Introduction

As companies make agentic AI a core part of their business, recurring tensions emerge around outsourcing human cognition. Should they optimize for speed, or safeguard their workforce against cognitive debt? Should they trust their AI system's outputs, or continue to question them? Should they allow the system to operate unmonitored, or remain vigilant for its rare failures?

These tensions all appear to force a hard decision: a dilemma in which business leaders must pick a side and pay the price. For example, should a company let its people race ahead independently with agentic AI, or preserve a shared understanding across the team? This thesis examines these tensions and tests whether they are dilemmas at all. A dilemma can be settled by choosing one pole. A paradox cannot. It must be managed by holding both poles.

Five such tensions recur. To sort each one, a seven-step Dichotomy Probe, built on the Dynamic Equilibrium Model (Smith and Lewis, 2011), tests both poles against documented evidence and leaves a trail that another researcher can re-check.

One underlying structure explains the pattern. Laid out as Equilibrium Mappings, the five paradoxes converge on a single mechanism at the human level: cognitive passivity. Each paradox spends a scarce human capacity in two directions at once. Lean on either pole, and you wear down the capacity the other pole depends on. Calibration, scrutiny, vigilance, practice, and shared understanding come under pressure together, and their costs compound.

This is the agentic paradox. Handing cognition to a system does not remove human judgment from the loop; it relocates the tension. The business leader shifts from doing the work to holding the balance, and the practitioner's question changes from picking a pole to holding both.

Clayton Christensen's The Innovator's Dilemma (1997) distilled what is now considered an indisputable business truth: past success becomes the very thing that blinds market leaders to disruptive threats. What elevated the book beyond this diagnosis, however, was its framing. By presenting the problem as a dilemma, he articulated a more general truth: that the hardest problems in business are rarely failures of execution but irreducible tensions that pull the business in opposite directions.

The study of business dilemmas has long been a recurring theme in management scholarship, producing a solid body of literature dedicated to understanding organizational decision-making challenges: from the classic short-term profit versus long-term value to exploration versus exploitation (March, 1991). Yet these tensions are not all of one kind. In Toward a Theory of Paradox: A Dynamic Equilibrium Model of Organizing, Smith and Lewis (2011) draw the line this thesis builds on: some tensions are dilemmas that force a choice between competing goals, while others are paradoxes whose poles must be held together rather than resolved.

Dilemmas are not a purely theoretical construct. They teach us a lot, most of the time the hard way. Outside academia, they don't arrive as theory; they arrive as a sudden shock, as blockers in the path to our business goals. They harden quickly into choices we are unprepared to weigh, overwhelming us with a decision tree whose ramifications we barely grasp. By the time we do, a new clue emerges, more ambiguous than the last, and we're trapped in a dilemma we can't escape.

The uncomfortable truth is that dilemmas also put everything around us under a harsh light. Dilemmas have the power to expose tensions between conflicting goals and stakeholders. They reveal cracks in our business processes, organizational design, and management structures. Dilemmas put our ability to drive change to the test. And most profoundly, they question our roles, values, and sometimes our own identity.

And they are stubborn. In organizational decision-making, they are as persistent as they are difficult to reconcile. They are often viewed as trade-offs: challenging, mutually exclusive choices that come at each other's expense. Resolving a dilemma means confronting an irreducible either-or situation. And it comes at a high cost: not only for what saying yes implies, but also for what saying no to the opposing pole leaves behind.

Not all of these tensions are the same, however. The Dynamic Equilibrium Model (Smith & Lewis, 2011) separates dilemmas from paradoxes, two words often conflated in everyday usage but worlds apart in how they are resolved in practice. A dilemma always forces a choice. Precisely because their poles are competing choices, detrimental to each other, the only possible course of action is to choose one side and reject the other. A paradox, by contrast, persists no matter which side you choose: its tensions keep resurfacing after a decision is made. Because both poles of a paradox are so deeply interrelated, embracing both is not only possible but, in many cases, the preferable course of action. The literature review returns to this distinction in detail.

