The Judgment You Cannot See

Abstract

Two claims now dominate the enterprise AI conversation, and both are close to consensus. The first is that governance is the bottleneck: the constraint on scaling AI is control, permission, and audit. The second, arriving fast behind it, is that judgment is the scarce resource: as intelligence becomes abundant, the human capacity to decide what matters becomes the prize. Neither claim is wrong. Both are now so widely held that neither can guide a decision. When every vendor, regulator, and keynote says the same thing, the sentence has stopped being a thesis and become a background hum.

This essay argues for the claim the consensus skips. Not all judgment survives contact with cheap machine prediction. Some judgment is a complement, and its value rises. Some judgment is a substitute, and its value collapses. The two look identical on an org chart and on a dashboard, which is precisely the problem. The judgment that complements machine intelligence is disproportionately tacit, contextual, and uncodifiable, which means it is invisible to every instrument institutions use to allocate people and money: the productivity metric, the governance framework, the AI training pipeline, the restructuring plan. So institutions are cutting the expertise they most need because they cannot see it on the ledger, while protecting and automating the codified judgment that was already becoming worthless. The governance boom does not correct this error. It accelerates it, because governance can only govern what can be written down.

The two consensuses

Start by naming what everyone already agrees on, so we can stop arguing about it.

The governance consensus is now total. Gartner warns that uniform governance will cause agent failures. Regulators have moved from principle to enforcement, with the EU AI Act phasing in high-risk obligations and NIST organizing its risk framework around govern, map, measure, and manage. A dozen vendors sell control planes, gateways, and audit layers. The message is uniform: AI deployment has outrun governance, and the fix is more governance. This is true, and it is also the safest thing an executive can say in 2026, which is a warning sign. A thesis that cannot be disagreed with is not doing any work.

The judgment consensus is newer but hardening just as fast. Erik Brynjolfsson argued this summer that as execution becomes abundant, the scarce contributions become framing the right problem and judging whether an answer is any good. The World Economic Forum has built a season of programming around trust and human connection as the durable sources of value in an age of abundant intelligence. PwC’s 2026 Global AI Jobs Barometer found that jobs being reshaped to require more human judgment are growing roughly twice as fast as those AI simply makes easier, with substantially higher wage growth (PwC, 2026). The economics have a name and a pedigree: Agrawal, Gans, and Goldfarb (2018, 2022) recast AI as a collapse in the cost of prediction, which raises the value of prediction’s complements, and they identified judgment as the central one.

If the essay stopped here it would be one more voice in the choir. The interesting part is a detail buried in the same body of work that the choir has skipped.

Not all judgment is a complement

Agrawal, Gans, and Goldfarb did not say judgment is scarce and leave it there. In their 2019 formal treatment they show something sharper and less comfortable: better prediction affects the returns to different types of judgment in opposite directions. Their words, not a paraphrase of the vibe: not all human judgment will be a complement to AI (Agrawal, Gans, & Goldfarb, 2019). Some judgment becomes more valuable as prediction gets cheap. Some judgment becomes worthless, because it was only ever a workaround for prediction being expensive.

Consider the difference concretely. A credit officer who developed a feel for which thin-file borrowers were worth a second look was doing something valuable when prediction was scarce. Much of that specific skill is now a substitute for a model, and its market value is falling toward the model’s price. But the same officer’s ability to sense when a portfolio’s incentives are quietly drifting toward disaster, to know which exception is the one that matters, to read the room in a workout negotiation, none of that is prediction. It is judgment in the strict sense: a statement of what we want and what we are willing to risk, applied to a situation the training data never saw. That judgment becomes more valuable as machines get better, not less.

Side by side, the two kinds of judgment behave differently.

So the real question is not whether judgment is scarce. It is which judgment is the complement and which is the substitute, and whether institutions can tell them apart before they cut the wrong one. The evidence says they cannot, and the reason is old.

