Roger Hunt

The IBM Finding and What It Actually Tells Us

IBM's recent security research found shadow AI present in 43% of incident reports, a figure now being used by insurers to set exclusion clauses and by EU regulators to impose new board-level disclosure requirements. The business press has framed this as a cybersecurity story. That framing is wrong, or at least incomplete. Shadow AI is not primarily a threat vector. It is evidence of a coordination failure between what organizations formally sanction and what workers actually need to do their jobs effectively.

The distinction matters because the policy response changes entirely depending on which diagnosis you accept. A security framing produces access controls, monitoring software, and insurance riders. A coordination framing produces a different question: why are workers routing around sanctioned tools in the first place, and what does that tell us about the gap between organizational AI governance and actual task environments?

Why Workers Adopt Shadow AI: The Competence Inversion Problem

Classical organizational theory assumes that governance structures are designed for workers who already understand the tools they are being given. The hierarchy provides rules; workers apply them. This assumption breaks down entirely in algorithmically-mediated environments, and shadow AI is one of the clearest demonstrations of that breakdown in recent corporate news.

When workers adopt unsanctioned AI tools, they are not primarily acting recklessly. They are solving a competence problem that the organization has not solved for them. Sanctioned enterprise AI tools frequently arrive with procedural training: here is how to open the interface, here is the prompt structure, here is the approved use case. What that training does not provide is any structural understanding of how the underlying model behaves, where it fails, or how to adapt when the approved procedure produces bad output. Workers who develop that structural understanding independently - often through unsanctioned experimentation with consumer tools - become more capable. The organization then classifies their superior performance as a security risk (Kellogg, Valentine, & Christin, 2020).

This is the competence inversion: the workers with the most accurate mental models of AI tools are frequently the ones operating outside sanctioned boundaries, because sanctioned training did not produce those models.

The Awareness-Capability Gap at the Board Level

The EU disclosure requirements emerging from this insurance data create an interesting pressure on boards. Directors are now required to acknowledge AI-related risks in governance documentation. But acknowledgment is not understanding. This is precisely the awareness-capability gap that algorithmic literacy research has documented at the worker level, now appearing at the governance level (Gagrain, Naab, & Grub, 2024).

A board that discloses "shadow AI risk" without any structural understanding of why shadow AI adoption occurs is performing governance, not exercising it. The disclosure requirement makes the awareness gap visible and official. It does not close it. Boards that treat this as a compliance checkbox will find that their disclosed risk models are wrong in the same way that workers' folk theories about algorithms are wrong: they identify that a system exists without understanding how it actually behaves (Hancock, Naaman, & Levy, 2020).

What Cannot Be Priced Cannot Be Governed

The headline framing - that companies cannot price the shadow AI risk they cannot see - is accurate but points in the wrong direction. The invisibility is not a detection problem. Organizations cannot see shadow AI risk clearly because they have no adequate schema for understanding why AI adoption diverges from sanctioned pathways. Insurance exclusions price the outcome of that ignorance. They do not address the ignorance itself.

Hatano and Inagaki (1986) distinguished routine expertise from adaptive expertise: routine expertise follows procedures effectively within known parameters, while adaptive expertise applies underlying principles to novel configurations. Corporate AI governance, as it currently exists in most organizations, is a routine expertise operation trying to manage an adaptive expertise problem. The EU disclosure mandates and insurer exclusion clauses are procedural responses. They will produce compliance behavior without producing the structural understanding that would actually reduce the underlying risk.

The Organizational Implication

The IBM data, read carefully, suggests that the 43% figure is not a measure of employee recklessness. It is a measure of the distance between how organizations have chosen to govern AI adoption and how workers have actually experienced the task demands that AI tools address. That distance is a coordination failure, and it will not be closed by tighter access controls or better disclosure language in annual reports. It requires organizations to reckon with the fact that sanctioned training has been producing awareness without capability, and that workers have been filling that gap on their own, outside organizational boundaries, in ways that are now showing up in insurance incident reports (Rahman, 2021).

The security framing gives organizations something to monitor and exclude. The coordination framing gives organizations something harder: a structural problem in how they develop and deploy AI competence at scale.

References

Gagrain, A., Naab, T. K., & Grub, J. (2024). Algorithmic media use and algorithm literacy. New Media & Society.

Hancock, J. T., Naaman, M., & Levy, K. (2020). AI-mediated communication: Definition, research agenda, and ethical considerations. Journal of Computer-Mediated Communication, 25(1), 89-100.

Hatano, G., & Inagaki, K. (1986). Two courses of expertise. In H. Stevenson, H. Azuma, & K. Hakuta (Eds.), Child development and education in Japan (pp. 262-272). Freeman.

Kellogg, K. C., Valentine, M. A., & Christin, A. (2020). Algorithms at work: The new contested terrain of control. Academy of Management Annals, 14(1), 366-410.

Rahman, H. A. (2021). The invisible cage: Workers' reactivity to opaque algorithmic evaluations. Administrative Science Quarterly, 66(4), 945-988.

A New Organizational Unit That Admits Something Uncomfortable

EY announced this week that it is creating what it calls an "AI Value Realization Office," a dedicated organizational unit whose explicit mandate is to ensure that EY's substantial AI spending actually produces measurable returns. The framing is notable: EY's leadership concluded that no existing department was capable of owning this function. Finance could not own it. IT could not own it. Strategy could not own it. The firm's response was to create an entirely new structural entity. This is not a rebranding exercise. It is an organizational admission that AI governance sits outside the competence boundary of every current function in the firm.

That admission deserves more theoretical scrutiny than it has received in the coverage so far.

When Existing Schema Structures Fail

Classical organizational theory treats the division of labor as a solution to competence allocation problems. You assign tasks to the unit best equipped to handle them. The problem EY is diagnosing is that no unit has the schema to handle AI governance - not because of resource constraints, but because the structural features of AI-mediated workflows do not map onto the categorical systems that existing departments were built around. Finance reads cost-benefit ratios. IT reads system specifications. Neither reads the feedback loops between model behavior, workflow adaptation, and output quality that determine whether AI spending generates value.

This is precisely what Kellogg, Valentine, and Christin (2020) identified in their review of algorithmic work: organizations routinely underestimate how much algorithmic systems invert the standard assumption that competence precedes deployment. Firms invest in AI systems before developing the organizational capacity to interpret what those systems are doing or why outputs vary. EY's new office is, functionally, an attempt to build that interpretive capacity retroactively, after the investment has already been made.

The Structural Problem Is Not Coordination, It Is Schema

The coverage of EY's announcement frames this as a coordination problem. Who is responsible for AI ROI? The new office will coordinate across functions to answer that question. But coordination assumes that the relevant information exists somewhere in the organization and simply needs to be routed correctly. The more difficult possibility is that the information does not yet exist in a usable form inside EY, and that no amount of coordination will produce it until someone builds an accurate structural model of how their AI deployments actually work.

Gentner's (1983) structure-mapping theory is useful here. Analogical reasoning transfers well when the relational structure of a source domain maps cleanly onto the target domain. EY's existing departments are trying to apply schema from domains - cost accounting, system administration, project management - whose relational structure does not transfer to AI value assessment. The variables that predict AI value (model behavior under distribution shift, user adaptation patterns, feedback loop dynamics) are structurally different from the variables those departments were designed to track. The new office, if it succeeds, will have to build a genuinely new schema rather than adapt an old one.

What Wall Street's AI Investments Suggest by Comparison

Recent reporting on Wall Street banks including JPMorgan, Citi, and Goldman Sachs shows a parallel pattern. These institutions are investing billions in AI while simultaneously acknowledging that workflows and organizational culture are being reshaped in ways that are not yet fully legible. The governance question - who evaluates whether this is working and by what criteria - remains largely unresolved across the industry. EY's move to create a dedicated office is one response to that unresolved question. It is structurally more honest than the alternative, which is to pretend that existing reporting lines are adequate.

Hatano and Inagaki (1986) drew a distinction between routine expertise and adaptive expertise that applies directly here. Routine expertise performs well on familiar problem types. Adaptive expertise generates solutions when the problem type itself is novel. Every large organization deploying AI at scale right now is facing a novel problem type: how do you evaluate a system whose outputs are probabilistic, whose behavior shifts as users adapt to it, and whose value is distributed across workflows in ways that resist clean attribution? Routine financial and operational expertise will not answer that question. EY is, perhaps unintentionally, institutionalizing the recognition that adaptive expertise is required.

