Before foundation models, building an AI system for a specific task usually meant training a model from scratch for that task alone. Foundation models flip that: one very large, generally capable model is trained once, on broad data, and then adapted, through prompting, retrieval, or fine-tuning, to many different downstream tasks without starting over each time.
This is why the same underlying model can draft a contract clause, summarise a meeting, and write code, without three separate systems being built. The tradeoff is that “generally capable” isn’t the same as “correct for your specific, regulated, high-stakes use case” out of the box.
For a buyer, the decision that matters isn’t usually which foundation model is largest; it’s which one, combined with the right retrieval, guardrails, and evaluation, actually performs on your task at an acceptable error rate and cost.