A foundation model is a very large AI model trained on a huge, varied body of data so it can handle many tasks without being built for any one. Large language models such as those behind popular chat assistants are the best-known examples, and image and multimodal models also fit the label.
Training one costs vast amounts of computing power, so only a handful of organisations build them. Everyone else adapts them through prompting, retrieval or fine-tuning, or accesses them through an API. The same underlying model can write copy, summarise a contract and answer questions.
Most AI products a retailer or restaurant uses, including chatbots, forecasting add-ons and POS assistants, are built on top of a foundation model from a major provider. Understanding this helps you see that the vendor's product is usually the layer around the model: your data, rules, integrations and support.
Dependence has consequences. If the underlying model changes, is retired or its pricing shifts, the product built on it can change too. Ask vendors which models they use, whether they can switch providers, where data is processed and how updates are tested before reaching customers.
Foundation models are why AI capability has become accessible to small businesses: you rent a general capability rather than building one.