The Deeper Implication

What EY's decision reveals is that the organizational theory of AI governance is still being written in real time. The firm is not implementing a known solution. It is creating a structural experiment. Whether the AI Value Realization Office develops genuine schema for understanding AI-generated value, or whether it becomes a rebranded cost-tracking function with new terminology, will depend on whether its staff can build structural understanding that does not yet exist in standard management education. That is a harder problem than the announcement suggests, and it is the problem that organizational theory should be focusing on right now.

The Study and What It Actually Shows

A study reported this week found that one in four business executives cannot explain the outputs their AI systems produce, while a majority of these same leaders rely on AI for consequential financial tasks including expense management and payments. This is not a finding about junior employees experimenting with new tools. This is a finding about decision-makers who have formally adopted AI into core operational workflows while lacking the structural understanding to evaluate what that AI is doing. The gap between adoption and comprehension here is not incidental. It is, I would argue, the defining organizational problem of this particular moment in AI deployment.

This Is Not an Awareness Problem

The instinctive response to this finding will be to call for more AI literacy training, and that response will largely miss the point. The executives in this study are almost certainly aware that AI systems exist, aware that those systems can produce errors, and aware that responsible oversight is expected of them. Awareness was never the constraint. The constraint is the absence of what I would call structural schema: an accurate internal model of how the system produces outputs, what kinds of errors it is prone to, and under what conditions its outputs should be treated with suspicion. Kellogg, Valentine, and Christin (2020) identified precisely this distinction in their work on algorithmic management. Workers in algorithmically governed environments regularly develop surface awareness of the systems controlling their work without developing the deeper structural understanding needed to respond adaptively when those systems behave unexpectedly.

Folk Theories Are Not Schemas

The Algorithmic Literacy Coordination framework I am developing in my dissertation draws on Gentner's (1983) structure-mapping theory to distinguish between folk theories and structural schemas. A folk theory is an individual's working impression of how a system operates, assembled from partial observations and inference. A schema is an accurate representation of the system's underlying relational structure. The 25% figure from this study suggests something important: senior executives, even those who have formally integrated AI into their workflows, are operating on folk theories. They have built intuitive impressions of what the AI generally does, impressions sufficient for routine use, but insufficient for the kind of adaptive judgment required when outputs are anomalous, stakes are high, or the system is operating outside its training distribution.

This distinction matters because folk theories are self-confirming under normal conditions. When AI outputs are reasonable, there is no signal that the underlying mental model is inadequate. The deficit only surfaces under pressure, which is precisely the wrong moment to discover it.

The Organizational Theory Angle

There is an organizational structure problem layered on top of the cognitive one. Rahman (2021) described the "invisible cage" dynamic in platform work, where workers are governed by algorithmic systems whose logic they cannot inspect and whose criteria they cannot directly observe. That analysis was developed in the context of gig workers, but the structure applies here with uncomfortable precision. When a senior executive approves a financial output they cannot explain, they are operating inside a governance structure they do not understand. The formal authority runs upward through the organizational hierarchy. The actual epistemic authority runs through the model. These two authority structures are misaligned, and that misalignment is invisible as long as the outputs appear reasonable.

Why Procedural Training Will Not Fix This

The training response that most organizations will deploy in response to findings like this one is procedural: checklists for reviewing AI outputs, approval workflows, flagging criteria. This is the equivalent of teaching platform workers which buttons to press rather than teaching them why the platform responds the way it does. Hatano and Inagaki (1986) distinguished between routine expertise, competence in executing known procedures, and adaptive expertise, the ability to respond effectively in novel and uncertain conditions. Procedural AI governance training produces routine expertise. The conditions under which that expertise is most needed, namely anomalous or high-stakes outputs, are precisely the conditions under which routine expertise fails.

What This Finding Demands

The study's finding is not primarily a training story. It is a governance story about how organizations have structurally separated the authority to adopt AI from the competence to oversee it. Closing that gap requires schema induction at the executive level: not courses about what AI is, but structured learning experiences that develop accurate internal models of how specific systems produce specific outputs. Gagrain, Naab, and Grub (2024) found that algorithmic literacy, properly understood, requires engagement with structural features of systems rather than surface familiarity. One in four executives cannot explain their AI's outputs. That number should be read as a baseline measurement, not a headline.

The Specific Event

Two stories broke in close succession this month that deserve to be read together. Business Insider reported that OpenAI has now lost twelve executives in 2026 alone, including Brad Lightcap and Fidji Simo. Separately, reporting on what is being called OpenAI's "rogue agent hack" described it as a watershed moment for AI safety - one that also surfaced internal questions about the organizational culture that produced the conditions for it. These are not separate stories. They are the same story told from two angles.

Turnover as Organizational Signal

Executive turnover at this scale is rarely random. In organizational theory, voluntary departure patterns function as revealed preferences. When twelve senior figures exit a single organization within a single year, the standard human resources framing - "pursuing new opportunities" - loses explanatory power. What remains is a structural question: what organizational conditions make sustained leadership participation untenable? The answer, in OpenAI's case, appears to involve a collision between safety culture norms and the commercial velocity the organization has committed to. The rogue agent incident did not create this tension. It made it visible.

Rahman's (2021) concept of the invisible cage is useful here. Rahman describes how algorithmic control systems constrain worker behavior in ways that are real but difficult to articulate or contest. OpenAI's departing executives are not platform workers in the gig economy sense, but the structural dynamic is analogous. When the rules governing acceptable professional conduct inside an organization are opaque, contested, or shifting, high-competence individuals with outside options leave. Those without outside options stay and adapt. The result is adverse selection at the leadership level.

Safety Culture as a Coordination Problem

The internal questions surfaced by the rogue agent hack are, at their core, coordination failures. Safety culture is not simply a set of policies. It is a shared schema about what risks are worth taking, how uncertainty should be handled, and who has standing to slow down a release. When that schema is absent or inconsistently held across an organization, individual actors substitute their own folk theories - local, impressionistic, and unverifiable by others (Kellogg, Valentine, & Christin, 2020).

This is the precise distinction my ALC framework draws between folk theories and structural schemas. A safety culture built on folk theories produces the appearance of coordination without the substance. Engineers develop individual impressions about what is safe enough. Product managers develop different impressions. Leadership develops a third set. No one is lying. Everyone is operating from incomplete structural understanding of the same constraint space. The rogue agent incident is what happens when those misaligned impressions converge on a decision point.

The Adaptive Expertise Deficit at the Organizational Level

Hatano and Inagaki (1986) distinguish between routine expertise - the capacity to execute known procedures reliably - and adaptive expertise - the capacity to respond effectively to novel situations. Most organizational safety training produces routine expertise. It generates checklists, review processes, and sign-off chains that function well when the threat landscape is familiar. Generative AI development does not present a familiar threat landscape. It presents a genuinely novel one.

OpenAI's reported internal culture problems are, in part, a consequence of deploying routine safety expertise against adaptive safety problems. The procedures exist. The capability to reason about novel failure modes in real time apparently did not exist consistently across the organization. That gap - between knowing the safety review process and understanding why it exists - is structurally identical to the awareness-capability gap I study in platform workers. Knowing a constraint exists is not the same as knowing how to respond when it is violated in a way no prior procedure anticipated.

What the Exodus Actually Costs

The practical consequence of losing twelve executives in a year is not primarily the loss of individual talent, though that matters. The primary cost is schema erosion. Organizational schemas - shared structural understandings of how decisions get made, what values take priority under pressure, and how disagreement is handled - are carried by people, not documents. When the people who constructed or contested those schemas leave, the schemas do not remain intact in a policy manual. They degrade.

Schor et al. (2020) describe how platform dependence creates structural precarity for workers. The inverse dynamic is less studied: when organizations become dependent on a small number of senior individuals to carry and transmit structural knowledge, voluntary departure produces institutional precarity. OpenAI is currently experiencing both the safety consequences of that precarity and the market visibility of it. Whether the organization treats this as a signal worth decoding or as a personnel problem to be managed will determine what comes next.

References

Hatano, G., & Inagaki, K. (1986). Two courses of expertise. In H. Stevenson, H. Azuma, & K. Hakuta (Eds.), Child development and education in Japan (pp. 262-272). Freeman.

Kellogg, K. C., Valentine, M. A., & Christin, A. (2020). Algorithms at work: The new contested terrain of control. Academy of Management Annals, 14(1), 366-410.

