In The Judgment Reserve, we proposed that organizations may be quietly consuming the human capacity that allows them to distinguish a plausible answer from a correct decision.
The first question was generational: what happens if we use artificial intelligence to replace the very tasks through which young professionals developed the judgment they will need when they become senior?
But that is only part of the problem. The next question matters much more for management: if the Judgment Reserve is an asset, how is it created, preserved, depleted and governed?
We are discovering a paradox. Machines are becoming more capable. Precisely for that reason, their errors are becoming harder to detect.
The problem is no longer an AI that answers badly
Early generative models often made obvious mistakes. They fabricated references, produced inconsistent arguments or displayed limitations that a trained professional could identify quickly.
Current models are radically different. They can write software, analyze extensive documents, produce financial models, work on contracts, develop strategies, research information and execute increasingly complex sequences of actions.
Stanford's 2026 AI Index captures the scale of the advance: on OSWorld, a benchmark that evaluates agents performing computer tasks, performance rose from roughly 12% to 66.3%.
But the same figure can be read another way: even after that extraordinary improvement, agents still fail roughly one in three attempts on this kind of structured benchmark.
METR adds an even more important warning. Its time horizon does not mean that an AI can autonomously receive any project equivalent to that number of human work hours. For activities whose outcome is hard to verify or whose errors have serious consequences, success probabilities above 98% may be necessary. As complexity grows, human interventions may become less frequent but much more expensive, because finding what failed can demand more work.
Error is not necessarily disappearing at the same pace as capability increases. It is changing its appearance.
From obvious error to plausible error
The evolution can be summarized as follows:
- Early AI: mediocre answer, visible error and relatively easy correction.
- Advanced AI: excellent answer, coherent structure, a hard-to-detect false assumption and a persuasive but mistaken conclusion.
The second situation may be more dangerous than the first. Not because the model is worse, but precisely because it is much better.
When a tool produces one hundred elements and ninety-nine are correct, the main problem is no longer producing the hundred. It is identifying the one that is wrong.
If that element is the financial assumption driving an entire model, a contractual clause interpreted incorrectly, a mistaken technical premise or a non-existent regulatory reference, the error can propagate through extraordinarily sophisticated work.
Net AI productivity = useful output − supervision cost − expected cost of residual error.
The third variable is the hardest to measure. It may become one of the most important.
The Judgment Reserve is not accumulated knowledge
An organization can store millions of documents and possess very little Judgment Reserve. It may have perfectly written procedures, complete databases, project histories and advanced knowledge-management systems.
That is information. Judgment is something else.
The Judgment Reserve is the distributed capacity of an organization to determine when an apparently correct answer deserves trust, when it requires verification and who has sufficient competence to perform that verification.
It includes technical knowledge, but also experience, organizational memory, contextual understanding, the ability to recognize exceptions, the skill to identify assumptions, knowledge of prior errors and the capacity to formulate alternative hypotheses.
It also contains something especially difficult to formalize: knowing that something does not fit before being able to explain exactly why.
That capacity is usually the product of years of exposure to real problems. It cannot be downloaded from a database.
An asset that does not appear on the balance sheet
An organization can think about its Judgment Reserve in a way similar to any other productive capital:
JRt+1 = JRt + training + experience + transfer − exits − obsolescence − deskilling.
This is not yet intended as a mathematical formula. Its purpose is to show that judgment has dynamics. It accumulates. It transfers. It can become dangerously concentrated in a few people. It can become obsolete. It can also deteriorate through lack of use.
The last point will become increasingly important. The risk is not only that organizations stop training junior staff. There is another, less visible risk: experienced professionals may lose part of their critical capacity by systematically delegating to AI the activities through which they exercised that judgment.
A company could then experience two processes at once: failing to create enough new judgment and ceasing to exercise the judgment it already has. That would be genuine intellectual decapitalization.
Productivity and decapitalization can coexist
This is one of automation's most important business paradoxes. A company can reduce headcount, increase output, shorten lead times, lower costs and raise profits while destroying future capacity.
It would not be entirely new. A factory can temporarily improve its results by cutting maintenance. While the machines keep running, the decision looks excellent and profit rises. The problem appears years later, when accumulated degradation exceeds the savings achieved.
