The Five Layers of the Gen AI Landscape and What They Mean for Business

The five layers of the gen AI landscape are infrastructure, models, platforms, agents, and applications. Think of them as a stack: each layer consumes what the layer below provides and adds something the layer above needs. Infrastructure powers models, models are made usable by platforms, platforms are where agents are built, and agents power the applications end users actually touch.

For a business leader, the stack is more than a technical diagram. It is a map of choices. An organization rarely needs to operate at every layer. A media company might only adopt applications. A bank with strict data requirements might invest in the platform layer to build custom solutions. A frontier research lab operates at the infrastructure and model layers. Knowing the map tells you where your money, people, and risk should go.

The exam expects you to place any offering or scenario onto this map quickly, so the rest of this lesson walks through each layer in turn.

The infrastructure layer provides the physical and virtual resources that make gen AI possible: specialized processors, high-speed networking, storage, and the data centers that house them. Training a foundation model requires enormous amounts of parallel computation, and serving one to millions of users requires reliable, low-latency capacity at scale. No other layer works without this one.

Google's offerings here include TPUs (Tensor Processing Units, Google's custom-designed AI accelerator chips), GPUs, and the AI Hypercomputer architecture that integrates compute, networking, and software into a system optimized for AI workloads, all running in Google's global data centers. Because this is delivered as cloud computing, an organization rents this capability on demand instead of buying hardware.

Two details are easy to misplace. First, AI Hypercomputer is more than chips: it also bundles optimized software (frameworks and runtimes tuned for the accelerators) and flexible consumption models, such as on-demand, committed-use, and Spot capacity, plus scheduling for short-lived jobs. Those are still infrastructure: they decide how compute runs and how it is paid for, not how a team builds a solution on a model, which is the platforms layer. Second, TPUs are rented as cloud capacity. No company has to buy TPUs, or any hardware, to use Gemini; when you call a managed model, Google runs it on its own infrastructure and you pay for usage.

The business implication: very few organizations should invest directly at this layer. Building AI infrastructure means massive capital expenditure, scarce specialist talent, and long timelines. For almost every enterprise, the right decision is to consume infrastructure through a cloud provider and let the economics of shared, purpose-built capacity work in their favor. What a leader should evaluate is not whether to build infrastructure, but whether their chosen provider's infrastructure is performant, reliable, and cost-efficient enough to support everything they build above it.

The models layer contains the foundation models themselves: large models trained on broad datasets that can generate and understand text, images, audio, video, and code. A foundation model is raw capability. It can do remarkable things, but on its own it is not a product; it has no connection to your data, your workflows, or your customers.

Google's models include Gemini, the flagship multimodal foundation model family; Gemma, a family of lightweight open models organizations can download and run on their own terms; Imagen for image generation; and Veo for video generation. These are covered in depth in the next lesson.

The business implication at this layer is model choice, not model creation. Training a competitive foundation model from scratch costs more than almost any enterprise can justify, so the practical decision is which existing model to use for each use case, weighing modality, context window, cost, performance, security, and how much customization or fine-tuning the business needs. A related strategic question is openness: a managed model like Gemini offers the strongest capability with no operational burden, while an open model like Gemma gives more deployment control. Organizations often use different models for different jobs rather than standardizing on one.

Openness and where the model runs are separate questions. Gemma's open weights make it possible to run a model on your own servers, but a requirement that data never leaves the organization's premises is met only when the model actually runs there. Calling managed Gemini over a private network connection, or serving Gemma from a managed endpoint deployed through Model Garden, still processes the data in Google Cloud. The reverse assumption is also wrong: a managed model is not automatically cloud-only, because Google offers Gemini on Google Distributed Cloud for organizations that need it in their own data centers or air-gapped environments.

The platforms layer turns raw model capability into something a business can actually build with. A platform provides managed access to models, tools for customization and grounding models in company data, deployment and scaling, and the monitoring, governance, and security controls an enterprise requires. Without a platform, every team would have to solve these problems from scratch.

Google Cloud's offering here is Agent Platform, which includes Model Garden (a curated catalog of Google, open, and partner models to discover and use), Agent Search for building search over enterprise data, and tooling to tune models and build agents. Google AI Studio serves as a lightweight environment for developers to prototype with Gemini quickly.

Classify an offering by its role, not by the words in its name. Model Garden contains models, but it is a catalog and deployment tool for builders, so it sits in the platforms layer; the models it lists (Gemini, Gemma, partner models) are the models layer. It is not something end users license either: staff who need gen AI in their daily work get it through an application such as Gemini Enterprise, not through the builder catalog. Likewise, Agent Search has "agent" in its name, but it is platform tooling that developers use to build search and retrieval over enterprise data, not an agent that pursues goals on its own.

Google AI Studio and Agent Platform are both builder tools, but they serve different stages:

A prototype that proves value in AI Studio is moved onto Agent Platform before it handles company data at scale, rather than shipped to production as-is.

The business implication: the platform layer is where most enterprises that want custom gen AI should invest. It is the sweet spot between control and speed. You get to use your own data, enforce your own security and governance, and shape solutions to your workflows, without owning infrastructure or training models. For example, an insurer that wants a claims-summarization tool grounded in its own policy documents does not need its own model; it needs a platform where its developers can combine a foundation model with company data safely. Grounding connects an existing managed model to company data at request time; it does not require owning, training, or self-hosting a model. The cost is that you need a development team; platforms serve builders, not end users.

