The rapid integration of artificial intelligence into daily life and professional workflows has introduced a critical new managerial challenge: agency decay. This phenomenon, identified and elucidated by Dr. Cornelia C. Walther, a visiting scholar at Wharton and director of global alliance POZE, describes the gradual erosion of a person’s fundamental capacities – their ability and willingness to observe carefully, think independently, choose deliberately, and act responsibly. Dr. Walther, a humanitarian practitioner with over two decades of experience at the United Nations, whose current research focuses on leveraging AI for social good, posits that this decay is not an inherent flaw of AI, but rather a consequence of convenience becoming the default setting for cognition. What began as a tool to assist human thought now risks thinking around us, then for us, ultimately diminishing our cognitive faculties without conscious notice.
The trajectory of AI’s ascent has been remarkably swift. When OpenAI unveiled ChatGPT on November 30, 2022, it arrived as a research preview, captivating millions with its conversational prowess, ability to draft fluent text, and capacity to answer follow-up questions. The initial public reception was a blend of experimentation, awe, apprehension, and widespread sharing of screenshots. It was a novelty, a curiosity. Yet, in just over three and a half years, this novelty has metamorphosed into indispensable infrastructure, embedding itself deeply within search engines, office software suites, customer-service platforms, coding environments, writing tools, and complex corporate workflows. This pervasive integration, while offering unprecedented efficiency, has simultaneously laid the groundwork for agency decay to manifest across individuals and organizations.
Defining the Erosion: What is Agency Decay?
Agency decay fundamentally represents a subtle, insidious shift in human-machine interaction. It’s a process where the convenience offered by AI bypasses the active engagement of human intellect. Instead of merely augmenting human capabilities, AI begins to supplant them, leading to a diminished capacity for critical observation, independent thought, deliberate decision-making, and accountable action. Dr. Walther highlights that the danger isn’t in AI being inherently harmful, but in how its seamless assistance can lull humans into cognitive complacency. The tool that initially helped us think can subtly begin to think around us, then for us, and eventually, we may find ourselves operating without fully realizing the extent to which our own cognitive muscles have atrophied. This erosion poses significant questions for individual development, organizational resilience, and societal progress, prompting an urgent need to understand its mechanisms and develop countermeasures.
The Four Stages of Human-AI Interaction: A Chronology of Decay
Dr. Walther’s research outlines a "Scale of Agency Decay" comprising four distinct stages: experimenting, integrating, relying, and depending. Each stage marks a deepening entanglement with AI, often accelerating faster than the corresponding governance, training, and cultural adaptations can keep pace. This progression is not exclusive to specific individuals or organizations; its reach is universal, affecting anyone engaging with generative AI.
Stage 1: Experimentation and Initial Gains
The journey into AI engagement typically commences with experimentation, a phase characterized by playful exploration and practical application. Users initially leverage tools like ChatGPT for discrete, often mundane, tasks: drafting emails, summarizing lengthy articles, generating creative titles, translating text, explaining complex concepts, debugging code snippets, or producing preliminary drafts of documents. The immediate business appeal of these applications is undeniable, promising significant boosts in productivity and efficiency.
Empirical evidence from studies has bolstered this initial excitement. A prominent study published in The Quarterly Journal of Economics found that providing customer-support agents with access to a generative AI assistant led to an average productivity increase of 15%. Notably, the most substantial gains were observed among less experienced and lower-skilled workers, suggesting AI’s potential to democratize high-quality output. A junior employee could articulate thoughts with greater clarity, a non-native speaker could write more fluently, and an overwhelmed manager could transform disparate notes into coherent prose, all thanks to AI’s assistance. At this initial stage, the human operator largely remains in control, using AI as an expansive tool to extend their reach. However, a subtle yet critical risk emerges here: the early, tangible gains can foster a broader, potentially dangerous, assumption that faster output invariably equates to better work. This assumption proves fragile and potentially misleading when tasks demand nuanced judgment, deep contextual understanding, ethical deliberation, or direct accountability.
Stage 2: Integration and the Shifting Locus of Judgment
The second phase, integration, signifies a pivotal shift where generative AI transitions from an occasional, standalone experiment to an embedded layer within everyday workflows. AI ceases to be a separate window opened sporadically and becomes an intrinsic component of routine tasks and decision-making processes. This deep integration fundamentally alters the human-AI relationship.
A growing body of research underscores this transformation, illustrating that integration is not merely about increased frequency of use but about weaving AI into core operational frameworks. Workers increasingly perceive AI systems not just as tools, but as active collaborators, directly incorporating AI-generated outputs into their drafts, analyses, and recommendations. In enterprise environments, AI is now a first point of contact in areas such as customer service, software development, legal research, and financial analysis, rather than a secondary aid. This shift is accompanied by a concerning rise in trust; experiments have shown that users frequently defer to AI suggestions, even when they contradict their own judgment, particularly if the system has performed well in prior tasks. An intriguing related finding suggests a correlation between lower trust in human counterparts and higher trust in AI systems, pointing to deeper psychological underpinnings.
