Intelligence Is Becoming Abundant. Judgment Becomes the Scarce Resource.

The cost of producing another recommendation is falling. The consequences of acting on the wrong one are not necessarily falling with it.

For an organization, that difference changes the meaning of analytical capacity. When preparing a serious analysis requires substantial professional effort, the ability to produce it is a constraint. When systems can generate alternatives, explanations, and proposed actions rapidly, the constraint can move toward deciding which question matters, which evidence is sufficient, and which consequences are acceptable.

Normatic's view is that this shift makes judgment a more consequential source of organizational advantage. Greater intelligence expands the range of possible actions. Judgment determines which of those possibilities the organization should pursue, under what conditions, and with whose authority.

Our first two Perspectives examined the workflow as the unit of transformation and the organizational changes that follow. This third argument concerns the design of decision-making within that organization: how evidence, analysis, authority, and responsibility combine to produce sound action.

Abundance changes the constraint

By intelligence, we mean capabilities such as processing information, recognizing patterns, generating options, reasoning, and producing recommendations. By judgment, we mean deciding what matters in a particular situation, weighing competing objectives, assessing consequences, and determining whether there is an adequate basis to act. In an organization, exercising judgment also requires clarity about who answers for the decision.

The capabilities overlap. AI can contribute to evaluating trade-offs, and human judgment can be inconsistent or poorly informed. The distinction is useful because producing a plausible answer and authorizing a consequential action are different functions, even when one person has traditionally performed both.

There is evidence that some forms of cognitive production can require substantially less human effort. In a preregistered experiment involving 453 professionals, Shakked Noy and Whitney Zhang found that ChatGPT reduced the average time spent on specified professional writing tasks by 40%, while assessed output quality increased by 18%. These were bounded writing tasks, not a test of executive judgment or a measure of total enterprise costs. Experimental evidence on the productivity effects of generative artificial intelligence

The strategic significance is the possibility of producing more useful cognitive work with the same human capacity. Organizations can explore questions that previously received little analytical attention and prepare alternatives that would have been too costly to develop routinely.

Abundance remains relative. Frontier capability, proprietary information, infrastructure, model quality, and access still differ materially. Verification, integration, and operation continue to carry costs. The claim is that the marginal availability of sophisticated cognitive capability is increasing, making it important to identify which constraints remain after an answer becomes easier to produce.

One of those constraints is the organization's capacity to choose responsibly among the answers.

More options require a stronger basis for choosing

Analysis is useful partly because it reduces uncertainty. Yet the ability to generate more analysis does not guarantee that the uncertainty relevant to a decision has been reduced.

Ten scenarios may illuminate a genuine range of possibilities, or they may express variations on the same unexamined assumption. A detailed forecast may improve planning, or it may give an unstable premise the appearance of precision. Several recommendations can agree because they share a source, rather than because independent evidence supports the conclusion.

As producing material becomes easier, deciding what deserves attention becomes part of judgment itself. Leaders need to distinguish questions whose answers could change a decision from questions that merely generate further discussion. More analytical activity can coexist with a weaker understanding of the choice at hand.

The issue becomes sharper when the objectives are contested. A recommendation to accelerate a product launch may reflect an assumption that speed matters more than reliability, profitability, or the effect on an existing customer relationship. Better estimates can clarify the consequences of those priorities. They cannot establish that the priorities have been legitimately chosen on the organization's behalf.

Our interpretation is that organizations must become more explicit about the principles that govern decisions. Strategic priorities and risk appetite cannot remain entirely implicit in the experience of a few senior people if AI-supported work is to operate consistently at scale. Some judgments need to be expressed as boundaries within which others can act; others need to remain open to examination because the situation is unusual or the objectives conflict.

This is also why confidence in an answer and permission to act should be treated separately. An estimate can be well supported while the proposed action remains unacceptable because of its consequences. Conversely, an uncertain estimate may still support a small, reversible experiment. The decision depends on the relationship between evidence, exposure, and the ability to learn or recover.

Human involvement must earn its place in the decision process

It is tempting to resolve these concerns by requiring human approval. That can preserve formal authority, but it does not establish that the combined process produces better decisions.

