Organizations are not just consumers of capability; they are producers of it. Every organization runs two systems simultaneously: one that produces outputs, and one that produces the people capable of producing those outputs. Most have become very good at measuring the first. Almost none are measuring the second. That is the gap AI is quietly opening up, and closing it is well within reach.
This article is about the second system.
Most leaders reading this will recognize something in their organization right now. The work is getting done faster, and the outputs are genuinely better. By almost every visible measure, productivity is up.
And yet something feels slightly off about the pipeline. Senior hires are coming from outside more often than expected. The succession plan looks thinner than it did five years ago. People who should be ready for bigger roles are technically capable but reluctant to take a position, own a room, or make a call when the situation is genuinely ambiguous.
The rest of this article names that dynamic, traces how it works, and offers four concrete practices to address it.
What expertise debt is
Expertise debt is the risk that occurs when the conditions that historically produced human capability change in ways that productivity metrics cannot see. Capability, as used here, refers to the human capacities organizations depend on but do not directly purchase; i.e., judgment, coordination, leadership, creativity, resilience, and abilities that emerge through experience rather than instruction.
This is not a new name for human capital erosion, knowledge loss, or succession pipeline weakness. Those phenomena exist independently of how work is organized. What is new is the mechanism: AI-driven productivity improvement that simultaneously removes the developmental substrate those capabilities depend on, while every standard metric reads the situation as improvement. What makes expertise debt worth naming is not that capability declines; organizations have always faced that risk. It is that expertise debt accumulates invisibly behind rising output numbers. A succession problem caused by attrition or underinvestment shows up in the data. One caused by AI-altered developmental conditions, in an organization whose productivity metrics are all moving in the right direction, does not. Expertise debt is the first and most visible instance of this dynamic. Judgment is where it surfaces earliest because it is most directly connected to the analytical work AI is changing first.
Capability debt is the parent theory. Expertise debt is its first measurable instance, the one closest to the surface, where the signal is clearest and the intervention most direct. The relationship between them matters for how this framework is used: capability debt will eventually surface in leadership pipelines, coordination capacity, and original idea generation as AI reshapes those domains. Expertise debt is where it becomes visible first because judgment is most directly exposed to the analytical work AI is changing now. Defined precisely, expertise debt is a function of two variables: how much developmental work has been removed from roles by AI, and how much developmental capacity has been built to replace it. Organizations that remove significant developmental work without rebuilding the conditions that produced it accumulate expertise debt. Those that invest deliberately in developmental substitutes, structured AI use, peer learning, decision narration, and real ownership are paying down debt as they go. The gap between removal and replacement is the exposure. This also explains why the four practices later in this article are not optional add-ons: they are the replacement side of the equation. The good news is that closing the gap does not require slowing down AI adoption. It requires designing for both systems at once.
This is not an argument that AI is bad for organizations. It is an argument that, as AI changes the nature of work, it may be changing the conditions that historically produced human capability in ways that will not appear in any standard metric until the cost is already significant.
An example
Consider a professional services firm. In 2019, people spent significant time on research: reading sources, synthesizing information, identifying patterns, and forming initial views. The work was slow, and much of it was routine. But it was also where people built their mental models of the domains they worked in, developing a sense of which sources were reliable, which patterns repeated across situations, and where conventional analysis tended to miss important nuance. That capability compounded over the years and became the foundation for the senior judgment the firm depended on.
By 2025, AI handled most of that research. People produced better work faster. Every visible performance metric has improved.
What has not been measured is whether people are still building the same mental models through different work, or whether that building is happening less because the work that produced it has changed. If it is the latter, the firm will not know for several years. The first signal will be people who are technically strong but struggle with the contextual judgment that distinguishes excellent senior leadership. The second will be a succession plan that is harder to fill from within. By the time those signals arrive, the conditions that created them are years in the past.
This pattern has been documented in fields with long training pipelines. Radiology programs reported in the early 2020s that residents who trained heavily on AI-assisted reads were faster and more accurate on standard cases but slower to develop the visual intuition that senior radiologists use to catch unusual presentations, the kind of judgment that does not transfer from a model’s confidence score. Military training research has shown similar dynamics: simulation-heavy programs improve procedural performance while sometimes reducing the adaptive decision-making that develops through high-stakes, unscripted exposure. In both cases, the same structure: outputs improve, a specific developmental pathway changes, and the cost appears later and elsewhere.
The same question applies wherever the work being automated was also doing developmental work. In product organizations, AI-generated user research is faster and more comprehensive than manual synthesis, but doing it manually builds the researcher’s intuition for which questions matter and which patterns are meaningful in context. In legal practices, AI document review is more thorough than manual review, but reading cases and contracts built the contextual understanding that later became judgment about what matters. The question in each case is the same: is the work being changed also the work that was building something?
