Homo agenticus in the age of agentic AI: Agency loops, power displacement, and the circulation of responsibility

You’re listening to “Homo agenticus in the age of agentic AI: Agency loops, power displacement, and the circulation of responsibility,” by Paul M. Leonardi. Published in 2025.

Abstract.

A defining characteristic of humanity is our relentless pursuit of control over our environment and our own destinies. We are, in essence, homo agenticus—humans defined by our search for agency. Philosophers have grappled with questions of voluntarism and determinism at least since the time of Aristotle, creating entire ethical frameworks predicated on a belief in human agency. Psychologists routinely investigate agency and locus of control, arguing that they are foundational to human motivation. Economists build models of markets based on assumptions of human agency and self-maximization. Legal systems distribute punishment based on evaluations of intentional, goal-oriented action. And religious traditions debate divine determination versus human agency.

Within our own studies at the intersection of organizational theory and information systems, we have moved from discussing the agency of technological systems in deterministic ways to arguments that human agents shape technological outcomes, to approaches that treat agency as materializing out of sociomaterial practices.

Today, scholars from across the fields of information systems, organization studies, communication, and beyond have begun to treat agency not as something to aid in explanation, but as the thing to be explained. It has become central to theorizing. As some examples, Murray et al. (2021) responded to the evolving capabilities of machine learning models by theorizing about “conjoined

agency,” Chandra et al. (2022) asked what it means to be “human” in a world filled by machine agency, and Vanneste and Puranam (2024) proposed that the ways we perceive AI to have agency affect whether and how people trust the technologies they use. Suddenly, discussions of agency are everywhere. This swift embrace of theorizing about agency at the dawn of the agentic AI era makes sense. With their conversational user interfaces and their generative capabilities, technologies powered by AI certainly have a capacity for action that seems to mirror, though less fully, the capacity that humans have. Companies are now racing to build AI agents that can reason and act autonomously. As the tech giant Amazon defines it, an AI agent is a “software program that can interact with its environment, collect data, and use the data to perform self-determined tasks to meet predetermined goals.

Humans set goals, but an AI agent independently chooses the best actions it needs to perform to achieve those goals.”1 To the casual observer, that certainly sounds like a technology that has agency.

Why such an interest in agency, generally, and in the context of AI agents specifically? Across all of these different areas of work, it seems that scholars are responding to the same human imperative: our need to understand and assert control within our changing environment. The multiplicity of approaches to the conceptualization and study of agency reflect our need to make sense of agency from every possible angle. As the philosopher Berlin (1969: 178) noted, “The desire to be self-directed... is a desire to be somebody, not nobody; a doer—deciding, rather than being decided for.”

Decades of research demonstrate that feeling in control is desirable and beneficial for human psychological health and functioning. When nursing home residents are given just a bit of control over their daily schedules, their health outcomes dramatically improve and mortality rates decline. When workers lose autonomy, they experience heightened stress, reduced motivation, and diminished well-being. Infants as young as eight months old exhibit agentive frustration when their actions fail to produce expected outcomes.

Yet for all this scholarly attention to defining, measuring, and theorizing agency, we have largely overlooked an important question: what happens when we attribute agency? While researchers continue to debate the ontological foundations of agency and seek ever more precise conceptualizations, I argue that we must shift our focus to the performative effects of agency attribution itself. This shift is crucial because the act of attributing agency does something beyond description—it actively creates the conditions for action and responsibility.

When we attribute agency to ourselves, we claim the capacity to act, accepting both the possibilities and burdens that come with such attribution. When we attribute agency to others—whether human colleagues, organizational structures, or increasingly, AI agents—we simultaneously divest ourselves of that capacity. This divestment represents a transfer of both power and responsibility. The entity to whom we attribute agency becomes, in our understanding, the locus of intentional action and the bearer of consequences. In the emerging landscape of agentic AI systems, these attributions take on heightened significance as we delegate decision-making authority to algorithmic agents, effectively ceding control over outcomes that profoundly affect human lives.

This process of attribution is neither static nor unidirectional. Instead, attributions of agency shift dynamically throughout sociotechnical systems, creating what I term “agency loops”—recursive patterns where responsibility for action circulates among human and non-human actors. A manager attributes agency to an AI recommendation system, which in turn bases its suggestions on data patterns that reflect previous human decisions, which were themselves influenced by earlier algorithmic outputs. These loops create a distributed form of power that defies traditional notions of centralized control or clear lines of accountability. Understanding agency attribution as the conferral of power reveals why questions of agency have become so pressing in the age of agentic AI.

As organizations increasingly deploy AI systems that exhibit goal-directed behavior, make autonomous decisions, and adapt to changing circumstances, the attribution of agency to these systems has the potential to reshape the distribution of power. The danger lies in how our attributions of agency to these systems redistribute our own capacity for action and accountability.

This paper explores how agency loops emerge and evolve within contemporary organizational contexts, tracing the circulation of power through patterns of attribution that connect human and artificial agents. By focusing on the performative effects of agency attribution rather than its definitional boundaries, we can better understand how power is negotiated, transferred, and sometimes lost in the complex assemblages that characterize modern organizational life. In doing so, we move beyond the question of what agency is to examine what agency attribution does—and what it means for human autonomy in an age of increasingly agentic machines.

1. A sense of agency.

While philosophers, sociologists and organizational scholars continue to debate the ontological foundations of agency, a different research tradition offers a more pragmatic approach to understanding how agency operates in human experience. The psychological literature on “sense of agency” examines how people experience agency and how those experiences shape their understanding of causation, responsibility, and control. Most critically, this research reveals that agency operates as an attribution process that actively distributes power and responsibility throughout social systems.

A sense of agency has been defined by Haggard and Tsakiris (2009: 242) as “the experience of controlling one’s own actions and, through them, events in the outside world.” It encompasses both low-level, pre-reflective feelings of action-ownership (“this movement is mine”) and higher-order judgments about causation and responsibility (“I caused that outcome”). Without a sense of agency, external control becomes meaningless because what matters to us is our conscious experience of influencing our environment. Our sense of agency is deeply integrated into our neurological architecture, as evidenced by brain imaging studies showing distinctive neural signatures in the angular gyrus and supplementary motor area when we experience control over our actions.

The sense of agency is so intrinsic to normal human functioning that its disruption characterizes severe psychiatric conditions—patients with schizophrenia often report feeling that their actions are controlled by external forces, while those with alien hand syndrome experience intentional movements they feel they did not initiate.

A sense of agency is crucial for human well-being and functioning. As Bandura (1989) demonstrates, our beliefs about our capacity to produce effects—what he terms “self-efficacy beliefs”—profoundly influence how we think, feel, motivate ourselves, and behave. These beliefs affect our goal setting, our analytical thinking, our motivational processes, our affective states, and our selection of environments and activities.

To understand how sense of agency operates, consider what happens when you press a key on your keyboard and see a letter appear on your screen. You experience a sense of agency—you feel that you caused the letter to appear. But how does your brain create this feeling? Wegner and Wheatley (1999) demonstrate that we experience agency when three conditions are met: priority (our thought about the action occurs before the action), consistency (our thought matches what actually happens), and exclusivity (no other obvious cause is present). Building on this research, Wegner developed a theory of “apparent mental causation” to explain this phenomenon, arguing that “people experience conscious will when they interpret their own thought as the cause of their action,” independent of any actual causal connection between thought and action (Wegner, 2004: 654).

