GAIL exam domains at a glance
| Section | Weight | What it really tests |
|---|---|---|
| 1. Fundamentals of gen AI | ~30% | AI vs ML vs gen AI; ML approaches and lifecycle; data types and quality; choosing a foundation model; the five landscape layers; Gemini, Gemma, Imagen, Veo |
| 2. Google Cloud's gen AI offerings | ~35% | Google Cloud strengths; Gemini app, Gemini Enterprise, Gemini for Workspace; Agent Search; Customer Engagement Suite; Agent Platform, Model Garden, RAG offerings; agent tooling and pre-built APIs; Agent Studio vs Google AI Studio |
| 3. Techniques to improve model output | ~20% | Model limitations; grounding, RAG, prompt engineering, fine-tuning, HITL; monitoring; prompting techniques; sampling parameters |
| 4. Business strategies for a successful gen AI solution | ~15% | Steps to adopt gen AI, measuring impact; secure AI and SAIF; Google Cloud security tools; responsible AI, privacy, bias, accountability |
Source: the official GAIL exam guide. Section 2 is the biggest and the most product-heavy — most last-week review time should go there.
Generative AI key terms for the GAIL exam
| Term | One-line definition |
|---|---|
| Artificial intelligence (AI) | Broad field of systems that perform tasks normally needing human intelligence. |
| Machine learning (ML) | Subset of AI: models learn patterns from data instead of explicit rules. Supervised (labeled data), unsupervised (unlabeled, find structure), reinforcement (learn from rewards). |
| Generative AI | ML that creates new content (text, images, code, audio, video) rather than only classifying or predicting. |
| Foundation model | Large model pre-trained on broad data that can be adapted to many tasks. Gemini is Google's flagship example. |
| Large language model (LLM) | Foundation model specialised in understanding and generating language. |
| Multimodal model | Accepts and/or produces more than one data type (text, image, audio, video). Gemini is multimodal. |
| Diffusion model | Generates images or video by iteratively removing noise. Imagen and Veo are Google examples. |
| Token | The unit a model reads and writes (a word, part of a word, or symbol). Pricing, limits, and output length are counted in tokens. |
| Context window | Maximum tokens a model can consider at once (prompt plus history plus retrieved content). A model-selection factor. |
| Prompt engineering | Designing the input (instructions, examples, context, format) to get better output without changing the model. |
| Prompt tuning | Learning a small set of soft-prompt parameters for a task while the base model stays frozen. Lighter than full fine-tuning. |
| Fine-tuning | Further training a pre-trained model on task-specific labeled data to change its behaviour, style, or specialisation. |
| Grounding | Connecting model output to verifiable sources (enterprise data, third-party data, or world data such as Google Search) so answers are factual and citable. |
| RAG | Retrieval-augmented generation: retrieve relevant documents at query time and add them to the prompt. A way to ground on private, current data without retraining. |
| Hallucination | Output that is fluent and confident but false or unsupported. Reduced by grounding/RAG, clear prompts, and human review. |
| Knowledge cutoff | The date after which the model has no training data. Grounding addresses it; fine-tuning only moves it. |
| Embeddings | Numeric vectors representing meaning, used to find semantically similar content during retrieval. |
| Agent | A system that uses a model to reason, plan, and take actions with tools to complete a goal, not just answer a prompt. |
| Tools (for agents) | How agents act on the world. The guide names extensions, functions, data stores, and plugins. |
| Human in the loop (HITL) | A person reviews or approves model output before it takes effect. Used for high-stakes or regulated decisions. |
| Drift | Model performance degrading over time as real-world data changes. Caught by monitoring and evaluation. |
Data quick hits: structured data fits rows and columns (sales tables, CRM records); unstructured does not (emails, PDFs, images, call audio). Labeled data has the answer attached (supervised learning); unlabeled does not. The guide's data-quality attributes: completeness, consistency, relevance, availability, cost, format. ML lifecycle: ingestion → preparation → training → deployment → management.
