I have been writing for several months about what I see more and more as the central problem in enterprise AI. I’m seeing it not just from an academic perspective: Of course, I’m a university professor with more than 30 years of experience, but I’m also the director of innovation of an artificial intelligence startup, and that’s teaching me a whole lot of new skills. Traveling from theory to code and back, day in and day out, is proving to be an amazingly enriching journey.
First of all, we should know by now that large language models are extraordinarily capable, but they were not designed to run companies. They are great at generating answers. But organizations require other things: persistent state, formal structures, permissions, feedback loops, measurable objectives, and the ability to learn from outcomes.
The more these ideas travel from essays and diagrams into working software, the clearer the problem appears. The concepts that sound convincing in prose tend to collapse when someone has to express them in code. “Memory” turns out not to be a data model. “Autonomy” means little without permissions. “Learning” is not the same thing as simply accumulating context. And “optimization,” as we all know, becomes dangerous unless somebody defines exactly what the system is being asked to optimize.
This is not simply a technical problem. It is an epistemological one: The AI industry has built Aristotelian machines. What enterprises need now is a Baconian approach.
The most powerful Aristotelian machines ever created
For more than 2,000 years, Western thought has been shaped by Aristotle’s extraordinary influence on logic and deduction. Start with premises, reason correctly from them, and jump to a conclusion.
The syllogism we study in Philosophy 101 is the classic example: All humans are mortal; Socrates is human; therefore, Socrates is mortal. If the premises are correct and the reasoning is valid, the conclusion follows.
All the large language models we know so far operate in a surprisingly similar way. They start by absorbing vast quantities of recorded human knowledge that was on the internet, transform this into vectors, and generate the most statistically coherent continuation from what they have learned. In doing so, they can reason, compare, summarize, infer, explain, and combine ideas across domains with quite remarkable sophistication.
However, they remain enclosed within the information available to them. They do not inherently observe what happens after an answer is produced. They do not test whether a recommendation worked. They do not revise their operating structure when a customer leaves, a sales campaign fails, or an apparently efficient policy causes an unexpected problem elsewhere.
Think hallucinations. When a model hallucinates, it is not necessarily malfunctioning; it is producing a plausible conclusion from imperfect premises, but without any structural mechanism to check that conclusion against reality. OpenAI’s own research connects hallucinations to next-word prediction and to training systems that reward guessing, rather than acknowledging uncertainty.
That is a profoundly Aristotelian condition.
Bacon was the one who changed the architecture of knowledge
Francis Bacon’s contribution, in the sixteenth and seventeenth centuries, was not simply to argue that observation mattered. This has already been noted by many people who had been observing the world and drawing conclusions from it long before Francis Bacon.
His deeper, and phenomenal, contribution was to place observation inside a repeatable learning structure: Formulate an idea, test it against reality, observe the result, revise the hypothesis, and repeat. His method rejected the primacy of syllogistic demonstration in favor of a process that moved from observations to principles and back again to new experiments and practical results.
The scientific revolution did not emerge because Bacon was necessarily more intelligent than Aristotle, but because the Baconian method organized intelligence in a different way. It allowed discoveries to be tested, errors to be exposed, results to accumulate, and knowledge to compound across people and generations. The decisive innovation was the loop.
This distinction is extremely important for AI. Today’s frontier models are our Aristotles: brilliant individual engines capable of extraordinary, superhuman inference. But the next step is not just to build a larger Aristotle with more parameters, more training data, and a longer context window, but to build the Baconian structure and loop around it.
Companies need experiments, not just answers
Most generative AI still operates as an open-loop system. A person asks for something, the model produces it, and the interaction ends right there.
That works extremely well for individual tasks such as drafting an email, summarizing a report, explaining a concept, generating some ideas, translating a document . . .
However, as anyone with a minimum of corporate experience can understand, companies are not collections of isolated requests. They are systems of actions and consequences. As I previously argued, enterprise AI must move from answers to outcomes, from prompts to constraints, and from copilots to systems of action.
A model may recommend changing a price, modifying a sales script, prioritizing certain customers, reorganizing a support workflow, or altering an approval process. But the quality of that recommendation cannot ultimately be judged by how persuasive it sounds. It must be judged by what happens next.
After the model’s recommendation, did conversion improve? Did churn increase? Did margins rise? Did customer trust decline? Did the new process save time in one department while creating additional work elsewhere? Every enterprise action is, whether management recognizes it or not, an experiment.
A real Baconian enterprise AI system would treat it that way. It would connect actions to outcomes, outcomes to objectives, and objectives to future behavior. It would not merely generate a recommendation, but also observe its consequences and, more importantly, learn from them.
As many are starting to realize, the unit of value would no longer be the answer; it would be the loop.
Why implementation changes the theory
This is where translating ideas into software becomes intellectually useful. As I realize when I move from academia into real-world implementation, code is much less tolerant of ambiguity than prose.
It is easy to write that an AI system should “learn from the company.” But a software architect must ask what constitutes an observation, where it is stored, which events matter, how outcomes are attributed to previous decisions, and what happens when you have multiple conflicting objectives.
It is easy to say that an agent should act autonomously. An implementation must define which objects it may access, which functions it may invoke, what it may modify, when human approval is required, and who’s going to remain accountable.
It is easy to advocate continuous optimization. But optimize for what? A customer service system rewarded for reducing handling time may learn, for instance, to terminate all conversations as fast as possible. A hiring system rewarded for retention may tend to select simple conformity. A sales system rewarded for conversion may discover techniques that work commercially, while eroding trust.
A loop can be wrong and compound, over and over again. That is why enterprise AI cannot be separated from governance. Every reward function encodes a theory of what matters, every constraint expresses an institutional decision, every permission boundary allocates authority. These are not simply engineering choices.
These are corporate choices expressed in software. The AI risk management framework from the National Institute of Standards and Technology similarly treats governance as a continuous, cross-cutting activity that must connect technical design to organizational policies, values, responsibilities, monitoring, and measurable outcomes throughout the system’s lifecycle. We are clearly well above the pay grade of our usual chatbots.
From generation to learning: The real transition
The industry continues to focus its attention on model capability: which model reasons better, codes faster, uses more context, or achieves the highest benchmark score. And those improvements matter . . . but they are just improvements in individual intelligence.
The enterprise opportunity lies somewhere else: in collective intelligence, in the architecture that coordinates models, people, data, processes, objectives, and feedback so that the organization itself becomes better at achieving outcomes. That is exactly why the model should not be the company’s durable asset. It should be the learning loop surrounding it.
While generative AI produces, Baconian AI learns. The first one creates content from accumulated premises, but the second one acts within a defined environment, observes what changed, and incorporates the result into the next decision. Feedback is all you need.
And when you think about it, this is not an incremental feature to add to a chatbot, but a different way of understanding what enterprise AI is for. Ultimately, it requires companies to become not simply “automated,” but formally represented and optimizable.
This future will not be built by eliminating the Aristotelian machine; deduction still remains enormously valuable. These large language models will help us generate hypotheses, interpret situations, propose actions, and make complex systems accessible through language. But they need to operate inside an architecture in which reality, instead of just eloquence, becomes the final arbiter.
Enterprise AI does not need another Aristotle.
It needs its Bacon.