This domain is about 30% of the exam. It is vocabulary and judgment: can you define the terms, tell the kinds of AI apart, pick a model for a business need, and say why data quality matters. Questions are written for business leaders, so they ask “which concept fits this situation” and never ask for maths or code. Google renamed Vertex AI to Gemini Enterprise Agent Platform (short form Agent Platform) in 2026, and Vertex AI Studio became Agent Studio. The current exam guide uses the new names, and older study material still says Vertex AI. They are the same platform.
Core terms
| Term | Meaning |
|---|---|
| Artificial intelligence (AI) | Systems that do tasks that normally need human intelligence: understanding language, seeing, deciding, planning |
| Machine learning (ML) | A branch of AI where a system learns patterns from data instead of following hand-written rules |
| Natural language processing (NLP) | AI that works with human language: classifying, translating, summarising, answering |
| Generative AI (gen AI) | ML that creates new content (text, images, code, audio, video) rather than only classifying or predicting |
| Foundation model | A very large model trained on broad data that can be adapted to many tasks without building a new model each time |
| Large language model (LLM) | A foundation model specialised in language. It predicts the next token (piece of a word) over and over |
| Multimodal foundation model | A model that takes in and/or produces more than one type of data, such as text, images, audio and video, in one model |
| Diffusion model | A model that creates images (and video or audio) by starting from noise and removing it step by step until it matches the prompt |
| Prompt | The input you give a model: instructions, context, examples, a question |
| Prompt engineering | Writing and refining prompts to get better output without changing the model |
| Prompt tuning | Training a small set of added values (a “soft prompt”) while the base model stays frozen. Cheaper than full fine-tuning |
| Token | The unit a model reads and writes, roughly a word or part of a word. Cost and limits are counted in tokens |
| Context window | The maximum number of tokens (input plus output) a model can consider in one request |
| Inference | Using a trained model to produce an output |
| Agent | A system built around a model that can plan, call tools and take multi-step actions toward a goal |
The nesting is AI > ML > deep learning > generative AI. Predictive AI forecasts or classifies (will this customer churn?). Generative AI makes something new (write the retention email).
Machine learning approaches
| Approach | Learns from | Typical use | Example |
|---|---|---|---|
| Supervised | Labeled examples (input plus the right answer) | Classification and regression | Spam detection, price prediction |
| Unsupervised | Unlabeled data, finding structure by itself | Clustering, anomaly detection, topic discovery | Grouping customers into segments |
| Reinforcement | Trial and error with rewards and penalties | Sequential decisions, games, robotics, tuning model behaviour from feedback | A robot learning to grasp objects |
Large language models are first trained on huge unlabeled text (self-supervised: the text itself supplies the next word to predict). They are then often refined with human feedback, which is a reinforcement-style step. Do not overthink this: if a question says “labeled data”, the answer is supervised; “no labels, find groups”, unsupervised; “reward”, reinforcement.
The machine learning lifecycle
The guide lists five stages and wants you to link each to Google Cloud tools.
| Stage | What happens | Google Cloud tools |
|---|---|---|
| Data ingestion | Collect data from sources into one place | Cloud Storage, Pub/Sub (streaming), Dataflow, BigQuery Data Transfer Service |
| Data preparation | Clean, transform, label, split into training and test sets | BigQuery, Dataflow, Managed Service for Apache Spark (formerly Dataproc), Dataprep-style tools, Cloud Data Fusion |
| Model training | Teach the model, tune and evaluate it | Agent Platform training, AutoML, custom training, BigQuery ML, TPUs and GPUs |
| Model deployment | Serve the model so applications can call it | Agent Platform endpoints, Cloud Run, GKE |
| Model management | Version, monitor for drift, retrain, govern | Agent Platform Model Registry, Model Monitoring, Pipelines, Feature Store |
The cycle repeats. Monitoring in production usually sends you back to ingestion and preparation with fresh data. Most of the effort in real projects is data work, not model work.
Choosing a foundation model
No model is best at everything. Match the model to the requirement.
| Factor | The question to ask |
|---|---|
| Modality | What inputs and outputs are needed: text only, images, audio, video, or a mix? A text-only model can’t read a photo |
| Context window | How much material must the model read at once (long contracts, hours of video, a codebase)? |
| Security and privacy | Where does the data go? Is it used to train the model? What certifications and controls exist? |
| Availability and reliability | Does the provider offer a service level, regional capacity and stable versions? |
| Cost | Price per token or per call, multiplied by expected volume. Smaller models cost much less |
| Performance | Quality of answers and latency (speed). Bigger models are usually better and slower |
| Fine-tuning and customization | Can the model be adapted to your data and terms when prompting is not enough? |
Rules of thumb that match how Google frames the choice:
- Start with the smallest, cheapest model that meets the quality bar, and move up only when tests show it falls short.
- Use a high-end model for complex reasoning and a fast, low-cost model for high-volume, simple tasks.
- Model Garden in Agent Platform lets you try Google, open and third-party models side by side, so the choice can be tested instead of guessed.
- Bigger is not always better. Speed, cost and fit often decide.
What gen AI does for a business
The guide groups use cases into four verbs.
| Verb | Meaning | Examples |
|---|---|---|
| Create | Produce new content | Marketing copy, product images, code, video, synthetic data |
| Summarize | Condense long material | Meeting notes, contracts, support tickets, research papers |
| Discover | Find and answer from large collections | Enterprise search, Q&A over company documents, multimodal search |
| Automate | Carry out steps of a process | Drafting replies, routing requests, agents completing workflows |
By output type: text generation (drafts, chat, translation), image generation (concept art, ads), code generation (assist, explain, test, migrate), video generation (short clips, product demos), data analysis (ask a question of data in plain language), and personalized experiences (recommendations and messages tailored to each user).
