Reskilling For What?
AI has made doing tasks cheap. The scarce skill now is governing the work.
Everyone is asking how to reskill people for AI. But reskill to do what? Which skills are newly important that require teaching at scale?
The obvious answers are useful but insufficient. Teach people AI literacy. Teach them how the tools work and how to integrate them into their workflows. Teach them how to prompt, connect to data sources, and build autonomous agents. All of that matters. But it mostly aims at capturing the efficiency gains from AI making task execution cheap.
The deeper reskilling question is what humans should become better at now that AI has made task execution easy.
Research, drafting, summarizing, comparison, analysis, scenario generation, and first-pass synthesis used to consume real time. That time expense disciplined what could be asked, how much could be explored, and when a team had to stop producing analysis and move toward decisions.
The economics have changed. The activities that used to be expensive are now fast and cheap. The activities that used to feel free — problem definition, evaluation, synthesis, judgment about what the analysis actually means — are now the entire game. They were always important. They were just hidden behind the higher real cost of executing the tasks.
The task execution layer is becoming cheap. The judgment layer is now the hard constraint.
The scarce skill is no longer producing the analysis. It is deciding what analysis is worth producing, steering the inquiry while it unfolds, evaluating what comes back, selecting and integrating what matters, and preserving the state of understanding so the next cycle of work starts higher.
The scarce skill has a name: governance.
Take something ordinary: a competitive market scan.
Before AI, the work had a familiar shape forced by the economics. A team might compare the major competitors, build a horizontal table of size, share, products, capabilities, and positioning, write a few vertical profiles, summarize recent strategic moves, and then synthesize where the market seems to be and where the gaps are.
Every additional inquiry cost analyst resources and calendar time. You could not run every possible comparison. You could not test every scenario. You could not re-architect the scan every time the first pass revealed a better question. Even if another line of inquiry might have been useful, it often was not worth the time before the signal went stale.
So the old “good market scan” was the best inquiry that fit the time, cost, and attention budget.
Now imagine the same market scan in an AI-native workflow.
The old scan is still possible. But now, in the same general decision window, you can ask much more.
You can compare twelve months of company announcements against analyst commentary. You can separate common market moves from genuinely differentiated moves. You can ask which competitor is saying one thing and investing behind another. You can test what happens if one competitor adopts another’s strategy. You can model how the market changes under regulatory pressure, shifting customer taste, or a new entrant unconstrained by incumbent economics. You can compare the market to an adjacent one that has already gone through a similar transition. You can use customer reviews and complaints to ask whether the official product categories are hiding the real basis of competition.
And, most importantly, you can use the results of one execution pass to redesign the next.
The task did not simply get faster. The inquiry space expanded. What is economically feasible expanded. And the judgment skills required to govern the inquiry become newly important. Not because the AI models are bad, but because they are good enough that what is missing becomes obvious.
When the cost of asking questions falls, the value of knowing what to ask and how to evaluate the answers rises.
This is where “human in the loop” needs to be reclaimed.
The phrase has been cheapened. In many settings, it now means something like: the model does the work, and a human checks yes or no. The human becomes a validation gate. The loop is treated as a production line with an approval checkpoint at the end.
That is not enough for serious knowledge work. The human is not in the loop to approve the output. The human is in the loop to govern the loop.
That governance has three parts.
First: frame the work and design the task sequence.
Define the real problem. Understand what questions we care about and what decision the work serves. Decide what kind of inquiry would actually change the view. Select and order tasks to complete. Choose what to delegate to AI, what to hold human, what to run first, and what the first pass should be allowed to change.
AI makes all of this more important because “cheap output” has a consequence: it is now cheap to do the wrong work excellently. A model can produce a polished market scan around the wrong competitor set. It can write a persuasive synthesis around a bad frame. It can answer the question that was asked while the real question sits one inquiry underneath or one assumption adjacent.
The more capable the model, the more dangerous the wrong frame becomes.
Second: evaluate outputs and steer the inquiry.
Once the first output comes back, the human’s job is not merely to edit the prose. The human has to ask: did this answer the real question? What did it assume? What did it miss? What is generic? What is surprisingly useful? How does this pass inform what questions to ask in the next pass? Where is the model smoothing uncertainty? Where is it solving a nearby problem? Which branch should stay alive one more pass, and which branch should die?
This is especially hard when the output is good. Bad output is easier to reject. Strong, fluent, plausible output is harder. It creates a false sense of completion and a false feeling of goodness.
The human is not simply stewarding tasks and checking for valid completion. The human is steering the whole inquiry.
Third: preserve and compound the state of understanding.
A lot of organizations will use AI to produce more artifacts. More memos. More summaries. More dashboards. More options. More market maps. More recommendations.
But artifacts do not automatically create learning and understanding.
A market scan can be produced, circulated, discussed, and forgotten. The next team starts again. The next executive asks the same question. The next deck recreates the same context. That is output, not compounding.
Compounding means the work leaves behind something that improves the next cycle: a sharper market understanding, a competitor typology, a decision heuristic, a known uncertainty, a branch worth revisiting, a better evaluation rubric, a reusable frame, or a checkpoint that preserves what actually changed in the state of understanding and that can be used to re-enter the work at altitude.
The point is not to store more documents, but to make the next return to the problem start from a higher altitude.
