A person can now obtain more analysis than they can reasonably inspect. A model can explain several paths, draft a convincing argument for each, and produce a neat recommendation. The abundance creates a new practical burden: deciding which question matters, which evidence deserves weight, and when to act.
That burden is judgment. It includes recognizing a bad premise, distinguishing a measured result from a plausible scenario, and accepting responsibility for a choice that remains uncertain. AI can assist these activities. It cannot make responsibility disappear by producing a confident answer.
Why does cheaper analysis make selection and judgment more valuable? Because the limiting resource can shift from producing alternatives to choosing among them. The shift is conditional: many people still need better access, training, and evidence. But where analysis becomes plentiful, good selection increasingly determines whether that analysis produces a useful result.
Judgment begins before the answer
A decision can fail because the question was wrong. A business may ask how to create more promotional material when its actual problem is an unclear offer. It may ask which model to buy when its records are inconsistent. It may ask how to increase orders when fulfillment is already beyond capacity.
A fluent answer can make the wrong question feel productive. The owner receives a plan and begins executing it. The underlying constraint remains untouched, while the plan creates additional work.
Judgment begins by inspecting the framing. What outcome is sought? What evidence suggests that this is the obstacle? Which other explanation could account for the problem? What would happen if the proposed intervention succeeded at its immediate task?
AI can help prepare those questions and offer counterexamples. The person needs to decide which fit the actual environment. A generic challenge is less useful than a specific observation about the business’s records, customers, or capacity.
This is why judgment is not merely choosing the best answer from a list. It includes deciding whether the list addresses the right problem at all.
Fluent reasoning is not independent evidence
A model can give a coherent explanation for a claim whose factual basis is weak. The explanation may be useful as a hypothesis, but its coherence does not establish that the claim is true.
Consider a recommendation to reduce prices. The assistant may argue that lower prices attract customers. The relevant decision also depends on contribution, demand response, capacity, customer expectations, and the reason sales are weak. A general mechanism cannot supply the missing local facts.
Ask what the recommendation relies on. Which facts are measured? Which are inferred? Which assumptions would reverse the conclusion? A useful answer should make those dependencies visible rather than use confidence to conceal them.
A second model can help challenge the reasoning, but agreement is not necessarily independent confirmation. The models may rely on similar material or reproduce the same familiar argument. Authoritative records, direct observations, and appropriately designed tests provide a different kind of evidence.
Judgment therefore includes source selection. It decides what would count as a reason to believe the claim, not just whether the wording sounds reasonable.
Evidence changes over time
A dated result can be accurate and still be a poor estimate of current conditions. Models, interfaces, tasks, and user practices change. A careful reader needs both the original result and relevant updates.
METR’s February 2026 developer-productivity update explains that its later study faced selection and timing problems, making the data unreliable as a current uplift estimate. It explicitly cautions against treating the early-2025 slowdown as a settled picture of early-2026 tools. The update’s participant impressions also do not establish a precise current effect.
That example shows judgment operating on evidence quality. The responsible conclusion is not to select whichever number supports enthusiasm or skepticism. It is to preserve the measured historical result, recognize the limitations of the newer evidence, and avoid claiming a current effect that the method cannot establish.
For a business, the same discipline applies locally. A workflow tested before a model update may need rechecking. A supplier term verified last year may be stale. A customer’s earlier preference may have changed.
A source date is therefore part of the claim’s meaning. Judgment decides when an earlier finding remains applicable and when the changed conditions require a new test.
Expertise helps identify the expensive omission
An expert may notice a missing qualification that a novice does not know to ask about. A technician can recognize when a procedure assumes the wrong equipment. An experienced operator can see that a schedule ignores setup time. A qualified reviewer can identify a consequential exception in an agreement.
AI can make introductory explanations easier to obtain. That access is useful, but it does not give every user the same ability to detect omissions. Training and access to appropriate expertise remain important complements.
A practical design should not require the least experienced person to independently validate every consequential claim. It can separate routine checks from questions that require a qualified reviewer. It can provide sources, clear uncertainty, and examples that help the user recognize the boundary.
The organization can also build expertise through repeated, checked work. A person who sees why a recommendation failed can learn a useful distinction. A person who only accepts the final answer may become more dependent without becoming more capable.
Judgment is therefore a capacity that can be developed. The goal is not to reserve useful tools for experts, but to give people a trustworthy way to learn and obtain help where their knowledge is insufficient.
