Organizational Judgment: The Danger of Doing the Wrong Thing Exceptionally Well

Organizational judgment in AI is not simply the ability to retrieve the right policy before taking action. It is the ability to recognize when a sensible action, performed efficiently and repeatedly, is no longer serving the organization. AI makes that distinction more consequential. A company can become faster, cheaper and more consistent at doing precisely what it should have stopped doing.

Consider an established business process. It has survived years of management reviews, generated familiar metrics and accumulated a library of exceptions that experienced colleagues know how to handle. Its performance may be perfectly respectable. Now the company introduces AI to prepare the inputs, process the requests, resolve the exceptions and produce the reports. Throughput improves. People spend less time on repetitive work. The project looks successful.

What has actually been validated? The efficiency of the existing process, certainly. But not necessarily its continued relevance. A process can be improved without being questioned. In fact, making it cheaper to operate may remove one of the few pressures that previously forced people to reconsider it.

I think this is an underappreciated challenge in the move toward AI-native organizations. We spend considerable effort asking what context an agent needs to execute work correctly. We spend less time asking who will challenge the context, the objective and the work itself. If AI becomes exceptionally good at reproducing our organizational habits, the quality of those habits matters more than ever.

An organization can become exceptionally efficient at reproducing a decision it should have reconsidered.

Organizational judgment in AI begins with what the company knows

Most companies have a formal account of how decisions happen. Strategies explain priorities. Policies establish boundaries. Process maps show responsibilities. Systems contain customer, financial and operational data. None of these is dispensable. But anyone who has led teams across functions knows that the formal picture is incomplete.

These are not mystical human abilities. AI can help detect patterns in communications, workloads or outcomes, subject to appropriate privacy and employment safeguards. It can help a manager prepare for a difficult conversation. But identifying a signal is not the same as participating responsibly in the relationship that gives the signal meaning. The colleague may need trust, discretion and a genuine conversation, not an automatically generated intervention.

The same distinction applies to organizational knowledge. When I have led product teams, I did not want every meaningful decision to come back to me. I wanted colleagues to understand enough of the product’s purpose, customer realities and commercial constraints to make good choices closer to the work. That understanding came from experience, shared reasoning and conversations, not simply from access to a document repository.

The argument in A Task Brief Is Not an Organizational Brief is that a legitimate instruction may still be incomplete. This becomes even more important when we consider not just an isolated task but the institutional habits surrounding it. A better-informed agent can recognize a missing stakeholder or an unusual precedent. It does not follow that the agent should settle every disagreement that emerges.

Alignment is something people create together

Suppose Sales wants a commitment to an important customer, Operations wants stability, and Finance wants to protect margins. Each objective is legitimate. A technically elegant optimization might identify the option with the strongest projected combined return. Yet the decision is not finished when the spreadsheet produces an answer.

In my experience, good cross-functional decisions often require people to meet, make their constraints explicit and negotiate what they are willing to sacrifice. Sales may accept a smaller near-term commitment. Operations may tolerate a carefully bounded exception. Finance may agree that a strategic relationship justifies a cost that looks unattractive in isolation. The participants do not simply discover a result; they develop an understanding of why the result is worth supporting.

This is a distinction I would not want to lose as AI improves. A system might prepare a more complete scenario analysis than any individual participant. It could identify dependencies, expose inconsistent assumptions and propose compromises. Those are valuable contributions. But an analytically attractive compromise does not automatically create the legitimacy, trust or shared ownership required to carry it out.

There is also no reason to romanticize the human process. Workshops can become political theatre. Seniority can overwhelm evidence. People can protect departmental targets at the expense of the business. Better AI analysis may expose exactly those weaknesses. The point is not that human judgment is always better. It is that a contested organizational choice has consequences for people who must understand, accept and remain accountable for it.

That is why objective architecture for agentic AI matters. Locally reasonable targets can contradict one another. Making those contradictions visible is part of the work. Deciding which sacrifice is legitimate, who may authorize it and when the choice must be reconsidered remains an organizational responsibility.

The efficiency trap: when success protects an outdated process

Now return to the process that AI has made faster. Its owners have good reasons to celebrate. Repetitive preparation has been reduced. The backlog is shrinking. Service levels are improving. Employees may be spending less time on tasks they never particularly valued. These are real benefits when measured honestly.

But a subtle change can follow. The organization becomes increasingly invested in the process because it has become a showcase for automation. The original operating assumptions are no longer debated. New requests are treated as opportunities to extend the same automation. The process becomes cheaper to repeat, and therefore easier to defend.

This is the scenario that concerns me more than an AI agent occasionally making an obvious mistake. Mistakes are visible when they trigger an exception. A consistently efficient process can conceal a deeper failure for longer: it may be answering a question the business should no longer be asking.

