructure
GPUs, TPUs · AI Hypercomputer
Higher layers give faster time to value, lower layers more control: start high and go down only when evaluation shows a gap.

Look-alike products, one line each

  • Gemini for Google Workspace helps one user inside Gmail, Docs, Sheets, Slides and other Workspace apps. Gemini Enterprise is search, an assistant and agents across the company's connected data, with permissions respected. The exam guide still names Gems (custom versions of Gemini in the Gemini app); in the app itself Gems are being replaced by "skills", reusable custom shortcuts (gemini.google/overview/gems now describes skills). Learn both names.
  • Agent Search finds and answers from approved content. Conversational Agents hold a dialogue to get something done and hand over to a person when needed.
  • Agent Assist helps a human representative who is already on a call or chat. Conversational Insights (Customer Experience Insights in the current docs) analyses conversations that have already finished.
  • Model Garden is a catalogue for discovering and deploying models. AutoML on Agent Platform trains a predictive model on your own labelled data.
  • Google AI Studio is for quick prototypes in a browser. Agent Studio is for designing enterprise agents inside a governed Google Cloud project.

Beginner tip: when two offerings both look right, ask one question before anything else: is the user a customer, an employee, or a developer? Customers point to Agent Search, Conversational Agents or the customer engagement products; employees to Gemini for Google Workspace or Gemini Enterprise; developers to Agent Platform. Write the answer in the margin of your scratch sheet for each practice question you get wrong, and you will see that most confusions disappear once the user is named.

Map 2: problem to fix

The techniques section (20%) is a matching exercise. Each remedy fixes a specific limitation, and each wrong option is a real remedy for another one.

The problemThe fix
Facts are stale (knowledge cutoff) or must come from company documentsGrounding or RAG
Invented facts, fake citations (hallucination)Grounding, clearer prompts that allow "I don't know", human review where errors matter
Right content, wrong shape or formatFew-shot examples in the prompt
A stable house style or domain language that prompts cannot captureFine-tuning, with good labelled examples
Answers too varied for a policy documentLower temperature or top-p
Answers stop mid-sentenceRaise maximum output tokens
Multi-step reasoning goes wrongChain-of-thought; ReAct when tools are needed; prompt chaining for separate stages
Quality falls months after launchMonitor for drift; re-evaluate and update
Going deeper: why fine-tuning is the wrong fix for facts that change

Fine-tuning adjusts the model's weights using your examples. Whatever it learns is fixed at the moment training ends, exactly like the original knowledge cutoff, so a price list tuned into the model on Monday is out of date when prices change on Tuesday, and the only way to refresh it is another training run. It also gives no record of where an answer came from. RAG works the other way round: the weights never change, a retrieval step finds relevant passages in an index at question time and adds them to the prompt, and the model writes from that supplied text. Updating knowledge means updating the index, new documents are usable as soon as they are indexed, and the answer can cite its sources. That is why the exam pairs "changing facts" and "must cite sources" with grounding, and "consistent style or format" with tuning.

Business strategy: start from the problem

The smallest section (15%) rewards one habit: define the business problem and measurable KPIs before choosing technology. Then match the need to a pattern (documents with citations: grounding; content creation: a generative model; many customer conversations: conversational agents; routine office work: a ready-made assistant). Know the security tools by job: IAM for access, Model Armor to screen prompts and responses, Security Command Center AI Protection to inventory AI assets and find threats and risks to them, Sensitive Data Protection to de-identify data. And keep the privacy pair straight: anonymisation is meant to be irreversible, pseudonymisation replaces values with tokens, and two-way tokens can be reversed with the key.

A study order

  1. Fundamentals: AI vs ML vs gen AI, model types, the ML lifecycle, data quality.
  2. The five layers and Google's products on each, with the look-alike pairs.
  3. Techniques: prompting styles, grounding and RAG, sampling settings, limitations and remedies.
  4. Business strategy, security and responsible AI, then mixed sets with the 90-minute clock.

Practise

Practise in the question bank: every question explains each option, right and wrong. Start with the domain you just studied, then take a timed mock exam.

Section weights as listed in Google's Generative AI Leader exam guide. Google renames gen AI products often; the exam guide lists the product names the exam uses.