Foundation Models

A practical guide to understanding what foundation models are and why they matter for AI product development.

What foundation models are, how they differ from traditional software, and what product and design teams need to know when building on top of them.

22 May 20264 min read

What it is

A foundation glossaryModelA model is a system or representation used to process data and generate outputs, often trained to perform specific tasks.Open glossary term is a large AI model trained on a broad range of glossaryDataData is raw information collected and stored for analysis, processing, or decision-making.Open glossary term that can be adapted and applied to a wide variety of tasks.

Rather than training a specialised glossaryModelA model is a system or representation used to process data and generate outputs, often trained to perform specific tasks.Open glossary term for each specific use case, a foundation model provides a general-purpose base that can be prompted, fine-tuned, or connected to external tools to handle many different applications.

GPT-4, Claude, Gemini, and Llama are examples of foundation glossaryModelA model is a system or representation used to process data and generate outputs, often trained to perform specific tasks.Open glossary term. Most AI products you interact with are built on top of one of these, rather than training an entirely new model from scratch.

Foundation glossaryModelA model is a system or representation used to process data and generate outputs, often trained to perform specific tasks.Open glossary term represent a fundamental shift in how software is built. Rather than writing explicit rules for every glossaryBehaviourBehaviour refers to how users interact with a system, including actions, patterns, and responses.Open glossary term, you are building on a glossarySystemA system is a collection of interconnected components that work together to achieve a specific function or outcome.Open glossary term that has learned from an enormous range of human-generated content.

That power comes with limitations — including inherited glossaryBiasBias is a systematic distortion in thinking or data that affects the accuracy of research or decision-making.Open glossary term, hallucination risk, and glossaryBehaviourBehaviour refers to how users interact with a system, including actions, patterns, and responses.Open glossary term that can be difficult to predict or fully control.

When to use it

Understand when building on a foundation glossaryModelA model is a system or representation used to process data and generate outputs, often trained to perform specific tasks.Open glossary term is the right approach.

It is most relevant when:

You are building AI-powered features and evaluating which model to build on

You are assessing the capabilities and limitations of an existing AI product

You need to understand the difference between the model and the product built on top of it

You are making decisions about model selection, switching, or evaluation

Key takeaway

Understanding foundation models helps you understand both the power and the inherited limitations of the AI features you build or use.

How it works

Understand the basic mechanism. Foundation glossaryModelA model is a system or representation used to process data and generate outputs, often trained to perform specific tasks.Open glossary term are trained on enormous glossaryDatasetA dataset is a structured collection of data used for analysis, training models, or processing.Open glossary term — often hundreds of billions of words of text — using a glossaryProcessA process is a defined sequence of steps used to achieve a specific outcome.Open glossary term called self-supervised learning. This produces a model with broad general capabilities.

From there, the glossaryModelA model is a system or representation used to process data and generate outputs, often trained to perform specific tasks.Open glossary term can be used directly through an API, fine-tuned on specific glossaryDataData is raw information collected and stored for analysis, processing, or decision-making.Open glossary term to improve glossaryPerformancePerformance refers to how quickly and efficiently a system responds to user actions and processes tasks.Open glossary term on a target task, or extended with tools and retrieval systems to handle specialised use cases.

The glossaryModelA model is a system or representation used to process data and generate outputs, often trained to perform specific tasks.Open glossary term provider — OpenAI, Anthropic, Google, and others — is responsible for training, maintaining, and updating the foundation model. Product teams glossaryBuildA build is the process of compiling and packaging code into a runnable application.Open glossary term on top of it, often without full visibility into exactly how it was trained or what glossaryDataData is raw information collected and stored for analysis, processing, or decision-making.Open glossary term it contains.

What this means for designers and product teams. Building on a foundation glossaryModelA model is a system or representation used to process data and generate outputs, often trained to perform specific tasks.Open glossary term means inheriting both its strengths and its limitations. The model's knowledge cutoff, glossaryBiasBias is a systematic distortion in thinking or data that affects the accuracy of research or decision-making.Open glossary term, and failure modes become your product's knowledge cutoff, biases, and failure modes unless you actively design around them.

glossaryModelA model is a system or representation used to process data and generate outputs, often trained to perform specific tasks.Open glossary term providers update foundation models over time, which can change glossaryBehaviourBehaviour refers to how users interact with a system, including actions, patterns, and responses.Open glossary term. A product that works reliably today may behave differently after a model update — making monitoring and evaluation an ongoing responsibility.