This distinction is far from academic. Whether a tension is a genuine dilemma or a paradox in disguise shapes how much corporate energy it consumes, and whether that energy is well spent. Organizations burn enormous cycles debating, quantifying, reframing, shelving, and revisiting the same tensions, often because they have misdiagnosed the problem in front of them. Take return-to-office vs. hybrid flexibility: is it a forced choice between competing goods, or a paradox whose poles only reveal their value when held together?

Questions like this are not hypothetical, and the cost of getting them wrong is high. When tensions of either kind are not identified, understood, and managed correctly, they can trap entire teams in analysis paralysis, derail strategic plans, or, even worse, polarize them into competing factions.

For most of the last century, these tensions surfaced slowly. A firm could feel a trade-off building over the years, argue about it, and still have time to choose. Artificial intelligence has taken that time away. No force in recent memory has pressed these tensions on business leaders faster, because AI advances along two tracks at once: a track of astonishing capability and a track of equally astonishing fragility. Every breakthrough arrives with its own yet.

Knowledge workers report real productivity gains from AI, yet those who lean on it recall less of their own work and engage with it less deeply—an effect Kosmyna and colleagues (2025) named cognitive debt. Companies now ship code their own engineers never wrote, with machines authoring 20 to 30% of the code at Microsoft and 11% of the live backend at Uber (Nadella, 2025; Naga, 2026), yet that code carries 75% more logic errors than human-written code (CodeRabbit, 2025). Worse, the reviewers meant to catch those errors slowly lose the skill to notice them. Frontier models outscore PhDs on the hardest science exams, yet still fail to read an analog clock more than half the time (Safar, 2025). Karpathy (2024) named the pattern jagged intelligence: superhuman skill sitting right beside a child’s mistake

The collision is genuinely unprecedented; the tensions underneath it are not. For nearly fifty years, a lineage of human-automation research has studied how people and machines share control, attention, and judgment, running from Sheridan's levels of automation (1978) to Endsley's work on the out-of-the-loop problem (Endsley & Kiris, 1995). That is half a century of hard-won thinking we can leverage. What is new is the speed and the scope.

The latest turn raises the stakes again. AI has moved from answering questions to taking actions. Agentic systems now plan, execute, and ship work that people used to do themselves. And like every dilemma before it, this shift does not arrive as theory. It arrives as a blocker in the path to a business goal, a forced choice a team is suddenly unprepared to weigh. Consider the questions a leader now has to weigh before handing cognitive work over to an agentic system:

• Should a company's agentic systems take on tasks people currently do, or rotate those people back through them to keep their competence sharp when those systems fail?
• Should the company tune the agentic system for maximum speed and scale, or slow it down with verification steps to keep employees checking its output?
• Should the company push the agentic system to deliver fast, finished deliverables, or make teams do enough of the work themselves to keep them understanding what they ship?

Nobody has definitive answers to these questions yet, and this paper doesn't pretend otherwise. The data is still emerging. What this thesis offers is a lens for classifying the kind of tension each question raises, so business leaders can choose a course of action based on fifty years of human-automation research rather than on intuition, vendor promises, and hype.

The paper is divided into six chapters, each with a single job.

Chapter II, Theoretical Framework, lays out the prior work this paper stands on. It develops the human-automation canon, the Dynamic Equilibrium Model of paradox theory (Smith and Lewis, 2011), and the agentic-era patterns that make the older mechanisms newly visible at scale. It also introduces the How Might We question as a design-thinking technique, so that the subsequent analysis can apply a method already established in this chapter.

Chapter III, Methodology and Results, justifies the research design and the data collection, then reports the answer to RQ1. It describes the five recurring tensions of outsourcing human cognition as findings, presented but not yet judged.

Chapter IV, Analysis and Transfer, answers RQ2. For each cluster, it gives the dilemma-or-paradox verdict, the Equilibrium Mapping that renders the tension, a How Might We question that opens the design space, and the transfer to business practice.

Chapter V, Discussion, draws out the derivatives and the recommendations for decision-makers, embeds the findings back into the theory of Chapter II, and gives a critical appraisal of the limits.

Chapter VI, Conclusion, summarizes the work, answers both research questions, states the agentic-paradox payoff, and names the outlook for future work.

Reference material sits outside these six chapters. The front matter carries the abstract and the lists, and the bibliography and appendix at the end hold the full sources and the overflow tables.