An Insynergy Insight in response to David Duncan's "How Do Workers Develop Good Judgment in the AI Era?" (Harvard Business Review, February 2026)
There is an observation that many senior leaders will recognize but few have paused long enough to examine.
When generative AI entered their organizations, it was the experienced people — the partners, the directors, the senior product managers — who benefited most. They moved faster. Their output improved. AI seemed to amplify something already present in how they worked.
For junior staff, the experience was different. Output arrived quickly and looked polished, but something was off. The work lacked the quality of conviction. When pressed on a recommendation, the person behind it often could not explain why it was right — or even whether it was. The tools had generated an answer. The person had not yet formed the capacity to evaluate it.
Most organizations noticed this asymmetry and treated it as a training problem. A gap in AI literacy. A need for better prompting skills. Some introduced review layers. Others created escalation protocols, routing ambiguous decisions upward to people with enough experience to judge.
These responses are reasonable. They are also insufficient, because they address the surface without examining the structure beneath it.
Something has shifted in how organizations produce people who can decide.
For decades, the development of judgment was not a program. It was a byproduct. Junior consultants wrote flawed analyses and received feedback. Junior product managers drafted rough specifications and defended them. Junior marketers built campaigns from nothing and watched them succeed or fail. The work was slow, often imperfect, and rarely efficient — but it placed people inside the consequences of their own choices.
That structure was never designed with particular deliberateness. It simply emerged from the economics of work: organizations needed humans to perform tasks, and in doing so, those humans accumulated the experience from which judgment eventually grew. Ownership was embedded in the workflow, not because anyone planned it, but because there was no alternative.
AI changed the economics. It did not change the need.
Now the tasks that once served as developmental scaffolding — research synthesis, first-draft analysis, pattern identification, option generation — can be completed in minutes by someone who has never done them before. The output is structurally competent. It arrives with a tone of confidence. And the person reviewing it is placed, for the first time, in the position of an editor who has never written.
David Duncan, writing recently in Harvard Business Review, articulates this tension with unusual precision. He describes a paradox that most organizations have yet to fully confront: AI simultaneously increases the need for judgment while eroding the experiences that historically produced it. The need grows because AI-generated output requires constant evaluation — what Duncan calls the iterative cycle of prompting, assessing, and redirecting. But the experiences that once built evaluative capacity — the slow, formative repetition of doing the work itself, under real ownership — are exactly what AI displaces.
Duncan's five dimensions of judgment — evaluative, contextual, tradeoff, anticipatory, and ownership — are useful not because they create a taxonomy, but because they reveal what is actually at stake. These are not skills that can be taught in a workshop. They are capacities that emerge when someone is placed inside a decision, made responsible for it, and allowed to experience its consequences over time.
The paradox, then, is not merely an inconvenience. It is a structural condition. And structural conditions require structural responses.
There is a way of seeing this problem that clarifies where intervention is most needed.
Every organization, whether it recognizes it or not, operates with an implicit set of boundaries that determine where AI's contribution ends and where human judgment must begin. These boundaries are not technical thresholds. They are organizational choices — sometimes deliberate, more often inherited — about who decides, who owns the outcome, and who encounters the consequences of being wrong.
When AI accelerates the production of an analysis, it moves work up to a certain point with remarkable speed. But judgment, responsibility, and the capacity to evaluate do not cross that point automatically. They require a person who has been placed, intentionally, on the other side of it — someone whose role includes not just reviewing output but owning the interpretation, bearing the uncertainty, and learning from the result.
This is the boundary that matters: not the line between what AI can do and what it cannot, but the line between where work is produced and where judgment is exercised. When that boundary goes unexamined, organizations default to arrangements that concentrate judgment among those who already have it and remove it from those who need to develop it. Escalation protocols, while protective, often teach junior staff that ambiguity is something to hand off. Review layers, while valuable, can reduce a person's relationship to their own work to one of curation rather than authorship.
Duncan observes this dynamic clearly. He notes that "human in the loop" designs, however well-intentioned, tend to optimize for control rather than for the development of the humans who are in the loop. The protocols manage risk. They do not produce the next generation of people capable of managing it.
To see the problem this way is to recognize that what is needed is not a better training program but a more deliberate architecture of decision-making itself.
If judgment forms through exposure to consequence, then organizations must design where that exposure occurs. If ownership sharpens evaluative capacity, then roles must be structured so that ownership is distributed with intention, not concentrated by default. If repetition under real stakes is what transforms competence into judgment, then some portion of that repetition must be preserved — not as inefficiency, but as infrastructure.
This is the work of designing decisions: not automating them or merely overseeing them, but shaping the conditions under which people learn to make them well. It means asking, for each role and each workflow, where the boundary sits between AI's contribution and human responsibility — and whether that boundary has been placed where it serves not only today's efficiency but tomorrow's organizational capacity.
Duncan points toward this when he draws on medicine and the military as domains where judgment has long been built deliberately — through simulation, graduated responsibility, case-based learning, and structured reflection. These are not coincidental features of those professions. They exist because the cost of learning through unstructured experience is too high. The insight for every other industry is that AI has created the same condition: the natural apprenticeship is no longer sufficient, and what replaces it must be designed.
Return, for a moment, to the observation at the beginning.
AI helped the senior partner more than the junior analyst. The explanation that most organizations reached for was that the junior person lacked skill. But the deeper explanation is that the junior person lacked the conditions under which skill becomes judgment — conditions the organization had never deliberately created, because it had never needed to. The work itself once provided them.
AI did not remove judgment from organizations. It removed the conditions under which judgment was incidentally formed. And because those conditions were never the product of conscious design, their disappearance went largely unnoticed — until the consequences began to surface in thinning leadership pipelines, in work that looks complete but cannot be trusted, in a growing dependence on a narrowing group of people who learned to decide in a world that no longer exists.
The question Duncan raises is the right one. The answer, we believe, is structural. Not "how do we train for judgment?" but "have we designed the places where judgment is meant to develop?" The distinction is not semantic. It is the difference between treating a symptom and redesigning the system that produced it.
Organizations that take this seriously will not simply add AI literacy to their development curricula. They will examine where, in their decision architecture, the boundary between AI's work and human responsibility has been drawn — and whether it has been drawn at all. They will ask whether their junior staff are placed inside decisions or merely adjacent to them. They will consider whether their review structures develop judgment or merely verify output.
These are quiet questions. They do not lend themselves to announcements or transformation programs. But they are the questions that will determine whether the next generation of leaders arrives prepared to decide — or merely prepared to review what a machine has already decided for them.