Sampling parameters: temperature, top-k, top-p
| Setting | What it does | Lower → | Higher → |
|---|---|---|---|
| Temperature | Controls randomness when choosing the next token | Focused, consistent, repeatable (extraction, classification, factual Q&A) | Varied, creative (brainstorming, marketing copy) |
| Top-k | Samples only from the k most likely next tokens | Fewer candidates, safer output | More candidates, more variety |
| Top-p (nucleus) | Samples from the smallest set of tokens whose probabilities add up to p | Narrower, more predictable | Wider, more diverse |
| Max output tokens / output length | Caps response length (and cost) | Shorter, cheaper, may truncate | Longer responses allowed |
| Safety settings | Thresholds for blocking harmful categories of content | Stricter blocking | More permissive |
Exam rule of thumb: none of these settings adds knowledge. Lowering temperature does not fix outdated or missing facts — that is a grounding problem. The guide names token count, temperature, top-p, safety settings, and output length; top-k is included here because it appears in Google's model docs and most study material.
Prompt engineering techniques
- Zero-shot: instruction only, no examples. Fine for simple, common tasks.
- One-shot / few-shot: include one or several input–output examples. Best for enforcing a format, tone, or classification scheme.
- Role prompting: “You are a compliance analyst…” — sets perspective, vocabulary, and tone.
- Prompt chaining: split a complex job into sequential prompts where each output feeds the next (extract → summarise → draft email).
- Chain-of-thought (CoT): ask the model to reason step by step. Improves multi-step reasoning, maths, and logic.
- ReAct (reason + act): the model alternates reasoning with actions such as tool calls or searches, then observes results. The pattern behind agents.
- Good prompt anatomy: clear task, relevant context, constraints, examples, and the output format you want.
When to use which technique: prompting vs grounding vs RAG vs fine-tuning
| The problem in the scenario | Best first answer | Why not the others |
|---|---|---|
| Output is vague, badly formatted, or off-tone | Prompt engineering (instructions, few-shot examples, role) | Cheapest and fastest; no data pipeline or training needed |
| Needs recent public facts or events after the knowledge cutoff | Grounding with Google Search | Fine-tuning cannot keep up with world events |
| Must answer from private, frequently changing company documents, with citations | RAG (pre-built RAG with Agent Search, or RAG APIs) | Data stays current at query time; no retraining; sources traceable |
| Needs a consistent specialised style, format, or domain behaviour that prompting cannot achieve | Fine-tuning (or prompt tuning for a lighter option) | Costlier, needs quality labeled data, and the knowledge goes stale |
| High-stakes or regulated decisions (medical, legal, credit, HR) | Human in the loop | No technique removes the accountability requirement |
| Task needs multi-step reasoning | Chain-of-thought prompting | Before reaching for a bigger model or fine-tuning |
| Task needs to take actions in other systems | An agent with tools (ReAct pattern) | A plain LLM can describe actions but cannot perform them |
| Quality is degrading after launch | Monitoring and evaluation (KPIs, drift monitoring, versioning) | Measure before you retrain |
Grounding is the goal (answers tied to sources); RAG is one way to do it. The guide distinguishes grounding on first-party enterprise data, third-party data, and world data. Deeper dive: retrieval-augmented generation explained.
Google Cloud gen AI product map (2026 names)
On April 22, 2026, Google renamed Vertex AI to Gemini Enterprise Agent Platform. The current exam guide uses the new names; older courses and videos may not. Former names are shown in brackets.