The agents layer is where gen AI stops just generating content and starts accomplishing tasks. An agent combines a foundation model with reasoning, memory, and tools, connections to systems like databases, search, and business applications, so it can plan a multi-step task, take actions in the real environment, and return a completed outcome rather than a paragraph of text.

On Google Cloud, organizations build custom agents on Agent Platform, use Agent Studio to design them, and deploy prebuilt conversational capability through offerings like Conversational Agents in Google's Customer Engagement Suite. An agent might look up a customer's order in a database, check a refund policy, issue the refund, and confirm it, all from one request.

The business implication: agents change the economics of automation. Traditional automation handled rigid, rule-based work; agents handle ambiguous, language-heavy work that previously required a person, such as customer support triage, research, and internal service desks. Leaders should identify processes with high volume, clear goals, and access to the systems an agent would need. The trade-off is trust: because agents take actions, they demand more governance, permissions design, and human oversight than a tool that only drafts text. Start with agents that assist people before agents that act alone.

A good early agent candidate passes all of these tests:

• High volume and a clear goal: the work repeats often and success is easy to recognize.
• Live system access: the agent can read and act on current data through the systems themselves, typically via APIs. Data that arrives only as a periodic file dump, such as a weekly spreadsheet, is a snapshot, not the live access an agent needs to act on today's state.
• Low-risk, easily corrected actions: early autonomy goes to actions that are cheap to undo, such as updating a ticket or answering a routine question. A process can be high-volume and goal-clear and still be a poor first candidate for autonomy if its actions move money or are hard to reverse, such as issuing refunds or changing credit limits; there the agent prepares the work and a person approves it.

Permissions follow the same logic: give an agent only the access its task requires, log what it does, and widen its autonomy as it earns trust.

The applications layer is where gen AI reaches end users inside finished products. No building is required: the AI is embedded in the software people already use, designed around specific jobs like writing, searching, analyzing, and collaborating.

Google's offerings here include the Gemini app for direct assistant-style use, Gemini for Google Workspace, which brings gen AI into Gmail, Docs, Sheets, and Meet, and Gemini Enterprise, which gives employees a secure front door to gen AI grounded in company knowledge, with multimodal search and custom agent capabilities.

Each of these has a defined audience. The Gemini app is a general-purpose assistant; it is not grounded in a company's own catalog or systems. A Gemini Enterprise deployment for employees serves staff: grounding it in the same product data a website uses does not turn it into a public shopper or member experience. For customer-facing use, Google offers a separate product, Gemini Enterprise for Customer Experience (announced January 2026), alongside custom solutions built on the platform and agent layers.

The business implication: applications are the fastest path to value and the lowest barrier to entry. Adoption is measured in days, requires no development team, and immediately raises productivity across broad functions. For many organizations, especially those without engineering capacity, this layer is the right starting point, and for every organization it is the layer where the most employees will touch gen AI. The limits are differentiation and fit: an off-the-shelf application serves common needs well but cannot encode your unique workflows or become a competitive moat. A secondary, manageable limitation is control over change: the provider decides when the product and its underlying model are updated. Governance, by contrast, is usually not a reason to rebuild a common capability yourself: Gemini for Google Workspace runs under the organization's existing Workspace admin and data controls, so a security or compliance concern is normally addressed through those controls rather than by building a replacement on a platform. Leaders typically adopt applications broadly for workforce productivity while reserving platform and agent investment for the processes that set the business apart.

Decide by matching each layer's trade-off, control versus speed, to the business need. Operating lower in the stack means more control, more differentiation potential, more cost, and more required expertise. Operating higher means faster time to value, lower cost, less specialized talent, and less customization. This is the classic build versus buy decision, applied layer by layer rather than as a single yes or no.

A practical way to frame it:

• Applications: buy when the need is common to every business, such as employee productivity, meeting summaries, and drafting.
• Agents and platforms: build here when the need is specific to your data, workflows, or customers, and the process is core to how you compete.
• Models: choose rather than create; select and, where justified, tune an existing foundation model.
• Infrastructure: consume from a cloud provider; almost never build.

Most enterprises end up operating at two layers at once: applications for the broad workforce, and the platform or agent layer for a small number of high-value custom solutions. The mistake the exam scenarios often probe is mismatching layer to need, for example, commissioning a custom-built platform project for a need an off-the-shelf application already meets, or expecting a productivity application to deliver a differentiated, data-grounded customer experience.

Consider a mid-sized retailer with three goals: raise employee productivity, improve customer support, and personalize product discovery on its website. Each goal lands on a different layer.

For productivity, the retailer adopts Gemini for Google Workspace, the applications layer. Merchandisers summarize supplier documents, marketers draft campaign copy, and store operations teams write briefings faster. No development work, value in the first week.

For customer support, it deploys a conversational agent through Google's Customer Engagement Suite, the agents layer, connected to its order-management system so the agent can check delivery status and process routine returns, escalating complex cases to people. For personalized discovery, its developers build on Agent Platform, the platforms layer, choosing Gemini from Model Garden and grounding it in the retailer's own product catalog so shoppers can search naturally, for example, a waterproof jacket for spring hiking under a set budget.

What the retailer does not do is just as instructive: it does not train its own foundation model and it does not buy AI hardware. It consumes the models and infrastructure layers through Google Cloud. One company, five layers, but direct investment in only the layers where its needs are unique.