AI models access to a company’s software, data, and workflows is not enough to make the business truly optimizable. AI may understand general concepts such as management, marketing, and logistics, but it usually lacks a detailed representation of a specific company’s customers, dependencies, internal processes, exceptions, permissions, risk tolerance, and the consequences of changing one part of the organization. The key distinction is between memory and a real data model: memory can recall what happened, while a model can represent entities, relationships, rules, constraints, and valid states.
The author says that context is not the same as understanding. Companies currently try to improve AI performance by giving models more context, but context only tells an AI what information to consider at a particular moment; it does not explain how the organization actually works. The proposed solution is an organizational ontology—essentially a structured map of the company containing its “nouns” and “verbs.” Employees, customers, factories, and orders become entities, while actions such as changing prices, launching promotions, or executing workflows become operations. This could function like a digital twin of the organization.
The next step is a world model that goes beyond describing the organization and learns how it behaves. Such a model could predict the consequences of decisions and help AI reason about cause and effect. For example, reducing customer-service call times might appear to lower costs, but it could also increase customer churn. The article argues that AI-driven optimization therefore needs causal reasoning rather than simply finding correlations. Companies could continuously feed decisions and their outcomes back into the model, allowing the organization to accumulate its own institutional learning over time.
Ultimately, the article suggests that the competitive advantage of corporate AI may not come from having access to the same frontier models as everyone else. Two companies might use the same ChatGPT, Claude, Qwen, or other model, but the company with the better proprietary representation of itself—its rules, history, relationships, decisions, and accumulated learning—could have a major advantage. The future, therefore, is not simply about making companies “optimizable”; companies first need to become understandable to AI.