He suggests that the experience of will emerges from the same inferential processes that govern all causal perception. Just as we observe a billiard ball striking another and infer causation from their temporal relationship, we observe our thoughts preceding our actions and construct a sense of personal agency from this conjunction. The feeling of conscious will, in this view, operates through the mind’s interpretation of apparent causal relationships rather than direct perception of genuine causal force.

This attribution process operates automatically, but it can be remarkably flexible and fallible. Moore (2016) describes how people feel agency over placebo buttons—pedestrian crossing buttons and elevator “close door” buttons that actually do nothing. Despite having no real causal power, people continue to experience control because the temporal relationship between button press and eventual outcome (lights changing, doors closing) creates apparent causation. Similarly, Henslin (1967) observed cab drivers playing craps who threw dice harder when they needed higher numbers and softer for lower numbers, demonstrating how people attribute agency even to objectively uncontrollable events.

These examples illustrate that our tendency to experience agency is based on inferential processes that can be systematically deceived by coincidental timing, consistency, and the absence of obvious alternative causes.

If our most intimate experience of personal causation operates through inference rather than direct perception, then all attributions of agency—whether to ourselves, other humans, or artificial systems—emerge through constructive rather than perceptual processes. We construct attributions of agency based on observable patterns of behavior, temporal relationships between apparent intentions and outcomes, and assessments of consistency between presumed goals and observed effects. These attributions then become the foundation for understanding responsibility, power, and control within any system where multiple potential agents might be operating. When we attribute agency to ourselves, we claim ownership over outcomes and accept responsibility for consequences.

When we attribute agency to others—whether people or systems—something consequential happens: we simultaneously reduce our own sense of control and transfer responsibility to the entity we perceive as causal. Empirical research demonstrates this transfer in action. Studies using intentional binding—the tendency to perceive voluntary actions and their outcomes as closer in time—reveal systematic patterns in how we distribute agency. Yoshie and Haggard (2013) found that people showed reduced intentional binding for actions with negative outcomes compared to positive ones, demonstrating that we distance ourselves from negative consequences by attributing them elsewhere.

This attribution process also extends to technological interactions. Studies comparing different computer interfaces find that people experience varying levels of agency depending on the interaction method: stronger agency when using “skinput” (tapping on their own skin to control computers) compared to traditional keyboards, and weaker agency with speech interfaces compared to keyboard input. Most remarkably, research on brain-machine interfaces reveals that people can experience agency even when controlling devices purely through neural signals without any overt behavior, though visual feedback becomes more influential when internal motor cues are absent.

As Moore (2016: 2) observes, this flexibility is both a strength and a vulnerability: “The flexibility that might make us vulnerable to agency errors in things like placebo buttons and voodoo dolls, can also allow our experience of agency to extend into new domains.” Our attribution system can adapt to new technologies and causal relationships, but this same adaptability makes us susceptible to attributing agency where none exists—or failing to recognize it where it does. In landmark experimental work, Wegner and Wheatley (1999) made participants feel agency over actions they never performed, simply because their thoughts preceded and matched those actions at appropriate intervals.

This convergence of evidence points toward a reconceptualization of agency. Agency is an attribution—a cognitive and social construction that emerges from our ongoing interpretation of causal relationships in our environment. This attribution process operates through the same inferential mechanisms whether we are attributing agency to ourselves, other humans, or artificial systems. We observe patterns of behavior, note temporal relationships between intentions and outcomes, assess consistency between goals and effects, and evaluate the exclusivity of potential causal factors. Based on these observations, we construct attributions of agency that become the basis for understanding who or what is responsible for outcomes.

The critical insight is that these attributions are performative rather than merely descriptive. When we attribute agency to any entity—human or artificial—we actively construct the conditions under which that entity will be treated as a responsible actor. We grant it the status of causal agent and simultaneously modify our own relationship to the capacity for action and responsibility within that domain. A sense of agency plays a key role in guiding attributions of responsibility, and these responsibility attributions directly determine where power resides within social systems. When we experience reduced agency in the presence of AI systems, we simultaneously transfer responsibility for outcomes to those systems, effectively redistributing power within human-AI assemblages.

Agency is what we attribute. To say that agency is attributed rather than embodied or enacted means that there is no inherent, objective property of “agentness” that resides within entities or emerges from the interaction between them. Instead, agency materializes through ongoing processes of interpretation and assignment. We observe patterns of behavior, note temporal relationships between apparent intentions and outcomes, assess consistency between presumed goals and observed effects, and based on these observations, we construct attributions of agency that become the foundation for understanding who or what is responsible for outcomes.

That does not mean that agency is merely cognitive, ephemeral, or simply a social construction. While agency attributions operate through interpretive processes, they have material consequences that reshape the physical and social world. When we define something as having agency—whether a human manager, an organizational procedure, or an AI system—we change our actions in response to that entity. We defer to its decisions, hold it accountable for outcomes, grant it authority over resources, and modify our behavior to accommodate its presumed intentions. These behavioral changes create real consequences: budgets get allocated, careers advance or stall, lives are saved or lost, and power structures are reinforced or transformed. Agency attributions thus operate as performative acts that bring into being the very realities they appear to describe.

Understanding agency as attribution rather than capacity reveals that the question concerns how our attributions of agency create the very realities of power and responsibility they appear to describe. In the age of agentic AI, these attribution processes become the primary mechanism through which power circulates between humans and machines, making the study of agency attribution crucial for understanding the distribution of power in increasingly automated societies.

2. Rethinking AI through the lens of agency attribution.

When we examine artificial intelligence through the traditional lens of capabilities and ontology—asking whether AI systems truly possess agency or intelligence—we miss the more pressing process that shapes human-AI interaction. The psychological research on sense of agency reveals that what matters concerns how humans experience and attribute agency to these systems. This attribution process actively redistributes power and responsibility in ways that have implications for human autonomy and organizational dynamics. Each attribution of agency to an AI system represents a moment where humans cede some portion of their sense of control and responsibility to that system. Understanding this process requires examining how humans experience their interactions with these systems and how those experiences reshape power relationships.

The language we use reveals this attribution process in action. When a physician declares “The AI caught a tumor I completely missed,” they are attributing diagnostic agency to the system while simultaneously reducing their own sense of responsibility for the detection. When a financial advisor explains “The algorithm determined the portfolio based on your risk tolerance,” they delegate judgment to a statistical model while positioning themselves as conduits rather than decision-makers. When a programmer admits “I don’t fully understand why the model made this prediction, but its accuracy rate is impressive,” they acknowledge opacity while still attributing competence and agency to the system. Neuroscientific research confirms that these are genuine perceptual experiences rather than merely linguistic conventions.

Humans activate similar neural pathways when observing actions by humans, robots, or even animated shapes that appear to act intentionally. As AI systems increasingly display behaviors that trigger our evolved agency-detection mechanisms, we experience them as genuine agents. This perceptual reality shapes how we interact with these systems and how we understand our own role in human-AI assemblages.

2.1. Attributions of agency to AI displace human power.

A growing body of empirical research demonstrates that when humans attribute agency to AI systems, they experience systematic displacement of power—understood as the combination of control and responsibility—from themselves to the technology. This displacement occurs even when humans retain objective authority over the systems and even when the AI systems have no actual autonomous capabilities. Ciardo et al. (2020) provide perhaps the clearest evidence of this power displacement. In their study, participants performed tasks alongside a humanoid robot named Cozmo that was presented as an intentional agent. Even when the robot never actually acted, participants reported significantly lower sense of agency over task outcomes compared to working alone.

