Quanta Magazine published an essay this week asking whether “attention,” the mechanism behind every large language model, is the essence of intelligence. Its central example shows how a model resolves the word it in “the animal didn't cross the street because it was too tired.” It weighs every other word in the sentence to work out that it means the animal.

Attention is remarkable at working out what words refer to. What it doesn't do on its own is ask whether a request makes sense for you. The essay is careful not to claim these models understand anything the way we do.

That gap is where most enterprises are stuck today.

The rain monitor in the desert

Ask a general-purpose model how to set up a rain monitor in a desert where it hasn't rained in decades and none is forecast. It will identify “rain monitor” as the entity, ask which model you have, and give you a clean step-by-step guide.

An expert would ask a different question first: why do you need one? With no rain and none expected, the right answer is eliminate. Don't do the work.

Now add one fact. The site runs an artificial irrigation system, and natural rain, though rare, would waste water if irrigation kept running. The same request now means automate: detect real rainfall and shut off the irrigation valve. And if the valve closes but the rain wasn't enough for the crop, that is human assist: someone needs to make a judgment call.

The words in the request never changed. The right action changed three times. Three things decided it, and none of them were in the language:

  • Intent: why the work exists at all.
  • Context: the facts surrounding it, such as the irrigation system.
  • Specificity: what is true for this site, this crop, this business.

Every enterprise is full of rain monitors. Consider a duplicate payment that is really an intentional split, a maintenance alert on equipment already scheduled for retirement, or a compliance task that one entity must do and another is exempt from. Fluency handles the words. Only intelligence specific to the enterprise gets the action right.

How we learned this

We started with classic machine learning. We classified work into three dispositions: eliminate, automate, human assist. The models got it right when new requests resembled the training data and failed when the wording changed. We retrained, and the cycle repeated.

Then LLMs arrived, and we expected them to close the gap. They didn't. They understood the words but not the work, at least not the way a subject-matter expert does.

So we changed what we trained on. Beyond what the request says, we trained on why it should be eliminated, automated or assisted. Then we used the same training loops for the what, when, where and how that turn a decision into action.

The same pattern held when we applied it to software delivery, which I wrote about earlier. Intelligence came from the organization's own history, not from the model's general knowledge.

The industry is compensating from the outside

The industry is starting to see this gap. A recent HackerNoon simulation by Rick Bahague of a governed enterprise agent shows the pattern well. It starts from one business question and uses a knowledge graph for topology, a vector index for policy context, certified data, and an evidence trail behind every number. I agree with almost all of it, especially starting from the question rather than the data lake.

But look at how it handles meaning. A hand-curated ontology and a metric registry tell the agent what each term means. The build describes this as resolving the user's intent. It is the intent of the question, not the intent of the work. It tells the agent what the user is asking. It does not tell the system why a pattern exists in this enterprise, or whether the work should be eliminated, automated or handed to a human. The actions it recommends are retrieved from policy documents, not learned from the enterprise's own outcomes.

That is the same wall our disposition classifiers hit years ago. Rules and definitions hold until the wording or the context shifts. Understanding has to be learned from the enterprise's own history, and it has to end in a decision, not just an answer.

Why this has to be built, not bought

A frontier model is trained on the internet. It knows what is generally true. Your enterprise runs on what is specifically true: your exceptions, your history, your reasons. That knowledge sits in people's heads, in decades of decisions, and in data nobody has connected.

This is why open-weight models matter. You can train them on your own intent, context and specificity. The result appears to reason because it has learned your causal structure, validated against your outcomes.

A rented model attends over the world. Enterprise intelligence has to attend over your business.

What this means for leaders

Frontier models give every enterprise the same fluency. What separates enterprises is whether their AI knows why the work exists, what surrounds it, and what is true only for them. That isn't a feature you buy. It is intelligence you build.

The most valuable question in enterprise AI isn't “how do I do this?” It is “why are we doing this, here, at all?”

This is the next article in my enterprise intelligence series. Earlier: how we built the platform and why enterprise intelligence hasn't started yet.