Consider a scene that is becoming increasingly plausible inside a modern security operation.

A junior analyst brings an incident assessment into review. It is excellent. The timeline is clean. The probable cause is clearly stated. The supporting evidence is organized. The recommended actions are reasonable. The document is better than what many experienced analysts would have produced only a few years ago.

Then the senior reviewer asks a different question:

What would have to be true for your conclusion to be wrong?

The room becomes quiet.

The analyst knows the answer the AI produced. But he has not fully worked the problem underneath it.

Advertisement - [email protected]

That distinction matters.

The report may be better. The analyst may not be.

Those two things are no longer guaranteed to improve together.

Artificial intelligence is rapidly moving beyond copilots that summarize documents or help draft reports. Agentic systems can increasingly perform complex workflows: ingest an alert, retrieve supporting logs, correlate threat intelligence, search historical incidents, recommend a disposition, prepare an incident summary, and initiate the next action.

From an operational perspective, this is extraordinary progress.

But there is a hidden cost that will not appear on most productivity dashboards.

Much of the work we are automating was never just work. It was training.

Often bad training, admittedly. Nobody should romanticize hours spent manually chasing meaningless false positives, building repetitive spreadsheets, or performing work a machine can now complete in seconds. We should eliminate unnecessary drudgery.

The problem is not that we are destroying a wonderful apprenticeship system.

The problem is that we are eliminating a bad school without building a better one.

Human Judgment Is Built From Accumulated Scars

The difference between a junior professional and an experienced one is not simply access to more information.

It is judgment.

Experienced people have seen things fail. 

They have followed the normal procedure and discovered that this case was not normal. They have dismissed something that later mattered. They have escalated something that turned out to be harmless. They have trusted a technically correct recommendation that failed because the surrounding context was wrong.

Over time, these experiences leave what I think of as accumulated scars. Those scars become pattern recognition.

A senior security analyst does not merely know what an attack looks like. The experienced analyst also knows what a convincing false positive looks like. More importantly, that person sometimes notices the small inconsistency that says, This looks like the same thing I have seen a hundred times before—but something about this one is different.

That kind of judgment is difficult to teach through a procedure manual. It is also difficult to acquire from an AI-generated answer.

For generations, organizations developed expertise almost accidentally. Junior employees performed lower-risk work. Seniors reviewed it. Juniors made mistakes. They were corrected. They tried again. Eventually, repetition turned into intuition and intuition matured into judgment.

It was inefficient.

It was uneven.

Sometimes it was unnecessarily painful. But it produced experienced people.

Now we are rapidly removing that friction.

The immediate benefits are measurable. The long-term consequences may not be.

Agentic AI Changes the Problem

When generative AI first entered the enterprise, a human generally remained inside the workflow.

The analyst researched, then asked AI for help. The engineer wrote code, then used AI to debug it. The manager developed an idea, then used AI to improve it.

That model accelerated human work. Agentic AI changes the architecture.

Imagine an alert entering a security environment. An AI agent collects endpoint data, queries identity logs, checks network activity, compares indicators against threat intelligence, examines similar historical events, assigns a probable cause, drafts the investigation, recommends remediation, and closes or escalates the case.

The junior analyst may no longer save three hours. The junior analyst may never enter the cognitive loop. That is a different problem entirely.

We can easily measure how much faster the incident was processed. We can calculate reduced workload, improved mean time to resolution, lower operating cost, and increased analyst capacity.

But where is the metric showing that a future senior analyst just lost another opportunity to develop judgment?

There probably isn’t one.

This is why the apprenticeship crisis can remain invisible for years.

The organization becomes faster. The dashboards improve. Senior people, augmented by increasingly capable AI, become extraordinarily productive.

Everything appears to be working.

Until those senior people begin to leave.

They retire. They move. They burn out. They get promoted.

Then the organization looks beneath them and discovers that the next generation has produced thousands of sophisticated outputs but has made surprisingly few independent decisions.

The organization automated the work faster than it regenerated the expertise. That is a strategic risk.

Every Automation Decision Owes a Development Tax

This does not mean we should preserve inefficient work simply to give junior employees something to do.

Quite the opposite.

Automate it.

But every time we remove work from a human workflow, leadership should ask another question:

What judgment was someone learning by doing this work?

If the answer is truly nothing, automate it and move on.

But sometimes tedious activity contained a hidden developmental function.

The analyst chasing false positives was learning the boundary between normal and abnormal.

The engineer debugging repetitive failures was learning how systems behave outside the architecture diagram.

The manager preparing an imperfect recommendation was learning to defend a position in front of people who disagreed.

If AI removes that experience, we need somewhere else to build the capability. I think of this as the development tax of automation.

Every automation decision that eliminates a developmental experience creates an obligation to rebuild that learning somewhere else.

The tax should not be paid through classroom training alone. Judgment requires experience.

Fortunately, AI can help us create a better apprenticeship system than the one we are eliminating.

1. Require the Defended Decision

The easiest way to misuse AI is to allow the quality of the output to hide the weakness of the reasoning behind it.

So stop reviewing only the artifact. Review the decision.

A developing employee should be able to explain what they initially believed, what evidence supported that belief, what AI changed in their thinking, what recommendation from AI they rejected, and what evidence would cause them to reverse their conclusion.

This does not mean banning AI until the employee finishes the assignment. That will quickly become artificial and counterproductive.

It means making sure the human remains intellectually responsible for the outcome.