Polanyi’s Paradox is doing the opposite of what it used to

Michael Polanyi’s observation from 1966 is that we know more than we can tell. A radiologist sees the tumor before articulating why. A senior engineer feels a pull request is wrong before locating the flaw. A grower reads a field. This tacit knowledge cannot be fully reduced to rules, because the expert does not consciously possess the rules (Polanyi, 1966).

For decades this was a source of comfort to workers. Economists called it Polanyi’s Paradox and treated it as the moat: machines could do what we could specify, humans kept what we could not specify, and the unspecifiable part turned out to be the valuable part. The paradox is now cutting the other way, and the reversal is the whole point. If the complementary judgment is precisely the judgment that is tacit, then the judgment AI makes more valuable is exactly the judgment that cannot be written down. And an institution can only manage, measure, reward, govern, and train on what can be written down.

Look at the instruments a modern enterprise uses to decide who is valuable.

Every one of these instruments is blind in the same direction. Each can see codified judgment perfectly and tacit judgment not at all. So when the pressure comes to cut, the tacit expert, whose value never entered a form field, looks like overhead, and the codified specialist, whose value is legible right up until the model absorbs it, looks like a keeper. Institutions are firing the complement and promoting the substitute, and their dashboards are congratulating them for it.

Ford paid three years to relearn this

The cleanest proof on record is recent and specific. Ford spent years leaning on automated quality systems, including roughly 900 AI-assisted inspection cameras, on the theory that ingesting design requirements into AI would produce a high-quality product. It did not. Over three years the company hired, promoted, or brought back about 350 veteran engineers, internally called gray beards, to fix quality problems the automated systems could not solve. Charles Poon, Ford’s vice president of vehicle hardware engineering, was blunt about the misjudgment: AI is only as good as the information used to train it, and the company had not paid enough attention to the experience of its most knowledgeable engineers, some of whom left before their knowledge was captured. Chief Operating Officer Kumar Galhotra put it plainly: the company had relied more and more on automated quality systems and been let down, so it brought the specialists back to hunt for failure points before a part reached the plant floor. The correction worked. Ford topped the latest J.D. Power Initial Quality Study, its best result in years.

Read the Ford story through the thesis and every detail lines up. The gray beards held tacit judgment, the kind that lets a veteran feel a defect before instruments flag it. That judgment never fully entered the design requirements, which is exactly why the AI could not be trained on it and why the org could not see its value on any dashboard. So the expertise was allowed to walk out the door, treated as legible cost rather than illegible complement. The bill came due as a quality crisis, and the fix was to repurchase, at a premium and after a three-year delay, the precise judgment the company had failed to price. Ford did not discover that AI is useless. It discovered that it had cut the complement and kept the substitute, and that the ledger never warned it.

Governance makes the blindness worse

Here is where the two consensuses collide, and why joining the governance chorus is a mistake rather than merely a cliché. Governance is a codification engine. Its entire function is to convert action into artifacts: policies, permissions, logs, audit trails, approvals. That is genuinely necessary for machine action, and I am not against it. But notice what it does to the judgment problem. A governance framework can only recognize, reward, and protect judgment that produces a trace. The tacit judgment that is the true complement to machine intelligence produces no trace by definition, so the more an institution invests in governing what it can see, the more invisible the uncodifiable complement becomes by comparison.

This is why the AI-native restructuring wave is more dangerous than it looks. When Coinbase eliminated what its CEO called pure managers and reorganized around AI-native pods, some staffed by a single person directing fleets of agents, the framing was that management overhead was being removed. But management is where a great deal of tacit judgment lives: knowing which exception to escalate, which quiet signal means trouble, which junior needs cover. Fold that into a single overloaded role and the supervision does not disappear. It relocates onto a person whose dashboards will never show the judgment they are now carrying. The Glean Work AI Index found workers already spend an average of 6.4 hours a week on the uncounted labor of supervising, correcting, and cleaning up after AI, close to a full working day that appears on no productivity metric (Glean, 2026). Governance will faithfully audit the agents’ actions. It will not see the human judgment quietly holding the whole arrangement together, right up until that judgment is cut and the quality crisis arrives, on schedule, the way it did at Ford.