| Product | Who it's for | Pick it when the scenario says… |
|---|---|---|
| Gemini app (Gemini Advanced, Gems) | Individuals | Personal productivity assistant; Gems are reusable custom versions of Gemini for a repeated task |
| Gemini for Google Workspace | Employees in Gmail, Docs, Sheets, Meet | “Help our staff write, summarise, and analyse inside the tools they already use” |
| NotebookLM | Knowledge workers | Research and summarise across uploaded sources, with answers grounded in those sources |
| Gemini Enterprise (absorbed Agentspace) | Whole organisation | Company-wide search across enterprise data, multimodal search, and prebuilt or custom agents for employees |
| Agent Platform (Vertex AI) | Developers, data teams | Build, tune, deploy, and govern custom models and agents on Google Cloud |
| Model Garden | Developers | Choose from Google, third-party, and open models (the “open approach”) |
| Agent Studio (Vertex AI Studio) | Enterprise developers | Prototype and test prompts and models inside Google Cloud with enterprise security and governance |
| Google AI Studio | Individual developers | Fastest free way to try Gemini models and get an API key; not the enterprise-governed option |
| Agent Search (Vertex AI Search) | Developers | Google-quality search over your own data; pre-built RAG for apps and agents |
| RAG APIs | Developers | Custom RAG pipelines where the pre-built option is not flexible enough |
| Agent Platform AutoML (Vertex AI AutoML) | Teams with little ML expertise | Train a custom model on your own data with low or no code |
| Customer Engagement Suite | Contact centres | Conversational Agents (customer-facing bots), Agent Assist (real-time help for human agents), Conversational Insights (analytics), Contact Center as a Service |
| Pre-built AI APIs | Developers | Single well-defined task: Speech-to-Text, Text-to-Speech, Translation, Document AI, Vision, Video Intelligence, Natural Language |
| AI Hypercomputer, TPUs, GPUs | Infrastructure | The infrastructure layer: Google's custom TPUs plus GPUs for training and serving at scale |
Google's models: Gemini (flagship multimodal reasoning), Gemma (lightweight open models you can run and customise yourself), Imagen (text-to-image), Veo (video generation). Other renames to recognise: Vertex AI Agent Engine → Agent Runtime; Vertex AI Feature Store → Agent Platform Feature Store.
Gen AI landscape layers and agents
- Infrastructure: compute and hardware (AI Hypercomputer, TPUs, GPUs, data centres).
- Models: the foundation models themselves (Gemini, Gemma, Imagen, Veo, plus third-party models in Model Garden).
- Platforms: tools to build, tune, deploy, and manage (Agent Platform).
- Agents: systems that reason, plan, and act with tools toward a goal.
- Applications: what end users touch (Gemini app, Gemini for Workspace, NotebookLM, Gemini Enterprise).
Agent tooling (from the guide): extensions, functions, data stores, and plugins — backed by services such as Cloud Storage, databases, Cloud Run, Cloud Functions, and the pre-built AI APIs. Agent Studio vs Google AI Studio: enterprise, governed, inside Google Cloud vs quick, individual experimentation with Gemini.
Responsible AI, security, and SAIF
- SAIF (Secure AI Framework): Google's framework for securing AI systems. Six core elements: expand strong security foundations to the AI ecosystem; extend detection and response to AI threats; automate defences; harmonise platform-level controls; adapt controls with faster feedback loops; contextualise AI risks within business processes.
- Security across the ML lifecycle: protect training data, models, and endpoints (risks include data poisoning, prompt injection, and data leakage through outputs).
- Google Cloud security tools named in the guide: secure-by-design infrastructure, Identity and Access Management (IAM, least privilege), Security Command Center (central risk and threat view), and workload monitoring tools.
- Privacy: anonymisation removes identifiers irreversibly; pseudonymisation replaces them with tokens that can be re-linked under control. Pseudonymised data is still personal data.
- Fairness and bias: biased or unrepresentative training data produces biased output. Fix at the data and evaluation level, not with sampling settings.
- Transparency, accountability, explainability: people should know AI is involved, be able to understand why it produced an output, and a named human owner remains accountable.
Business strategy essentials (Section 4)
- Start from the business problem, not the model. Factors: business requirements and technical constraints (data, budget, latency, skills).
- Buy before build: a prebuilt app (Gemini for Workspace, Gemini Enterprise) usually beats a custom build when it meets the need — faster time to value, less operational overhead.
- Choosing a model: modality, context window, security, availability and reliability, cost, performance, and customisation options.
- Integrating gen AI: identify use cases, pilot small, measure, iterate, then scale with governance and change management.
- Measuring impact: define KPIs before launch — productivity, cost, quality, customer satisfaction, adoption — and track them over time.
- Gen AI can create, summarise, discover, and automate: map each scenario to one of these four verbs to find the right solution type.
GAIL trap-question patterns
- “Fine-tune it” for changing data: if the facts change often, the answer is grounding or RAG, not fine-tuning.