It is at this stage that the transition from cognitive offloading to belief offloading truly begins. Cognitive offloading, which involves delegating memory or routine tasks to AI, might seem benign. However, as integration deepens, individuals start outsourcing not just effort but also critical evaluation. Research on automation bias and "algorithm appreciation" reveals a propensity for individuals to accept AI-generated answers as correct, even in the face of contradictory evidence. The guiding question subtly shifts from "Is this task suitable for AI?" to "What did the AI say?" This seemingly minor change is profound, as it repositions the locus of judgment from human intellect to algorithmic output.
The real-world consequences of this shift are becoming increasingly apparent and difficult to ignore. In the notable case of Mata v. Avianca Inc., lawyers faced severe sanctions for submitting legal filings that contained fabricated case citations generated by ChatGPT, underscoring a profound failure in human verification. Similarly, in Canada, the Moffatt v. Air Canada decision held the airline responsible after its chatbot provided a customer with inaccurate information regarding bereavement fares, leading to financial penalties. These incidents are not isolated anomalies but illustrative examples of a broader pattern where integrated AI systems are trusted beyond their actual reliability, and where the critical step of human verification is smoothly, often unconsciously, displaced by uncritical acceptance. The legal and ethical implications of such failures demand a re-evaluation of how AI is integrated and overseen in professional practices.
Stage 3: Reliance and the Cost of Unexamined Trust
Reliance marks the stage where AI becomes the default starting point for many tasks. While the human user still typically reviews the AI’s output, the initial generative impulse has fundamentally shifted from human thought to machine production. This stage is characterized by AI becoming the first port of call, setting the agenda for subsequent human interaction.
Adoption data robustly demonstrates the rapid spread of this reliance. Studies by the Pew Research Center indicated that by 2025, 34% of U.S. adults had used ChatGPT, a figure approximately double that of 2023. Among employed adults, a significant 28% reported using AI for work-related tasks. Further reinforcing this trend, Gallup’s findings from the third quarter of 2025 revealed that 45% of U.S. employees utilized AI at work at least a few times annually, with 23% engaging with it several times a week, and 10% daily. These statistics highlight a widespread, ingrained presence of AI in the modern workplace.
Crucially, the problem here is not merely the frequency of AI usage itself, but rather the unexamined reliance that often accompanies it. A 2025 Microsoft Research study, which surveyed 319 knowledge workers and collected 936 examples of generative AI use in professional settings, yielded a disconcerting insight: higher confidence in AI was directly associated with less critical thinking, while conversely, higher confidence in one’s own task ability correlated with more critical thinking. The uncomfortable lesson is clear: the more individuals trust the AI tool, the more deliberate and vigilant they must be in actively protecting and cultivating their habit of questioning its outputs.
The costs of this unexamined reliance are increasingly manifesting within organizations. BetterUp Labs, in collaboration with the Stanford Social Media Lab, has coined the term "workslop" to describe polished-looking, AI-generated work that, despite its superficial presentation, lacks genuine substance and effectively shifts the burden of critical thought onto colleagues. Their survey of 1,150 full-time U.S. desk workers revealed that 40% had encountered "workslop" in the preceding month. Each incident reportedly required approximately two hours to resolve, accumulating an estimated monthly cost of $186 per employee. This phenomenon is a tangible illustration of agency decay operating at a team level: one individual saves time by outsourcing cognitive effort, while another subsequently loses time attempting to reconstruct the missing intellectual substance. This not only impacts productivity but can also foster resentment and undermine collaborative trust within teams.
Stage 4: Dependence and Broadening Implications
The final stage, dependence, emerges when individuals or teams find themselves struggling significantly to perform, decide, or verify tasks without the direct assistance of AI. While this stage is still considered reversible, the warning signs are undeniably visible and point to a profound alteration in human cognitive processes and organizational capabilities.
A compelling 2025 study from the MIT Media Lab, which utilized electroencephalography (EEG) to measure brain activity during LLM-assisted essay writing, provided concerning insights. Participants who used large language models (LLMs) to generate essays exhibited weaker neural connectivity compared to those who wrote without such tools. Furthermore, these LLM-users reported lower ownership of their essays and experienced greater difficulty in quoting their own work. While a subsequent scholarly comment raised valid concerns about the study’s sample size, reproducibility, and specific aspects of the EEG analysis, the core question it poses remains profoundly relevant for educational institutions, employers, and leaders worldwide: what are the long-term implications for skill formation and the development of deep cognitive abilities when the crucial process of generating a first draft is increasingly outsourced? The potential for a generation to grow up with diminished capacity for original thought and expression is a serious consideration.