More recent evidence illustrates why the arrangement matters. In a study published in March 2026, Mustafa Akben, Vinayaka Gude, and Haya Ajjan compared GPT-5 mini responses generated in November 2025 with existing human responses on two standardized emotion and mental-state recognition tests. The model performed strongly against individuals. Combining independently collected human and AI answers through voting, however, produced higher accuracy than aggregating either source alone. Collective and augmented intelligence outperform artificial intelligence on emotion recognition tests

This was a statistical aggregation study, not a trial of people interacting with AI in a live workflow. Its results do not establish a general ranking of humans, AI, and their combinations, or measure organizational judgment. They demonstrate a specific way that complementary contributions can improve an answer.

The management implication we draw is that collaboration must be designed and evaluated for the task and models being used. Human involvement alone does not establish superior accuracy, and accuracy alone does not settle responsibility. The organization must establish what contribution the person is expected to make and whether the arrangement enables that contribution.

A reviewer may add essential context, challenge an objective, or authorize a commitment that the system has no right to make. A reviewer may also introduce an error or accept a recommendation without examining it. These possibilities require a more precise account of the human role than a box labelled approval.

Research on overreliance illustrates the design problem. In an experiment with 199 participants, Zana Buçinca, Maja Barbara Malaya, and Krzysztof Gajos found that interventions intended to induce more deliberate engagement reduced acceptance of incorrect AI advice compared with simpler explanation-based designs. Participants also gave less favorable subjective ratings to the designs that reduced overreliance most. This was a specific experimental setting, not evidence that every decision should include more friction. To Trust or to Think

For Normatic, this suggests that ease of use and quality of judgment can sometimes pull in different directions. The right design helps people think where thinking changes the outcome, while allowing routine work to proceed within established limits. Optimizing only for rapid acceptance risks confusing a smooth interaction with a sound decision.

Judgment can be strengthened by changing how it is exercised

Judgment is often described as a quality possessed by experienced individuals. Organizations can also improve the conditions under which it is exercised.

AI can help assemble the relevant evidence, identify contradictions, surface precedent, and expose assumptions behind a recommendation. It can make alternative interpretations cheaper to explore and show which missing information would matter most. These contributions are valuable when they make the basis of a decision more open to examination, rather than simply making one proposed answer more persuasive.

This requires care. Asking the same model to critique its recommendation does not provide independent verification. A generated explanation may be coherent without being a reliable account of why an answer is correct. Support for judgment should preserve access to source evidence and leave room for uncertainty, disagreement, and the possibility that no available option is yet acceptable.

Consider an illustrative supplier disruption. A system identifies orders at risk, estimates delays, and proposes an alternative source. The recommendation favors the option with the lowest expected interruption cost. A decision-maker knows that changing suppliers may also affect a strategic relationship and that the alternative source's ability to meet a particular requirement has not been confirmed.

A useful decision process would make those issues visible. AI could compare alternatives under different assumptions, assemble relevant records, and identify the evidence needed to resolve the uncertainty. An authorized person could weigh the competing objectives and decide whether to proceed, gather more information, or accept a delay. This is an illustrative scenario, not a Normatic client case.

The human contribution is not limited to correcting an erroneous number. It may involve deciding which consequences belong in the analysis, whether the objective is appropriate, and whether the available evidence justifies the commitment. AI can strengthen that contribution without being granted authority to make the commitment itself.

Judgment therefore becomes something the organization can develop through its decision process: better evidence, clearer alternatives, meaningful challenge, and a deliberate connection between conclusions and action.

Separate the production of advice from the authority to act

A decision process contains several functions that need not be assigned to the same participant. Information can be gathered by one combination of people and systems, analysis produced by another, and recommendations tested against independently maintained evidence or rules. Decision authority can sit with a named person or be exercised automatically within a previously authorized boundary. Execution, monitoring, and exception handling require their own responsibilities.

In the supplier example, preparing an alternative order does not have to confer permission to place it. A system might collect evidence and draft the transaction, while an authorized manager approves the commercial commitment. An integration could then execute the approved action, with monitoring that detects whether delivery conditions are being met. Other routine purchases might proceed automatically under established limits, with exceptions sent to a responsible person.

The useful design question is how intelligence, judgment, authority, and accountability should be distributed across that process. Comparing human and AI performance on a task helps inform this decision, but does not settle the entire allocation.