Rahman, H. A. (2021). The invisible cage: Workers' reactivity to opaque algorithmic evaluations. Administrative Science Quarterly, 66(4), 945-988.

Schor, J. B., Attwood-Charles, W., Cansoy, M., Ladegaard, I., & Wengronowitz, R. (2020). Dependence and precarity in the platform economy. Theory and Society, 49(5-6), 833-861.

The Specific Event

A notable shift is underway in corporate HR practice. Companies are actively revising performance review frameworks to include explicit evaluation criteria around employees' AI use. This is not a fringe experiment at a handful of tech firms. According to recent reporting, organizations across sectors are scrambling to operationalize what "good AI use" looks like in a formal appraisal context. The scramble itself is the signal worth analyzing. When organizations cannot define the competence they are trying to measure, the measurement instrument reveals more about institutional confusion than about employee performance.

The Measurement Problem Is a Schema Problem

The core difficulty is that most organizations are attempting to evaluate AI competence through behavioral proxies: Did the employee use the tool? How frequently? Did output volume increase? These are topographic measures. They describe the surface of the behavior without capturing the structural logic underneath it. This distinction, between topology and topography, is one I have written about in relation to platform work, but it applies here with equal force. An employee who uses an AI assistant daily to generate first drafts of reports is doing something categorically different from an employee who understands why a particular prompt structure reduces hallucination rates in a specific model architecture. Both show up identically in a usage log.

This maps directly onto what Hatano and Inagaki (1986) identified as the distinction between routine and adaptive expertise. Routine expertise produces consistent performance under stable conditions. Adaptive expertise produces diagnostic flexibility when conditions change. Performance reviews built around frequency metrics are, at best, measuring routine competence. They are calibrated to reward employees who have learned the topography of AI tools, not those who have developed transferable structural understanding of how these systems behave across contexts.

Why Organizations Cannot Easily Fix This

The reason organizations are struggling is not primarily technical. It is theoretical. Most HR frameworks were built on an implicit assumption that competence is largely stable and observable prior to task performance. You hire for demonstrated skills, then deploy those skills. Platforms and AI systems invert this. As I argue in my dissertation research on the Algorithmic Literacy Coordination framework, these environments generate competence endogenously through participation. You cannot fully assess AI competence before the employee has worked with the specific combination of tools, organizational data structures, and task contexts that define their role. The competence emerges from the interaction, not from prior training alone.

Kellogg, Valentine, and Christin (2020) document this dynamic in their review of algorithmic work arrangements, noting that workers develop what amount to folk theories about how systems behave, theories that are often partially accurate but structurally incomplete. The same phenomenon will appear in corporate AI adoption. Employees will develop impressionistic accounts of when AI helps and when it does not, and those folk theories will be mistaken for genuine competence during performance reviews because the evaluator often holds the same folk theory.

The Awareness-Capability Gap, Reproduced at Organizational Scale

Research on algorithmic literacy consistently demonstrates that awareness of a system's existence does not translate into improved outcomes (Gagrain, Naab, & Grub, 2024). Knowing that an algorithm governs content distribution does not tell you how to respond to it effectively. The same structural gap is now appearing at the organizational level. Companies are aware that AI use matters. They are aware that variation in AI competence produces variation in output quality. But awareness of this variance does not translate into the organizational capacity to measure it accurately or develop it systematically.

Hancock, Naaman, and Levy (2020) raise a related concern in their treatment of AI-mediated communication: when AI is embedded in consequential human processes, the legibility of human agency becomes contested. Performance reviews are exactly that kind of consequential process. If an employee produces a high-quality strategic memo with significant AI assistance, the review system has to answer a question it was never designed to answer: what exactly is being evaluated, and at what layer of the work?

What This Actually Requires

Organizations that want to build valid AI competence metrics need to move away from usage tracking and toward schema-based assessment. The relevant question is not whether an employee used an AI tool but whether that employee can transfer their understanding of AI behavior to a novel task context they have not encountered before. Gentner's (1983) structure-mapping theory provides a useful anchor here: genuine competence is demonstrated when structural relations, not surface features, transfer across domains. A performance review framework built on this principle would look less like a usage audit and more like a structured diagnostic interview. Most organizations are not close to that yet, and the scramble visible in current reporting suggests they may not know what they are actually trying to build.

References

Gagrain, A., Naab, T. K., & Grub, J. (2024). Algorithmic media use and algorithm literacy. New Media & Society.

Gentner, D. (1983). Structure-mapping: A theoretical framework for analogy. Cognitive Science, 7(2), 155-170.

Hancock, J. T., Naaman, M., & Levy, K. (2020). AI-mediated communication: Definition, research agenda, and ethical considerations. Journal of Computer-Mediated Communication, 25(1), 89-100.

Hatano, G., & Inagaki, K. (1986). Two courses of expertise. In H. Stevenson, H. Azuma, & K. Hakuta (Eds.), Child development and education in Japan (pp. 262-272). Freeman.

Kellogg, K. C., Valentine, M. A., & Christin, A. (2020). Algorithms at work: The new contested terrain of control. Academy of Management Annals, 14(1), 366-410.

The Specific Event

Saber Interactive's CEO recently issued a public denial following claims by Stella Sacco, the former lead writer on Rideshare Stimulator, that Saber had replaced her with ChatGPT. The denial is structurally interesting not because it settles the factual dispute, but because of what it reveals about how organizations now communicate around AI labor decisions. The CEO did not deny that AI writing tools are in use at Saber. The denial was narrower: no writer was replaced by AI. This is a meaningful distinction, and the narrowness of it is the point.

What the Denial Actually Communicates

When an organization denies a specific operational claim rather than a categorical one, it signals that the categorical version of the claim is not deniable. Saber's response functions as a kind of organizational hedge. It leaves open the question of whether AI tools have reshaped writing workflows, reduced headcount needs, or altered how creative labor is scoped and contracted. Kellogg, Valentine, and Christin (2020) identified that algorithmic systems at work often restructure tasks in ways that are invisible at the level of formal job titles, making straightforward substitution claims both easy to deny and empirically difficult to falsify. The framing of "replacement" is itself a contested schema. If a contractor is not renewed because AI tooling changed the scope estimate for a project, that is functionally substitution without being classifiable as replacement in any legal or HR record.

The Organizational Theory Problem Underneath

What this case illustrates is a coordination failure at the level of organizational communication, not just a labor dispute. Sacco's claim and Saber's denial share almost no common vocabulary for what it means to substitute AI for human creative work. This is not unusual. Rahman (2021) describes how algorithmic control often produces invisible cages, structures of constraint that are real in their effects but difficult to name through conventional employment categories. Game studios, like gig platforms, are increasingly mediating creative output through tools that reshape what a writer's job actually consists of, independently of whether any individual writer is formally terminated. The absence of shared schema for these transitions is not an accident. It serves organizations well to keep the grammar of AI labor substitution ambiguous.

The Awareness-Capability Gap in Reverse

Most discussions of algorithmic literacy focus on workers who are aware that algorithms govern their outcomes but lack the structural understanding to respond effectively (Gagrain, Naab, and Grub, 2024). The Saber case presents something adjacent but distinct: a worker with direct, first-person knowledge of an AI substitution event who nonetheless lacks the organizational vocabulary to make that claim legible to external observers. Sacco knows what happened to her workflow. What she does not have is a schema that her claim can map onto in a way that forces a substantive denial rather than a narrow technical one. This is not an individual failure. It reflects a structural gap in how creative industries have developed frameworks for naming AI-mediated labor change. The asymmetry is significant. Organizations have legal and communications infrastructure for denying replacement. Workers have personal testimony. These are not equivalent epistemic resources in a public dispute.

What This Means for How Organizations Communicate About AI

The Saber incident is a preview of a communication pattern that will intensify. As AI tooling becomes normalized in creative and knowledge work pipelines, organizations will increasingly face public claims that are true in spirit and deniable in letter. The strategic response is already visible: issue narrow denials, avoid categorical statements, and let the ambiguity of "replacement" do the work. Hancock, Naaman, and Levy (2020) noted that AI-mediated communication introduces new asymmetries in how agency is attributed and perceived. In corporate communication, that asymmetry favors whoever controls the schema. Right now, organizations control the schema, and workers like Sacco are left arguing in folk-theory terms against institutional responses built on definitional precision.

