What AI Changes, and What It Doesn't

Every major technology changes the price of something. The internet made distribution dramatically cheaper. Cloud computing lowered the cost of starting and scaling software. Smartphones made computing ambient. Artificial intelligence is now changing the price of execution.

A growing number of tasks that once required hours of specialist effort can now begin with a prompt. Writing, analysis, research, coding, design, planning and synthesis can all move faster. A small team can attempt work that previously demanded a larger one. An individual can cross disciplinary boundaries with less friction than before.

That is not a small change. But I think we misunderstand AI when we treat faster execution as the whole story.

The bottleneck moves upstream

When execution becomes easier, the scarce part of the system changes. The question is no longer only, “Can this be done?” Increasingly, it becomes, “What should be done, why, and according to whose judgment?”

An AI system can give you ten strategies in seconds. It does not automatically know which strategy belongs to your actual situation. It can generate a persuasive argument without knowing whether the premise deserves belief. It can help produce a polished product without deciding whether the product solves an important problem.

This means capability can increase while direction remains weak.

I think that is one of the central tensions of the AI era. We are rapidly improving our ability to produce outputs while still relying on people to define objectives, supply context, decide what matters and accept responsibility for consequences.

The more powerful the tool becomes, the less trivial those human functions become.

Judgment does not disappear because answers arrive faster

Before generative AI, friction naturally limited how much we could produce. You could not explore fifty versions of an idea without paying for them in time. Now abundance is cheap. The cost of generating another answer is close to zero.

Abundance creates its own problem: selection.

When there are many plausible answers, judgment becomes the ability to distinguish between what is fluent and what is useful, what is possible and what is appropriate, what sounds right and what survives contact with reality.

This is why I do not think the most valuable AI skill is prompt engineering in the narrow sense. The deeper skill is framing. Can you describe the problem accurately? Can you identify the constraints that matter? Can you recognize when the system has optimized the wrong thing? Can you ask a better second question after receiving the first answer?

Those are not merely interface skills. They are thinking skills.

Context becomes a source of leverage

AI systems are broad. Human lives are specific.

The model may know the general structure of a market, but you know the customer who keeps refusing to pay. It may understand operations in theory, but you know which employee workaround is keeping the business alive. It may know agricultural best practices, but you know the local constraints around water, transport, labor and purchasing power.

That local context changes the answer.

People sometimes treat context as something AI will eventually make irrelevant. I think the opposite is often true. As general capability becomes more accessible, specific knowledge can become more valuable because it tells us where that capability should be applied.

This is especially important in emerging markets. A solution that looks perfect on paper can fail because of payment behavior, infrastructure, trust, distribution or informal workarounds that are invisible in a generic analysis. The person closest to the operating environment still has something the model does not automatically possess.

AI can extend agency or quietly replace it

There is another distinction I care about: assistance versus abdication.

There is nothing inherently wrong with delegating work to a machine. Delegation is one of the foundations of progress. We delegate arithmetic to calculators, memory to notebooks, navigation to maps and repetitive work to software.

The problem begins when we stop noticing which decisions we have delegated.

If I use AI to generate options and then choose deliberately, the system has increased my agency. If I accept the first confident answer because evaluating it feels inconvenient, the same system may have reduced my agency.

The difference is not the tool. The difference is whether I remain the author of the objective and the judge of the result.

This idea sits behind much of what I am exploring in Unabdicated. Becoming AI-native should not mean becoming passive. It should mean learning to use increasingly capable tools without quietly surrendering the human responsibilities that give those tools direction.

The future of work is also a future of responsibility

Much of the conversation about AI and work focuses on replacement: which tasks disappear, which professions shrink and which new roles emerge.

Those questions matter, but there is another layer. When machines participate in more of the execution, responsibility does not evaporate. Someone still chose the objective. Someone still deployed the system. Someone still decided that the output was good enough to act on.

That means future professionals may be judged less by how much they can personally produce and more by the quality of the systems they direct.

A manager may oversee human and machine contributors. A designer may spend less time manually producing variations and more time defining taste. A developer may write less routine code and spend more time deciding architecture, constraints and trade-offs. A founder may be able to test more ideas but face even greater pressure to choose the right problem.

Execution becomes leverage. Judgment becomes accountability.

What I think AI actually changes

AI changes who can attempt difficult work. It changes the speed of iteration. It changes the cost of crossing disciplines. It changes the minimum team size required for certain kinds of products. It changes the amount of experimentation available to a curious person.

Those are enormous shifts.

What it does not automatically change is the need for intent. It does not remove the need for standards. It does not resolve conflicting values. It does not decide what deserves attention. It does not make responsibility disappear.

The opportunity, then, is not to resist machine capability. It is to pair it with stronger human direction.

That combination interests me much more than automation for its own sake: humans who can think clearly about objectives, use AI aggressively where it creates leverage and remain responsible for what gets built.

That is the relationship I am trying to learn.