Author - Juan Morán
What if the biggest limitation of artificial intelligence is not how much it knows, but its inability to learn from what happens when it acts?
For centuries, Western thought was dominated by Aristotle's approach to knowledge: start with accepted premises, apply logical reasoning, and reach conclusions. Francis Bacon changed that by insisting that conclusions had to be tested against reality. Knowledge was no longer just about reasoning, but about experimenting, observing results, and correcting mistakes.
I think this distinction offers a useful way to understand where we are with AI, particularly in the corporate world.
We have spent years and hundreds of billions of dollars building increasingly powerful language models. They can reason, write, analyze, and explain almost anything. But when companies deploy them, something seems to be missing. They can produce extraordinarily sophisticated outputs, but they have no inherent mechanism for determining whether those outputs actually improved the business.
An individual using ChatGPT provides that missing mechanism almost unconsciously: we evaluate answers, correct mistakes, reformulate questions, and decide whether the result is useful. But a company cannot rely on someone manually steering millions of simultaneous interactions. It needs objectives, feedback, and continuous improvement built into its operating architecture.
The comparison with AlphaZero is particularly revealing. DeepMind's system achieved superhuman performance in chess without studying centuries of human strategies. It learned by playing, evaluating outcomes, and improving. Meanwhile, language models trained on enormous amounts of human knowledge about chess can explain sophisticated strategies while still struggling to execute them reliably.
The difference is not simply intelligence. It is the ability to optimize toward a measurable objective.
Now consider what happens inside a company. Every customer interaction, sales conversation, supplier negotiation, pricing decision, or support request produces information about what works and what doesn't. Yet most organizations simply process these interactions and move on. Even when AI agents are involved, they usually execute workflows rather than systematically improve them.
What if every interaction became an experiment, and its outcome helped improve every subsequent interaction?
This would require something fundamentally different from adding more agents or choosing a more powerful model. The company would need a persistent representation of its operations, structured records of actions and outcomes, clearly defined reward functions, and an optimization mechanism capable of learning continuously. Models and agents would become replaceable components of a larger system that accumulates organizational experience.
There is also an important managerial implication. Optimizing conversion, revenue, or efficiency is relatively straightforward to describe, but what happens when those objectives conflict with profitability, privacy, fairness, or long-term relationships? Defining what constitutes success, and what must never be sacrificed to achieve it, remains a fundamental leadership responsibility.
The most valuable AI infrastructure a company builds may not be the model it uses, or the number of agents it deploys, but the system that allows it to learn from its own operations.