In an incident review, I care far less about whether the presentation looks polished than whether the analyst can answer questions such as:

What assumption is carrying the most risk? What evidence contradicts your conclusion? What would make you change your mind?

What did AI recommend that you chose not to accept? Why?

If the person cannot answer those questions, the organization received an output. It did not necessarily develop capability.

2. Make AI the Adversary

AI does not have to eliminate practice. It can create more of it.

A developing analyst can be given a simulated incident and asked to make the call. Then the AI can be instructed to attack the conclusion.

Introduce conflicting evidence. Change one critical variable.

Create a lookalike false positive.

Remove an assumption the analyst relied upon. Introduce a business constraint.

Make the normal control inappropriate for this particular case. Then require the analyst to revise the decision—or defend it.

This is where a safe learning environment becomes important. People need somewhere they can be wrong.

They need somewhere to try an idea that turns out to be foolish. Somewhere to misunderstand the situation, push too far, miss the signal, discover the exception, and learn why the mistake happened.

We understand this intuitively in high-consequence professions. We do not want the first truly difficult situation someone encounters to be the one in which failure produces catastrophic damage.

Businesses need to adopt the same principle for an AI-native workforce.

If machines are removing naturally occurring practice from day-to-day work, we must manufacture higher-quality practice deliberately.

And AI itself may become one of the best tools we have ever had for doing it.

3. Give People Bounded Real Consequences

Simulation is not enough.

Eventually, someone has to make a real decision.

Real environments contain pressure, personalities, uncertainty, incomplete information, organizational politics, customer impact, competing objectives, and consequences that even sophisticated simulations cannot completely reproduce.

People need hands-on responsibility. But responsibility can be progressive.

Give the developing professional ownership of a genuine decision with a deliberately bounded blast radius. Keep an experienced person close enough to prevent unacceptable damage, but not so close that the junior employee never feels the weight of the decision.

Then conduct an after-action review. Not simply:

Was the answer correct?

Ask instead:

What did we believe at the time? What assumptions did we make? What information did we miss? What did AI see that we did not?

What did the human see that AI did not? What would we do differently next time?

This is how scars become useful rather than merely painful. The objective is not to manufacture failure.

It is to manufacture learning.

The New Apprenticeship Is a Human-AI Development System

The future apprenticeship model should not attempt to recreate the workplace of the past. It should be better.

We now have an opportunity to build development systems in which humans and AI improve together.

AI can expose a junior employee to more scenarios in a month than that person might naturally encounter in a year. It can create counterfactuals, challenge assumptions, simulate difficult stakeholders, introduce unexpected evidence, and provide immediate feedback.

But the human still needs to reason. The human still needs to decide.

The human still needs to experience consequences.

And the human still needs experienced leaders who can explain not only what happened, but why it mattered.

The real unit of development is therefore no longer just the individual. It is the human-AI system.

Organizations need to develop three things simultaneously:

the AI, the human, and the relationship between them.

A powerful model paired with someone incapable of challenging it can still produce dangerous decisions.

An experienced person who refuses to use AI effectively will increasingly operate below potential.

And a junior employee who allows AI to do all the reasoning may become extraordinarily productive without ever becoming truly experienced.

The goal is not simply human plus AI.

The goal is a human-AI system in which each makes the other more capable.

The Failure Will Arrive Late

This may be the most dangerous part of the apprenticeship crisis. The damage is delayed.

If a company shuts down an important production system today, operations feel the consequences immediately.

If a company destroys part of its human-development pipeline today, the consequences may not appear for five or ten years.

In the meantime, the numbers may look excellent. More output.

Fewer people.

Faster decisions.

Lower costs.

Higher productivity.

That makes this risk particularly difficult for leaders to see. Senior expertise is a slowly renewing resource.

An organization can spend it for years without realizing that it is not replenishing it.

Then one day the experienced people are gone, and the organization discovers that the people underneath them know how to operate the machinery but have never accumulated enough judgment to know when the machinery should not be trusted.

By then, rebuilding the missing layer may take years.

Automate the Work. Do Not Automate the Learning.

AI should eliminate unnecessary work.

Agentic systems should absorb repetitive workflows. Organizations should pursue the productivity gains aggressively.

But leaders must stop assuming that productivity and capability automatically move together.

Sometimes they do. Sometimes they separate.

And when they separate for long enough, an organization can become remarkably efficient while becoming increasingly fragile.

The old apprenticeship system was imperfect. We should not preserve it out of nostalgia.

We should replace it with something better: defended decisions, deliberate adversarial practice, bounded real-world responsibility, and after-action learning in which AI participates without removing the human from the cognitive loop.

If AI becomes the apprentice, we must redesign apprenticeship itself. Because organizations will still need masters.

And that leads to one question I believe should be added to every serious automation review:

Before you automate the work, name the judgment it was teaching—and tell me where that judgment will be built next.

If the room cannot answer both, the automation decision is not finished.

About the author

A U.S. Army veteran and former Major in the Infantry, Kevin brings a mission-driven mindset to corporate leadership—applying battlefield-tested decision frameworks to complex technology environments. His approach unites resilience, intelligence, and disciplined execution to align security with business outcomes.

Kevin is widely recognized for his thought leadership on the intersection of AI, security, and organizational excellence. He regularly writes and speaks on how trust, governance, and visionary leadership enable enterprises to harness AI responsibly and transform securely in an era defined by intelligent systems.

*The views and opinions expressed by the author do not necessarily state
or reflect the views or positions of Hyperight.com or any entities they
represent.