A hypothetical purchasing decision
Imagine an owner choosing equipment for a small service operation. An assistant compares three options and ranks the least expensive first. The comparison includes advertised capacity and purchase price. This is an illustrative case, not a report of a Salars purchase.
The owner notices that the preferred option requires a facility condition the business does not meet. Another option has a higher purchase price but fits the available space and the actual workload. A third offers capacity the business is unlikely to use.
The relevant judgment is not a mystical intuition. It is recognition of a constraint missing from the initial comparison. The owner then asks for verified installation requirements, maintenance terms, and total usable cost.
A qualified technician may need to inspect the facility before the commitment. The assistant can organize the questions and evidence. The owner can compare the feasible options after the unsupported one is removed.
The final decision may still contain uncertainty about future demand. That uncertainty should remain visible. The purchase should not be justified by a model-generated growth forecast that the business has not independently supported.
This example shows how judgment makes analysis productive: it identifies the important omission, obtains the right evidence, and directs the comparison toward the actual decision.
Selecting the metric is a judgment call
A system can optimize response time while increasing inaccurate promises. It can optimize clicks while attracting people who do not need the offer. It can optimize generated output while increasing review work.
The metric is an expression of the objective. Choosing it requires understanding the desired result and the ways a proxy can become detached from that result. A convenient measure should not silently replace the purpose.
For a support workflow, the objective might be a resolved problem with an appropriate handoff for unsupported cases. Response speed can be one measure, but it should sit beside resolution and error indicators. For an estimate workflow, the business may value feasible, accurate commitments more than a fast first draft.
A person should ask what behavior the metric rewards. Could the system improve the score by doing something the organization would reject? If so, the evaluation needs a guardrail or a better outcome measure.
This judgment remains necessary even when the calculation is automated. The machine can compute the chosen score accurately while the organization pursues the wrong result.
Good selection includes deciding what to ignore
When analysis is inexpensive, the supply of possible tasks grows. A person can investigate every minor question, compare every tool, and revise every document. The opportunity cost becomes attention.
A useful filter asks whether the question affects a decision or serves a chosen purpose. Some research is valuable for curiosity or learning. It should be allowed without pretending that every inquiry is required for the business’s next action.
For operating decisions, identify which unanswered question could change the step. If none can, more analysis may have little immediate value. The person can act within the supported scope or deliberately set the issue aside.
This filter also applies to generated recommendations. A long list may contain several reasonable ideas that the owner lacks capacity to implement. Choosing one useful improvement and rejecting the rest for now can produce a better result than opening ten projects.
When Decision-Making Becomes Search explains how to structure the comparison. Judgment gives the search a stopping point and protects the person from an endless queue of plausible work.
Authority should follow consequence
A person can use AI to prepare a low-consequence draft with lightweight review. A public promise, financial commitment, or sensitive disclosure requires stronger evidence and a clear authority boundary.
Judgment includes recognizing this difference. The same model output may be acceptable as an internal hypothesis and unacceptable as a message sent to a customer. The consequence changes the standard.
The boundary should be enforced where possible by the tools and downstream system. A process that lacks sending permission cannot accidentally turn an internal draft into an external promise through the same interface. A spending cap limits exposure even when the recommendation is wrong.
Human approval is useful when the reviewer has the evidence and capacity to make the decision. An overloaded reviewer clicking through vague requests provides weaker control. The workflow should prepare a reviewable result and identify the specific uncertainty or commitment.
The Permission Architecture develops these mechanisms. Here they show judgment becoming part of system design rather than a generic instruction to “keep a human involved.”
Accountability is a relationship, not an output feature
A generated answer cannot itself accept the obligations created by a business decision. Customers and partners need to know who stands behind the promise and how a problem will be corrected.
The owner may delegate preparation and routine execution within defined rules. The owner still needs a process for understanding and handling the results. A system that produces an explanation after an incident is not the same as an organization that can repair the incident.
This is especially important when several tools or agents contribute. The business should be able to identify the source evidence, the accepted decision, the action taken, and the person responsible for the supported workflow. Otherwise responsibility can become diffused among components.
The value of judgment includes deciding which obligations the business should accept in the first place. A profitable-looking opportunity may require a level of availability or expertise the organization cannot provide.
