The Expertise Paradox

Why AI raises the value of judgment — precisely as organisations are told they need less of it

This essay is part of a continuing series on leadership in complexity. Earlier posts traced how specialisation built the modern world (The Age of Mastery), how complexity broke the logic of control (The Great Unravelling), and what The Specialist Dilemma costs leaders who built their identity on mastery. This essay turns to a newer pressure: what AI does to expertise — and to the people responsible for developing it.

Here is the uncomfortable thing nobody in the AI briefing is saying out loud.

The people in your organisation getting the most from these tools are, almost certainly, your most capable people. Not your most enthusiastic. Not the ones who attended the lunch-and-learn and promptly built a personal Copilot. Your most capable — the ones with enough accumulated judgment to recognise when the output is wrong, redirect it usefully, and take genuine responsibility for what comes out the other end.

This is not what the pitch deck said. The pitch deck said AI would compress the gap between your best people and your average ones. That a junior analyst armed with the right prompts could produce work that rivalled a senior one. That organisations could scale expertise they didn't have, develop capability they couldn't afford, and close skill gaps they hadn't managed to close in the previous decade of trying.

And here is the part worth sitting with: the pitch deck is not wrong, exactly. All of that is, in some meaningful sense, true. The gap does compress. The tool does help. A junior analyst can produce something a senior one would recognise as passable.

The question is what that costs.

The evaluator problem

There is a structural asymmetry at the heart of AI assistance that receives far less attention than the capabilities it unlocks — which is, one suspects, not an accident.

To use a tool well, you need to know what good looks like. Not approximately. Not in outline. The kind of knowing that is felt rather than listed — the experienced clinician's instinct that something in the presentation doesn't fit, the seasoned finance director's unease with a set of numbers that pass every explicit test but still feel wrong. That knowledge is not a nice-to-have for AI-assisted work. It is the precondition for it.

The model does not know when it is wrong. It produces its output — fluent, structured, confident — with precisely the same surface quality whether it is brilliant or fabricated. It cannot signal the edges of its competence because it has no access to them. The responsibility for knowing where the model ends and error begins sits entirely with the person receiving the output.

Strip that person of sufficient expertise and you have not closed the capability gap. You have hidden it behind a polished surface. Which, it turns out, is worse.

This is the expertise paradox in its plainest form. AI is most useful to people who know enough to evaluate it. It is most dangerous to people who don't.

What organisations are actually saying

Organisations are, quite reasonably, excited. The tools are remarkable. The time savings are real. The use cases multiply faster than anyone can pilot them. And in the middle of all of this genuine enthusiasm, a set of conclusions is being drawn that deserve more scrutiny than they are getting.

The first is that expertise can now be developed faster. That the years of deliberate practice, accumulated failure, and hard-won pattern recognition that used to constitute professional capability can be meaningfully compressed by giving people better tools earlier. There is something to this. But what gets quietly dropped is that the tools are only as useful as the judgment of the person wielding them — and that judgment cannot be borrowed from the tool. The tool will not tell you when to trust the tool.

The second conclusion is that you need fewer experts than you used to. That a smaller number of highly capable people, augmented by AI, can do the work that once required a larger team of domain specialists. This is probably true in some contexts and dangerously false in others, and the organisations making this calculation are not, in the main, working hard to distinguish between the two.

The third — the most seductive, and the most worth interrogating — is that the investment case for deep expertise has gone down. Why develop expensive, slow-to-build human capability when the model can approximate it on demand?

The answer is that the model cannot approximate it on demand. It can produce the form of expertise on demand. The structure. The vocabulary. The confident surface. What it cannot produce is the independent judgment required to know whether what it just produced is right.

The borrowed judgment trap

Consider what is actually happening when an under-experienced professional uses AI to produce work beyond their current capability.

They are borrowing judgment they have not yet earned. The output may be serviceable — it may even be good. But if they cannot evaluate it, if they lack the domain knowledge to assess the reasoning, catch the errors, and stand behind the conclusions, then the work is not theirs in any meaningful sense. It carries their name but not their accountability. And when something goes wrong, as eventually it will, they will not know why. Because they were never fully inside the thinking.

This is not a new problem. Organisations have always had people producing work above their actual comprehension. What AI does is industrialise it — make it faster, smoother, and far easier to miss until it matters.

The surface has never looked this good. The gap between surface and substance has never been this easy to paper over.