Artificial intelligence may produce a similar effect on intellectual capital.
A company can appear more efficient while silently consuming its Judgment Reserve.
Financial statements will not show that depreciation until the missing capacity is needed.
Education must also change its objective
For centuries, a significant part of education was designed around the scarcity of knowledge. The teacher possessed information. The student had to acquire it. The student then demonstrated an ability to reproduce or apply it.
AI changes that relationship. A machine can produce an explanation, essay, computer program, translation, commercial strategy or mathematical solution in seconds.
A growing share of educational value must therefore shift from producing an answer to evaluating an answer.
Instead of “solve this problem,” we can ask: “This is the solution proposed by an AI. Find its assumptions, determine which ones require verification and explain the circumstances under which the solution would be wrong.”
Instead of “draft this contract,” we can ask: “Identify the risks contained in this perfectly written contract.”
Instead of “build this financial model,” we can ask: “Determine which assumptions could make this financially coherent model produce the wrong decision.”
Artificial intelligence should not eliminate cognitive effort from education. It should allow us to raise the level at which that effort is exercised.
Learning to distrust correctly
There is an opposite danger. We do not need generations that distrust every result produced by AI. That would destroy the productivity gain we are trying to obtain.
Judgment does not mean questioning everything. It means knowing what deserves to be questioned.
A strong Judgment Reserve should make it possible to distinguish processes that are reliable enough to automate, processes that require sampling, processes that demand systematic review and decisions that necessarily require human accountability.
The mature organization will not be the one that keeps a person watching every AI operation permanently. It will be the one that knows exactly where that person is needed.
Can the Judgment Reserve be measured?
Perhaps not with a single number. Trying too early would probably be a mistake. But we can begin to build a multidimensional map of judgment risk.
- Redundancy: how many people can independently validate each critical decision?
- Concentration: which essential knowledge depends on one person?
- Replacement time: how long does a professional need to achieve genuine autonomy in each function?
- Audit capacity: can someone verify AI-generated work without depending on the same system that produced it?
- Transfer: is there an effective chain of knowledge transmission between senior, middle and junior staff?
- Independence: do reviewers reach their own conclusions, or merely validate the machine's answer?
- Exercise: do specialists still solve enough problems themselves to keep their judgment current?
- Error memory: does the organization preserve knowledge of its wrong decisions and why they were wrong?
This could eventually become a Judgment Reserve Index, although its first version should be a map rather than a simplified score. The goal would not be to rank companies. It would be to detect fragility before the lost judgment is needed.
A new function of corporate governance
The Judgment Reserve may become part of what we now consider risk management, business continuity and corporate governance.
Boards currently ask whether liquidity is sufficient, what happens if a supplier fails, which people are critical and what the succession plan is.
It may soon be equally reasonable to ask: who inside this organization is capable of determining that our artificial intelligence is wrong?
The second question will be even more uncomfortable: who will be able to do so in ten years?
Value is shifting
During the first years of the AI revolution, we mostly talked about increasing capabilities: more parameters, more data, more compute and more automation.
But if the cost of producing analysis, documents, software and applied knowledge continues to fall, scarcity may move elsewhere.
When producing an answer costs almost nothing, determining whether that answer deserves trust becomes valuable.
When everyone has access to extraordinarily intelligent machines, access to intelligence will no longer constitute a competitive advantage by itself. The difference may lie in who knows how to use it, challenge it and stop it when necessary.
That is the true nature of the Judgment Reserve: not a nostalgic defense of human work, but the human infrastructure required to automate far more without losing the ability to know when we should not.
In the age of artificial intelligence, preserving judgment will not be a way of protecting the past. It may become one of the most important investments in building the future.
Sources and conceptual note
- Stanford HAI: 2026 AI Index Report, technical performance and agents.
- METR: limitations of time-horizon measures and reliability requirements.
The expanded definition of the Judgment Reserve, its dynamic equation and the possible Judgment Reserve Index are the author's conceptual proposals. They are presented as a management framework that should be validated before becoming a comparative metric.
© 2026 Javier F. Pérez Bernabé. All rights reserved. Published by Quórum 12.
ONLINE SUBSCRIPTION
Receive every new analysis and invitation to open events.
This is a separate public category: it does not confer membership, admission or access to the private portal.