Foundation Model
What a foundation model is
The term foundation model was introduced to describe a new class of AI system that emerged from training at unprecedented scale. Rather than being built for a specific task, a foundation model is trained on a vast and varied dataset, developing general capabilities across language, reasoning and knowledge that can then be applied to a wide range of downstream tasks.
GPT-4, Claude, Gemini and Llama are examples of foundation models. Each was trained on enormous quantities of text, code and other data, and each can perform competently across a broad range of tasks without being specifically designed for any of them. This generality is what makes them foundational: they provide a capable base from which specialised systems are built, either through fine-tuning, through carefully designed prompts, or through integration with retrieval systems and other tools.
The foundation model concept represents a significant shift in how AI systems are developed. Previously, building a capable AI system for a specific task required a substantial development effort specific to that task. With foundation models, much of the underlying capability already exists. The development effort becomes one of adaptation and integration rather than construction from scratch.
Choosing between foundation models
For organisations building AI systems, the choice between foundation models is a technical decision that should follow the architectural design, not precede it. Different foundation models have different strengths: some perform better on reasoning tasks, some on code generation, some on following complex instructions, some on operating in specific languages or domains. The right model for a given application depends on what the application actually requires, which can only be determined once the application has been properly designed.
Foundation models are also developing rapidly. A model that is clearly superior today may be superseded within months. This is one of the reasons why the architectural layer, which defines what the system needs to do, is more durable than the model selection layer, which defines which tool currently does it best. A well-designed architecture can accommodate a change in foundation model without requiring a redesign of the system it supports.
Foundation models and organisational knowledge
A foundation model has no knowledge of your organisation, your clients, your documents or your internal processes. Its training data is general and has a cutoff date. This is not a limitation of any particular model. It is an inherent characteristic of how foundation models are built. Connecting a foundation model to organisational knowledge through RAG, or specialising its behaviour for a specific domain through fine-tuning, is how it becomes genuinely useful rather than generically capable.
Insight
For professional services organisations, the practical implication of foundation models is that the general capability they provide is a starting point, not a finished product. A foundation model that can draft a contract is not the same as a system that can draft a contract according to your firm's house style, drawing on your precedent library, flagging clauses against your current risk thresholds, and producing an output that meets your quality standards.
The foundation model provides the language capability. The architecture around it provides the specificity that makes it professionally useful.
Misconceptions
Foundation models are not interchangeable. Different models have different strengths, weaknesses and cost profiles. The choice of foundation model is a technical decision that should be informed by the specific requirements of the application being built.
A more capable foundation model does not automatically produce better outcomes. The quality of the system built around the model matters as much as the model itself. A well-designed system using a less capable model will often produce better results than a poorly designed system using a more capable one.
Foundation models do not remain state of the art for long. Building AI systems that are architecturally independent of any specific model ensures they can be updated as better options become available without requiring a full rebuild.