Task 1.3: Identify the core layers of the gen AI landscape and the business implications.

• AThe infrastructure layer

  Incorrect

  Infrastructure is not the layer that offers APIs and deployment tools to bridge models and agents.

• BThe model layer

  Incorrect

  The model layer is the trained models themselves, such as Gemini; the platform is what gives access to them and deploys them.

• CThe application layer

  Incorrect

  Applications sit on top of the platform; they do not provide its APIs and deployment tools.

• DThe platform layer

  Correct

  The platform layer, such as Gemini Enterprise Agent Platform, gives teams the tools to build, deploy and manage models and agents without running the underlying hardware themselves.

The platform layer is the bridge that makes models usable for building agents and applications.

• AThe application layer, such as Gemini in Gmail and Docs

  Correct

  Ready-made applications put gen AI inside familiar tools with no development work, which gives the fastest time to value when the need is standard productivity.

• BThe agent layer, by building a custom agent with Agent Development Kit

  Incorrect

  A custom agent can be powerful, but it needs developers and a project, which this team does not have.

• CThe model layer, by fine-tuning Gemma on past campaigns

  Incorrect

  Tuning an open model needs ML skills and infrastructure, and it would still leave the team without an interface in its tools.

• DThe platform layer, by calling the Gemini API through Agent Platform

  Incorrect

  Calling a model through a platform is how developers build their own applications; it is not usable by the team on its own.

The higher the layer, the less you build and the less you control. For a common need with no technical staff, buy at the application layer.

• AThe application layer

  Incorrect

  Applications are what end users interact with; the lab is creating a model, not a user-facing tool.

• BThe agent layer

  Incorrect

  Agents come after a model exists; they are not where the cost of training from scratch sits.

• CThe infrastructure layer

  Correct

  Training a model from scratch is a compute problem: the lab needs the hardware, such as TPUs or GPUs, and the systems to run training at scale, which is the infrastructure layer.

• DThe platform layer

  Incorrect

  A platform can schedule and manage training jobs, but the cost and scale of training from scratch sit in the compute beneath it.

Training from scratch is the most infrastructure-heavy path in gen AI. Most organizations start higher up the stack and use existing models instead.

• AThe model layer

  Incorrect

  A model provides the reasoning, but on its own it does not call the inventory system or send email; the agent wraps it with tools and a goal.

• BThe infrastructure layer

  Incorrect

  Infrastructure runs the workload; it does not decide which supplier to order from.

• CThe application layer

  Incorrect

  Applications are the user-facing layer, such as a chat interface; this process runs without a person interacting with it.

• DThe agent layer

  Correct

  An agent is software that pursues a goal by reasoning about its inputs and using the tools available to it, such as an inventory system, a pricing API and email.

The agent layer is where software works toward a goal over several steps and acts on other systems, without a person driving each step.

• AThe application layer

  Incorrect

  The application layer is where users interact with the assistant; the team is choosing what powers it.

• BThe infrastructure layer

  Incorrect

  Infrastructure is the compute that runs models, such as TPUs and GPUs.

• CThe model layer

  Correct

  The model layer is the trained models themselves, such as Gemini, Gemma and partner models offered in Model Garden; choosing among them is a model decision.

• DThe agent layer

  Incorrect

  The agent layer adds goals, reasoning loops and tools on top of a model; the team has not reached that design step.

Each layer has its own choices: which model, which platform, which agent design, which application. This one is about the model.

• AOn the application layer, by enabling Gemini in Gmail

  Incorrect

  Gemini in Gmail helps people with email; it cannot implement the insurer's claims workflow.

• BOn the infrastructure layer, by training a new model on GPUs

  Incorrect

  Training a new model is rarely needed for a workflow problem and adds the highest cost and effort.

• COn the application layer, by licensing a generic chatbot

  Incorrect

  A generic chatbot is quick to adopt but offers little integration or control, which the insurer needs most.

• DOn the platform layer, with Gemini Enterprise Agent Platform

  Correct

  A platform such as Agent Platform gives developers models, data tools, agent frameworks and deployment, so a team can build a custom solution with control over each part.

Building on a platform trades some speed for control and customization; it makes sense when the process is specific and the skills exist.

• AThe Gemini app, an assistant people chat with

  Correct

  The application layer is the user-facing part of gen AI; the Gemini app is an assistant people chat with directly.

• BModel Garden, a catalog where teams pick models

  Incorrect

  Model Garden is part of the platform, where teams discover, test, tune and deploy models; end users do not work in it.

• CAgent Development Kit, a framework for coding agents

  Incorrect

  ADK is a framework developers use to build agents; the agents may serve users, but the kit itself is a builder's tool.

• DCX Agent Studio, a console for designing agents

  Incorrect

  CX Agent Studio is where teams design and manage customer-facing agents; it is used to build agents, not by the end users who talk to them.

Applications sit on top of everything else; users see them, and the layers where teams pick models and build agents stay underneath.

• AServing and scaling are handled for you, and you pay per use

  Correct

  On a managed platform such as Agent Platform, Google handles model deployment and serving, and pay-as-you-go options suit variable traffic, so the company avoids buying and running hardware.

• BOwned servers give full control over hardware and its tuning

  Incorrect

  Control is a real benefit of owning hardware, but it argues for the other route and needs a team to run the servers, which the company does not have.

• COwned servers carry a fixed cost that is simple to budget

  Incorrect

  A fixed cost is predictable, but hardware sized for peaks sits idle between them; with uneven demand, pay-per-use fits better.

• DA Provisioned Throughput commitment keeps the bill flat

  Incorrect

  A flat commitment pays for the same capacity every day; with uneven, hard-to-predict demand, much of what is paid for would go unused.

The lower you build in the stack, the more you own and manage. For uneven demand and no hardware team, consuming models through a platform with pay-per-use pricing gets value faster.

• AA model, a vector database and a set of retrieved documents

  Incorrect

  Retrieval helps a model answer from the company's documents, but nothing in this set lets it look up an order or issue a refund.

• BA model, tools for taking actions, and a reasoning loop

  Correct

  An agent uses a model to reason, tools to interact with its environment, and an iterative loop in which it reasons, acts and observes the result.

• CA model, a system prompt and memory of past conversations

  Incorrect

  A persona and memory help an agent stay consistent and keep context, but without tools it still cannot reach the order system or issue refunds.

• DA model, a tuning dataset and an evaluation dataset

  Incorrect

  These are ingredients for adapting and testing a model; they do not give the system a way to act.

Tools and a reasoning loop are what turn a model that answers into an agent that acts. Retrieval, memory and tuning can make it better, but they do not let it take actions.