The Structural Implication

From an organizational theory standpoint, what the Saber case demands is not sympathy for either party but a clearer conceptual vocabulary for AI labor transitions in creative industries. Schor et al. (2020) argued that dependence and precarity in platform economies are structurally produced, not individually experienced. The same logic applies here. The inability to name what happened to Sacco is not her interpretive failure. It is a gap in the organizational theory of AI labor that the field has not yet filled, and that gap is already being exploited in real-time corporate communication.

The Specific Event

Ribbon, a voice-based AI recruiting platform, has announced that its tools improve hiring accessibility by allowing candidates to interview at any hour, removing the scheduling constraints of traditional human-led interviews. The pitch is straightforward: flexible timing reduces friction, and reduced friction improves access. This is a reasonable surface-level observation. It is also, from an organizational theory standpoint, a serious misdiagnosis of where inequality in hiring actually originates.

Accessibility Is Not the Same as Equity in Outcomes

The framing Ribbon is using conflates procedural access with substantive competence. Being able to complete an interview at 2 a.m. does not change what the interview is evaluating or how algorithmically mediated scoring systems weight candidate responses. This distinction matters because the academic literature on algorithmic labor environments is consistent on a specific point: access parity does not produce outcome parity. Kellogg, Valentine, and Christin (2020) documented extensively how workers operating within identical algorithmic systems produce dramatically different outcomes, a variance that access alone cannot explain. Ribbon's accessibility argument addresses the topography of the hiring process - the scheduling surface - while leaving the topology entirely intact.

The topology here is the underlying structure of how an AI system evaluates a candidate's spoken responses, what features it weights, what schemas it applies, and how those schemas interact with candidate communication styles that were not equally represented in the system's training data. None of that changes because the interview window is now 24 hours.

The Folk Theory Problem in AI Hiring

There is a more subtle issue embedded in Ribbon's announcement that connects directly to the awareness-capability gap I study in platform coordination contexts. Candidates who know they are being evaluated by a voice-based AI system will develop folk theories about how that system works. They will adjust their pacing, their vocabulary, their sentence structure. Some will do this effectively. Many will not, and crucially, the ones who do it effectively will not necessarily be the most qualified candidates for the role. They will be the candidates with the most accurate structural schemas about how AI voice evaluation systems operate.

Gagrain, Naab, and Grub (2024) distinguish precisely between this kind of folk theorizing and genuine algorithmic literacy. Folk theories are individually constructed impressions, often partially correct, rarely systematically accurate. They emerge from experience but do not reliably generalize. Candidates who have encountered AI screening tools before will bring folk theories from those encounters. Whether those theories transfer productively to Ribbon's specific system depends on whether the structural features are shared - which is exactly the transfer question my dissertation research is trying to answer in a different context.

What Organizational Theory Predicts Here

Sundar (2020) identified a core tension in AI-mediated communication: the machine agency attribution problem. When humans interact with AI systems that produce human-like outputs - spoken evaluation, natural language prompts, conversational interview formats - they apply social heuristics designed for human interaction. They try to read the room, mirror conversational energy, build rapport. These strategies are not merely irrelevant in an AI evaluation context. They may actively penalize candidates who deploy them if the system is optimizing for different features entirely.

This is the competence inversion that platforms routinely produce, and that classical hiring theory does not anticipate. Traditional interview training builds routine expertise: here is how to answer a behavioral question, here is how to structure a response using the STAR method. That training was designed for human interviewers who bring interpretive flexibility to the evaluation. AI voice scoring systems do not. Hatano and Inagaki (1986) drew this line clearly between routine expertise, which fails in novel contexts, and adaptive expertise, which requires understanding the structural principles of a domain. Candidates trained on human interview norms are bringing routine expertise to a novel structural context.

The Governance Question Ribbon Is Not Asking

What is absent from Ribbon's announcement is any discussion of what the system is actually measuring, how it weights responses, whether its scoring has been audited for demographic bias, and what recourse candidates have when the system produces an outcome they cannot interpret or contest. Rahman (2021) described this configuration as the invisible cage: systems that shape behavior through opaque constraints that workers cannot see, contest, or adapt to without structural knowledge they are not given.

Framing 24/7 availability as an equity intervention is a distraction from these harder governance questions. The scheduling constraint was never the primary barrier. The primary barrier is the structural opacity of how algorithmically mediated evaluation systems work, who they were built to evaluate, and what competencies they actually surface. Until those questions are answered publicly, accessibility claims from AI recruiting platforms should be read as marketing, not reform.

References

Gagrain, A., Naab, T., & Grub, J. (2024). Algorithmic media use and algorithm literacy. New Media & Society.

Hatano, G., & Inagaki, K. (1986). Two courses of expertise. In H. Stevenson, H. Azuma, & K. Hakuta (Eds.), Child development and education in Japan (pp. 262-272). W. H. Freeman.

Kellogg, K. C., Valentine, M. A., & Christin, A. (2020). Algorithms at work: The new contested terrain of control. Academy of Management Annals, 14(1), 366-410.

Rahman, H. A. (2021). The invisible cage: Workers' reactivity to opaque algorithmic evaluations. Administrative Science Quarterly, 66(4), 945-988.

Sundar, S. S. (2020). Rise of machine agency: A framework for studying the psychology of human-AI interaction. Journal of Computer-Mediated Communication, 25(1), 74-88.

The Specific Event

California's SB 903, currently moving through the state legislature, would prohibit companies from marketing AI chatbots as therapy and would require licensed clinician review of AI therapeutic decisions. This is not a speculative regulatory proposal. It is a direct legislative response to an observable market condition: companies have been deploying conversational AI systems in clinical-adjacent contexts without the governance infrastructure that licensed practice requires. The bill draws a hard boundary between a chatbot that provides emotional support and one that is advertised as performing therapy. That distinction sounds clean in a legislative summary. It is considerably messier in practice.

Why the Therapy-Chatbot Boundary Is a Communication Problem, Not Just a Legal One

The framing of SB 903 as a consumer protection bill is accurate but incomplete. What the bill is actually trying to regulate is a specific kind of communication failure: the collapse of the distinction between a system's functional outputs and a user's interpretation of those outputs. Hancock, Naaman, and Levy (2020) identified AI-mediated communication as a category where the perceived source of a message fundamentally alters how that message is received and acted upon. When a user believes they are in a therapeutic relationship, they disclose differently, they interpret responses differently, and they rely on continuity of care assumptions that a chatbot cannot structurally honor. Banning the marketing of chatbots as therapy is an attempt to regulate the user's schema before interaction begins, because the interaction itself may be too late to correct it.

The Awareness-Capability Gap in Clinical Contexts

My dissertation research focuses on the gap between algorithmic awareness and effective response to algorithmic systems. The same structural problem appears in the SB 903 context, but with higher stakes. Algorithmic literacy research consistently shows that knowing a system is algorithmically mediated does not translate into knowing how to respond to it appropriately (Gagrain, Naab, and Grub, 2024). A user who has been told, in a terms-of-service disclosure, that they are interacting with an AI and not a licensed therapist has awareness. They do not necessarily have the schema to modulate their reliance behavior accordingly. SB 903 attempts to solve this by restricting what companies can claim, but it does not address what users will infer regardless of what companies claim. The marketing prohibition is a supply-side intervention for what is partly a demand-side cognition problem.

The Clinician Review Requirement and the Limits of Procedural Oversight

The bill's second provision, requiring licensed clinician review of AI therapeutic decisions, raises a distinct organizational question. Review requirements assume that the reviewing professional can accurately assess what the AI system did, why it did it, and whether that action was appropriate. This is the topology versus topography problem I have written about in other contexts. A clinician reviewing an AI-generated response can assess the surface content of that response. They cannot necessarily assess the structural logic by which the system generated it, what training signals shaped that output, or how the system would respond to the same user in a slightly different context. Kellogg, Valentine, and Christin (2020) documented that workers operating alongside algorithmic systems frequently develop folk theories about system behavior that do not accurately represent system structure. There is no reason to expect licensed clinicians to be immune to this pattern simply because their professional training is rigorous in other domains.

What This Means for Organizational Governance of AI Systems

The deeper issue SB 903 surfaces is that current organizational governance frameworks for AI were not designed for systems that communicate. Most AI governance literature focuses on decision-support systems where a human retains visible decisional authority. Therapeutic chatbots are different because the communication itself is the intervention. The system is not helping a clinician decide; the system is doing the thing that, in a licensed context, would constitute practice. Rahman (2021) described the organizational dynamics of algorithmic control as an invisible cage, where workers adapt to constraints they cannot fully see. The SB 903 scenario inverts this: users are not workers adapting to a platform, but vulnerable individuals potentially adapting their self-disclosure and help-seeking behavior to a system that has no professional accountability structure beneath it.