Critically, this reduction did not occur when participants worked alongside a passive air pump that could influence outcomes but was perceived as intentional. The study reveals that the mere attribution of intentionality—actual capability or intervention—is sufficient to displace human sense of agency and, by extension, felt responsibility and control.

Zanatto et al. (2021) demonstrate how this displacement occurs along a continuum of AI involvement. Their manipulation of three automation levels reveals that power displacement begins even with minimal AI participation. In the System Warning condition, humans still performed the physical action but were prompted by the computer. Despite retaining motor control, participants showed significantly reduced intentional binding—a reliable indicator of diminished sense of agency. This suggests that once humans perceive AI systems as initiating or directing action, they begin to cede responsibility and control even for actions they physically perform. The full automation condition (System Decision) produced the most complete power displacement, with intentional binding becoming weakest or entirely absent.

But the critical insight is that displacement begins as soon as humans attribute any causal role to the AI system. The locus of initiation—who decides when and what to do—proves psychologically more important for maintaining agency than who physically executes the action.

Berberian et al. (2012) extended this finding to professional contexts where maintaining human oversight is crucial for safety. In their aviation simulation study, both implicit and explicit measures of agency declined monotonically with increasing automation levels. As AI systems assumed more control over aircraft functions, pilots felt progressively less responsible for outcomes, even when they retained supervisory authority and legal accountability. This represents a dangerous misalignment between legal responsibility and psychological experience of control.

Power displacement reveals itself most clearly when things go wrong. Jia et al. (2022) found that participants consistently credited themselves for successful outcomes when working with AI systems but blamed the AI for failures. This asymmetric attribution pattern demonstrates the motivated nature of power displacement: humans strategically attribute agency in ways that maximize their sense of control over positive outcomes while minimizing responsibility for negative ones. Yet even this strategic attribution represents a form of power displacement, as it requires acknowledging that AI systems can be causal agents capable of producing undesired effects. When humans attribute agency to AI, they actively reconstruct the distribution of power within the assemblage.

3. Agency loops: power displacement and the circulation of responsibility.

Such power displacement is dynamic rather than permanent. Most AI-human interactions follow recursive patterns that I call “agency loops”—cycles where power circulates through processes of delegation, attribution, contingency, reassertion, and reconfiguration. These loops represent the mechanism through which agency moves through sociotechnical systems, constantly redistributing responsibility and control as humans attribute agency to different entities in different circumstances. Agency loops occur because the attribution of agency to AI systems is inherently unstable. While humans readily attribute agency to AI during smooth operations, this attribution becomes problematic when systems fail, produce unexpected results, or face accountability demands.

At these moments, the comfortable fiction of autonomous AI agency collapses, and humans must reassert their own agency to address consequences. This reassertion leads to reconfiguration of future attribution patterns, setting up new conditions for subsequent cycles. The cyclical nature of these interactions means that agency is always moving through a system rather than residing permanently in any single location. Power circulates through predictable patterns that shape how responsibility and control are distributed over time. Understanding these patterns reveals how complex human-AI assemblages actually operate and where opportunities for intervention might emerge.

Agency loops unfold at multiple temporal scales, from split-second interactions to multi-year organizational transformations. They can nest within each other, with micro-loops of individual correction embedded within macro-loops of institutional learning. Most importantly, they are evolutionary processes where each cycle potentially alters the conditions for subsequent interactions. The five phases of an agency loop are illustrated in Fig. 1, summarized in Table 1, and explained in detail below.

3.1. Phase 1: delegation - establishing the conditions for attribution.

Every agency loop begins with an act of human delegation where people design, implement, configure, or deploy AI systems with

Agency loop framework.

specific parameters and capabilities. During this phase, humans retain maximum sense of agency because they are actively shaping the system’s possibilities and constraints. They are the architects making deliberate choices about what the AI can and cannot do. Consider a data scientist developing a machine learning model for medical diagnosis. She selects training data, chooses features, defines the objective function, and sets performance thresholds. Throughout this process, she experiences complete agency—every decision reflects her judgment, expertise, and intentions. The model exists as an extension of her analytical capabilities rather than an independent agent. Her sense of control and responsibility remains absolute.

Similarly, when a manager implements an automated hiring system, she configures which criteria it evaluates, how it weights different factors, and what thresholds determine candidate advancement. She experiences this as exercising managerial authority, making strategic decisions about organizational priorities. The system serves as a tool for implementing her vision of effective hiring practices.

During delegation, humans often explicitly acknowledge their agency and responsibility. Documentation reflects human ownership: “I designed this system to optimize for accuracy while minimizing bias.” “We implemented these parameters based on business requirements.” “The team configured the AI to align with company values.” This phase establishes both the technical conditions and psychological expectations that will shape subsequent attribution patterns.

3.2. Phase 2: attribution - the shift to perceived AI agency.

Once AI systems begin operating within established parameters, users and observers gradually shift from experiencing them as tools to perceiving them as agents. This transition represents the critical moment where power begins to displace from humans to AI through attribution processes. Users interact with the system and experience its responses as autonomous, purposeful, and intelligent rather than as predetermined outputs of human programming. For example, AI used in radiology begins analyzing patient scans and identifying potential abnormalities. Radiologists start to experience the system as having diagnostic capabilities: “The AI flagged this suspicious area.” They begin attributing perceptual and analytical agency to the system, experiencing it as a colleague who notices things they might miss.

Their language shifts from describing the system as implementing their diagnostic approach to crediting it with independent insights. Or consider AI tools used in reviewing job applicants. The AI starts screening applications and ranking candidates. Hiring managers experience the system as making judgments: “The algorithm selected these top applicants based on their qualifications.” They attribute evaluative agency to the system, experiencing it as having opinions about candidate quality rather than simply executing their predefined criteria. The system begins to feel like a hiring partner with its own assessment capabilities.

During this phase, humans experience progressively reduced sense of agency as they attribute causal power to the AI system. They begin to see outcomes as resulting from AI decisions rather than human choices. Their sense of responsibility shifts accordingly—successes and failures increasingly feel like products of AI agency rather than human judgment. This attribution requires belief that the AI acts as a causal agent producing effects in the world.

3.3. Phase 3: contingency - when AI attributions of agency become problematic.

The attribution phase continues until an unexpected event disrupts the smooth operation of the system and reveals the limitations or problems with treating AI as an autonomous agent. These contingencies take various forms: technical failures, unexpected outcomes, ethical violations, legal challenges, or public controversies. They represent moments when the attribution of agency to AI systems becomes untenable or dangerous.

Maybe the AI used to examine radiological scans misses an obvious tumor while flagging numerous false positives, leading to patient harm and malpractice concerns. The comfortable attribution of diagnostic agency to the system suddenly becomes problematic when explanations are demanded: Who is responsible for the misdiagnosis? How should the error be explained to patients and families? What does this mean for future diagnostic decisions? Or the AI used to screen job applicants exhibits systematic bias, discriminating against qualified candidates from underrepresented groups in violation of employment law. The attribution of evaluative agency to the system becomes legally and ethically untenable when regulators investigate and demand accountability: Who decided to implement discriminatory practices? How will the organization explain these outcomes to affected candidates?

What changes are required to prevent future discrimination?

These contingencies shatter the illusion of autonomous AI agency and create urgent pressure for human intervention. They reveal that the comfortable delegation of agency to AI systems carries risks and responsibilities that cannot be avoided through attribution. Humans discover they cannot escape accountability by claiming the AI acted independently—they remain liable for consequences regardless of their subjective attribution patterns.

