What is Foundation Model?

Foundation models are the building blocks of modern AI applications. Rather than training a model from scratch for each task, which would require enormous data and compute resources, businesses start with a foundation model and adapt it to their needs. This is what makes AI accessible to mid-market firms rather than only the largest enterprises.

The foundation model landscape is evolving rapidly. OpenAI, Anthropic, Google, Meta, and Mistral all offer models with different strengths, pricing structures, and deployment options. For regulated UK businesses, the choice of foundation model involves more than capability. You need to consider where the model runs, what happens to your data, what contractual protections exist, and whether the provider offers deployment options that meet your compliance requirements.

A key distinction is between proprietary and open-weight models. Proprietary models like GPT-4 and Claude are accessed through APIs and run on the provider infrastructure. Open-weight models like Llama and Mistral can be downloaded and run on your own infrastructure, giving you complete control over data handling. Both approaches have valid use cases in regulated industries, and many firms use a combination.

For client-affecting applications in financial services or healthcare, firms often prefer models that can run within their own cloud tenancy, ensuring data never leaves their controlled environment. For internal productivity tools with lower data sensitivity, API-based access to proprietary models may be appropriate and more cost-effective.

The practical advice for mid-market firms is to avoid betting on a single foundation model. The technology is moving fast, and today’s leading model may be superseded in six months. Build your applications with an abstraction layer that allows you to swap the underlying model without rebuilding the entire system. This gives you flexibility to adopt better models as they emerge and protects your investment in the surrounding workflow, governance, and integration work.