Over the past several months, I have used AI to develop a much deeper understanding of the systems surrounding data centers. That has meant learning about power, cooling, fiber, construction, computing equipment, financing, customer demand, power generation at gas fields, and even the possibility of developing computing capacity in places such as Venezuela.

AI made an extraordinary amount of information accessible to me. But access to information was not what produced the growth.

The growth came from wanting to understand what made an answer true.

I challenged assumptions. I separated the building, the electrical infrastructure, and the computing equipment into different systems with different requirements. I questioned whether development in Venezuela would actually be less expensive than comparable development in the United States. That led back to the price of gas, land, permits, materials, fiber, political risk, and finally the question that sits beneath the entire data center expansion: Who is the customer for all this computing capacity?

AI helped me reach the information. Questioning, testing, and connecting the information turned it into knowledge.

The development of this essay followed the same pattern.

The first draft was coherent. It sounded credible. It also was not right.

The examples were too long. Some of the language was too soft. It suggested that an experienced employee might merely sense that an AI answer did not look right when the real requirement is much higher: the employee must know that the answer is wrong and be able to explain why.

More importantly, the draft treated thoughtful use of AI as an individual responsibility. It did not establish management’s obligation to define the purpose of the system, its boundaries, and the process through which people would continue to learn.

AI organized the original idea. Human judgment found what was missing. AI then helped reconstruct the argument. That collaboration led to a conclusion neither the first draft nor the original idea had stated clearly enough:

The system supported by AI must continue improving, and the humans responsible for it must continue developing. Without growth on both sides of the equation, each will eventually fail the other.

The Condition We Are Overlooking

Both of my experiences depended on the same condition: I wanted to understand, not merely obtain an answer.

AI can supply an answer. It cannot guarantee that the person receiving it will ask which assumptions produced it, where its boundaries lie, what evidence supports it, or what would make it wrong.

That distinction is becoming more important as AI becomes more polished. A crude error is often easy to detect. A fluent, reasonable answer that is almost correct is much more dangerous. The more convincing the technology becomes, the more knowledgeable the human supervising it needs to be, not the less.

This cannot be left entirely to personal curiosity. Management creates the environment in which curiosity either has value or becomes an obstacle.

If employees are measured only by speed and output, they will naturally use AI to obtain the answer, complete the task, and move on. Taking time to understand how the answer was produced will appear inefficient. Asking one more question may be treated as hesitation rather than discipline.

The danger is not that everyone will suddenly stop thinking. The danger is that organizations will build systems in which understanding is no longer required.

Work at the Entry Level Was Doing More Than Its Assigned Tasks

Much of the discussion about AI and employment focuses on how many jobs may disappear. That matters, but it misses part of the operational risk.

Work at the entry level is not simply inexpensive labor. It is where people encounter repetition, exceptions, mistakes, difficult customers, incomplete information, and the difference between what the procedure says and what the operation actually requires. Those experiences form the knowledge and judgment that allow someone to become a capable supervisor, manager, or executive.

If AI removes the work but the organization does not deliberately replace the learning, it has not merely reduced headcount. It has removed part of its development system.

The savings will appear quickly. A reduced payroll shows up in the next financial report. The loss of expertise may remain hidden for years, until experienced employees retire or leave and the organization discovers that it no longer has people capable of recognizing when the system is wrong.

A company can therefore report a successful return on AI while becoming less capable of managing itself.

This is not an argument for preserving unnecessary work. It is an argument for understanding what the work was producing before eliminating it.

Management Must Define the System

The first responsibility of management is not to purchase an AI tool or calculate how many positions it might eliminate. It is to define what the organization expects the combined human and AI system to accomplish.

That begins with purpose and boundaries.

What problem are we asking AI to solve? What decisions may it make? Where must human authority remain? What degree of uncertainty is acceptable? Who owns the result when the system is wrong?

Those answers shape the working process. Management must then decide how employees will use AI in the actual workflow, what knowledge they must continue to acquire, how outputs will be challenged and verified, and how errors will be identified, explained, corrected, and incorporated into future work.

It must also decide how junior employees will develop when the traditional learning path has been automated.

Only then can the organization define success. Lower cost and faster completion belong in that definition, but so do stronger employee capability, sounder decisions, retained institutional knowledge, resilience when the technology fails, and the ability to know, not merely suspect, when AI has produced the wrong answer.

The sequence matters:

Purpose creates boundaries. Boundaries shape the process. The process determines whether people learn. Learning produces judgment. Judgment makes meaningful oversight possible. Meaningful oversight improves the combined system.

Or more simply:

Purpose → Boundaries → Process → Learning → Judgment → Correction → Improvement

The failure path is equally clear:

Cost reduction → Fewer people → Less experience → Weaker judgment → Less effective oversight → Greater dependence on AI

Return on investment measures whether AI produced an economic benefit under management’s selected assumptions. It does not determine whether the organization became more knowledgeable, capable, resilient, or able to recognize when the technology got something wrong.

Three Systems Are Converging

The scale of current data center development helped me see the larger structure. AI depends on three connected systems.

The first is the physical system: land, power, cooling, fiber, buildings, and equipment.

The second is the computing system: chips, models, data, training, inference, and applications.

The third is the human system: purpose, knowledge, judgment, authority, and accountability.

The physical system creates computing capacity. The computing system converts that capacity into capability and output. The human system gives the capability purpose, establishes its boundaries, judges its output, and converts it into operational value.

The first two can create capacity and output. They may even create revenue in the short term. Only the third can turn them into purposeful and durable value.

This is why the human system is not merely one more component. It is the component that gives the others meaning.

A data center can be constructed, energized, and technically successful while failing economically because the customers or workloads never materialize. In the same way, an AI system can function exactly as designed while failing organizationally because management never defined what it was supposed to accomplish or which human capabilities had to remain around it.

Paper megawatts do not run servers.

Installed AI does not create organizational intelligence.

Technology creates capability. Purpose, integration, and management turn capability into value.

Human Judgment Is Infrastructure

We understand that an AI model must be trained, tested, corrected, and continually improved before it can be trusted with serious work. What we have not fully recognized is that the humans supervising it must be developed just as deliberately, and neither development process is ever complete.

Human judgment may not appear on a balance sheet, but an operation depends upon it just as a data center depends upon power, cooling, fiber, and physical access. If management removes the process that develops judgment, it is allowing critical infrastructure to deteriorate.

That is why AI cannot belong solely to IT, human resources, finance, legal, or operations. It changes workflows, staffing, training, incentives, authority, accountability, and risk across the organization. It must be managed as one operating system.

The relevant question is not simply, “What can AI do?”

It is, “What kind of organization and people will this way of using AI produce five years from now?”

If the answer is an organization that completes more work with fewer people but no longer develops the knowledge required to judge that work, the technology has not made the organization intelligent. It has made the organization dependent.

AI can make knowledge more accessible. It can accelerate analysis, expose connections, test ideas, and help people develop faster than many of us once thought possible. I know that because it has done those things for me.

But it produces growth only when the human and the process are designed for growth.

Management’s responsibility is therefore larger than adopting AI responsibly. It must define the purpose, establish the boundaries, preserve the path through which people develop judgment, and measure the capability of the combined human and AI system over time.

Without continued growth on both sides of the equation, each will eventually fail the other.

⚓

References