A summary preserves what was said. A checkpoint preserves what now matters.
Frame the work. Steer the inquiry. Compound the understanding.
Those skills are not soft or abstract. They are core operating skills in the AI era.
They determine whether AI produces more noise or better decisions. They determine whether task savings become cost takeout, output inflation, or capability multiplication.
The irony is that while AI has made judgment the most important human skill, it may also have made judgment structurally harder to acquire.
This is the apprenticeship problem.
In the old model of knowledge work, judgment formation was often embedded inside task execution.
A junior analyst learned by building the market scan. A junior lawyer learned by doing the research and drafting the memo. A junior consultant learned by building the model, writing the slides, getting corrected by the partner, and slowly discovering what the client actually cared about.
Those tasks did not just produce usable outputs for the project, they were also the reps through which future judgment formed.
If AI removes the junior task layer without replacing its learning function, organizations may save time while weakening the pipeline that produces future senior judgment.
This is not a sentimental argument for preserving inefficient work forever. AI can accelerate learning in some structured domains. Customer-support studies, for example, found that AI assistance disproportionately helped newer and lower-skill workers, suggesting that AI can sometimes help people move down the experience curve faster by exposing them to better practices.
But ambiguous professional judgment is different from a structured support workflow.
The feedback is slower. The signal is noisier. The answer is not always knowable. The judgment often lives in the framing, not the task output. And the developmental value of the work may be hidden inside the struggle to produce it.
If AI strips away the task reps without replacing the learning function, organizations will not just have fewer juniors. They will have fewer future seniors with the right judgment to carry the work.
The answer is not to make people do manual work for its own sake, but to redesign the learning rep.
A junior person may no longer need to spend twenty hours producing the first market scan from scratch. But they should still learn how to design the inquiry. They should compare possible task architectures before running them. They should predict which inquiry will be most useful before asking the model. They should evaluate why a particular AI-generated synthesis is plausible but wrong. They should see senior edits not as copy changes, but as judgment compressions. They should ask what the output changed, what it failed to see, and what the next pass should test.
And they should be coached, evaluated, and rewarded on their ability to do this sort of governing of the work.
But that is not the whole story, because it treats the AI model as a neutral absence — as if the senior used to be in the room, and now no one is.
That is not what is happening.
The model is in the room. By default, the model is what the junior is being shaped by, hour after hour, day after day. The model is fluent, confident, smoothing, hallucinating. It rushes to the artifact. It defends loosely established priors. It rounds the exact distinction down to the nearest readily available one. It cannot tell the junior when its output is plausible but wrong, because it does not know.
An ungoverned model interacting with a junior all day is not the absence of an apprentice’s teacher. It is an active mis-apprenticeship under a teacher unqualified to teach. And because every junior in the industry is being shaped by the same model or model family, an entire cohort of future seniors is being trained against the same correlated errors. The next generation of seniors will share the same blind spots — confidently, articulately, with no internal way to detect it.
The answer is to govern the model so it becomes fit to apprentice. That is a discipline. It is teachable. It is the same discipline this essay has been describing — framing the work, steering the inquiry, compounding the understanding — applied now to the specific purpose of shaping the development arc of someone’s judgment rather than producing a deliverable.
This brings us back to the organizational choice.
When AI compresses the task layer, companies have options. Three of them.
They can capture the savings as margin and return them to shareholders through dividends, buybacks, or reduced headcount expense. In some cases that will be economically rational. If the freed human capacity cannot be redeployed into higher-value work, the case for fewer humans becomes hard to avoid.
They can reinvest the savings into the next round of task automation. That may feel like productivity, but it accelerates the same race-to-the-bottom dynamic and does nothing for the firm’s capability at the modes where genuine advantage lives.
Or they can redeploy human capacity into higher-value work: better problem framing, better inquiry design, better evaluation, better synthesis, better judgment transfer, better apprenticeship, better decision checkpoints, better institutional memory.
Only the third option creates lasting advantage. The first two have a finite floor — you can only cut so much, and you can only automate so much before the marginal returns collapse. The third has a much higher ceiling, because what you are buying is not reduced cost but multiplied capability — what the institution can become capable of doing that it could not do before.
This is not a humanist supplement to fiduciary analysis. It is what the fiduciary math actually requires when you stop modeling automation as a single number and start modeling it as a choice about where the freed capital actually goes.
IBM has placed this bet explicitly. Against the backdrop of the largest single-year wave of technology-sector layoffs on record, IBM announced in early 2026 that it would triple US entry-level hiring, specifically into the jobs the industry chorus had been declaring AI-replaceable. The reasoning, stated publicly by CHRO Nickle LaMoreaux, was that if firms stop investing in entry-level capability now, the senior pipeline collapses in three to five years and the well dries up. Most of IBM’s peers have placed the opposite bet.
We will see, over the next decade, who the winners and losers will be. The winners will not necessarily be the organizations that automate the most tasks. They will be the organizations that compound the most judgment through how they teach their workers to frame the work, steer it while it is alive, and compound what matters. And how they shape the next generation against a teacher worth learning from.
That is what reskilling must be for.
Bud Bhattacharyya is the founder of re:compound, where he works with senior leaders and expert-led organizations on operationalizing governance for serious human–AI collaboration. If this argument lands and you are working on how it actually gets implemented in your firm, write to elissa@recompound.ai.