Sense of Control

The first thing Tom Martinez noticed was that he had started asking permission for things he used to simply decide.

Martinez—a composite drawn from the kind of mid-level manager who now populates customer-service floors across the insurance industry—had run his team for eight years. He hired the agents, trained them, wrote the workflows, took the ugly escalations himself. The job was hard, but it was his. When a policyholder whose house had just burned down needed an exception to a filing deadline, Martinez made the exception. When an agent was having a bad week, he knew it from the sound of her voice, not from a dashboard.

In early 2024 the company installed an AI customer-service platform. It triaged inquiries, drafted replies, scored risk, and rated each agent in real time. At first Martinez treated it as a smart assistant—useful suggestions he was free to overrule. Eighteen months later, the relationship had quietly inverted. Overriding the system now required a written justification logged into a field that upper management reviewed. A high override rate got flagged: not "Martinez knows something the model doesn't," but "Martinez isn't leveraging AI effectively." Changing a workflow, once an afternoon's work, now needed sign-off from IT, compliance, and operations. Evaluating his own people meant deferring to the platform's numbers—resolution time, containment rate, satisfaction score—over what he could see with his own eyes.

Martinez still had the title. He still made decisions. But he had become, as he put it, a layer of insulation between the algorithm and the humans it governed—responsible for outcomes he no longer controlled. He was stressed, resentful, and slowly checking out of work that had once meant something precisely because he got to decide how it was done.

That last sentence is the subject of this chapter. Something is happening to the human experience of control at work, and it has a name.

What agency decay is

Agency decay is the gradual erosion of a person's capacity for autonomous decision-making, and of their felt sense of independence from the systems around them. It is not the loss of a job. It is the loss of authorship within a job—the slow narrowing of the space in which a worker's own judgment carries weight. The word "decay" is deliberate. It does not describe a single dramatic act of disempowerment but a slow half-life: each week the discretionary zone shrinks a little, each concession feels reasonable on its own, and only in aggregate does the worker realize how little room is left.

The scale is what makes this worth a chapter rather than a footnote. By 2025 roughly three-quarters of workers—around 75%—were using AI in some form on the job (Knowledge Worker, 2025). Usage is not the same as saturation, so a second figure matters more: by early 2025, workers in at least 36% of occupations were using AI for a quarter or more of their tasks (Stanford, Future of Work with AI Agents, 2025). That is the threshold at which AI stops being an occasional convenience and starts shaping the rhythm of the day. When AI touches a quarter of your tasks, it is no longer a tool you pick up; it is an environment you work inside. Thirty-six percent of occupations crossing that line simultaneously is not a scattering of early adopters. It is a structural shift in who—or what—sets the terms of work.

Agency decay is the psychological signature of that shift. And it is not an inevitable byproduct of the technology. It is produced by specific, identifiable mechanisms, each of which can, in principle, be designed differently.

The autonomy paradox

Start with the deepest of these mechanisms, because it explains why well-intentioned deployments still hollow out control.

AI systems are built to optimize, standardize, and reduce variance. Those are their virtues. But autonomy is, almost by definition, the freedom to introduce variance—to deviate, improvise, and exercise judgment in ways an algorithm cannot anticipate or pre-approve. The two are structurally at odds. This is the autonomy paradox: the very properties that make an AI system valuable to an organization are the properties that constrain the humans working under it.

The mechanism by which a worker learns to stop exercising judgment is worth watching closely, because it is subtle. Nobody forbids Martinez from making the exception for the policyholder whose house burned down. He is free to do it. But the system does not see a compassionate judgment call; it sees a policy deviation. So it flags the action, demands a written justification, and logs the event as a potential performance risk. The exception still gets made—this time. But the worker has learned that judgment now carries a tax: friction, documentation, exposure. Do this a few dozen times and the lesson lands in the body before it reaches the conscious mind. The path of least resistance is to defer to the system. Autonomy is not removed; it is made expensive, and workers rationally economize on it until it withers. That is decay—not prohibition, but the quiet retirement of a faculty that costs too much to use.

