Where intelligence lives

Most conversations about AI in organizations begin with the choice of a model: which one reasons best, which one costs least, which one the company should standardize on. The question is reasonable, and the answer will keep changing. A model adopted this year will be replaced, upgraded, or combined with others, often more than once, before the decisions it helped make have fully played out.

That rate of change raises a prior question: where should the intelligence of an organization live? Not the capacity to generate an answer, which can be rented, but what makes an answer belong to this organization: what its words mean, who is allowed to decide what, what was done before and what came of it. If those live inside a particular model, the organization's intelligence is replaced whenever the model is.

My answer is that the durable layer is not the model but something closer to an operating system: a layer that preserves meaning, causality, authority, and memory across models. This is not an argument for building a foundation model. It is an argument about what should not be rented.

Models generate; agents act

It helps to separate three roles that are easily collapsed into one. A model generates: given context, it produces text, code, a plan, a classification. An agent acts: it takes a goal, uses tools, and changes something in the world. Neither of them, by default, keeps what an organization needs to remain itself over time—the identity of the things it deals with, the context in which a choice was made, and the consequences that followed.

An operating system is a useful analogy for the third role, as long as it is taken precisely. An operating system does not compute a program's answer. It gives programs a stable environment: identities, permissions, memory that persists after a process ends, and interfaces that let hardware and programs change without everything else being rewritten. Its value lies in what it holds constant while its components change.

The same division applies here. Models are closer to processors: indispensable, improving quickly, and interchangeable in principle. Agents are closer to programs. What most organizations lack is the layer beneath both, the one that says what things are, who may do what, and what happened. Without it, every new model and every new agent starts from a prompt, and the prompt quietly becomes the place where the organization's meaning is stored.

The swap test

A simple test shows where an organization's intelligence currently lives. Suppose the model underneath every workflow were replaced tomorrow with another of similar capability. What would survive unchanged? The definitions of key terms, the boundaries of authority, the record of past decisions and their outcomes? Or would all of that have to be rediscovered from the behavior of the new model?

Consider a hypothetical company that uses an agent to prepare contract renewals. Over a year, what counts as “ready for renewal,” which discounts a manager may approve alone, and which earlier discounts preceded a customer leaving all end up written into a long prompt, tuned by trial and error to one model. Then the model is replaced. The new model reads the same prompt and interprets the discount rule slightly differently. Nothing fails loudly. Proposals change, approvals continue, and no one can say which meaning moved, because the meaning was never stored anywhere except in the interaction between one prompt and one model.

Now suppose the same company had kept those things as records the agent consults rather than instructions it absorbs: a defined status for renewal readiness, an explicit authority boundary for discounts, and a linked history of past renewals and their outcomes. Replacing the model becomes an experiment instead of a reset. The question is no longer whether the new model can be coaxed into behaving like the old one, but whether, given the same meaning and the same authority, it proposes better renewals. The organization can compare, because what it knows did not move with the model.

What the operating layer holds

The layer has to hold four things, and each has been the subject of an essay on this site. The first is meaning: the entities, relationships, statuses, and evidence that carry operational consequence. Meaning Before Automation argued that this must be a small, living agreement rather than a perfect ontology. In the operating layer, that agreement becomes the environment every model reads from, instead of a paragraph each prompt restates.

The second is authority: which actions may be observed, prepared, executed reversibly, or only authorized by a named person. Human Authority in the Age of Agents argued that authority is a system, not a button. A system of authority cannot be re-derived from whichever model happens to be running; it has to sit outside every model and apply to all of them. The third is memory: decisions with their rationale, owner, execution, and outcome, including the questions that remain open. Building Companies That Learn called this memory with consequence.

The fourth is causality: not only what happened, but which action it followed and how that action may have shaped what was observed. Does a Prediction That Allocates Observe the Future? showed how an allocation can manufacture the very evidence later counted as accuracy. A model cannot tell this from its training data or its context alone. The operating layer can at least record which decision changed which allocation, so that the next model is not trained on the previous one's intervention as though it were the world.

When the model is enough

The strongest objection is that model providers will absorb this layer. Models already come with longer context, persistent memory, and built-in tools. If that continues, a separate operating layer may look like duplicated infrastructure: expensive to build, and likely to fall behind what the providers ship.

In some cases the objection wins. When the work carries little consequence, when the meanings involved are generic rather than specific to one organization, when switching models is unlikely, or when mistakes are cheap to reverse, the model and its built-in memory are enough. Building an operating layer there would be ceremony.

But the argument is not about where the data is stored. It is about who owns the meanings, and whether they can be inspected, corrected, and carried elsewhere. If a provider's memory lets an organization see what it holds, change it deliberately, attach authority to it, and take it along when it leaves, then that memory is already functioning as the operating layer; it is simply hosted by someone else. The question to ask of any platform is not how much it remembers, but whether what it remembers remains the organization's own.

A layer that can be corrected

An operating layer can fail in its own ways. It can harden into a bureaucratic ontology that no one updates, so that the organization's official meanings drift away from its actual work. It can become a new kind of lock-in, merely moved from the model provider to the layer. And by centralizing meaning, it can erase the local differences through which people notice what shared definitions miss. The Evolution of Machines and the Human Future described a common ground that makes distinctiveness possible rather than demanding sameness; an operating layer that imposes one meaning on every team works against it.

So the layer has to be built for correction. Each meaning should record where it came from and who may change it. Disagreements should stay visible as disagreements instead of being settled by a default. Where consequences differ, local variants should be allowed. And the layer itself should be observable, so that it can be tested and replaced as honestly as a model can.

Models will keep improving, and the organizations that use them should welcome each improvement. That is exactly why the part that must not change with them should be built deliberately. What an organization means, who may decide, what it did, and what that caused are not features of a model. They are the organization's own intelligence, and they belong in a layer that outlasts any model running on top of it.

Back to writing