A direct caveat is warranted here. The longitudinal evidence that AI-heavy organizations produce weaker future leaders does not yet exist; the timelines are too short. The analogies from radiology, military training, and flight simulation are the strongest available evidence, not proof. They establish that the mechanism is real and has precedent, not that it will manifest in professional organizations at scale in the same way. This article is a leading indicator argument: the conditions for expertise debt are being set now, and the organizations that act early will be ahead of a problem most others will not see until it is already expensive to fix. That is not a reason for alarm. It is a reason for deliberate design.
What AI may specifically be changing
These patterns existed before AI. Organizations have always had knowledge that was hard to transfer, pipelines thinner than ideal, capable people who were better at executing than deciding. What AI may be changing is the mechanism that used to keep those problems in check.
For most of the history of professional work, capability developed as a natural byproduct of doing the job. Nobody designed it as development. It happened automatically because the work required it. Someone working through a research question was not just producing a deliverable. They were building pattern recognition, encountering edge cases, and developing the tacit understanding that later became judgment. AI may be changing that dynamic: outputs improve, developmental conditions shift, every visible metric goes up, while something harder to see changes.
The evidence from fields with longer experience, such as surgery simulation, flight training, and radiology, points to sequence as the determining factor: whether AI is used after the human has formed an independent view, or instead of forming one. When AI substitutes for the cognitive struggle, the developmental work disappears. When it follows that struggle, it can sharpen and accelerate it. AI may itself become a more powerful developmental tool than the work it replaces: faster feedback loops, transparent reasoning, exposure to edge cases no junior person would encounter for years. That is a genuinely exciting possibility. But it requires deliberate design. The organizations that capture it are the ones that treat sequence as a conscious choice rather than a default.
When it becomes expensive
The costs emerge slowly and become significant when organizations try to do things that the pipeline can no longer support.
Early on, outputs are better, and productivity is higher. There is no obvious pain. The developmental conditions may be changing, but nothing in standard metrics reflects this.
In the middle stage, several signals appear together: succession planning shows fewer ready candidates than expected; people moving into demanding positions take longer to reach full independence; external hiring for senior roles increases; institutional knowledge begins concentrating on those who built their capability before the work changed.
Later, the consequences become structural. Leadership roles are harder to fill from within. Organizational adaptation slows. Strategic decisions require more external input. Correction is expensive because the root cause is years in the past.
Pipeline depth is the most visible place where expertise debt manifests. If productivity is rising while candidates ready for senior roles across multiple time horizons are declining, that gap is expertise debt made measurable. But pipeline depth is not the only signal. Time to independence, the proportion of senior roles filled externally, the concentration of critical knowledge in a small number of people, and the rate at which people meaningfully challenge AI-generated recommendations are all indicators of the same underlying condition.
Not all organizations are equally exposed, and understanding where you sit matters. Vulnerability tends to be higher where training pipelines are long, and the formative work is primarily analytical; where AI adoption has been fast and broad rather than targeted; and where the ratio of junior-to-senior roles means few people have built capability under pre-AI conditions. Professional services, knowledge-intensive industries, and organizations that grew rapidly on AI tooling from the outset face the steepest exposure. Organizations with shorter pipelines, stronger apprenticeship cultures, or deliberate developmental infrastructure have more room to maneuver: not immunity, but more time and more options. Knowing which category you are in shapes where to start.
Diagnosing it in your organization
The following questions are a starting diagnostic. Risk signals indicate conditions where expertise debt is likely accumulating. Protective signals indicate conditions that counteract it. The pattern across the full set matters more than any single answer.
| Diagnostic question | Signal |
| Has developmental work been removed from the role? | Risk |
| Is pipeline depth declining across time horizons? | Risk |
| Is external hiring for senior roles increasing? | Risk |
| Is critical knowledge concentrating in a few people? | Risk |
| Are people challenging AI outputs with their own views? | Protective |
| Are leaders narrating live decisions in team settings? | Protective |
| Do promotions require a developmental track record? | Protective |
An organization with several risk signals and few protective ones is accumulating expertise debt regardless of what its productivity metrics show. The protective signals are not just indicators; each one corresponds directly to a practice that can be built.
What to do about it
Organizations do not produce capability directly. They produce the conditions from which capability emerges: pattern recognition built through repeated exposure, confidence built through genuine ownership of consequential decisions, resilience built through failure and recovery, and social understanding built through navigating real organizational dynamics.
Four suggested changes might address those conditions directly. Each modifies something that already exists in every organization. None requires a new budget, significant process change or new systems.
Narrate live decisions in existing meetings.
Once a month, a senior leader spends a few minutes in an existing meeting describing a real decision they are currently facing, outcome unknown. They share the options, what they are weighing, and where they are genuinely uncertain. The team asks questions. What this produces is tacit knowledge transfer that used to happen naturally through proximity, pattern recognition built by watching experienced reasoning operate in real time, and psychological safety around uncertainty that changes how the whole organization approaches ambiguous situations. The key is genuine uncertainty rather than finished thinking, which requires the most senior people to go first and make it normal.
