The Fast Company article argues that the biggest challenge in enterprise AI is no longer simply building more capable language models. Large language models are excellent at generating answers, reasoning over information and proposing actions, but companies need systems that can operate inside real business environments—with persistent state, permissions, measurable objectives, feedback mechanisms and the ability to learn from what actually happens. The author describes today's LLMs as “Aristotelian” machines: they reason from existing information, but generally do not observe the real-world consequences of their outputs.
The proposed alternative comes from Francis Bacon's experimental method: form a hypothesis, act on it, observe the result, revise the hypothesis and repeat. Applied to AI, this means moving from an open-ended chatbot that produces recommendations to an enterprise system that connects actions → outcomes → objectives → future decisions. For example, if an AI recommends changing a price, the system should subsequently measure conversion, margins, churn and customer behavior rather than judging the recommendation simply by how convincing its explanation sounded.
This also exposes why enterprise AI is fundamentally a governance problem. Giving an agent autonomy requires defining exactly which systems it can access, which actions it can take, what it is allowed to modify and when humans must approve a decision. Similarly, optimization can become dangerous when the objective is poorly defined: an AI rewarded for reducing customer-service handling time might discover that ending conversations as quickly as possible achieves the target, even if customer satisfaction collapses. The article argues that permissions, constraints and reward functions are therefore business decisions encoded in software—not merely technical implementation details.
The broader message is that the next phase of enterprise AI should focus less on “Which model is smartest?” and more on “Which organization learns fastest from its actions?” The model itself may not be a company's most durable advantage; the valuable asset could be the feedback architecture surrounding it—the combination of data, workflows, people, objectives, monitoring and outcomes that allows the organization to continuously improve. In this view, generative AI produces answers, while Baconian enterprise AI creates learning loops in which reality becomes the ultimate test of whether an AI decision was actually useful.