Trap: gen AI is not the right tool for everything. A fixed calculation, an exact lookup, or a task with zero tolerance for error is better served by normal software or by gen AI that is grounded and checked by a person.
Data types and why they matter
| Pair | Meaning | Examples |
|---|---|---|
| Structured | Fits rows and columns with a defined schema | Sales tables, transactions, inventory |
| Unstructured | No fixed shape | Emails, PDFs, images, audio, video, chat logs |
| Semi-structured | Tagged but flexible | JSON, XML, logs |
| Labeled | Each example carries the right answer or tag | Photos tagged “cat” or “dog”, emails marked spam |
| Unlabeled | Raw examples with no tags | A folder of unlabeled photos, raw text |
Most business data (often cited as around 80%) is unstructured. Gen AI’s business impact is largely that it can finally read, search and summarise that material. Supervised learning needs labeled data, which is expensive to produce. LLM pre-training works on unlabeled text.
Business implications: more data types means richer solutions (a model that reads text plus images plus audio) but also more work on storage, privacy and cost.
Data quality and accessibility
A model is only as good as its data: poor data in means poor output out (and confident-sounding wrong output from gen AI).
| Characteristic | What it means | Failure example |
|---|---|---|
| Completeness | No important gaps in records or coverage | Missing regions skew the model |
| Consistency | Same thing recorded the same way everywhere | “NY”, “New York” and “N.Y.” treated as different |
| Relevance | Data actually relates to the task | Training a support bot on marketing brochures |
| Availability | The data can be reached when needed | Data stuck in silos or behind slow processes |
| Cost | Collecting, storing, labelling and moving data costs money | Labelling millions of items by hand |
| Format | Data is in a form the tools can use | Scanned images of tables instead of text |
Also remember accuracy, freshness, representativeness (does it cover all the groups the model will serve?) and governance (who may use it for what). Poorly representative data is the usual source of bias.
The layers of the gen AI landscape
Think of five layers, bottom to top. Each layer has providers, and a business can buy at any layer.
| Layer | What it is | Google examples | Business meaning |
|---|---|---|---|
| Infrastructure | Chips, servers, networks, data centres that run training and inference | AI Hypercomputer, TPUs, GPUs | Cost, speed and scale. Needed by model builders, hidden from most buyers |
| Models | The foundation models themselves | Gemini, Gemma, Imagen, Veo, plus partner and open models | Capability, price per token, openness |
| Platforms | Tools to find, customise, evaluate, deploy and govern models | Agent Platform, Agent Studio, Model Garden | Lets developers build on models without running infrastructure |
| Agents | Systems that reason, use tools and act | Agents built with Agent Platform, Conversational Agents, Gemini Enterprise agents | Automation of multi-step work |
| Applications | Finished products people use | Gemini app, Gemini for Workspace, Gemini Enterprise, Customer Engagement Suite | Fastest time to value, least control |
Trade-off: moving up the stack means faster and easier but less customisation. Moving down gives control but needs skills and budget. A typical leader question is “which layer fits a company with no ML team that wants value this quarter?” Answer: applications, or prebuilt agents.
Google’s foundation models
| Model | What it is | Strengths and use cases |
|---|---|---|
| Gemini | Google’s flagship multimodal family. Takes text, images, audio, video and code, with very large context windows. Comes in tiers: larger “Pro” models for hard reasoning, “Flash” models for speed and cost, and smaller still for high volume | Chat, summarising long documents and video, code, data analysis, agents, reasoning, image understanding |
| Gemma | Open-weights lighter models built from the same research as Gemini. Can be downloaded, run on your own hardware or on laptops and phones, and fine-tuned | When you need control, on-device or offline use, low cost, or a model you can customise and host yourself |
| Imagen | Google’s text-to-image model family | Marketing visuals, product mock-ups, image editing, high-quality images with legible text |
| Veo | Google’s text-and-image-to-video model, which can generate video with sound | Ads, storyboards, product demos, short film clips |
Other names that may appear: Lyria (music generation) and Chirp (speech models). Model versions change often, so the exam tests the family and the use case, not version numbers.
Quick matching:
- Needs text plus images plus video in one prompt, long context: Gemini.
- Must run on its own servers or device, open weights: Gemma.
- Make a picture from a description: Imagen.
- Make a video clip from a description: Veo.
Common traps
- “Foundation model” does not mean “text only”. Multimodal models are foundation models too.
- Prompt engineering changes the input. Fine-tuning changes the model. Prompt tuning trains a small add-on while the base model stays frozen.
- Gemma is open weights; Gemini is not. Open weights does not mean “no cost” to run.
- Labeled vs unlabeled is about the presence of answers. Structured vs unstructured is about the shape of data. They are different questions.
- A larger context window helps with long inputs but costs more and can still miss details.
Exam checklist
- Define AI, NLP, ML, gen AI, foundation model, multimodal, diffusion, LLM, prompt engineering and prompt tuning in one line each.
- Match supervised, unsupervised and reinforcement learning to a scenario.
- List the five ML lifecycle stages with a Google Cloud tool for each.
- Name the seven selection factors for a foundation model and apply them to a scenario.
- Sort use cases into create, summarize, discover and automate.
- Tell structured from unstructured and labeled from unlabeled, with examples.
- Name the data quality traits: completeness, consistency, relevance, availability, cost, format.
- Name the five layers: infrastructure, models, platforms, agents, applications.
- Match Gemini, Gemma, Imagen and Veo to their jobs.