At the enterprise level, dependence often manifests as an illusion of scale without corresponding depth. McKinsey’s 2026 global survey on AI reported that a staggering 88% of organizations were utilizing AI in at least one business function. However, the survey also highlighted that many of these organizations were still in the nascent stages of scaling AI effectively and capturing genuine enterprise-level value. Significantly, the same report detailed negative consequences arising from AI use, with inaccuracy being one of the leading issues, underscoring the risks of uncritical dependence.
Beyond individual and organizational impacts, dependence also carries a critical planetary dimension, often overlooked due to AI’s seemingly immaterial nature at the point of use. The reality is far from immaterial. The International Energy Agency (IEA) projects that the electricity demand from data centers, driven significantly by AI workloads, could more than double by 2030, potentially reaching around 945 terawatt-hours. This surge in energy consumption has profound environmental implications. Furthermore, the U.S. Government Accountability Office (GAO), in its report on generative AI’s environmental and human effects, explicitly notes that generative AI systems consume substantial energy and water resources. The report critically points out that companies often fail to provide sufficient detail regarding these impacts, hindering proper accountability. Alarmingly, as of now, there are no robust legal obligations compelling these companies to be transparent about their environmental footprint. This broader context underscores that managing agency amid AI now extends beyond merely protecting human attention and judgment; it necessitates a holistic approach that links personal convenience, institutional design, and the often-hidden planetary costs.
Reversing the Trend: Cultivating Co-Agency with the A-Frame
Agency decay, while tangible and pervasive, is not an irreversible condition. The imperative now is to transition from passive adoption of AI to a deliberate cultivation of co-agency – a synergistic relationship where humans and AI collaborate in ways that actively strengthen human judgment, skill, responsibility, and a broader care for the complex systems upon which life depends. Dr. Walther proposes the A-Frame as a practical starting point for this reversal, comprising Awareness, Appreciation, Acceptance, and Accountability.
Awareness: Noticing the Handover
The journey to reasserting agency begins with heightened awareness, specifically, noticing the precise moment of delegation to AI. Before engaging an AI tool, individuals must pause and ask themselves: "What exactly am I handing over?" This introspection should extend to identifying whether the task involves drafting, complex reasoning, nuanced judgment, empathetic communication, rigorous verification, or ultimate accountability. For teams and organizations, this translates into a systematic mapping of AI’s entry points within workflows. Critical identification of areas where human review, intervention, or ultimate decision-making is not merely beneficial but absolutely essential is paramount. This proactive mapping ensures that AI is used as a tool for enhancement, not abdication.
Appreciation: Valuing Human Capacity
Appreciation means strategically leveraging AI where it genuinely expands human capacity, rather than replacing it. AI excels at tasks such as summarizing vast amounts of information, translating languages, comparing complex datasets, simulating various scenarios, and revealing cognitive blind spots that humans might otherwise miss. The human role, however, must be rigorously preserved for the higher-order cognitive functions: framing the problem accurately, setting the underlying values and ethical boundaries, deeply understanding the contextual nuances, and ultimately defining what "good" or "successful" truly means in any given situation. By consciously distinguishing between tasks where AI provides leverage and those requiring uniquely human insight, we can ensure that human intellect remains the driving force.
Acceptance: Confronting AI’s Limits
Acceptance demands a profound honesty about AI’s inherent limitations. AI systems, despite their sophisticated outputs, can project an aura of certainty while being fundamentally incorrect or generating "hallucinations." Similarly, individuals can feel highly productive while simultaneously becoming less skilled due to over-reliance. Organizations, too, can rapidly scale AI tools without adequately scaling the human judgment and critical thinking required to oversee them. Therefore, every AI-enabled workflow must be designed to incorporate robust verification standards, clear escalation paths for problematic outputs, and crucially, designated moments where individuals are required to practice the underlying skills without automation. This deliberate practice helps maintain human proficiency and prevents skill atrophy.
Accountability: Embedding Responsible Governance
Accountability transforms individual agency into systemic governance. On a personal level, fostering a "human first draft" habit for any important thinking or creative work ensures that original thought is exercised before AI assistance is sought. Within teams and organizations, this translates into concrete policies: requiring rigorous source checks for AI-generated content, transparently disclosing meaningful AI use in reports or decisions, assigning clear human ownership for all AI-supported decisions, and expanding procurement and strategic planning to include the energy, water, and infrastructure impacts of AI technologies. This holistic approach ensures that the benefits of AI are weighed against its broader costs and responsibilities.
The forthcoming phase of AI development and adoption will undoubtedly favor organizations that proactively recognize and treat human agency as a paramount business asset. Autonomy, discerning judgment, earned trust, and a conscious commitment to planetary responsibility are not merely "soft" concerns; they represent the foundational infrastructure for sustainable performance and long-term viability. Companies that strategically protect and cultivate these human attributes will be uniquely positioned to deploy AI with greater discipline, imaginative foresight, and an undeniable legitimacy, ultimately leading to more robust, ethical, and sustainable innovation.
About the Author
