Every enterprise now has an LLM strategy. Copilots for the engineers, chat over the wiki, agents triaging tickets. All of it useful, and almost none of it touching the questions the business actually runs on: Which customers are about to leave? Which loans will default? Which claims are fraudulent? What will demand look like in March?
That's not a maturity gap that the next model release will close. It's a data gap, and it's structural.
The signal isn't in the text
LLMs are pretrained on public text, so what they encode is the world's written knowledge: language, code, reasoning patterns, facts everyone shares. But the information that differentiates your company from every other company was never written down as prose. It lives in structured event streams, transactions, orders, claims, sessions, payments. Millions of rows, each one a decision a customer made.
A bank's competitive knowledge is not in its documents; it's in ten years of transaction histories. A retailer's is in orders, carts, and returns. No amount of clever prompting gives a text model access to patterns it has never seen, and that data has never been on the internet, so no frontier model has ever trained on anything like it.
Your moat was never going to be a model everyone can rent.
Why RAG and agents don't close the gap
The standard answer is retrieval: connect the LLM to your data warehouse and let it look things up. That works for questions whose answers already exist as records, "what did this customer buy last month?" It does nothing for questions whose answers must be inferred from patterns across millions of sequences, "which of my two million customers will churn this quarter?"
Retrieval fetches; it doesn't learn. An LLM reading one customer's history through a context window is doing the equivalent of a doctor meeting a patient with no medical training, the record is there, but the prior knowledge that makes the record meaningful is not. And running even that weak inference across an entire customer base, continuously, is computationally absurd. Prediction at enterprise scale is a training problem, not a prompting problem.
Text models and behavior models solve different problems
The point is not that LLMs are overrated. It's that "AI for the enterprise" is two different jobs:
- Language work, drafting, summarizing, answering, coding. General knowledge applied to your documents. LLMs, rented from a lab, are the right tool.
- Behavioral prediction, churn, risk, fraud, demand, lifetime value. Company-specific knowledge extracted from your event history. This needs a model that has actually been trained on that history.
The second job is what an enterprise foundation model does: a transformer pretrained with self-supervision on your own event sequences, learning your customers' grammar the way an LLM learned language's. Same architecture, same training objective, applied to the data that actually runs your business.
The irony of the LLM era
LLMs did enterprises one enormous favor: they proved, publicly and irrefutably, that pretraining works, that a single self-supervised model beats a thousand hand-built specialist systems. The irony is that most companies absorbed the lesson only for text, the one modality where their data adds the least.
The companies that internalize the full lesson will pretrain on the data where their advantage actually lives. Their churn models, risk models, and fraud models will share one brain that no competitor can rent, because it was trained on data no competitor has. That asset compounds: every new prediction gets cheaper and better as the foundation improves.
Rent your language model. Own your behavior model. The first makes your teams faster; the second is the one that knows your business.