The efficiency trap

The second mechanism concerns pace, and it turns AI's greatest selling point against the people it was sold to help.

Researchers at the Wharton School describe an AI efficiency trap: productivity tools that generate perpetual pressure rather than relief (Knowledge at Wharton, 2025). The logic is grimly simple. AI lets a worker or team do more in less time. But the time saved does not become slack—it becomes the new baseline. A team that handled 30 cases a day now handles 50; the target promptly becomes 50, then 60, then 70. The efficiency gain is converted, almost frictionlessly, into a proportionally higher quota.

Here the demonstration becomes self-defeating. By showing what it can do on a good day, the team has redefined the ordinary day. The AI has established the ceiling of the possible, management has installed that ceiling as the floor of the expected, and the worker has lost the one lever that used to regulate a sustainable pace: the ability to say this is a reasonable amount of work. That judgment now belongs to the throughput curve. The constant availability of the tools deepens the bind—what Wharton's researchers call digital omnipresent stress, the low-grade compulsion to exploit round-the-clock capability that, far from expanding freedom, colonizes the mental space where recovery used to live. In practice a team may absorb 40% more volume while every metric records them as simply more efficient. The metrics are not lying. They are just not measuring the exhaustion.

The micromanagement paradox

The third mechanism is the one workers describe most viscerally, because it is the one they can feel watching them.

AI was supposed to reduce micromanagement—to handle routine oversight so that managers could turn to strategy and mentoring. The opposite has often occurred. Algorithmic monitoring produces more surveillance than any human manager ever could, because it never blinks, never gets bored, and never extends the benefit of the doubt. It tracks keystrokes, response times, idle intervals, deviation from the optimal script, and it renders all of it as a live dashboard ranking each worker against the others.

The micromanagement paradox is that this pervasive oversight eliminates precisely the discretionary space that skilled work depends on. A human supervisor watching over your shoulder at least understands why you spent four extra minutes on a distraught caller; the system sees only a response-time flag. The room to think, to build rapport, to apply contextual judgment, compresses toward zero, because the algorithm optimizes for what it can measure and rapport is not a metric. And the erosion runs in both directions. The manager sees the same dashboards leadership sees. A manager who shields their team's autonomy by ignoring the numbers is read as failing at the job; a manager who enforces the numbers has become, functionally, an enforcement arm of the algorithm. Empirical work on AI and autonomy finds exactly this: workers enjoy less autonomy when reduced to executing machine-issued instructions, and monitoring capabilities can worsen work quality by stripping out the discretionary slack skilled work needs (Journal for Labour Market Research, 2025). Autonomy decays at the worker's desk and at the manager's, and the manager becomes both a victim of the mechanism and its instrument.

What workers actually want

If all of this were simply the price of progress that workers grudgingly accept, the story would be sad but stable. It isn't stable, because there is a large and measurable gap between what workers want and what they are given.

The clearest evidence comes from Stanford's Future of Work with AI Agents study, which built a tool called the Human Agency Scale and applied it across 104 occupations (Stanford, 2025). The scale runs from full automation at one end—AI acts, humans are absent—to full human control at the other, with an explicit midpoint the researchers labeled H3: equal partnership, in which humans and AI share decision-making authority and neither acts unilaterally.

Human Agency Scale	What it means
Toward full automation	AI executes the task; minimal or no human involvement
H3 — Equal partnership	Humans and AI share authority; neither acts alone
Toward full human control	Human executes; AI assists at most

Equal partnership was the single most common preference, chosen by workers in 47 of the 104 occupations. This is the substantive finding hiding inside a dry number: when asked how much say they want, the plurality of workers did not ask to keep AI out, and did not ask to hand it the wheel. They asked to share the wheel. The desire is not technophobia. It is a claim to co-authorship.