3.4. Phase 4: reassertion - reclaiming human agency and responsibility.

When contingencies arise, humans must dramatically reassert their agency to address consequences, provide explanations, and implement corrections. This phase involves a complete reversal of the attribution patterns that characterized the preceding phases. Humans can no longer claim the system was autonomous; they must acknowledge and exercise their authority over it.

The data scientist who developed the diagnostic AI must now debug the system, identify sources of error, and implement improvements. She cannot claim the AI made independent mistakes—she must take responsibility for its design, training, and deployment decisions. Her agency resurfaces as she analyzes model performance, adjusts parameters, and validates improvements. The system reverts to being a tool she controls rather than an agent with independent capabilities. The hiring manager must explain discriminatory decisions to affected candidates, implement remediation processes, and revise the system to prevent future bias. She cannot claim the algorithm acted on its own—she must acknowledge her role in configuring its criteria and deploying it without adequate oversight. Her managerial agency resurfaces as she takes responsibility for outcomes and implements changes.

During reassertion, humans often experience intensified sense of agency and responsibility as they work to address consequences of their previous delegation. They must exercise explicit control, make visible decisions, and accept accountability for both the AI’s past actions and their own corrective measures. This phase often involves significant stress as humans confront the gap between their subjective experience of having delegated agency and their objective responsibility for outcomes.

3.5. Phase 5: reconfiguration - adjusting future attribution patterns.

Based on their experience with contingency and reassertion, humans adjust their delegation practices and attribution patterns for future interactions with AI systems. This reconfiguration phase sets new conditions for subsequent loops, potentially altering how agency will be attributed and power will be distributed in future cycles.

Reconfiguration can take multiple forms. Humans might modify the AI system’s technical parameters, implement new oversight mechanisms, change their mental models of the system’s capabilities, or alter their willingness to attribute agency to it. They might develop more skeptical attitudes toward AI recommendations, require human approval for consequential decisions, or implement additional monitoring and auditing processes. The medical AI might be reconfigured with different sensitivity thresholds, additional validation requirements, or mandatory human review of all findings. The data scientist might implement new testing procedures, bias detection algorithms, or transparency mechanisms. These changes reflect updated understanding of the system’s limitations and appropriate boundaries for delegation.

The hiring AI might be reconfigured with bias detection algorithms, demographic parity constraints, or human review of all hiring decisions. The manager might implement new training requirements, auditing processes, or candidate feedback mechanisms. These changes reflect lessons learned about the risks of delegating evaluative agency to AI systems.

Crucially, reconfiguration creates new conditions that shape how future loops will unfold rather than simply restoring the original state. Each cycle through the loop changes participants’ understanding of AI capabilities, their strategies for managing human-AI collaboration, and their comfort with attributing agency to AI systems. The evolution of these attribution patterns over time represents the mechanism through which human-AI relationships mature and stabilize. Understanding agency loops in this detailed way reveals how power actually circulates throrememugh sociotechnical systems rather than simply transferring from humans to machines. It shows why the question concerns how agency attribution creates dynamic distributions of responsibility and control that constantly evolve through ongoing cycles of delegation, attribution, contingency, reassertion, and reconfiguration.

3.6. Recursive loops.

Agency loops are recursive systems that produce cumulative change. Each cycle through the loop—delegation, attribution, contingency, reassertion, and reconfiguration—shapes the next. Delegation establishes technical constraints and implicit expectations about what the AI system should do. Attribution defines how agency is distributed and signals how much responsibility humans are willing to transfer. Contingency reveals how those attributions hold up under pressure. Reassertion is a return to human control that forces clarification of roles and consequences. Reconfiguration sets new conditions for interaction, encoding lessons into modified practices, altered oversight structures, or revised system parameters. As humans move through these cycles, they gradually rework how they relate to AI systems.

Power and responsibility shift, stabilize, or fracture based on what happens during each loop. Over time, these shifts create new organizational patterns that reflect the history of these recursive adjustments.

Each loop demands something different from the humans involved. These demands are not only technical or procedural. They are cognitive, emotional, and social in nature, and they accumulate across repeated cycles. Delegation requires systems thinking and confidence in one’s judgment about what the AI should be allowed to do. Attribution demands interpretation, the calibration of trust, and careful impression management. Contingency requires rapid sense-making under uncertainty, often accompanied by fear and the need to preserve reputation. Reassertion brings with it accountability anxiety and the need to justify and explain decisions. Reconfiguration involves pattern recognition, reflection, and often political negotiation to align new practices with institutional realities.

These human demands shape how loops evolve and determine how easily responsibility is reclaimed or redistributed at each stage. Table 2 summarizes the forms of human labor that each phase draws upon.

Each loop changes the baseline. Humans do not return to the same conditions when they engage with the system again. They carry forward updated beliefs about the AI’s reliability, usefulness, and risk profile. They revise their practices and refine their expectations. This accumulation of experience shapes how people delegate in the future and how they attribute agency in new situations. A system that performed well during early loops may be granted more discretion. A system that created public backlash or operational disruption may face tighter control. These histories matter. They influence the institutionalization of norms, the design of safeguards, and the trust placed in human or machine actors. The loop becomes the mechanism through which human-AI relationships evolve through patterned interaction over time.

Agency loops operate at multiple time scales. Micro-loops unfold during everyday interactions. These may last seconds or minutes and involve quick shifts in how users assign and reclaim control. Meso-loops span weeks or months and include shared experiences that shape how teams use and interpret AI systems. Macro-loops stretch across years and involve organizational learning, policy redesign, and shifts in how technologies are governed. These loops are nested. What happens in a single moment of interaction can feed into team practices. Those practices can become templates for broader institutional responses. Organizational policies and oversight mechanisms shape how individual users are trained to work with AI systems. Through this nesting, agency becomes structured by time. It becomes layered and historical.

The organization accumulates loops of interaction that define what AI becomes in practice.

Understanding agency loops requires following how power and responsibility move over time. These movements are shaped by the structure of the loop and by the memory that each cycle preserves. As people delegate to AI, they shift authority away from themselves. When AI systems produce unexpected outcomes, people step back in to manage the consequences. When humans reassert control, they redefine who is accountable and what forms of oversight are necessary. Reconfiguration sets the terms for the next delegation. Through repeated cycles, power becomes redistributed across humans and machines in patterned ways. Some tasks become increasingly automated, while others are pulled back into human hands. The loop reflects where discretion resides, who is seen as competent to act, and whose decisions are subject to review. It also shapes how responsibility circulates.

With each cycle, humans adjust how much control they give to AI systems and where they believe accountability lies. Responsibility moves with attribution. When agency is assigned to a system, responsibility tends to follow. When attribution becomes unstable or breaks down, responsibility is recalled and reassigned. Over time, these adjustments establish norms for who owns outcomes, who must explain them, and who is positioned to fix them. The loop produces these norms through repetition. It embeds them in routines, policies, and organizational memory. The more a system is used, the more responsibility becomes distributed through these learned patterns. What appears to be a single delegation decision is part of a larger system through which power and accountability are continually made and remade.