The organisations most exposed are those deploying AI as a substitute for the development they should have been doing anyway. The ones where the promise of the tool has allowed leadership to defer the harder conversation about why the pipeline of capability is thinner than it should be. Where the dashboard shows impressive adoption numbers — prompts run, hours saved, documents generated — and nobody is asking whether the quality of judgment in the organisation is improving or quietly eroding.

Research on AI-induced deskilling puts the risk plainly: the capacity to oversee tasks effectively can itself be lost when AI enables people to operate outside their domains of expertise. A project manager who delegates technical decisions to a model, lacking the engineering background to evaluate what comes back, isn't supervising AI. They are deferring to it. The distinction matters, and it compounds.

That is a question worth asking before the answer reveals itself somewhere inconvenient.

The investment case, restated

Here is the argument that the current moment actually makes, if you are willing to follow it somewhere uncomfortable.

The value of expertise has not declined. It has shifted — from the production of work to the evaluation of it. The expert who used to spend most of their time generating analysis now spends more of it reviewing, questioning, and improving what the model generates. The skill set is similar. The leverage is higher. And the scarcity, if anything, is greater — because the premium now attaches specifically to people who can tell good from plausible, and there is no shortcut to producing them.

Which means the investment case for developing genuine expertise — the slow, expensive, effortful kind that produces people who actually know things, rather than people who can generate the appearance of knowing things — has not gone down.

It has gone up.

Not because AI isn't useful. It demonstrably is. But because AI has raised the stakes on getting the evaluation right, and evaluation requires the one thing the model cannot supply.

This is not an expertise problem. It's a sensemaking problem.

There is a temptation, at this point in the argument, to conclude that the answer is simply more expertise. Deeper specialists. More credentials. Double down on mastery.

That is the wrong conclusion, and the Weave series has been building the case against it for some time.

The Age of Mastery showed how specialisation, taken far enough, produces people who "master the threads but lose the fabric" — experts who optimise their own corner of the system while missing the emergent risks created by its interplay. The Specialist Dilemma showed what this costs individuals: identities so tightly bound to their domain that adaptation feels like betrayal.

The expertise paradox doesn't reverse this. It sharpens it.

What AI demands of leaders is not specialists who go deeper. It is people who can synthesise — who have the domain knowledge to evaluate the model's output and the contextual range to understand where that output lands. Who can tell when the analysis is technically correct but organisationally wrong. When the recommendation is logically coherent but practically unworkable. When the summary is accurate and the conclusion it implies is still misleading.

That capacity — what The Four Disciplines of Weave calls sensemaking — is exactly what AI cannot replicate and exactly what borrowed judgment erodes. The Weavist leader who can hold the expert's eye and the systems thinker's perspective simultaneously is not just better placed to lead in complexity. They are, as it turns out, the only person in the room capable of genuinely supervising what the model produces.

This is the resolution the expertise paradox points toward. Not more depth. Not more breadth. More integration — the capacity to move between knowing and evaluating, between the thread and the fabric, without losing either.

What this demands of leaders

The executives most at risk here are not the sceptics who haven't adopted AI. They are the enthusiasts who adopted it without thinking clearly about what it changes — and, more importantly, what it doesn't.

What it changes: the speed of production, the cost of first drafts, the accessibility of structured thinking about unfamiliar problems.

What it doesn't change: the requirement for independent judgment at the point of consequence. The accountability that attaches to decisions made in your name. The need for organisations to develop people who genuinely know things — not people who can efficiently retrieve the appearance of knowing them.

A Microsoft and Carnegie Mellon study of 319 knowledge workers found that higher confidence in AI is directly associated with less critical thinking — and that for 40% of tasks, workers reported applying no critical scrutiny to AI outputs whatsoever. The researchers also found that those with high self-confidence in their own domain expertise were significantly more likely to question and refine what the model produced. Which is a precise, empirical restatement of the paradox: expertise is not the problem AI solves. It is the condition AI requires.

The honest version of an AI strategy, for any organisation with high stakes attached to its decisions, is not how do we use AI to do more with less expertise? It is how do we use AI to extend the reach of our best judgment — and what do we need to do to ensure there is enough of that judgment to go around?

The second question is harder. It does not compress into a roadmap or a business case with a payback period. It is, in the end, a question about what kind of organisation you are building — and whether the people inside it are becoming more capable, or simply becoming faster at producing the surface of capability.

Both are possible. Only one is worth doing.

The tools will keep improving. The judgment required to use them wisely will not improve automatically. That part is still on you — and, more precisely, it is on the people you are responsible for developing.