What to look for

Focus on:

Model capabilities — what the model does well and where it struggles

Knowledge boundaries — what the model knows and what lies beyond its training cutoff

Inherited limitations — what biases or failure modes are present in the model

Provider reliability — the track record and roadmap of the model provider

Update policies — how model updates are communicated and managed

Where it goes wrong

Most issues come from: Treating a foundation glossaryModelA model is a system or representation used to process data and generate outputs, often trained to perform specific tasks.Open glossary term as a black box and hoping for the best is not a glossaryProduct StrategyProduct strategy defines how a product will achieve business goals by solving user problems in a focused and sustainable way. It sets direction, priorities, and trade-offs to guide decision-making.Open glossary term.

No ongoing evaluation after initial deployment

Assuming model updates will only make things better

No testing for inherited bias or hallucination risk

Choosing a model based on benchmark performance rather than real-world task performance

No contingency plan if the model provider changes access, pricing, or terms

What you get from it

Understanding foundation glossaryModelA model is a system or representation used to process data and generate outputs, often trained to perform specific tasks.Open glossary term gives you:

Clearer expectations about what AI features can and cannot do reliably

A better framework for model selection and evaluation

Awareness of the risks that come with building on models you do not fully control

More informed conversations with engineers and leadership about AI product strategy

Key takeaway

Foundation models are powerful but not neutral. Knowing what you are building on — and what it comes with — is the foundation of responsible AI product design.

FAQ

Common questions

A few practical answers to the questions that usually come up around this method.

What is a foundation model?

A foundation glossaryModelA model is a system or representation used to process data and generate outputs, often trained to perform specific tasks.Open glossary term is a large AI model trained on broad glossaryDataData is raw information collected and stored for analysis, processing, or decision-making.Open glossary term that can be adapted to a wide range of tasks. Rather than training a new model for each use case, most products are built on top of an existing foundation model from providers like OpenAI, Anthropic, or Google.

How is a foundation model different from the AI product built on top of it?

The foundation glossaryModelA model is a system or representation used to process data and generate outputs, often trained to perform specific tasks.Open glossary term is the underlying glossaryCapabilityCapability refers to an organisation’s ability to perform a specific function or deliver a particular outcome.Open glossary term — the trained glossarySystemA system is a collection of interconnected components that work together to achieve a specific function or outcome.Open glossary term that understands and generates language. The product built on top of it adds structure, tools, prompts, and design to make that capability useful for a specific purpose. The model provides the intelligence; the product provides the context and application.

Can I switch foundation models?

Technically yes, but it is not trivial. Different glossaryModelA model is a system or representation used to process data and generate outputs, often trained to perform specific tasks.Open glossary term have different glossaryCapabilityCapability refers to an organisation’s ability to perform a specific function or deliver a particular outcome.Open glossary term, glossaryBehaviourBehaviour refers to how users interact with a system, including actions, patterns, and responses.Open glossary term, and failure modes. Switching models requires re-evaluation of your prompts, system design, and quality standards. It is worth planning for the possibility from the start rather than building in a way that makes switching difficult.

What happens when a foundation model is updated?

The glossaryModelA model is a system or representation used to process data and generate outputs, often trained to perform specific tasks.Open glossary term provider glossaryReleaseA release is the point at which a product or feature is made available to users. It marks the transition from development to real-world use and often involves deployment, communication, and monitoring.Open glossary term a new glossaryVersionA version is a specific iteration of software or a product at a point in time.Open glossary term with different performance characteristics. This can improve outputs in some areas while changing behaviour in others. Monitoring your AI feature's performance after model updates and testing before switching to a new version are both essential.

Are smaller or open-source models a viable alternative?

For some use cases, yes. Smaller glossaryModelA model is a system or representation used to process data and generate outputs, often trained to perform specific tasks.Open glossary term can be faster, cheaper, and easier to deploy privately. Open-source models offer more control and transparency. The trade-off is typically glossaryCapabilityCapability refers to an organisation’s ability to perform a specific function or deliver a particular outcome.Open glossary term — larger proprietary models tend to outperform smaller alternatives on complex tasks, though that gap is narrowing.

Quick take

Most AI products are not building models — they are building on top of foundation models. Understanding what that means changes how you design, evaluate, and manage risk.

Related Services

Artificial Intelligence