Ask for a developmental record alongside the performance record in senior promotions.
When someone is promoted into a role requiring significant judgment, their sponsor provides a specific account of two or three decisions the candidate owned under genuine uncertainty: what was the situation, what were the options, what did they decide, why, and what happened. This reshapes what the whole organization optimizes for. When people know that reaching senior levels requires a track record of owning consequential decisions, they seek those situations out, document their reasoning, and ask for feedback on their thinking rather than just their outputs. One change to the promotion conversation changes developmental behavior across the organization.
Respond to problems with reasoning rather than answers.
In regular one-on-ones, when someone brings a problem or decision, the response is not ‘here is what I would do’ but ‘here is how I am thinking about this: what I am weighing, what I am uncertain about, what would change my view.’ In practice, this means pausing before answering, naming the competing considerations out loud, and sometimes ending without a conclusion: ‘I am not sure yet; here is what I would want to know before deciding.’ What this produces is reflection and metacognition, the habits that underlie good judgment. People can get answers from AI faster than from a colleague. What they cannot get from AI is a window into how an experienced person actually thinks through genuine uncertainty: the false starts, the competing considerations, the moment where the frame shifts. That is precisely what this practice preserves, and it costs nothing except the willingness to think out loud.
Assign real ownership of decisions to people ready to grow into them.
Some decisions that currently default to the most senior people available could be owned by others. Not the large irreversible ones, but ones consequential enough to matter and recoverable enough that an imperfect call is not catastrophic. The owner frames the options, makes the call, lives with the result, and reflects afterward on what they would do differently. This works because it changes the structural conditions of work, not just the behavior of an individual manager. Accountability and identity form through repeated ownership of difficult decisions. Pattern recognition deepens through direct exposure to how situations actually unfold. Both compound over time in ways that no amount of observation or instruction can replicate.
Two more practices worth building
Build structured peer learning around real decisions. A regular forum, monthly works, and bi-weekly is better, where people at similar levels share decisions they are currently navigating: what they are uncertain about, what they are leaning toward, what they would need to know to decide confidently. Others respond through their own reasoning rather than conclusions. No advice-giving, no right answers. This is the most scalable change in this list because it does not depend on any individual manager’s willingness to be vulnerable or any senior leader’s calendar. It distributes capability rather than concentrating it, and it creates the conditions for collective judgment to develop alongside individual judgment.
Build deliberate practice into how AI is used. When AI produces analysis or recommendations, create a step where people assess the situation independently before reading the AI output, then compare their view against it. This is not about distrusting AI. It is about ensuring AI is used in ways that develop capability rather than substitute for it. The difference between AI that builds judgment and AI that replaces it often comes down to this one structural choice about sequence.
How to know it is working
Three things worth tracking quarterly: whether senior leaders are narrating live decisions in team settings; whether people are genuinely owning decisions rather than merely contributing to them; and whether people are applying their own judgment to AI recommendations before acting on them. If the third is declining while output quality is improving, expertise debt is likely accumulating. Developmental work is being removed faster than it is being replaced, and that is the signal to act.
Annually: whether senior promotions include a developmental record; whether the pipeline is getting deeper or shallower across three time horizons; and whether time to independence in demanding roles, external hiring rates, and knowledge concentration are moving in the right direction. These signals together give a more complete picture than pipeline depth alone.
What to expect
Year one: different conversations, not different pipeline outcomes yet.
Year two: the pipeline starts looking different. More people are on a credible trajectory, even if the ready-now count has not changed.
Year three: the organizational signals become answerable with real data. The organization knows not just who is ready but who is developing and what they still need.
None of this requires structural change or significant budget. It requires some senior people to behave differently in meetings and conversations that are already happening, and a few deliberate choices about how work and AI use are structured.
Organizations are not just consumers of capability. They are producers of it. Most have become very good at measuring what AI helps them produce. The next step is measuring whether they are still producing the capabilities those gains depend on.
That measurement, built now, will matter more in five years than it does today. The organizations that build it early will have options that those who wait will not. Expertise debt is where that work begins. But the larger idea it points toward, that organizations are capability production systems whose developmental infrastructure deserves as much deliberate design as their operational systems, is the more important insight. That is the argument this series has been building toward. The good news is that it is entirely within reach.
—————————————————————————————————————————-
This is the final article in a three-part series on AI and the future of expertise. Part 1, AI is changing the value of expertise, examines how AI shifts what organizations are actually paying for when they hire. Part 2, AI is changing how expertise develops, traces the developmental pathways that are quietly changing as AI takes over more of the formative work. Part 3 is this article.
The Expertise Debt Framework v1.0, including the Expertise Demand Canvas, Expertise Debt Diagnostic, Judgment Development Model, maturity model, and organizational archetypes, is available as a standalone practitioner’s guide.
