AI Should Make Experts More Powerful, Not Less Necessary

For most of my career in software development, one assumption sat quietly underneath almost every planning conversation: skilled human effort was scarce. There were only so many engineers. Only so many hours in a sprint. Only so many things a team could investigate, design, build, test, document, and maintain.

We built processes around that constraint.

AI is beginning to break it.

The cost of producing a first draft of code, documentation, an analysis, a test plan, or a proposed solution has fallen dramatically. Problems that once required hours of mechanical work can increasingly be reduced to minutes. That is extraordinary.

But I think we're drawing the wrong conclusion from it. The most interesting future isn't one where AI makes experts less necessary. It is one where AI makes expertise dramatically more powerful.

Automation changes where the hard work lives

When technology automates a task, it tends to automate the predictable parts first. That is exactly what we should want.

If an experienced billing specialist spends hours locating supporting clinical documentation for a denied claim, that search is not the highest-value use of the specialist's expertise. If an engineer spends half a day creating repetitive scaffolding that follows patterns the organization already understands, typing the scaffolding isn't the valuable part of the work. If a compliance professional spends hours copying updates from regulatory websites into tracking documents, copying text isn't why we need the compliance professional.

The value lies elsewhere. It lies in recognizing the unusual claim where the standard payer interpretation doesn't apply. It lies in noticing the subtle architectural consequence hidden inside an otherwise reasonable software change—including the plausible-looking agent edit that passes tests and still embeds the wrong invariant. It lies in understanding that two regulations appear to conflict and knowing the issue requires investigation rather than another automated update.

As automation improves, these cases don't disappear. They become a larger percentage of the work humans still need to do. And they are usually the harder cases.

This creates a paradox

The better our automation becomes, the easier it is to imagine that deep human expertise matters less. I think the opposite is true. Automation allows an expert to operate across a much larger surface area. But when something falls outside the automated system's assumptions, that same expert must be capable of recognizing it. That requires more than knowing which button to press.

The person supervising the system needs enough understanding to ask:

• Does this result actually make sense?
• What assumption produced it?
• What evidence supports it?
• What happens if that assumption is wrong?
• Is this an ordinary case or an exception?
• When should I stop trusting the normal process?
• Can I recover if the automation fails?

Those questions describe the capacity to supervise—not a mandate to re-check every ordinary output by hand.

Risk, evidence quality, and process maturity should change how deep that involvement goes. A well-validated, low-risk path does not need the same human attention as a novel, high-impact decision. Human involvement should shrink, shift, or intensify with evidence—not stay frozen, and not default to universal manual review.

A person who cannot answer those questions when it matters isn't really supervising automation. They are depending on it. Those are very different things.

The goal should be cognitive leverage, not cognitive substitution

This distinction has increasingly shaped the way I think about AI. I don't want AI to replace the judgment of the people doing important work. I want it to remove everything that prevents those people from applying their judgment. That means removing search, repetitive translation, boilerplate, unnecessary coordination, and the need to keep reconstructing information the organization should already know.

In other words, remove friction before removing responsibility. That is a very different model of AI adoption. It asks:

What prevents this person from spending more of their time doing the part of the job we actually need their expertise for?

Sometimes the answer will be AI. Sometimes it will be traditional automation. Sometimes it will be better data. Sometimes it will be a better process.

And sometimes substitution is the right call: well-understood, low-risk, high-volume work with strong validation may not need deep expert involvement on every instance. That is not anti-expertise. It is refusing to waste scarce judgment on work the organization already knows how to check.

The technology is subordinate to the objective.

This changes what good leadership looks like

A large part of leadership is creating an environment where talented people can apply their expertise with as little unnecessary friction as possible. AI does not change that principle. It expands its possibilities—and it forces clearer operating choices. If an engineer can delegate repetitive implementation work to an agent, leadership should help ensure the engineer has the context, standards, validation systems, and authority necessary to use that capability safely. If an operations specialist can investigate ten times as many cases, leadership should not merely celebrate the throughput increase.

It should ask whether the specialist can still identify the one case in ten where the usual answer is wrong. If AI can execute work at machine speed, production capacity is no longer the scarce resource by default. Validation capacity must grow with production capacity, or the organization is amplifying mistakes as efficiently as it amplifies output. For leaders, that points to a short operating-model checklist:

• Who is accountable for outcomes when throughput jumps?
• Are incentives rewarding exception quality, recovery, and retained understanding—or only volume?
• Are the people closest to the work helping redesign what automation removes and what stays human-owned?

The limiting factor moves. Management needs to move with it.

Productivity alone is a dangerous measure

Imagine two organizations. The first introduces AI and doubles its output. People generate more code, process more documents, answer more requests, and close more cases. But gradually, fewer people understand the underlying systems.

Unusual cases become harder to judge. Problems propagate faster because production capacity has outgrown validation capacity.

The second organization also doubles its output. But it deliberately uses automation to eliminate repetitive work while keeping experts engaged in diagnosis, architecture, validation, exception handling, and continuous improvement. It measures not only throughput, but quality, failure rates, recovery speed, retained learning, and whether humans can still challenge the system. These organizations may initially look equally productive.

They are not equally resilient.

I do not assume every organization will choose the second path. Many will keep optimizing reported throughput because scorecards reward it—and that can win short-horizon metrics while hollowing expertise. The thesis is conditional: leverage without retained judgment and validation capacity is fragile.

Expertise should compound

The opportunity I find most compelling isn't simply that AI can make an expert faster. It is that the combination can make expertise itself more scalable—if we treat captured judgment as something to challenge, not something to obey. Imagine a senior specialist who solves a difficult problem. Today, the organization often gets one resolved problem—maybe a note, a memory, a procedure change, or nothing durable at all.

The better question is:

How do we make sure the organization is permanently better because this expert solved that problem—without turning that expert into temporary training data?

If hard-won judgment can be retained and reused the next time a similar situation occurs, expertise begins to compound. But the human remains essential. Experience is not infallible.

Circumstances change. Policies change. Technologies change. The lesson that was correct yesterday can become the mistake of tomorrow. Practitioners should remain accountable decision-makers. Systems that help expertise propagate still need people who can overturn yesterday's answer when the world has moved.

Better together

I think the AI conversation has been framed too often as a contest: human or machine, automation or employment, enthusiast or skeptic. That framing is unnecessarily limiting. Humans and AI have different strengths—as do AI and deterministic software, and specialists and formal processes.

The interesting question isn't which one wins. It is how we design systems that allow each to do the work for which it is best suited. My expectation—and it is a hypothesis, not a forecast—is that organizations that learn to do this well will not have less expertise.

They will have more leverage per unit of expertise. That phrase needs care.

I do not mean fewer experts carrying the same load. I mean each expert's judgment covering more ground because friction is gone, not because headcount was cut under an efficiency banner. If residual work is mostly hard exceptions, cutting the expert bench can increase load on whoever remains—and judgment atrophies if people only ever see rare failures and never practice it deliberately.

What would falsify this thesis for me? Throughput rising while validation capacity lags; expertise hollowing out as routine work disappears; incentives that punish slow, correct exception handling—if that pattern still produces durable results, the argument needs revision. Where the hypothesis holds, the best people spend less time digging, copying, formatting, coordinating, and reconstructing. They spend more time interpreting, challenging, designing, diagnosing, deciding, teaching, and improving.

That does not sound like the disappearance of the expert. It sounds like the conditions under which expertise becomes more powerful—if we build for it on purpose. And it leads to an important architectural question: If AI shouldn't do everything, how do we decide where intelligence is actually necessary?