The Structural Takeaway

SB 903 is a meaningful legislative step, but it is solving the legible part of the problem. Restricting marketing language addresses how companies describe their products. It does not address the communication topology that makes AI-mediated therapeutic interaction categorically different from other AI-mediated tasks. Governance frameworks that focus exclusively on claims and disclosures will consistently lag behind the actual reliance behaviors they are trying to manage. The structural features of how users form relationships with AI communication systems need to be part of the regulatory analysis, not an afterthought to it.

The Specific Convergence

Two stories surfaced this week that, read separately, seem unrelated. The first: workers are increasingly building dedicated savings funds to finance career breaks, not retirement, but near-term exits driven by unsustainable workplace pressure and rising job insecurity (Fortune, 2025). The second: a separate survey found that workers are growing anxious that AI tools are revealing gaps in their foundational job competencies, gaps they had successfully concealed or never recognized until AI made the comparison legible (Forbes, 2025). When you read these together, a specific organizational dynamic comes into focus that neither story names directly.

The burnout fund is not primarily a wellness story. It is a competence-exposure story. Workers are not just tired. They are building financial escape hatches because AI is functioning as an involuntary diagnostic, surfacing the difference between what they understood themselves to be doing and what the work actually required. That is a coordination failure with a precise theoretical description.

When Awareness Becomes Threatening Rather Than Useful

The algorithmic literacy literature has documented a persistent gap between awareness and capability. Knowing that an algorithm is evaluating your output does not translate into knowing how to improve that output (Kellogg, Valentine, and Christin, 2020). The AI-exposure dynamic reported this week is structurally similar but inverted. Workers are not discovering that an algorithm is judging them. They are discovering, through the comparative baseline that AI tools provide, that their own mental models of their work were inaccurate.

This is the distinction between folk theories and structural schemas that runs through my dissertation research. A folk theory is an individual's working impression of how something operates, assembled from personal experience without systematic verification. A schema is an accurate structural representation. Workers who believed they understood how to write a client brief, structure an analysis, or synthesize research were operating on folk theories of professional competence. AI outputs are now providing a reference point that makes the folk theory visible as a folk theory. That transition, from implicit confidence to explicit uncertainty, is destabilizing in ways that salary increases or flexible scheduling cannot resolve.

The Organizational Failure Is Structural, Not Individual

Organizations are responding to this dynamic poorly, largely because they are diagnosing it incorrectly. The instinct is to treat worker anxiety about AI exposure as an adoption problem, a communication failure, or a change management deficit. The more accurate diagnosis is that organizations built workflows around tacit competencies they never formally verified, and AI is now auditing those competencies in real time.

Hatano and Inagaki (1986) distinguish between routine expertise, the ability to execute familiar procedures, and adaptive expertise, the ability to recognize when a procedure no longer fits the situation and adjust accordingly. Most professional workers developed routine expertise in relatively stable task environments. AI tools are not replacing that expertise so much as making its limits visible. The workers who are most anxious, according to the Forbes reporting, are not those who lack intelligence. They are those whose professional identity was built on procedures that AI can now perform faster. That is a routine expertise problem, not a capability problem in any deeper sense.

Burnout Funds as Revealed Preference Data

The burnout fund trend provides something that survey data about "AI anxiety" does not: revealed preference evidence. Workers are not just reporting discomfort. They are reallocating savings, a costly behavioral signal. This is consistent with what Schor et al. (2020) describe as the precarity dynamic in algorithmically-mediated work, where workers experience high formal autonomy alongside high structural vulnerability. The combination of unpredictable evaluation criteria and asymmetric information about what actually drives performance outcomes produces exactly this kind of exit preparation behavior.

What organizations should be attending to is not the burnout fund as a wellness indicator but as an organizational signal. When workers systematically build financial buffers against their own employers, they are pricing in the probability of sudden competence reclassification. That is not a morale problem. That is a coordination breakdown.

The Implication for How Organizations Respond

The standard organizational response to AI-driven skill disruption is procedural retraining: new software tutorials, AI literacy workshops structured around platform-specific tasks. The research on schema induction suggests this is the wrong level of intervention (Gentner, 1983). Teaching workers how to use a specific AI tool does not resolve the underlying problem, which is that they lack structural frameworks for evaluating when AI output is reliable, when their own judgment adds value, and how to recognize the difference. Procedural training on new tools compounds the original problem by adding another layer of routine expertise on top of an already fragile foundation. What the data from this week actually calls for is schema-level intervention, and most organizations are not equipped to deliver it.

The Announcement and What It Actually Does

Adobe has moved every application in its Creative Cloud suite inside ChatGPT. This is not a minor feature update. As of this week, users working inside OpenAI's conversational interface can invoke Photoshop, Express, Acrobat, and the rest of the Adobe catalog without leaving the ChatGPT environment. The direction of the integration is worth noting: Adobe is not embedding ChatGPT inside its own products, which it has also done separately. It is placing its products inside OpenAI's product. Adobe is choosing to exist as a layer inside someone else's platform.

Why Platform-Inside-Platform Is a Coordination Problem, Not Just a Business Strategy

The standard read on this news is competitive: Adobe is threatened by Canva, Canva has been gaining ground with non-professional users, and embedding Adobe tools inside ChatGPT captures the growing segment of users who begin their creative workflows through conversational AI rather than through dedicated design software. That competitive framing is accurate but incomplete. What Adobe's decision also represents is a structural choice about where coordination happens. By embedding inside ChatGPT, Adobe accepts that OpenAI's platform mediates the initial interaction with users. The algorithm that surfaces Adobe's plugin, the conversational logic that decides when to invoke it, and the session context that shapes how users arrive at the tool all belong to OpenAI, not Adobe. Adobe retains execution capability but surrenders the upstream coordination layer.

This matters for organizational theory because it represents a concrete case of what Kellogg, Valentine, and Christin (2020) describe as algorithmic management extending beyond labor markets into firm-level strategy. The entity being managed by the algorithm is no longer just a gig worker deciding which tasks to accept; it is Adobe, a company with a market capitalization above $150 billion, deciding that its products should be discoverable and activatable on terms set by another firm's conversational model.

The Folk Theory Risk at the Organizational Level

My ALC framework draws a distinction between folk theories and structural schemas. A folk theory is an agent's working impression of how a system operates, formed through experience but not validated against the system's actual logic. A structural schema is an accurate representation of the constraints and affordances the system actually imposes. Workers who operate from folk theories about platform algorithms tend to misattribute outcomes and respond to surface-level signals rather than underlying structure (Gagrain, Naab, and Grub, 2024).

Adobe's integration decision may be built on a folk theory about how users interact with ChatGPT. The implicit model seems to be: users ask ChatGPT for help with creative tasks, the plugin surfaces Adobe tools, users adopt them, and Adobe retains the relationship. But that model assumes a static, legible routing logic inside ChatGPT. In practice, OpenAI's system decides which plugins to surface, when to recommend alternatives, and how to present options. Adobe does not control that decision, and its stated parameters are not public. If OpenAI changes the weighting logic - as platforms routinely do - Adobe's visibility inside ChatGPT can shift without notice. Rahman (2021) calls this the invisible cage: the constraints shaping outcomes are real and consequential, but not transparent to the agents operating within them.

Dependency Without Legibility

Schor et al. (2020) identified dependence and precarity as structural features of platform labor, but the same dynamics apply to firms that treat another firm's platform as a distribution channel. Adobe's revenue depends partly on user acquisition. If a meaningful share of new user acquisition now flows through ChatGPT, then Adobe has introduced a dependency on a system it cannot audit, cannot modify, and cannot exit without losing that acquisition channel. The switching cost compounds over time as more of Adobe's user base arrives through the ChatGPT interface.

What is absent from Adobe's announcement, as far as I can determine, is any public discussion of what structural knowledge Adobe has about how ChatGPT routes users to plugins. Adaptive expertise, as Hatano and Inagaki (1986) distinguish it from routine expertise, requires understanding why a system behaves as it does, not just how to operate within it. Adobe is executing a procedure - build the integration, publish the plugin - without public evidence of a structural understanding of the coordination layer it is entering.