The gap sharpens when you compare workers to the people designing the tools. On 47.5% of tasks, workers preferred a higher level of human agency than technologists judged necessary (Stanford, 2025). On nearly half of all tasks, in other words, the people who do the work want more control than the people who build the systems think the work requires. That gap is the entire problem in miniature. It is not a gap between what is technically possible and what workers can handle; it is a gap between lived experience and design assumption. The worker knows something about the texture of the task—the exception that matters, the caller who needs four extra minutes—that the design does not encode. The systems being deployed drift toward lower human agency anyway, because autonomous systems are cheaper, more scalable, and more predictable than ones that must stop to ask. The result is structurally produced: deployment optimizes for the properties that reduce agency, precisely on the tasks where workers most want to keep it.

What it costs a mind, and what it costs an organization

Why does any of this matter beyond worker grumbling? Because perceived control is not a soft preference. It is one of the most robust predictors of psychological well-being in the entire occupational-health literature.

Decades of research establish the relationship: workers with genuine autonomy over their work report better mental health, higher job satisfaction, and lower burnout, while low perceived control is reliably associated with elevated stress and anxiety, higher rates of depression and burnout, weaker organizational commitment, and rising intention to quit. This is among the most replicated findings in the field, predating AI by half a century—which is exactly why it matters here. AI has not invented a new pathology. It has industrialized a known one, applying a well-understood stressor at unprecedented scale. When control loss was idiosyncratic—a bad manager here, a rigid department there—its mental-health toll was real but scattered. When 36% of occupations cross the AI-saturation threshold at once, a large share of the workforce experiences the same erosion of control in the same period. A stressor that has always harmed individuals is being deployed as a population-level exposure.

The costs are real but diffuse, and diffuseness is dangerous. Agency decay does not spike; it accumulates. It rarely appears in any dashboard the organization actually watches, because the efficiency gains are immediate, quantifiable, and reportable while the psychological costs are slow, distributed, and hard to pin on any single system. What is not measured is not managed. So the organization often does not see the problem until it surfaces in the one place it cannot hide: turnover.

This is where agency decay becomes a governance failure with a price tag. When a skilled worker quits citing loss of autonomy, the organization loses more than a headcount. It loses institutional knowledge—the undocumented understanding of which exceptions matter and why. It loses relational capital—the trust that worker had built with clients and colleagues, which does not transfer with the role. And it loses adaptive capacity—the human ability to handle the case the system has never seen, which is the very thing the standardized system cannot replicate. None of this shows up in an AI performance dashboard. All of it shows up in the exit interview. The failure appears in turnover data long before, and far more clearly than, it appears in any metric the AI itself produces.

The gap between what leaders say and what they deploy

Organizational leaders are not, for the most part, villains in this story, and it is worth being precise about why the outcomes are bad anyway. By 2025 many companies had explicitly shifted their stated priorities from raw automation toward making work easier for people—employee experience, reliable information access, a sense of control over the digital environment (WORKAI, 2025).

Yet the same organizations frequently roll out wellness initiatives—mindfulness apps, flexible scheduling, mental-health resources—while simultaneously tightening the algorithmic monitoring that drains the very autonomy those initiatives are meant to restore. Workers experience this combination not as care but as performance: the organization solemnly acknowledges the stress while maintaining, and intensifying, the systems producing it. The predictable effect is not resolution but a deepening sense of powerlessness, which now attaches not only to the tasks but to the organization's whole posture. The problem is not hypocrisy so much as an asymmetry of incentives. The efficiency gains are legible and land on this quarter's report; the agency costs are illegible and land on next year's turnover. A wellness app is cheap, visible, and satisfies the felt need to do something. Rebuilding autonomy into the workflow is expensive, invisible on the dashboards leadership watches, and cuts against the efficiency case that justified the AI in the first place. Good intentions collide with an accounting system that cannot see what they are trying to fix.

Where this is heading

Is agency decay accelerating, or settling toward some equilibrium? The honest answer is that the trajectory points toward acceleration, for a structural reason. The arrival of agentic AI—systems that take sequences of actions and make decisions across extended workflows rather than answering one query at a time—pushes the default further toward machine authority. These systems are harder to audit in real time, their behavior is contextually variable, and each increment of delegated authority looks locally reasonable. Left to their own momentum, deployment patterns bend toward less human agency, not more, because that is the cheaper and more scalable path.