4. Agency loops in practice: empirical evidence.

To understand how agency loops operate in practice and the different stable configurations they can produce, I have chosen three recent papers that have presented data about people’s use of AI-powered technologies in the process of organizing. Although there are many papers showing AI’s effect on productivity and other papers discussing people’s reactions to using

What each phase demands of human.

new AI systems in their work, today there are still few empirical explorations of people actually using AI in the context of their work. I have chosen three studies that reveal distinct evolutionary trajectories for agency loops, examining these cases through the agency loop framework reveals that the critical factor determining outcomes concerns how humans manage the tension between delegation benefits and control imperatives as attribution processes unfold over time. Each case demonstrates different strategies for navigating this tension, producing stable configurations that preserve human agency while leveraging AI capabilities—though through very different mechanisms and with varying implications for power distribution and responsibility allocation.

4.1. Radiological diagnosis: distributed diagnostic agency through AI interrogation.

The first study, conducted by Lebovitz et al. (2022), focuses on why radiologists in three departments—breast imaging, chest imaging, and pediatric imaging—accepted or rejected AI diagnostic outputs at a large U.S. teaching hospital. Radiologists used AI tools to support diagnoses for breast cancer, lung cancer, and pediatric bone age assessment, consulting the AI after forming their own independent judgments. The findings revealed significant variation in how radiologists incorporated AI results.

In lung cancer diagnosis, radiologists consistently engaged with AI outputs by performing “AI interrogation practices” to reconcile divergent opinions, achieving “engaged augmentation.” In contrast, in breast cancer and bone age diagnoses, radiologists often dismissed or uncritically accepted AI results, reflecting “unengaged augmentation.” This case reveals how agency loops can create stable “distributed diagnostic agency” where human expertise and AI capabilities operate in productive tension, achieved through the development of systematic “AI interrogation practices” that transform opacity from a barrier into a domain for expert interpretation.

4.1.1. The logic of delegation.

The delegation emerged from radiologists’ acknowledgment of human perceptual limitations in life-or-death diagnostic contexts, creating conditions where professionals were willing to share analytical responsibility despite their traditional autonomy. The system was deliberately structured to preserve professional authority by requiring independent human judgment before AI consultation, allowing radiologists to experience delegation as extending rather than replacing their expertise. Initially, radiologists attributed minimal agency to the AI, viewing it as an advanced pattern recognition tool rather than an independent diagnostic entity. This careful positioning maintained radiologists’ formal control over diagnostic decisions while delegating only specific analytical tasks.

4.1.2. Attribution dynamics.

Through repeated interactions where the AI identified pathology that radiologists had missed, the system gradually acquired genuine agency attribution as the temporal relationship between AI analysis and diagnostic success triggered attribution of independent analytical capabilities. Radiologists began experiencing the AI as possessing independent analytical intelligence that could exceed human performance, particularly when the system detected abnormalities that human analysis had overlooked. This attribution pattern systematically displaced radiologists’ sense of sole diagnostic authority through “causal redistribution,” where successful diagnoses increasingly felt like products of AI insight rather than purely human expertise.

While radiologists remained legally responsible for final diagnoses, they experienced reduced control over the diagnostic process as AI recommendations became difficult to ignore.

4.1.3. Contingency and the collapse of comfortable attribution.

The attribution dynamics became unsustainable when the AI’s frequent disagreement with initial human assessments created uncertainty that couldn’t be resolved through normal professional practices, as the opacity of AI reasoning processes meant radiologists could neither fully trust nor adequately evaluate the system’s recommendations. This contingency revealed the problematic nature of attributing agency to an opaque system, with radiologists finding themselves caught between acknowledging the AI’s demonstrated capabilities and being unable to understand its reasoning processes. The situation created a power vacuum where neither human nor AI authority felt adequate, as radiologists retained legal responsibility but lacked the interpretive control necessary to evaluate AI contributions effectively.

4.1.4. Reassertion.

Rather than accepting or rejecting AI authority outright, radiologists developed sophisticated “AI interrogation practices” involving systematic examination of AI-flagged areas through multiple imaging techniques and historical data review, constituting active reconstruction of diagnostic authority within sociotechnical assemblages. Through these interrogation practices, radiologists reconstructed their relationship with AI agency, positioning themselves as expert interpreters of AI outputs rather than passive recipients of AI recommendations, thereby reclaiming interpretive authority over AI outputs without rejecting the system’s analytical value. They established human expertise as the ultimate arbiter of diagnostic meaning while preserving AI contributions, transforming the opacity problem from a barrier to agency into a domain for exercising professional judgment.

4.1.5. Reconfiguration.

The sustained practice of AI interrogation led to “distributed diagnostic agency”—a stable configuration where human expertise and AI capabilities operated in productive tension rather than hierarchical relationship, enhancing rather than diminishing both human and AI contributions to diagnostic work. This reconfigured system involved nuanced agency attribution where the AI retained attributed diagnostic capabilities while human agency was enhanced through new interpretive skills. Rather than zero-sum competition, the reconfiguration achieved dynamic power equilibrium where human expertise and AI capabilities reinforced each other, with radiologists gaining influence through their ability to integrate AI insights with clinical knowledge while the AI maintained authority as a specialized pattern recognition partner.

This reconfiguration did not mark a final resolution. Instead, it established new starting conditions for the next loop. As radiologists continued to work with the AI system, their interrogation practices evolved, and new contingencies emerged. Diagnostic challenges changed with new patient cases, system updates, and evolving performance expectations. In some instances, AI recommendations proved correct in edge cases the radiologists had previously dismissed, prompting further refinement of interrogation methods. In others, the AI introduced novel errors, leading to new doubts and temporary retrenchment of human authority. Each cycle of interaction modified the terms of delegation, recalibrated attribution, and triggered fresh episodes of reassertion and reconfiguration.

Over time, these loops created a path-dependent trajectory that shaped how diagnostic power and responsibility were distributed across humans and machines.

These loops unfolded across different time scales. On a daily basis, radiologists cycled through micro-loops as they interpreted AI outputs and adjusted their judgments in real-time. Across weeks and months, they engaged in meso-loops in which emerging patterns of system performance and professional adaptation influenced how AI tools were socially understood and organizationally governed. These micro- and meso-loops interacted: recurring individual decisions accumulated into collective norms, and those norms fed into broader reconfigurations of diagnostic practice. The result was a nested, recursive system of human-AI interaction where agency was continually redistributed, not through one-time decisions, but through the ongoing interplay of practice, expectation, and organizational learning.

4.2. AI scheduling assistants: managed autonomy through social mediation.

The second study, by Endacott and Leonardi (2022), examines whether professionals across a variety of industries choose to intervene in an AI’s scheduling tasks or allow it to operate autonomously. Through an inductive, comparative field study of two AI scheduling tools—one autonomous and conversationally fluent, and the other less autonomous and lacking conversational fluency—the authors investigated how users (principals) and their communication partners managed and interpreted the AI’s actions. The findings show that principals of the autonomous tool engaged in reactive practices such as diplomacy and staged politeness to manage impressions when the tool acted unpredictably. Communication partners often transferred their impressions of the AI to the principal, particularly when the tool made errors, creating reputational risks.

By contrast, principals of the less autonomous tool retained more control, focusing only on framing the tool as a convenience. This case demonstrates how agency loops can achieve “managed autonomy” where AI systems retain independence within human-monitored social boundaries, revealing how representational agency creates distinct attribution and control challenges that require ongoing social mediation rather than technical solutions.

4.2.1. The logic of delegation.

Delegation emerged from principals’ dual motivation to solve scheduling inefficiencies and demonstrate technological sophistication to clients, with AI systems deliberately configured with human names and conversational capabilities to maximize autonomy while serving both functional and reputational purposes. Principals strategically attributed limited social agency to their AI assistants, designing them to appear human-like to external observers while maintaining internal awareness of their artificial nature. This arrangement allowed principals to experience delegation as maintaining strategic power while outsourcing tactical execution, with the AI positioned as an autonomous social representative that could enhance both efficiency and professional image without threatening core managerial authority.