We have seen analogous problems long before generative AI. Businesses institutionalize practices because they once worked. They invest in reporting systems around those practices. They train people to optimize the associated indicators. Eventually the indicator becomes the job. I have explored that mechanism in Output vs Outcome: When the Number Becomes the Job. AI can make the feedback loop faster, but it did not invent the underlying management problem.

It is tempting to say that AI should simply challenge the process on its own. In some circumstances it can do exactly that: generate alternatives, identify anomalies, compare scenarios and surface evidence that established assumptions no longer hold. We should use those capabilities. But identifying a plausible alternative is not the same as authorizing a different organizational ambition. Somebody must decide which consequences matter, whose interests are affected and what the business is willing to change.

Similarly, a field experiment involving consultants found that AI assistance improved performance on some tasks but could worsen results on tasks outside the model’s competence. The lesson is not a permanent division between work that humans do and work that AI does. It is that the boundary changes and requires active assessment. The Harvard Business School research on the jagged technological frontier gives a useful empirical reason not to assume that successful automation in one activity validates adjacent activities.

Decision support is not the same as decision authority

The question of organizational judgment in AI becomes clearer when we stop treating every decision as the same kind of event. Imagine a routine return authorization. If a transaction falls below an agreed threshold, meets explicit conditions and can be reversed, an organization may sensibly delegate it to software. The meaningful human decision was made earlier, when the boundary and acceptable outcomes were defined. The system is carrying out that decision at scale.

Now imagine a major industrial supplier requesting a price increase. The immediate commercial impulse may be to reject the request or threaten to find another supplier. AI could substantially improve the purchasing manager’s preparation: estimate the exposure, identify alternative suppliers, compare switching costs, assess continuity risks and model different contract structures. That analytical contribution might change the negotiation entirely.

Yet the supplier is not a variable in a spreadsheet. It is a business relationship with its own constraints and incentives. A manager may need to understand why the increase is being requested, decide whether a temporary concession is warranted, and negotiate terms that preserve future cooperation. There may be no objectively correct answer that can be inferred from the purchasing target alone.

The question is not whether AI can produce sophisticated negotiation language or recommend a commercially defensible outcome. It often can contribute to both. The question is who has the mandate to commit the organization, bear the consequences and maintain the relationship when the recommendation proves difficult in practice.

I would therefore distinguish between bounded execution, informed recommendation and consequential commitment. These are operating-model choices, not a new taxonomy of AI capability. A system can be trusted to act autonomously within a narrow authorization while being required to prepare options for a different decision. Authority should be proportionate to consequences, reversibility, uncertainty and the organization’s ability to observe what is happening.

This aligns with the direction of established risk-management guidance. The NIST AI Risk Management Framework emphasizes context, accountability, ongoing monitoring and defined human oversight. The OECD AI Principles similarly address human agency, oversight and accountability. Neither source resolves the commercial negotiation for the manager. They reinforce the need to design responsibility around actual use and impact rather than assuming that technical capability settles the question.

Human oversight must question the goal, not only the method

A common design assumption is that an agent should escalate when it detects an exception. That is sensible but insufficient. The agent may not recognize the exceptional condition. More importantly, the organization may have changed while the task remained technically valid. An agent that continues to meet its objective can be doing exactly what it was instructed to do while contributing to an increasingly inappropriate outcome.

Consider a hypothetical system tasked with improving productivity and margin. Within a narrow scope it might propose valuable process improvements. Given broader authority and poorly specified constraints, it could also recommend measures that harm employees, customer relationships or long-term capability. The problem is not that AI inevitably makes such recommendations, nor that a current system can autonomously implement every one of them. The problem is that a legitimate local objective does not authorize every possible means of achieving it.

Oversight cannot therefore consist only of approving outputs that an agent has flagged as unusual. Someone must periodically ask what the system has actually been optimizing, which side effects are accumulating, whether the original target remains desirable and whether the activity should continue at all. This is not just an AI safety procedure. It is management doing its job.

A review meeting that merely confirms the dashboard is green does not meet that standard. Effective reassessment has to create the possibility of a different decision: revise the objective, change the authority boundary, stop the workflow, test an alternative or reallocate the people and resources supporting it. Without that possibility, review cadence becomes another efficient ritual that protects the status quo.

The hardest part is that humans also prefer continuity. We have incentives to defend projects we sponsored and processes we designed. An executive who keeps the final approval right but never challenges the assumptions behind a successful automation has retained formal control without exercising much judgment.