- “Turn down the temperature” for hallucinations about facts: temperature controls randomness, not accuracy. Missing facts need grounding.
- “Train a new model from scratch”: almost never right for a business scenario. Use an existing foundation model.
- Build when you could buy: if a prebuilt Google app solves the stated problem, a custom Agent Platform project is usually the distractor.
- Google AI Studio vs Agent Studio: enterprise data, governance, or security in the stem points to Agent Studio / Agent Platform.
- Agent Assist vs Conversational Agents: helping a human agent in real time is Agent Assist; talking directly to customers is Conversational Agents.
- Old vs new names: an option saying “Vertex AI Search” and one saying “Agent Search” describe the same product — read for the capability.
- Removing the human: for high-stakes decisions, an answer that removes human review for speed is usually wrong.
- Technical depth bait: GAIL is a leader exam. An option full of implementation detail is rarely better than the one tied to business value and responsible use.
Sample Question (Section 2 — Google Cloud's Gen AI Offerings)
A contact centre wants to cut average handling time. Supervisors want human agents to receive suggested answers and relevant knowledge articles in real time while they talk to customers. Which Google Cloud offering fits best?
- Conversational Agents
- Agent Assist
- Google AI Studio
- Gemini for Google Workspace
Answer: B. Agent Assist, part of the Customer Engagement Suite, supports human agents with real-time suggestions. Conversational Agents (A) talk to customers directly instead of helping staff. Google AI Studio (C) is a developer prototyping tool. Gemini for Workspace (D) helps in Gmail and Docs, not live contact-centre conversations. More like this: GAIL practice questions.
Frequently asked questions
Can I use this cheat sheet during the Generative AI Leader exam?
No. GAIL is proctored, whether you take it online or at a test centre, and you cannot use notes or outside material during the exam. Use this page to review before the exam: read it once early to see the shape of the syllabus, come back to individual tables as you study each section, and read it end to end in the last 48 hours.
Which Google Cloud product names does the GAIL exam use in 2026?
The current official exam guide uses the post-April-2026 names: Agent Platform (Gemini Enterprise Agent Platform, formerly Vertex AI), Agent Studio (formerly Vertex AI Studio), Agent Search (formerly Vertex AI Search), Agent Platform AutoML, and Agent Platform Feature Store. Gemini Enterprise now includes what was Google Agentspace. Learn both old and new names, because study material written before the rename still uses the Vertex AI names.
What is the most important table on this cheat sheet?
The "when to use which technique" table. Section 3 (techniques to improve output) and many Section 2 and 4 scenarios come down to picking the least expensive approach that solves the stated problem: prompt engineering first, then grounding or RAG for missing or changing facts, then fine-tuning for consistent style or specialised behaviour, with human in the loop for high-stakes outputs.
Does the GAIL exam test temperature, top-k, and top-p?
The official exam guide lists sampling parameters and settings such as token count, temperature, top-p (nucleus sampling), safety settings, and output length. Top-k is not named in the current guide, but it appears in Google's model documentation and in many study resources, so knowing it in one line is cheap insurance. Expect conceptual questions (which setting makes output more creative or more consistent), not number-tuning.
What is Google's Secure AI Framework (SAIF)?
SAIF is Google's conceptual framework for securing AI systems. It has six core elements: expand strong security foundations to the AI ecosystem, extend detection and response to bring AI into an organisation's threat universe, automate defences, harmonise platform-level controls, adapt controls with faster feedback loops, and contextualise AI system risks in surrounding business processes. The exam guide asks you to identify its purpose and benefits, not memorise implementation detail.
Is this cheat sheet enough to pass the Generative AI Leader exam?
No. It compresses the syllabus for review; it does not teach it. Pair it with the official Google Skills learning path, a structured study plan, and timed practice questions. The exam rewards applying these terms to business scenarios, which only practice questions train.
Turn This Cheat Sheet Into Real Practice
A cheat sheet reinforces what you know. 500 scenario-based Generative AI Leader questions show you what you don't. Try 30 free — free account, no credit card. $11.99 one-time for the full bank, or $9.99/month Pro for every certification.