4.2.2. Attribution dynamics.

Both principals and communication partners began experiencing the AI as human-like assistants with independent social judgment, as the AI’s human names, conversational fluency, and autonomous decision-making facilitated attribution of genuine social agency rather than mere task execution. What began as strategic attribution of limited social agency escalated into genuine attribution of independent social competence, with communication partners treating the AI assistants as autonomous social actors while principals found themselves responding similarly. This attribution process displaced principals’ control over professional interactions through “representational agency”—the AI was actively representing principals in social relationships, acquiring autonomous social authority that extended beyond the technical delegation principals had intended.

4.2.3. Contingency and social breakdown.

The attribution dynamics became unsustainable when AI systems made socially inappropriate scheduling decisions that embarrassed principals or frustrated communication partners, creating contingencies that threatened professional relationships and reputational standing rather than mere technical performance. The successful attribution of social agency became a liability when the AI failed to meet social expectations, as having convinced others that the AI possessed social competence, principals couldn’t easily dismiss failures as technical glitches without undermining the entire delegation premise. This contingency revealed misalignment between attributed social agency and actual social competence, with principals having delegated social representation but retained ultimate responsibility for relationship maintenance.

4.2.4. Reassertion.

Principals developed three sophisticated practices to reassert control while maintaining delegation benefits: interpretation (reframing AI actions to preserve professional identity), diplomacy (actively managing AI-human relationships), and staging politeness (performing appropriate social behavior toward AI in visible contexts). These practices involved sophisticated management of how others perceived AI agency while maintaining principals’ social authority, representing dynamic modulation of agency attribution rather than simple acceptance or rejection of AI capabilities. Through these social mediation strategies, principals developed new forms of power based on mediating between AI systems and human social networks, becoming influential as interpreters and managers of AI social behavior.

4.2.5. Reconfiguration.

The sustained practice of social mediation strategies achieved stable reconfiguration where AI retained attributed social agency while human control was maintained through ongoing social monitoring rather than complete independence or direct control. This reconfigured system involved balanced attribution where AI systems retained social agency for routine interactions while humans maintained superior social intelligence for complex situations. The resulting arrangement allowed principals to benefit from AI scheduling capabilities while developing competence in managing social consequences of AI agency, expanding professional competence to include deploying, monitoring, and socially mediating AI agents while maintaining responsibility for outcomes. The reconfiguration achieved through social mediation strategies did not bring closure.

Instead, it created new starting conditions for the next loop. As principals continued to deploy and monitor their AI assistants, new contingencies emerged. Some clients responded positively to the AI’s social fluency, prompting principals to experiment with expanding the assistant’s autonomy. Others reacted negatively to specific errors, leading principals to increase their monitoring or subtly revise how they framed the AI in future interactions. These shifts altered how future delegation was performed and how attribution was handled when new errors occurred. Over time, principals refined their interpretive, diplomatic, and politeness strategies, learning which worked best in different social and professional contexts.

Each loop became a learning cycle that restructured not only the technical use of the tool but the social positioning of the AI within the principals’ networks of reputation and influence.

These adjustments unfolded across multiple temporal layers. On a day-to-day basis, principals engaged in micro-loops as they responded to minor misunderstandings or improvised new ways to smooth over awkward exchanges. Over weeks and months, they entered meso-loops in which patterns of interaction and reputational outcomes began to accumulate. These patterns influenced how much autonomy they felt comfortable giving the AI and how they trained communication partners to interact with it. Organizational memory about what worked and what backfired began to take hold. The recursive nature of these interactions meant that each new loop carried forward not just individual preferences but a broader social learning process.

Over time, the balance of power between human and AI actors shifted, not because of a one-time decision, but through repeated, socially mediated recalibrations of agency and responsibility.

4.3. Police predictive analytics: curatorial authority through algorithmic substitution.

The third study, conducted by Waardenberg et al. (2022), examined the implementation of a predictive policing system, the “Crime Anticipation System” (CAS), within the Dutch police force. Intelligence officers emerged as “algorithmic brokers,” tasked with bridging the gap between data scientists and police managers by translating opaque algorithmic crime predictions into practical insights. Over time, these brokers enacted three evolving roles: messenger, interpreter, and curator. Initially, they struggled to relay predictions effectively due to the lack of understanding of the algorithm’s logic and police managers’ needs. Through iterative practices, they learned to interpret the data, add qualitative context, and tailor predictions to managerial decisions. However, the black-box nature of the algorithm eventually led brokers to substitute their own judgments for algorithmic outputs.

This case reveals how impassable knowledge boundaries created by machine learning opacity can lead to complete human substitution of AI outputs, resulting in a type of curatorial authority where humans dominate AI systems while maintaining technological legitimacy, representing agency loops pushed to their ultimate extreme.

4.3.1. The logic of delegation.

Delegation emerged from organizational imperatives to modernize policing through “scientific” crime prediction rather than individual professional needs, with police managers, overwhelmed by technical complexity, transferring AI interpretation responsibility to intelligence officers through hierarchical assignment rather than professional negotiation. Initial agency attribution was minimal, with the system understood as a sophisticated analytical tool rather than an independent agent. This hierarchical delegation pattern meant police managers retained formal authority while ceding practical responsibility for AI interpretation, creating power flows that moved downward through organizational structure.

4.3.2. Attribution dynamics.

Intelligence officers gradually experienced the system as an autonomous predictive agent rather than a statistical tool, as the system’s apparent ability to identify crime patterns invisible to human analysis strengthened attribution of genuine analytical intelligence and predictive capability. Officers progressively attributed greater analytical agency to the system, moving from viewing it as a mapping tool to experiencing it as an independent crime analyst with superior pattern recognition abilities. This attribution process displaced officers’ sense of analytical authority, with system outputs that conflicted with human intuitions experienced as decisions by an independent analytical agent, leading officers to question their own analytical capabilities.

4.3.3. Contingency and the opacity crisis.

The attribution dynamics became unsustainable when officers encountered the opacity of machine learning algorithms, creating what Waardenberg et al. term an “impassable knowledge boundary” that differed structurally from the manageable opacity seen in the radiology case or social incompetence in the scheduling case. This contingency revealed the impossibility of maintaining agency attribution to a system whose reasoning remained inaccessible, with officers facing the paradox of attributing analytical agency while being unable to evaluate that agency’s foundations or explain its logic to organizational stakeholders. The opacity crisis created a power vacuum where neither human nor AI authority felt adequate.

4.3.4. Reassertion.

Rather than accepting opacity or abandoning AI entirely, officers developed radical reassertion strategies involving construction of alternative analytical systems that preserved technological legitimacy while restoring human interpretive control, creating explainable tools that provided transparent analytical processes they could fully understand and defend. Through alternative system development, officers reversed the agency attribution pattern, positioning themselves as superior analytical agents capable of creating more reliable and explainable predictions than machine learning systems. This reassertion involved complete reclamation of analytical power, with officers establishing themselves as ultimate authorities on crime pattern analysis while maintaining technological legitimacy through their new systems.