The Capacity Dividend should finance the right to reconsider

This is where I see a more constructive opportunity. AI can reduce the effort required to keep existing operations running. What we do with that recovered capacity is a management choice. We can use it to produce more of the same, remove cost, absorb new demand, improve quality or create room to experiment. There is no automatic dividend that arrives in an innovation budget simply because a team adopts AI.

I have argued before that the Capacity Dividend needs an investment policy, and that its use should be treated as a portfolio of deliberate reinvestments. The organizational-judgment problem makes that argument more urgent. Some recovered capacity should be directed not toward accelerating the current operating model but toward examining whether it deserves to remain the current operating model.

Imagine an organization where a team spends most of its week producing recurring management information. AI reduces that burden materially. One response is to demand twice as many reports. Another is to reduce staffing. A third is to preserve part of the recovered capacity for the work the reporting system never left room for: questioning which decisions the reports improve, whether the metrics still reflect the business, which assumptions have become obsolete and what experiments could create more value than another reporting cycle.

The third option does not require romanticizing innovation. It requires a disciplined allocation. Give people time to investigate an alternative. Make the hypothesis explicit. Test it at limited cost. Compare outcomes. Stop when evidence does not support the idea. Then return what has been learned to the organization. AI can support every stage of that process, from research and prototyping to scenario comparison and documentation.

That is why the Capacity Dividend is not merely a personal productivity story. It can become an organizational capability for renewal. The business can keep operating while a portion of its human attention is deliberately assigned to asking whether the business should operate differently.

There is a tension here worth making explicit. AI might itself identify many of the better alternatives. It might be more consistent than an experienced manager at comparing options, less attached to a legacy process and faster at generating counterarguments. Good. The point is not to protect human pride from better analysis. It is to ensure that somebody with legitimate responsibility can evaluate the consequences, mobilize the organization and change direction when the evidence warrants it.

Shared knowledge makes better judgment possible, but cannot replace it

None of this works particularly well if the organization cannot explain its own processes. In many companies, people still learn essential operating details by asking a colleague which template to use, where a decision was recorded or whom to contact when the documented process fails. The problem is familiar to anyone who has tried to maintain a knowledge repository over several years. The pages exist; their completeness, relevance and discoverability vary.

AI makes that weakness more visible. If humans depend on informal networks to find the latest decision or exception, an agent operating from the official repository may act with confidence on incomplete information. Adding more documents is not enough. The organization needs reliable ownership of knowledge, visible updates, traceable decisions and a way to challenge material that no longer matches practice.

I am interested in the knowledge graph as part of that infrastructure: connecting objectives, processes, roles, customers, decisions, constraints and the reasoning that relates them. But I would resist the claim that putting these relationships into a graph makes organizational judgment fully machine-readable. A relationship can be documented without being legitimate. A precedent can be retrieved without being wise. A value statement can be accessible without resolving what to do when values collide.

The stronger ambition is shared organizational context for humans and AI. A colleague should be able to understand why a decision was made without knowing the one person who remembers the meeting. An agent should be able to find relevant constraints and explain which information is missing. A leader should be able to see where policy, precedent and actual behavior have drifted apart. This extends the idea of organizational cognition beyond making information available closer to the work: the organization also needs the capacity to revise what it thinks it knows.

The Human Premium is the willingness to change the answer

When people discuss the Human Premium in AI-enabled work, the conversation often settles on the human as the final reviewer. I think that is too small a role. A professional’s contribution can be recognizing that a customer is asking for a solution that will hurt other customers, explaining why the request should not be fulfilled literally and helping the customer arrive at a better outcome without damaging the relationship.

It can be the purchasing manager who understands a supplier’s difficult position but still protects the company’s interests. The team leader who notices a colleague disengaging and makes time for a conversation. The executive who tells two departments that neither local target is worth winning at the expense of the organization’s larger objective. Or the product leader who asks whether an impressively automated workflow should exist at all.

AI can improve the evidence available in each situation. It can propose alternatives and challenge complacency. It may also expose weaknesses in human reasoning that organizations have tolerated for years. That is a reason to strengthen human judgment, not to declare it infallible. Responsibility without curiosity is not much of a premium.

These commitments are connected. Shared knowledge improves the quality of analysis. Appropriate authority makes action safer and more legitimate. Deliberate reinvestment keeps the organization from treating its present form as permanent. None can compensate fully for the absence of the others.

The real test of an AI-native organization will not be how many of its existing processes it can automate. It will be whether the organization becomes better at deciding which processes, commitments and ambitions deserve to exist in the first place. We should not use AI merely to become more efficient versions of ourselves. We should use the capacity it creates to become better versions of ourselves.

Executive question: If AI made your most established operating process dramatically cheaper tomorrow, who in your organization would have the time, evidence and authority to ask whether you should still be running it?