The Broader Implication

Adobe is not making a mistake in any simple sense. The competitive pressure from Canva is real, and the move to be present inside ChatGPT is rational given where user attention is shifting. But the decision illustrates that platform dependency is no longer a condition that applies only to individual workers and small businesses. Large firms are now accepting positions inside algorithmically-governed environments they did not build and cannot fully observe. The organizational theory question this raises is straightforward: what does competent strategy look like when the coordination layer is opaque and not under your control?

That is the question my research is designed to address at the individual level. I suspect the organizational-level answer follows a similar logic: structural schema over folk theory, topology over topography, adaptive positioning over procedural execution. Adobe has made the move. Whether it understands the structure it has moved into is a different question entirely.

California Senate Bill 903, currently advancing through the state legislature, would prohibit chatbot developers from advertising their products as therapy and impose new disclosure requirements on AI tools deployed in formal mental health settings. The proximate cause is straightforward: chatbot usage for mental health advice has grown dramatically over the past two years, and regulators are responding to documented cases where users treated AI responses as clinical guidance. The bill is, on its face, a consumer protection measure. But the governance logic embedded in SB 903 reveals something more interesting, and more troubling, than its sponsors likely intended.

The Competence Assumption in Regulatory Design

Every governance framework rests on assumptions about what users already know. Classical consumer protection law assumes an informed party who can, in principle, distinguish a product claim from a clinical diagnosis. What SB 903 implicitly acknowledges is that this assumption has collapsed in AI-mediated communication contexts. Users interacting with conversational AI in moments of psychological distress are not operating as informed consumers evaluating product claims. They are engaging with what Hancock, Naaman, and Levy (2020) called AI-mediated communication, where the interface layer actively shapes the user's perception of the relationship itself. The chatbot does not just deliver information. It simulates a communicative role, and users respond to that simulation with the schemas they have for that role, not the schemas they have for software products.

This is precisely what makes SB 903 both necessary and insufficient. The bill targets advertising claims, which is the topography of the problem. The topology, the underlying structural shape of the issue, is that users lack schemas for distinguishing conversational competence from clinical competence. Banning the word "therapy" in a marketing headline does not install that schema. It removes one misleading signal without addressing the signal-processing deficit that made the misleading signal dangerous in the first place.

Awareness Without Capability

The research literature on algorithmic literacy has documented a consistent and uncomfortable finding: awareness of how a system works does not produce improved outcomes when interacting with that system. Gagrain, Naab, and Grub (2024) formalize this as the awareness-capability gap. Users can be told, and can accurately report, that an AI chatbot is not a licensed therapist. That declarative knowledge does not reliably change how they process the chatbot's responses during an emotionally activated interaction. Sundar (2020) provides a complementary mechanism here, noting that machine agency cues, specifically the experience of a system that responds, adapts, and appears to understand, trigger social processing heuristics that override analytical evaluation. The disclosure that SB 903 requires will be read cognitively at sign-up and ignored affectively during use.

This is not an argument against disclosure requirements. It is an argument that disclosure requirements address the wrong layer of the problem. The governance intervention that would actually move outcomes is schema induction, training users to recognize the structural difference between a system optimized for response fluency and a system accountable for clinical accuracy. These are not the same competence, and the former can mimic the latter with high fidelity. Gentner's (1983) structure-mapping theory predicts that surface similarity, in this case, conversational warmth and apparent understanding, will dominate structural dissimilarity, the absence of licensure, training, and accountability, when users lack schemas that organize around the structural dimension.

What This Reveals About Platform Governance More Broadly

The California legislature is doing what legislatures do: responding to visible harm with visible intervention. I do not fault the pragmatics of this. But SB 903 is a case study in what happens when governance frameworks inherit the competence assumptions of classical communication regulation and apply them to algorithmically-mediated environments where those assumptions no longer hold. Kellogg, Valentine, and Christin (2020) argued that algorithmic work environments generate forms of dependence that existing regulatory categories were not designed to address. AI therapy chatbots are an extreme instance of this: the dependence is not economic but epistemic, and the harm is not labor precarity but clinical substitution.

The more productive regulatory question is not "what can companies claim?" but "what do users need to understand structurally about this class of system before they encounter it in a vulnerable state?" That is a schema-induction problem, not a disclosure problem. California has identified the right domain. The bill as written operates at the wrong level of the governance stack.

References

Gagrain, A., Naab, T., & Grub, J. (2024). Algorithmic media use and algorithm literacy. New Media & Society.

Gentner, D. (1983). Structure-mapping: A theoretical framework for analogy. Cognitive Science, 7(2), 155-170.

Hancock, J. T., Naaman, M., & Levy, K. (2020). AI-mediated communication: Definition, research agenda, and ethical considerations. Journal of Computer-Mediated Communication, 25(1), 89-100.

Kellogg, K. C., Valentine, M. A., & Christin, A. (2020). Algorithms at work: The new contested terrain of control. Academy of Management Annals, 14(1), 366-410.

Sundar, S. S. (2020). Rise of machine agency: A framework for studying the psychology of human-AI interaction. Journal of Computer-Mediated Communication, 25(1), 74-88.

The Specific Event

A recent BBC report surfaced a pattern that deserves more analytical attention than it has received. Women returning to the workforce after career gaps are describing their experience with AI recruitment tools as something close to systematic elimination. One interviewee described having to "Botox her CV," deliberately obscuring dates and gaps to survive automated screening. This is not a story about bias in the colloquial sense. It is a story about what happens when algorithmic systems are trained on historical hiring data that reflects historical exclusion, and then deployed as neutral infrastructure.

Trained on the Past, Deployed in the Present

The mechanism here is specific. AI recruitment tools typically learn from prior hiring outcomes, which means they inherit the distributional properties of whoever was hired before. If career continuity correlates with hiring success in the training data, and if career continuity correlates with gender due to structural caregiving norms, then the model has effectively encoded a proxy for gender without any explicit instruction to do so. Rahman (2021) described algorithmic control as an "invisible cage," a system of constraint that operates without apparent agency. What makes the recruitment case particularly clean is that the constraint is not even visible to the people administering it. The hiring managers using these tools may have no idea what the underlying model weights.

The Awareness-Capability Gap, Inverted

My dissertation research focuses on what I call the awareness-capability gap: the finding that knowing an algorithm exists does not translate into knowing how to respond effectively to it (Kellogg, Valentine, & Christin, 2020). The recruitment case presents an interesting inversion of this dynamic. The women in the BBC report are aware of the algorithm. They are developing folk theories about what it screens for. They are modifying their CVs accordingly. But their adaptations - hiding dates, smoothing timelines - are reactive and individually improvised. They lack what Gentner (1983) would call a structural schema: an accurate model of the system's underlying logic that would allow principled, transferable navigation rather than case-by-case guesswork.

This is exactly the distinction between topography and topology that I have been developing in my framework. Topography is knowing that a particular gap in your CV is likely to trigger a filter. Topology is understanding the structural relationship between model training, proxy variable selection, and outcome distributions well enough to anticipate how different systems will behave. The women adapting their CVs are working at the topographic level. They are navigating specific terrain without a map of why the terrain is shaped the way it is.

The Organizational Theory Problem

What makes this a governance failure, not just a technical one, is that organizations have structurally separated the people who deploy these tools from the people who understand their statistical properties. Procurement teams buy the tool. HR teams operate it. Neither group has the schema required to audit what the model is actually doing. Kellogg et al. (2020) identified this pattern in their review of algorithmic management: the workers most affected by algorithmic decisions are systematically the least positioned to interrogate or contest them. That asymmetry is not accidental. It is a structural feature of how these tools are marketed and implemented.

Schor et al. (2020) made a related point about platform dependence: when workers must conform to systems they cannot inspect, precarity becomes structural rather than incidental. The recruitment context extends this logic upstream. The precarity begins before employment, at the screening stage, where the system's opacity is most complete and the worker's leverage is lowest.

What This Means Practically

The "Botox your CV" strategy is a folk theory in action. It is individually rational and collectively corrosive, because it trains future applicants to game a signal that may shift without notice as the underlying model is retrained. What would actually change outcomes is schema-level transparency: published documentation of what features these systems weight, what training data they used, and what demographic audits were conducted before deployment. That is not a request for radical disclosure. It is the minimum condition for informed organizational governance. Without it, companies are not making hiring decisions. They are delegating them to a statistical artifact of whoever they hired before, and calling that neutrality.

References

Gentner, D. (1983). Structure-mapping: A theoretical framework for analogy. Cognitive Science, 7(2), 155-170.

Kellogg, K. C., Valentine, M. A., & Christin, A. (2020). Algorithms at work: The new contested terrain of control. Academy of Management Annals, 14(1), 366-410.