4.3.5. Reconfiguration.

The sustained practice of alternative system development led to a kind of curatorial authority, such that intelligence officers became influential knowledge curators who shaped organizational understanding while maintaining technological legitimacy, achieving stable substitution of AI outputs with their own expert judgments rather than returning to pre-AI practices or accepting algorithmic authority. The reconfigured system involved officers attributing superior analytical agency to themselves while reducing the original AI system to a legitimizing tool rather than an independent analytical authority. Officers achieved enhanced power through their unique position as interpreters and managers of technological resources, becoming authoritative figures who could navigate between technological possibility and organizational reality.

The reconfiguration that granted intelligence officers curatorial authority did not mark the end of the agency loop. Instead, it established new preconditions for future cycles. As officers continued to work with alternative tools, they encountered new challenges in justifying the legitimacy of these homegrown systems to organizational stakeholders. When new predictive failures emerged or pressure mounted to reintroduce advanced machine learning, officers had to reassess how much authority to grant algorithmic outputs and how much to substitute with human judgment. These decisions triggered additional loops in which delegation, attribution, and reassertion played out again. Over time, officers refined the systems they had built, updated their communication strategies with police managers, and recalibrated the balance between algorithmic inputs and human sensemaking.

Each loop created further consolidation of curatorial power, anchoring the officers as the epistemic center of the organization’s data work.

These evolving agency relationships unfolded through nested loops at multiple temporal and organizational levels. On a daily basis, officers evaluated and sometimes quietly discarded CAS predictions while inserting insights from their own tools into reports. At the meso level, teams coordinated around how to structure these reports, justify substitutions, and navigate shifting expectations from

Case comparisons.

police management. At the macro level, these accumulated practices transformed the organization’s broader approach to algorithmic governance. Officers not only reclaimed interpretive control but also reshaped how predictive analytics were institutionally framed and deployed. The recursive nature of these loops reveals how opacity, once experienced as a breakdown, became the foundation for an enduring redistribution of agency and responsibility—one that was neither a rejection of AI nor a blind acceptance, but an institutional path forged through repeated cycles of technical boundary-pushing and social negotiation.

4.4. What agency loops reveal: collective insights from three divergent trajectories.

The comparative analysis of these three cases reveals several insights about how agency loops operate and what determines their evolutionary trajectories. Most importantly, these cases demonstrate that agency loops follow predictable patterns that can be understood and potentially guided toward beneficial outcomes. Table 3 provides a systematic comparison of how the five-phase agency loop framework manifested across three distinct organizational contexts—radiological diagnosis, AI scheduling assistants, and police predictive analytics. While each case involved sophisticated AI systems with genuine analytical capabilities, they produced dramatically different outcomes: distributed diagnostic agency, managed autonomy, and curatorial authority, respectively.

By examining these cases through the same analytical dimensions—from initial delegation logic through final power distribution—the table reveals both the universal structure of agency loops and the context-specific factors that shape their evolutionary trajectories. This comparative view demonstrates that successful navigation of agency loops depends less on the technical sophistication of AI systems than on how humans strategically manage attribution processes, develop appropriate intervention practices, and reconfigure power relationships over time. Four key insights emerge from this cross-case analysis that have important implications for both theory and practice.

4.4.1. The primacy of attribution over technology.

The cases demonstrate that the trajectory of agency loops depends more on attribution processes than on the technical characteristics of AI systems. All three cases involved sophisticated AI systems with genuine analytical capabilities, yet they produced dramatically different outcomes: distributed diagnostic agency in radiology, managed autonomy in scheduling, and curatorial authority in policing. These differences cannot be explained by technological sophistication alone—the radiological AI and predictive policing system were both highly advanced machine learning systems, yet they led to opposite outcomes regarding human agency. The critical factor was how humans navigated the attribution processes that emerged around these technologies.

Radiologists developed interrogation practices that allowed them to work productively with AI opacity, scheduling principals developed social mediation strategies that managed representational agency, and police officers created alternative systems that bypassed opacity entirely. These human responses, rather than technological features, determined whether agency loops evolved toward collaboration, mediation, or substitution.

4.4.2. The contingency structure: multiple types of breakdowns.

The cases reveal that different types of contingencies create different challenges for maintaining agency attribution, requiring different solutions. In radiology, the contingency involved analytical uncertainty—disagreement between human and AI assessments that couldn’t be resolved through normal professional practices. This created a knowledge problem that could be addressed through developing new interpretive capabilities. In scheduling, the contingency involved social breakdown—AI systems making decisions that violated social norms and damaged professional relationships. This created a reputation problem that required ongoing social mediation rather than technical solutions. In policing, the contingency involved an opacity crisis—the inaccessibility of AI reasoning that made attribution impossible to sustain.

This created an epistemological problem that required abandoning the original AI system in favor of alternatives. Understanding these different types of contingencies is crucial because they require different intervention strategies. Analytical uncertainty can be addressed through training and practice development, social breakdown requires social mediation capabilities, and opacity crisis may require technological substitution.

4.4.3. Power enhancement through strategic attribution management.

Perhaps most surprisingly, all three cases demonstrate that humans can actually enhance their power and influence through strategic management of agency attribution rather than simply losing control to AI systems. In each case, the final configuration involved humans gaining new forms of authority that were unavailable before AI implementation. Radiologists gained authority as expert interpreters who could integrate AI insights with clinical knowledge in ways that neither humans nor AI could achieve independently. Scheduling principals gained authority as social mediators who could manage complex human-AI relationships and navigate impression management challenges. Police officers gained authority as knowledge curators who could shape organizational understanding while maintaining technological legitimacy.

This pattern challenges common assumptions about AI implementation leading to human disempowerment. Instead, the cases suggest that thoughtful engagement with agency attribution processes can create opportunities for humans to develop new forms of expertise and influence that enhance rather than diminish their agency.

4.4.4. Temporal dynamics and path dependence.

Finally, the cases reveal that agency loops exhibit strong path dependence—early experiences with agency attribution shape subsequent patterns in ways that can lock human-AI relationships into particular trajectories. In radiology, early experiences of AI identifying missed pathology created conditions for productive interrogation practices. In scheduling, early experiences of social delegation created conditions for ongoing mediation strategies. In policing, early experiences of opacity created conditions for substitution approaches. This path dependence means that the early phases of agency loops are particularly crucial for determining long-term outcomes. Organizations cannot simply deploy AI systems and expect optimal human-AI relationships to emerge naturally—they must actively guide early attribution processes toward beneficial patterns.

5. Conclusion.

This paper has introduced three interconnected concepts that challenge how organizational theory and information systems research approaches human-AI relations: agency loops, power displacement, and the circulation of responsibility. Together, these concepts reveal that the relationship between humans and AI systems is far more dynamic and complex than existing theories suggest, requiring us to move beyond static conceptions of agency that have dominated organizational studies of technology.

Power displacement through agency attribution represents the mechanism through which human-AI relationships are negotiated. When humans attribute agency to AI systems, they experience systematic reduction in their own sense of control and responsibility, even when they retain formal authority. This displacement is an active reconstruction of causal relationships that redistributes power throughout sociotechnical systems—a process that existing IS theories of agency have not adequately captured. Importantly, this displacement is neither permanent nor unidirectional. The concept of agency loops reveals that power displacement operates as part of a cyclical process—delegation, attribution, contingency, reassertion, and reconfiguration—where responsibility and control circulate through human-AI assemblages rather than simply transferring from humans to machines.

This circulation challenges the linear progression from technological determinism to social construction that has characterized much IS theorizing about agency (see for review, Leonardi & Barley, 2010), revealing instead recursive patterns where technology and social practices continuously reshape each other.