From Worker to Governor explores these decision rights. Judgment becomes more valuable when the person directs a larger system and must remain able to answer for its actions.
A disagreement can improve the process
People may reasonably assign different weights to time, income, risk, and personal obligations. A model’s recommendation can conceal those differences by presenting one ranking as the answer.
Make the tradeoff explicit. If an option wins only when revenue growth receives a high weight, the owner should see that dependency. If another option preserves a shorter schedule, that benefit should remain visible rather than be treated as a weakness.
A useful disagreement concerns the objective, the evidence, or the acceptable uncertainty. It does not need to be resolved by asking the model to adjudicate the person’s values. The people responsible can choose after understanding the consequences.
AI can help by restating the alternatives and identifying where the disagreement actually lies. It should not invent consensus or hide the unresolved choice behind a numerical score.
This is one of the human consequences of cheaper analysis. More options can make priorities clearer, but only if the decision process permits the person to say what kind of outcome they want.
Judgment has failure modes too
Human judgment is not infallible. A person can favor a familiar idea, defend a sunk cost, ignore inconvenient evidence, or mistake confidence for expertise. Keeping a person in the process does not automatically make the decision sound.
Useful safeguards make the reasoning inspectable. Record the criteria before the result, seek a relevant counterexample, compare a simpler alternative, and preserve the evidence that would change the decision. Another qualified reviewer can challenge a consequential omission.
A decision record should distinguish what was known at the time from what became clear later. An unfavorable result does not necessarily prove the original decision was unreasonable. A favorable result does not prove the method was sound.
The organization can learn by examining the mechanism and the available evidence. Did the decision ignore a constraint? Did the environment change? Was the result within the uncertainty that had been acknowledged? These are more useful questions than praising success and blaming failure after the fact.
A Learning Ledger can retain this analysis. It turns judgment into a revisable practice rather than a claim of personal certainty.
The scarcity is contextual
Judgment becomes especially valuable where analysis is abundant and consequences are meaningful. Elsewhere, access to basic information or a suitable tool may remain the primary obstacle. The series title should not erase those differences.
A worker without training may need a clear explanation before they can evaluate alternatives. A business without reliable records may need data cleanup. A person facing a specialized issue may need qualified help. Telling them simply to exercise better judgment can be unhelpful.
The productive response combines access, training, evidence, and responsibility. It gives the person a clearer view of the task and a route to help. Good judgment then has something trustworthy to work with.
This also limits broad claims about economic scarcity. A useful operating principle for one organization does not establish a universal prediction about wages or society. The practical claim is that selection can become a larger determinant of value when producing analysis becomes easier.
That claim is enough to guide a better workflow without turning it into an unsupported forecast.
AI Leverage in Practice
Choose one consequential decision and prepare a short review brief. State the objective, the feasible alternatives, the authoritative evidence, the controlling assumptions, and the next commitment. Ask AI to identify an omission or counterexample rather than merely endorse the preferred choice.
Verify the challenge against actual conditions. Obtain qualified review where the task requires it. Identify what evidence would reverse the recommendation and whether that evidence can be obtained before acting.
Use a proportional authority boundary. Keep preparation separate from external commitments, and make the proposed result concrete enough for the responsible person to inspect. Record the accepted choice and the uncertainty that remains.
Today’s tools can support comparison, explanation, and challenge. Future systems may improve their ability to select actions, but the objective and acceptable consequences still belong to the people and institutions responsible. A more capable recommendation does not eliminate the need to know what it serves.
After the outcome, compare it with the original reasoning. Preserve a bounded lesson, including conditions where it does not apply. The goal is to improve judgment through evidence rather than accumulate confidence.
Selection that makes abundance useful
Cheaper analysis increases the value of judgment when the person has more plausible paths than time to inspect or implement them. Judgment chooses the question, weighs the evidence, recognizes constraints, and decides when the remaining uncertainty is acceptable.
Its purpose is not to defend human superiority. It is to make the complete decision more responsible and useful, including the parts where people themselves need better methods and qualified help.
The next part of The Age of AI Leverage builds a Personal AI Chief of Staff. That system needs the same foundation: clear objectives, reliable evidence, bounded authority, and a person who can direct the work toward a life worth serving.
Sources
- METR, We Are Changing Our Developer Productivity Experiment Design, February 24, 2026; current methodological limitations prevent treating either historical results or newer participant impressions as a universal present-day effect.
Loading comments…