Rahman, H. A. (2021). The invisible cage: Workers' reactivity to opaque algorithmic evaluations. Administrative Science Quarterly, 66(4), 945-988.

Schor, J. B., Attwood-Charles, W., Cansoy, M., Ladegaard, I., & Wengronowitz, R. (2020). Dependence and precarity in the platform economy. Theory and Society, 49(5), 833-861.

What Happened and Why It Matters

This week, the Trump administration shared the details of its AI cybersecurity framework with OpenAI, Anthropic, and other major AI laboratories while keeping the public entirely uninformed. The framework exists. Its contents are being acted upon by the most consequential AI developers in the world. And the public, whose infrastructure and institutions this framework ostensibly protects, cannot read it. That is not a minor transparency gap. It is a structural design choice with significant implications for how coordination between government and platform operators actually works.

Asymmetric Information as Governance Architecture

The standard critique of this arrangement focuses on democratic accountability, and that critique is valid. But I want to focus on something more specific: what it means when a small set of organizations receives structural information about governance constraints that everyone else is denied. This is not just a political story. It is an organizational coordination story.

When the White House briefs OpenAI and Anthropic on cybersecurity expectations while withholding those same expectations from the public, it creates a formal information asymmetry that mirrors the dynamic Kellogg, Valentine, and Christin (2020) identified in algorithmic governance at work. In that context, platform operators possess structural knowledge about how systems are designed to function, while workers possess only partial, often folk-theoretic impressions of those systems. The consequential actors are the ones who understand the actual topology of constraints. Everyone else is navigating topography they cannot accurately map.

What the White House has effectively done is grant a small cohort of firms access to the topology of regulatory intent while requiring everyone else, including researchers, civil society organizations, and competing developers without equivalent access, to construct folk theories from observable signals. That asymmetry compounds over time.

The Competence Gap This Creates

Rahman (2021) describes how platform-dependent workers operate inside what he calls an "invisible cage," where the rules governing their outcomes are structurally opaque even when the consequences of those rules are visible. The secrecy around this cybersecurity framework operationalizes something comparable at the level of national AI governance. Firms inside the briefing have adaptive capacity because they know which structural features of their systems the government considers high-risk. Firms outside the briefing are left developing routine responses to signals they can observe without understanding the underlying logic.

This distinction between routine and adaptive expertise matters here in a non-trivial way. Hatano and Inagaki (1986) argued that adaptive expertise requires understanding the principles behind procedures, not just the procedures themselves. A firm that knows the government's actual threat model can reason from principle. A firm reconstructing that threat model from public statements and enforcement actions is procedurally guessing. The gap between those two positions is not a knowledge gap in the ordinary sense. It is a structural advantage that regulatory secrecy manufactures and sustains.

Why Organizational Theory Should Pay Attention

There is a tendency in organizational research to treat government-industry coordination as a background condition rather than an object of study in its own right. That tendency is increasingly difficult to defend. When the government selects a small set of firms to receive structural information about regulatory architecture, it is not simply informing them. It is altering their competitive position in ways that are durable and compounding. Schor et al. (2020) documented how platform dependence produces precarity through information asymmetry at the worker level. The same mechanism operates here at the firm level, with the government functioning as the platform operator.

The firms that received this briefing now have something that cannot be redistributed simply by releasing the document later. They have lead time. They have the ability to align product development, security architecture, and compliance infrastructure to a threat model their competitors do not yet possess. In platform coordination terms, that is the equivalent of receiving algorithmic documentation before a major policy update. The advantage is not just informational. It is temporal and structural.

The Transparency Question Is Actually a Coordination Question

I am not arguing that all national security information should be public. That position would be indefensible. What I am arguing is that selective disclosure to commercial actors, rather than to independent researchers or regulatory bodies, is a coordination choice with predictable organizational consequences. It concentrates adaptive capacity in the firms that already possess the most market power. It converts a governance mechanism into a competitive instrument. And it does so invisibly, which is precisely what makes it worth examining carefully.

When the shape of the rules is known only to the parties with the most resources to act on them, the rules do not function as neutral constraints. They function as barriers. That distinction deserves more attention from organizational theorists than it is currently receiving.

References

Hatano, G., & Inagaki, K. (1986). Two courses of expertise. In H. Stevenson, H. Azuma, & K. Hakuta (Eds.), Child development and education in Japan (pp. 262-272). Freeman.

Kellogg, K. C., Valentine, M. A., & Christin, A. (2020). Algorithms at work: The new contested terrain of control. Academy of Management Annals, 14(1), 366-410.

Rahman, K. S. (2021). The invisible cage: Workers' reactivity to opaque algorithmic evaluations. Administrative Science Quarterly, 66(4), 945-988.

Schor, J. B., Attwood-Charles, W., Cansoy, M., Ladegaard, I., & Wengronowitz, R. (2020). Dependence and precarity in the platform economy. Theory and Society, 49(5-6), 833-861.

The Claim on the Table

A recent Forbes piece introduces what the author terms RHLCF - Reinforcement Human Learning from Computer Feedback - framing it as the natural successor to RLHF (Reinforcement Learning from Human Feedback). The argument is structurally straightforward: just as humans once provided feedback signals to train AI systems, AI systems will increasingly provide feedback signals to train humans. The piece positions this as evolutionary, even optimistic. I want to take that claim seriously, because the organizational theory implications are considerably less tidy than the framing suggests.

What Feedback Authority Actually Does in Organizations

Classical organizational theory treats feedback as a control mechanism that flows from principals to agents. Managers evaluate workers; workers adjust behavior. The authority embedded in that feedback loop is not incidental - it is the mechanism through which organizations socialize competence and enforce norms. When you invert the feedback direction, as the RHLCF framing proposes, you are not simply changing the communication channel. You are relocating the source of evaluative authority from human judgment to algorithmic output. Rahman (2021) describes a version of this dynamic in the platform economy as the "invisible cage" - workers who nominally retain autonomy but whose behavioral parameters are continuously narrowed by algorithmic feedback structures. RHLCF, if adopted at scale, does not limit this dynamic to gig workers. It extends it into conventional employment.

The Awareness-Capability Gap Gets Structurally Embedded

The RHLCF framing assumes that receiving AI-generated feedback will improve human performance. This assumption deserves scrutiny. Research on algorithmic literacy consistently finds what I have called the awareness-capability gap: knowing that an algorithm is shaping your environment does not translate into knowing how to respond effectively to it (Kellogg, Valentine, and Christin, 2020). The RHLCF model implicitly assumes workers will develop accurate schemas of what the AI system is optimizing for, then calibrate their behavior accordingly. But Gagrain, Naab, and Grub (2024) find that most workers develop folk theories rather than accurate structural schemas - plausible but often incorrect models of how algorithmic systems operate. If workers are receiving feedback from systems they fundamentally misunderstand, the feedback loop does not produce genuine competence development. It produces behavioral compliance shaped by misread signals.

Routine Versus Adaptive Expertise Under Inverted Feedback

Hatano and Inagaki (1986) distinguish between routine expertise - the ability to execute procedures reliably within a stable environment - and adaptive expertise - the ability to recognize when structural conditions have changed and adjust accordingly. The RHLCF model, as described in the Forbes piece, is optimized for producing routine expertise. AI feedback systems reward consistency with past patterns. They are not well-suited to signaling when a worker should deviate from those patterns because the underlying environment has shifted. This is not a minor limitation. It is a structural one. Organizations that build feedback architectures around AI evaluation may find they are systematically developing workers who are highly responsive to current algorithmic criteria and poorly equipped to transfer that competence when platform logic changes. The variance puzzle I track in my dissertation research - why workers with identical access produce dramatically different outcomes - is not solved by giving everyone the same AI feedback. It may in fact be amplified, as initial behavioral differences get locked in through differential reinforcement.

The Governance Question the Piece Does Not Ask

The Forbes framing treats RHLCF as a workplace evolution story. The governance story is harder. Hancock, Naaman, and Levy (2020) note that AI-mediated communication systematically alters the perception of message authenticity and authority. When feedback originates from an AI system rather than a human supervisor, workers may assign it different epistemic weight - either dismissing it as mechanical or, more problematically, over-crediting it as objective. Neither response is accurate, because AI feedback systems encode the values and priorities of whoever designed them. Sundar (2020) frames this as machine agency - the tendency for users to attribute autonomous judgment to systems that are, in fact, expressing human-designed optimization targets. RHLCF does not eliminate the politics of evaluation. It obscures them behind an interface that signals neutrality it cannot actually deliver.