The circulation of responsibility emerges as the connecting thread that links power displacement to agency loops. Responsibility flows through attribution processes that constantly redistribute accountability across the assemblage rather than simply residing in human decision-makers or AI systems. This circulation creates opportunities for strategic management of responsibility—humans can temporarily cede control in some domains while maintaining or enhancing it in others, developing new forms of influence through their ability to shape how responsibility flows through the system.

5.1. Agency loops.

The agency loop framework reveals that human-AI relationships unfold through predictable temporal patterns that challenge static conceptions of technology adoption and use that have dominated IS research. Unlike linear models that assume stable divisions of labor between humans and machines, or even sociomaterial approaches that emphasize ongoing entanglement, agency loops demonstrate that these relationships are characterized by cyclical patterns of delegation, attribution, contingency, reassertion, and reconfiguration. This temporal perspective addresses a significant gap in existing IS theories, which have often treated agency as either a property of entities or an emergent feature of practices without adequately explaining how agency relationships evolve over time.

Where studies like Lebovitz et al. (2022) have shown radiologists developing “interrogation practices” and Kellogg et al. (2020) have identified worker resistance strategies, agency loops explain these as predictable phases in recursive cycles that create path-dependent evolution in sociotechnical assemblages.

The cyclical nature of agency loops explains apparent contradictions in how humans experience AI systems that previous IS theories have struggled to address. Humans can simultaneously feel empowered by AI capabilities and disempowered by AI decisions, experience enhanced expertise through AI assistance while feeling reduced control over outcomes, and maintain formal authority while experiencing diminished agency. These seemingly paradoxical experiences reflect different phases of agency loops rather than inherent tensions in human-AI collaboration, providing a more nuanced understanding of human-technology relations than existing frameworks allow.

Understanding agency loops as temporal processes also reveals why the same AI technology can produce dramatically different organizational outcomes—a puzzle that has challenged deterministic and constructivist theories alike. The trajectory of agency loops depends on how humans navigate attribution processes, what organizational practices emerge to manage contingencies, and how responsibility circulation is guided through institutional design. This suggests that successful AI implementation requires attention to temporal dynamics and ongoing process management rather than one-time technical deployment or social negotiation. These divergent trajectories underscore the importance of early and ongoing intervention in the agency loop.

Table 4 outlines concrete intervention points across the five phases, highlighting how organizations can strategically influence power distribution and responsibility circulation by developing targeted capabilities at each.

5.2. Power displacement.

The concept of power displacement challenges how IS research has understood power in human-technology relations. Traditional IS theories have focused on how technology affects power by changing information flows, decision-making processes, or organizational structures. While valuable, these approaches have missed a more basic mechanism: the psychological process through which attributing agency to technology automatically reduces human sense of control and responsibility. Power displacement operates through attribution mechanisms that are psychological rather than structural in nature. When humans experience AI systems as autonomous agents, this attribution automatically reduces their sense of personal agency regardless of their actual control over the system or their position in organizational hierarchies.

This psychological reality means that power displacement occurs even when humans retain complete formal authority over AI systems, revealing the inadequacy of structural approaches to power that have dominated IS research.

This insight extends sociomaterial perspectives that have emphasized the inseparability of human and technological agency in practice. While sociomaterial approaches have shown that agency emerges from human-technology assemblages, they have not explained the specific mechanisms through which this emergence redistributes power. Power displacement reveals that the entanglement of human and technological agency systematically affects human experience of control in ways that can be predicted and potentially managed.

The temporary and strategic nature of power displacement also challenges zero-sum assumptions about human-AI relationships that have influenced both popular discourse and academic research. Rather than humans simply “losing power” to AI systems, power displacement reveals a more complex process where control and responsibility are redistributed in ways that can enhance the capacities of both humans and AI systems. This provides a more optimistic perspective on human-AI collaboration than either technological determinism or resistance-focused social construction approaches would suggest, aligning with recent work showing how humans can gain new forms of authority through strategic engagement with AI systems.

5.3. Circulation of responsibility.

The circulation of responsibility represents perhaps the most profound challenge to existing IS frameworks for understanding accountability in technological organizations. Traditional IS approaches to responsibility have assumed that accountability can be clearly allocated to specific decision-makers through appropriate design choices, implementation processes, or governance mechanisms. The circulation perspective reveals accountability as an emergent property of attribution processes rather than a fixed assignment, requiring reconceptualization of how responsibility operates in sociotechnical systems.

Responsibility circulation operates through the same attribution mechanisms that drive power displacement, but with distinct organizational and ethical implications that existing IS theories have not adequately addressed. As humans attribute agency to AI systems, they simultaneously transfer responsibility for outcomes to those systems, even when they retain legal and formal accountability. This creates systematic misalignments between felt responsibility and formal accountability that can have profound consequences for decision-making and organizational learning—consequences that traditional IS approaches to ethics and responsibility have not anticipated.

The dynamic nature of responsibility circulation means that accountability relationships are constantly being renegotiated through ongoing attribution processes rather than being established once through design or implementation decisions. Responsibility can be concentrated in human agents during some phases of agency loops, distributed across human-AI assemblages during others, and strategically shifted between participants depending on outcomes and circumstances. This fluidity creates both opportunities and dangers for maintaining meaningful accountability that go beyond the static frameworks that have dominated IS research on technology ethics.

Understanding responsibility circulation also reveals why traditional governance mechanisms may be inadequate for managing AI systems, extending critiques of technological governance that have emerged in technology research. Approaches that attempt to fix responsibility in particular locations may be undermined by attribution processes that continuously redistribute accountability. More effective governance may require developing institutional mechanisms that can work with rather than against responsibility circulation, creating stable frameworks for accountability within dynamic attribution processes.

Understanding agency loops, power displacement, and responsibility circulation opens new research agendas that both extend and redirect scholarship on technology and organizing. We need empirical studies that trace how these processes unfold across different organizational contexts, professional cultures, and technological configurations, building on recent work examining AI adoption and human-AI collaboration. We need design research that develops interventions for guiding agency loops toward beneficial configurations, extending IS work on participatory design and user-centered development to include attention to attribution processes and their effects on power and responsibility distribution.

Most importantly, we need theoretical work that develops new conceptual frameworks for understanding agency, power, and responsibility in systems characterized by circulation, emergence, and co-constitution rather than allocation, possession, and interaction.

It is tempting to say that as homo agenticus, we seek to attribute agency. But as I have argued, it is more appropriate to say that as homo agenticus we create agency and displace power through attribution. The framework of agency loops demonstrates that our relationship with AI systems is not a zero-sum battle for control but an ongoing negotiation over the circulation of responsibility and power—a negotiation we can learn to conduct with intention and skill. The empirical evidence from radiology, scheduling, and policing shows that when humans understand attribution as an active process rather than a passive recognition, they can guide agency loops toward configurations that enhance human authority while leveraging AI capabilities.

The organizations and individuals who will thrive are those who recognize that agency attribution is perhaps the most consequential skill of our technological age. In mastering this skill, we do not simply adapt to AI—we actively construct the conditions under which human agency can flourish alongside artificial intelligence, ensuring that our relentless pursuit of control serves not just our survival but our continued evolution as fundamentally agentic beings.

CRediT authorship contribution statement

Paul M. Leonardi: Writing – review & editing, Writing – original draft, Supervision, Project administration, Conceptualization.

Acknowledgements.

This paper was supported by a grant from the National Science Foundation (IIS-2211942). The author wishes to thank Michael Barrett for helpful comments that improved the manuscript in dramatic ways.