The countervailing question is whether governance is developing fast enough to shape how control is allocated before the norms harden. Here the picture is genuinely open. Traditional AI governance frameworks concentrate on accuracy, bias, legal compliance, and data security—all legitimate, none of which touch the question of how much discretion a human retains (Lexology, 2026). A different governance vocabulary is only beginning to emerge, and the window to establish it exists precisely because the norms are not yet locked in. Once "the system decides and the human documents exceptions" becomes the taken-for-granted shape of a job, it is far harder to renegotiate than it would be to design well now.

There is some evidence of movement in the better direction. A minority of organizations are experimenting with what gets called human-centered AI: dedicated override mechanisms with low friction and, crucially, no performance penalty for using them; regular reviews of AI parameters informed by worker feedback; and protected spaces for human judgment, such as case-review meetings deliberately held outside the algorithmic metrics. Where these are implemented consistently, they tend to lift morale and retain skilled workers. But they remain exceptions, and they will stay exceptions until the incentive structure changes—until, that is, agency erosion carries a cost that lands on the same report as the efficiency gain. The most plausible lever is making the invisible cost visible: building autonomy audits and retention-of-discretion metrics into the frameworks organizations already use to evaluate technology investments, so that human agency competes for attention on the dashboard rather than only in the exit interview.

Should agency be a design constraint?

This is finally a question of what we owe one another, not just what works. The book's position is that human agency should be treated as a binding design constraint on AI deployment, not an afterthought addressed through wellness programs after the damage is done—and the case rests on more than sentiment.

Consider the specific ethical shape of what agency decay does to someone like Martinez. He holds responsibility without authority: accountable for outcomes he can no longer determine, exposed to blame for the failures of a system he cannot overrule without penalty. In almost every other domain of employment ethics we recognize this as an unjust arrangement—responsibility and authority are supposed to travel together. AI deployment routinely severs them, and does so for workers who never chose the deployment in the first place. That combination—unchosen, and separating accountability from control—is what grounds a genuine obligation, not merely a nice-to-have.

Operationalizing the commitment does not require abandoning AI's real efficiency benefits. It requires making the trade-off transparent rather than letting it happen by default. Three governance elements would do most of the work. First, defined authority thresholds: explicit statements of what an AI system may do autonomously and what decisions require a human in the loop, decided deliberately rather than inherited from whatever the vendor shipped. Second, override rights without performance penalty: the worker's ability to overrule the system must be real, which means using it cannot be logged as a mark against them—an override tax is just autonomy decay with extra steps. Third, regular audits of worker discretion: periodic measurement of how much genuine judgment the system leaves in human hands, treated as a first-class risk alongside bias and accuracy, with the results reported where leadership will see them.

The efficiency gains of standardization are genuine, and this book does not pretend otherwise; a framework that ignored them would be as dishonest as one that ignored the psychological costs. The point is to weigh the two openly. Right now the trade is made silently and in one direction, because only one side of it appears on the books. A transparent framework would force the question into the open: this deployment buys us 20% more throughput at the cost of this much human discretion—is that a trade we mean to make? Organizations that ask the question out loud may still choose efficiency. But they will have chosen, rather than drifted, and the workers bearing the cost will at least be able to see the ledger.

How sure are we?

Intellectual honesty requires asking how much of this is solidly established and how much is inference, because the answer is not uniform across the chapter's claims.

The strongest link is the oldest one: the relationship between perceived control and psychological well-being is among the best-replicated findings in occupational health, established long before AI and across countless industries. What is newer, and therefore less certain, is the specific attribution of current control loss to AI specifically. This is a real epistemic difficulty and it should not be waved away. Many things are shifting in workplaces at once—remote and hybrid arrangements, organizational restructuring, post-pandemic churn—and each affects the sense of control. Isolating AI as the operative variable is genuinely hard. The best current research does it by comparing tasks and roles at different levels of AI saturation and by measuring worker preferences against system design directly, as the Human Agency Scale work does; the 47.5% preference gap, for instance, is a comparison internal to AI-mediated tasks and does not depend on separating AI from remote work. But the cleanest causal isolation is still emerging, and confident readers should hold the AI-specific attribution a notch below the general control–well-being link.

