Foundation Model
What vendors mean: a powerful general-purpose AI. What it actually means: the engine underneath their product, which they almost certainly didn't build.
The Technical Definition
A foundation model is a large AI model trained on a broad dataset that can be adapted to many downstream tasks. The name was popularized by Stanford’s CRFM in 2021 to describe a new pattern: instead of training a fresh model for each application, you train one massive general-purpose model and then specialize it through fine-tuning, prompting, or retrieval.
GPT-4, Claude, Gemini, Llama, and Mistral are all foundation models. They were trained at enormous expense — hundreds of millions of dollars in compute — on internet-scale text. Almost every enterprise AI product you’re being pitched is a thin layer on top of one of them.
What This Actually Means for Your Business
When a vendor demos their AI product, you’re looking at three things stacked together: a foundation model someone else trained, a layer of prompts and retrieval the vendor wrote, and a user interface. The vendor is selling you the second and third layers. The first one — the part that’s actually doing the heavy intellectual work — was built by Anthropic, OpenAI, Google, Meta, or Mistral.
This matters for several reasons that don’t show up in the sales deck.
First, the vendor’s product inherits the strengths and weaknesses of its underlying foundation model. If they’re using a cheap older model to keep their margins fat, your users get an experience that feels dated against ChatGPT’s free tier. If they’re using a top-tier model, the unit economics they showed you may not survive contact with your actual usage volume.
Second, “proprietary AI model” is one of the most overused phrases in enterprise software right now. In nine cases out of ten, it means one of three things: they’ve fine-tuned an open-source foundation model on their own data, they’ve wrapped a commercial foundation model with custom prompts, or they’ve stitched together calls to multiple foundation models behind a single interface. None of those is bad. But none of those is a moat. Their moat — if they have one — is in their data, their workflow integration, or their domain expertise. Not the model.
Third, foundation models change. Anthropic ships a new Claude. OpenAI ships a new GPT. Quality jumps, prices drop, capabilities shift. A vendor who has built carefully on top of foundation models can pass those gains through to you. A vendor who has hardcoded one model into their product cannot. Ask which one you’re buying.
The lock-in question is also real. If your vendor is built on GPT-4 and OpenAI changes its terms or pricing, your vendor’s economics break — and so do yours. If your vendor is built on Llama or Mistral and self-hosts, that risk shifts to infrastructure cost. Neither is wrong. But they’re different bets.
Reality Check
What the vendor says: “We’ve built a proprietary AI model trained specifically for your industry.”
What that means in practice: They’ve taken Llama or fine-tuned an OpenAI deployment with industry-specific prompts and a curated dataset. The base model is not theirs. Whether the fine-tuning is meaningful depends entirely on what data they used and how much of it — questions you should be asking.
What Operators Actually Do
The companies that buy AI well do two things upfront. They ask which foundation model sits underneath, and they ask what happens when a better one ships. The answer they want is “we can swap with minimal rework.” The answer that should worry you is “we don’t disclose that” or “our system is too tightly coupled to migrate easily.”
Smart operators also test foundation models directly before buying anything built on top of them. Spend a weekend with Claude and GPT-4 and a representative sample of your real work. You’ll learn more about what’s possible than from any vendor demo. You’ll also discover that, for a meaningful percentage of use cases, the vendor’s wrapper is doing less than you assumed.
The other pattern: keep model choice optional. Architectures that abstract the foundation model behind a service layer let you change vendors without rebuilding the application. That’s worth more than any single point-in-time benchmark.
The Questions to Ask
-
Which foundation model is underneath, and why that one? If they won’t say, that’s the answer. If they say “proprietary,” ask what was proprietary about its training and who paid for the GPU run.
-
What happens when a better model ships? Can they swap the underlying model without breaking the integration? How long would migration take, and who pays for it?
-
What did you actually build on top? The interesting layer is fine-tuning data, retrieval, prompts, and workflow logic — not the base model. Make them defend their actual contribution.