A lot of AI adoption is presented as a speed story. Write faster, summarize faster, code faster, answer faster. That framing is not wrong, but it is incomplete in a way that matters. In many teams the real change is not that people suddenly become more capable. It is that they spend less time inside the parts of work where capability is actually formed. The first draft, the rough analysis, the painful comparison between two weak options, the attempt to explain something clearly enough that another person can challenge it. When those steps disappear too early, output may improve for a while. Competence often does not.
Efficiency Can Hide Degeneration
Teams often confuse cleaner outputs with stronger thinking. If an AI system produces plausible briefs, decent copy, usable code, and polished summaries, a manager can easily conclude that the team is leveling up. Sometimes it is. Sometimes it is simply borrowing coherence from the model. The distinction is subtle, but important. A person who can improve a draft is not automatically a person who could have produced the thinking behind it. A team that can operate an AI layer is not automatically a team that understands the domain more deeply. This is why some organizations look sharper on the surface precisely when their internal understanding is getting weaker. The dashboards look better, the documents read smoother, and the underlying comprehension becomes thinner.
Friction Is Where Judgment Forms
Most valuable professional skills are not built in the clean, final version of work. They are built in the mess before it. Judgment develops when you have to decide what to ignore, what to keep, what trade-off is acceptable, what ambiguity is dangerous, and what apparently good answer is only well phrased. Taste develops through exposure, comparison, and correction. Clear writing develops by seeing your own vagueness, not by pressing regenerate until the sentence sounds respectable. This is why removing friction is not always the same as removing waste. Some friction is just bureaucracy. Some friction is where skill is formed. If a team cannot tell the difference, it may optimize away the very conditions that produce seniority.
Delegation Changes the Observer
There is another layer here that people tend to miss. When you delegate too much of the early cognitive work to a model, you are not only changing the work. You are changing the observer. The person reviewing AI output is not standing in the same epistemic position as the person who wrestled with the problem from the start. Review sounds safer than creation because it looks controlled. In practice it often reduces attention. People scan instead of thinking. They validate tone, structure, and plausibility, while missing the fact that the core framing may already be wrong. This matters in strategy, product decisions, hiring, customer research, and management communication far beyond copywriting. A polished mistake is often more dangerous than a clumsy one, because it bypasses the discomfort that would have triggered scrutiny.
Why Smart Teams Still Walk Into It
The appeal is obvious. Nobody gets rewarded for preserving healthy cognitive friction. People get rewarded for speed, scale, responsiveness, and visible throughput. AI helps with all of that, at least in the short term. It also flatters the user. It creates a feeling of leverage, which is often real, but not always in the way people imagine. In ambitious organizations there is also a deeper motive: nobody wants to feel slow while everyone else is talking about acceleration. So teams quietly move the boundary of what humans still need to understand firsthand. Not because they are lazy, usually. Because social pressure and local incentives make that move look rational. The problem appears later, when the team is asked to handle ambiguity without the supports it has grown used to.
Responsible Adoption Means Preserving Apprenticeship
I do not think the answer is to avoid AI or romanticize manual effort. That would be another form of laziness, just dressed up as principle. The more useful question is where AI reduces waste and where it erodes apprenticeship. Those are not the same category. In some workflows it makes sense to let the model handle repetitive formatting, synthesis of known material, or exploration of alternatives. In others, especially where judgment is still immature, it may be wiser to keep people close to the raw task for longer than feels efficient. Otherwise you end up with organizations full of people who can supervise outputs they no longer know how to produce, challenge, or repair. That may work in stable conditions. It becomes expensive the moment reality stops being cooperative.
This is why the discussion about AI and jobs often misses the more immediate issue. Replacement is dramatic, so it gets attention. Skill atrophy is quieter, so it slips through governance, hiring, and training decisions almost unnoticed. And yet it may shape the next few years of work more than the headline question ever will.