Execution speed makes the allocation more consequential. If an automated process can apply a mistaken rule repeatedly before anyone intervenes, the cost of an error depends on how much it can do before the organization detects and contains it. Authority limits, monitoring, and recovery therefore belong in the decision design, alongside the quality of the original recommendation.

Accountability must also be usable in practice. The person responsible for a process needs sufficient visibility, authority, and support to intervene. Naming an owner who cannot inspect failures, stop actions, or secure changes creates nominal accountability. A review trail should explain the evidence and permissions behind consequential actions so that responsibility can lead to correction and learning.

Human accountability can remain explicit even when no person approves each individual action. The relevant judgment may have been exercised when the organization set the operating boundary, accepted the evidence of reliability, and assigned responsibility for monitoring it. Those decisions must be revisited when circumstances or system performance change.

Allocate attention according to consequence and uncertainty

Blanket review treats all cases as if they make the same demand on human attention. They do not. A routine, reversible action supported by reliable evidence differs from an unusual commitment whose consequences are difficult to reverse.

A more deliberate organization would allocate attention according to those differences. Review requirements would reflect the action's consequences, the quality and completeness of evidence, and demonstrated reliability under comparable conditions. A model's statement that it is confident would not, by itself, establish that reliability.

This creates responsibilities beyond the review queue. Escalation criteria need to identify who receives an exception and what that person can decide. Recurring exceptions need to inform improvements to the workflow, rather than becoming a permanent pool of unexplained manual work. The knowledge used by people and systems needs owners who can correct obsolete guidance and resolve contradictions.

There are also limits to how far attention can be concentrated. If automation leaves a team handling almost exclusively difficult cases, the remaining workload may require more expertise and recovery time per case. A reduced case count alone does not demonstrate that fewer reviewers or wider managerial spans are appropriate.

Judgment can be improved through feedback as well. In experiments by Jonathan Baron and John Hershey, participants evaluated decision-making more favorably when they were told that the outcome was favorable, despite having the information available to the original decision-maker. Outcome bias in decision evaluation Comparing a decision's rationale with what subsequently happened can reveal weak assumptions and inappropriate boundaries. That comparison should distinguish the quality of the original decision from the outcome alone: a reasonable choice under uncertainty can still end badly, while a poorly supported choice can succeed by chance. The aim is to improve the decision process, including when to trust, challenge, or defer its recommendations.

These are design choices for particular operating contexts. They do not imply that every organization needs the same approval structure or degree of automation.

The advantage lies in converting capability into sound action

Abundant intelligence does not make models, data, expertise, or infrastructure unimportant. It changes the possibilities available to organizations and can expose weaknesses in how they choose among them.

The scarce resource is the institutional capacity to frame the right question, establish an adequate basis for action, allocate authority, and learn from consequences. That capacity is not fixed. AI can help expand it, provided the organization designs the relationship between analysis and judgment rather than assuming that one automatically supplies the other.

Normatic's strategic proposition is that advantage will increasingly depend on the quality of this relationship. An organization that generates more recommendations but cannot decide what to trust or who may act has expanded its output without resolving its constraint. An organization that combines stronger evidence with explicit authority and effective feedback can turn greater intelligence into a more dependable capacity to act.

The challenge is to design how intelligence becomes judgment, authority, and action—and to remain accountable for what follows.

Sources

• Shakked Noy and Whitney Zhang — Experimental evidence on the productivity effects of generative artificial intelligence. Science, 2023. A preregistered experiment on professional writing tasks; it measures task productivity, not executive decision quality.
• Mustafa Akben, Vinayaka Gude, and Haya Ajjan — Collective and augmented intelligence outperform artificial intelligence on emotion recognition tests. Scientific Reports, 2026. GPT-5 mini tested in November 2025 against existing human data; compares independent response aggregation on two recognition tests, not live workplace collaboration.
• Zana Buçinca, Maja Barbara Malaya, and Krzysztof Z. Gajos — To Trust or to Think: Cognitive Forcing Functions Can Reduce Overreliance on AI in AI-assisted Decision-making. Proceedings of the ACM on Human-Computer Interaction, 2021. Experimental evidence on deliberate engagement, overreliance, and usability trade-offs.
• Jonathan Baron and John C. Hershey — Outcome bias in decision evaluation. Journal of Personality and Social Psychology, 1988. Experiments examining how knowledge of outcomes influences evaluations of decisions made under uncertainty.