The allocation question, restated

If the complement to machine intelligence is tacit judgment, and tacit judgment is invisible to every standard instrument, then the durable advantage in the next economy is not the largest model, and it is not the tightest governance stack, both of which everyone will eventually have. The durable advantage is the ability to find, price, protect, and allocate the specific judgment that no form field can hold. That is a genuinely hard capability, which is why it is defensible. Anyone can buy a frontier model. Anyone can buy a control plane. Almost no one can reliably tell, before the quality crisis, which of their people hold the tacit complement and which hold the soon-worthless substitute.

This reframes the labor question too. The winners will not be the workers who use the most AI, and firms will not win by cutting the most headcount. The winners will be the institutions that build deliberate pathways for tacit expertise to form, transfer, and compound, through apprenticeship, participation, and situated practice, because tacit knowledge moves only through people in real consequence and never through documents alone. The firms racing to strip out the human layer are dismantling the transmission mechanism for the exact expertise their AI systems will depend on, and they are doing it because their instruments told them that layer was cost. New work will keep forming around machine intelligence, and it will keep carrying wage premiums for scarce expertise, as the labor economics predict. But those premiums will accrue to the institutions that can see the invisible complement, and they will be paid, late and at a markup, by the institutions that could not.

The test, and the discomfort

There is a way to check whether your own institution is making this error, and it is uncomfortable on purpose. Take the last round of cuts, or the next AI-native restructuring on the roadmap, and for each role removed or collapsed, ask one question: was this person’s real value ever captured in the metrics used to justify the decision? If the honest answer is that their contribution lived mostly in judgment that never made it into a dashboard, a document, or a training set, then you did not remove overhead. You removed the complement to the machine you are betting on, and you will repurchase it later at Ford’s price.

So the question I would not want to answer in front of my own board is this. Every instrument we own can see codified judgment and is blind to tacit judgment, and the tacit kind is the part AI makes more valuable, not less. Given that, how confident are we that our productivity metrics, our governance dashboards, and our restructuring plans are cutting the expertise we no longer need, rather than the expertise we cannot yet see we need? And if we cannot answer that with evidence, what exactly are we governing so carefully, other than the wrong half of the problem?

A note on AI-assisted research and intelligence provenance

The market-signal research and monitoring behind this essay drew on Astra, the Adaptive Strategic Thesis Reasoning Agent under development at 451 Labs. Source lineage and claim provenance were tracked with Clearstory, the knowledge and intelligence provenance system that is the primary realization of the Astra thesis at 451 Labs. Neither Astra nor Clearstory is a shipped product. Both are theses under active development. All analytical judgments, contestedness ratings, gap identifications, and prognosis claims in this essay remain the author’s own.

References

Agrawal, A., Gans, J., & Goldfarb, A. (2018). Prediction machines: The simple economics of artificial intelligence. Harvard Business Review Press.

Agrawal, A., Gans, J. S., & Goldfarb, A. (2019). Exploring the impact of artificial intelligence: Prediction versus judgment. Information Economics and Policy, 47, 1–6.

Agrawal, A., Gans, J., & Goldfarb, A. (2022). Power and prediction: The disruptive economics of artificial intelligence. Harvard Business Review Press.

Autor, D., Chin, C., Salomons, A. M., & Seegmiller, B. (2026). What makes new work different from more work? (NBER Working Paper No. 34986). National Bureau of Economic Research. https://doi.org/10.3386/w34986

Glean. (2026). Work AI index 2026. Glean Work AI Institute.

Polanyi, M. (1966). The tacit dimension. Doubleday.

PwC. (2026). 2026 global AI jobs barometer. PricewaterhouseCoopers.