A second question is whether to trust the workers. Self-reports of autonomy loss could be inflated by a general anxiety about AI, or by nostalgia for a past that was less autonomous than memory paints it; they could equally be understated by workers who have already adapted their expectations downward and no longer notice what they have given up. The useful check is corroboration: where organizations' own performance metrics can be examined, do they support or contradict the worker accounts? The override-rate data is telling here—the fact that high override rates get flagged as a problem is not a worker perception but an organizational artifact, and it corroborates the account of a shrinking discretionary zone. Objective monitoring intensity, likewise, is measurable independent of how workers feel about it. Where these organizational metrics have been examined, they tend to corroborate rather than contradict the subjective reports, which is the strongest reason to take the self-reports seriously.

The largest genuine unknown is reversibility. If an organization reverses course and explicitly designs for greater human agency, does perceived control recover—and on what timeline? Here the honest answer is that we do not yet know. The human-centered deployments are too recent and too few to have generated longitudinal evidence on recovery. It is plausible that agency, like a muscle, can be rebuilt once the friction on judgment is removed; it is also possible that once workers have internalized diminished expectations—reframing themselves from decision-makers into facilitators—the psychological recovery lags well behind the structural change, or does not fully arrive. This matters enormously for policy, because a reversible harm and a permanent one call for different urgency. Until the evidence exists, the prudent course is to treat the harm as potentially sticky and act before the norms, and the expectations, harden into place.

Summary

This chapter examined agency decay—the gradual erosion of autonomous decision-making capacity and felt independence among workers deployed alongside AI—and the psychological and organizational costs that follow.

Agency decay is a loss of authorship, not employment. It is the slow narrowing of the space in which a worker's own judgment carries weight, occurring at scale: roughly 75% of workers now use AI, and 36% of occupations use it for a quarter or more of their tasks—the threshold at which AI becomes an environment rather than a tool.
Three mechanisms produce it. The autonomy paradox: optimization and standardization are structurally at odds with discretion, so judgment gets taxed with friction until workers economize on it. The efficiency trap: productivity gains convert into proportionally higher targets, so the demonstration of capacity becomes self-defeating and workers lose control of their own pace. The micromanagement paradox: algorithmic monitoring surveils more relentlessly than any human could, eliminating discretionary space and turning managers into enforcers of the algorithm.
Workers want co-authorship. Across 104 occupations, "equal partnership" was the most common preference, and on 47.5% of tasks workers wanted more human agency than technologists deemed necessary—a gap between lived experience and design assumption, not between capability and preference.
The costs are real but diffuse. Low perceived control is one of the most robust predictors of poor mental health in occupational research; AI industrializes this known stressor at population scale. The damage rarely shows on efficiency dashboards—it surfaces first in turnover, where the departure of skilled workers costs institutional knowledge, relational capital, and adaptive capacity that AI cannot replicate.
Leadership intent is not the bottleneck; incentives are. Wellness programs deployed alongside tightening surveillance read as performative because the efficiency gains are legible and immediate while the agency costs are invisible and deferred.
The trajectory bends toward acceleration, driven by agentic AI, but the norms are not yet locked in—leaving a real window for governance to shape how control is allocated.
Human agency should be a binding design constraint, operationalized through defined authority thresholds, override rights without performance penalty, and regular audits of worker discretion—making the efficiency-versus-agency trade-off transparent rather than default. Organizations that never chose to erode their workers' autonomy still owe those workers remedies when they do, because separating responsibility from authority is unjust in any employment ethic.
The evidence is uneven, and honesty requires saying so. The control–well-being link is rock solid; the AI-specific attribution is strong but still being cleanly isolated from concurrent changes; worker self-reports are corroborated by organizational metrics like override-flagging; and reversibility remains genuinely unknown—which is itself a reason to act before expectations harden.

Sources

Last updated: 2026-08-31

V2 (in progress) Previous: V1