What This Means Practically

The question worth asking about RHLCF is not whether AI can provide useful performance feedback - it clearly can in bounded domains. The question is what organizational theory predicts happens when feedback authority migrates to systems that workers do not structurally understand, that optimize for historical patterns rather than adaptive capacity, and that represent human evaluative priorities as objective machine outputs. The answer, drawing on Schor et al. (2020) and Rahman (2021), is a form of platform-style dependence extended into the interior of conventional employment. That may or may not be the future of work. But it should be analyzed as a governance structure, not celebrated as an evolutionary step.

The Specific Event

OpenAI recently banned a cluster of China-linked ChatGPT accounts found to be running a covert influence operation targeting U.S. debates about data centers and artificial intelligence infrastructure. The operators used the service to generate English- and Chinese-language posts, memes, and coordinated messaging designed to shape public opinion on a specific policy question. OpenAI identified and removed the accounts. The incident is being framed primarily as a platform governance story, which is accurate but incomplete. The more analytically interesting problem is structural: the operation succeeded at the production layer before it was caught at the distribution layer, and that sequencing tells us something important about where competence gaps actually live in algorithmically-mediated environments.

Why Platform Governance Is Not the Same as Platform Literacy

The standard response to influence operations is to strengthen detection and enforcement. OpenAI banned the accounts; case closed. But this framing conflates two distinct problems. The first is a governance problem: platforms need better systems for identifying coordinated inauthentic behavior. The second is a schema problem: the people and institutions targeted by this content largely lack the structural understanding required to recognize when algorithmically-generated material is shaping their information environment. These are not the same problem, and solving one does not solve the other.

Hancock, Naaman, and Levy (2020) introduced the concept of AI-mediated communication to describe interactions where artificial agents modify, generate, or evaluate content in ways that recipients cannot directly observe. The ChatGPT influence operation is a near-perfect empirical instance of this dynamic. The recipients of those posts on X did not have access to the production conditions of the content they were reading. They had no structural cues to distinguish synthetically generated opinion from organically produced opinion. That is not a failure of individual intelligence. It is a predictable outcome of operating without an accurate schema for how AI-generated content enters and circulates through social media environments.

The Awareness-Capability Gap Appears Again

There is a reasonable objection here: most people now know that AI-generated content exists. Awareness of synthetic media is no longer a niche concern. But awareness is not schema, and this distinction is central to my dissertation research on Algorithmic Literacy Coordination. Gagrain, Naab, and Grub (2024) draw a similar distinction in their work on algorithmic media use, separating surface-level awareness of algorithmic systems from the deeper structural understanding required to actually adjust behavior in response to those systems. Knowing that AI-generated posts exist tells you approximately nothing about how to identify one, how to evaluate the credibility of a source that may or may not be synthetic, or how to reason about the incentive structures that produce coordinated campaigns.

This is what I mean by the awareness-capability gap. The gap is not a knowledge deficit in the colloquial sense. It is a schema deficit. Users exposed to the "Data Center Bandwagon" campaign were not ignorant of AI. Many of them were actively engaged in debates about AI policy. Their engagement did not protect them because engagement without structural schema does not produce the kind of adaptive expertise that Hatano and Inagaki (1986) describe. Procedural familiarity with a platform, knowing how to post, how to reply, how to read a thread, does not transfer to the novel problem of detecting synthetic influence at the point of consumption.

What This Means for Organizational Theory

Kellogg, Valentine, and Christin (2020) observed that algorithmic systems at work create visibility asymmetries: the platform sees worker behavior in detail, while workers have only partial visibility into the platform's logic. The influence operation case extends this asymmetry into the civic domain. The operators of the campaign had precise knowledge of what the platform could detect and what it could not, at least for a period of time. The targets of the campaign had no equivalent structural knowledge about the production conditions of what they were reading. That is not just a political problem. It is an organizational design problem, one that concerns how institutions develop and distribute the kind of schema-level competence that would make populations more resilient to this class of manipulation.

Rahman (2021) describes algorithmic control as an "invisible cage," a structure that shapes behavior without announcing itself. The influence operation case adds a layer to this metaphor: the cage can be operated by external actors, not just platform owners, and users who lack structural schema cannot tell the difference. The policy conversation following this incident will focus on what OpenAI did to stop it. The more durable question is what kind of cognitive infrastructure would have reduced the campaign's effectiveness before detection. Those are different interventions, and right now, almost all institutional energy is going toward the first.

The Transfer Problem, Stated Plainly

If schema induction, teaching people the structural features of how AI-generated content is produced and distributed, enables transfer across novel influence contexts, then general literacy training of this kind should outperform platform-specific warnings or one-time content removals. That is a testable claim. The ChatGPT influence operation gives us a natural reference point: a population exposed to synthetically coordinated content, most of whom had general awareness of AI but lacked structural schema for detecting it. The outcome was predictable under my framework. Whether targeted schema training would have changed that outcome is an empirical question, and one worth taking seriously as these operations become cheaper and easier to run.

The Claim Itself

DeepSeek founder Liang Wenfeng made a striking statement this week: his $60 billion AI company operates without KPIs, without mandatory overtime, and without direct management of individual contributors. The claim is circulating as a counterpoint to Silicon Valley's recent embrace of China's 996 work culture - 9am to 9pm, six days a week - which companies like Meta have publicly signaled admiration for. Wenfeng's position is that his researchers are internally motivated and structurally unconstrained. The coverage frames this as a management philosophy story. I think it is actually a coordination story, and a theoretically important one.

What Happens When You Remove Explicit Coordination Mechanisms

Organizational theory has a well-established position on this. Coordination does not disappear when you remove formal mechanisms - it migrates. When hierarchy is absent, coordination happens through shared norms, mutual adjustment, or what Thompson (1967) called reciprocal interdependence. When KPIs are removed, the behavioral signal that replaces them is rarely nothing. It is usually something harder to observe: reputational feedback, peer comparison, or in AI research environments specifically, publication output and benchmark performance. DeepSeek's researchers are not unmanaged. They are managed by the algorithmic and reputational structures of the research community itself. The absence of internal KPIs does not mean the absence of performance signals. It means those signals are externalized.

The Awareness-Capability Gap at the Organizational Level

This distinction matters for a reason that connects directly to the research I am currently developing. The Algorithmic Literacy Coordination framework argues that workers in algorithmically-mediated environments often develop awareness of the structures governing their behavior without developing the functional capacity to respond to those structures effectively (Kellogg, Valentine, and Christin, 2020). The awareness-capability gap is typically described at the individual level - a worker knows an algorithm exists but cannot act on that knowledge. What the DeepSeek case suggests is that this gap can operate at the organizational level as well. Liang Wenfeng may accurately perceive that his researchers are intrinsically motivated. What he may not be accurately modeling is the external algorithmic and reputational architecture that is doing the coordination work his internal systems deliberately avoid.

Folk Theory Versus Structural Schema in Management Claims

Gentner's (1983) structure-mapping theory draws a clear distinction between surface-level analogies and structural analogies. A folk theory of management, like "no one manages them," captures a surface observation - there are no explicit supervisors assigning tasks - without mapping the underlying relational structure that actually governs behavior. This is precisely the distinction the ALC framework makes between folk theories of platforms and structural schemas. Folk theories reflect individual impressions of how a system works. Structural schemas reflect accurate models of the system's constraint architecture. The claim that DeepSeek researchers have no performance pressure is a folk theory. The structural reality - that they operate inside a global AI research field with extremely legible output signals, citation counts, benchmark leaderboards, and peer recognition - is the schema that the folk theory obscures.

Why the Silicon Valley Versus DeepSeek Framing Is the Wrong Frame

The current media framing positions 996 culture against Wenfeng's no-KPI model as competing management philosophies. This framing has rhetorical appeal but limited analytical value. Both approaches share an assumption that the primary coordination problem in AI research organizations is motivational - how hard do you push people? The more interesting coordination question is structural: how do organizations develop and transfer the adaptive expertise required to operate effectively in environments where the performance criteria themselves are algorithmically defined and rapidly shifting? Hatano and Inagaki (1986) established that routine expertise - executing known procedures reliably - breaks down precisely when environmental conditions change. AI research is a domain where the relevant benchmarks, tools, and evaluation criteria change continuously. Neither overtime mandates nor KPI removal addresses that structural challenge.

The Practical Implication for Organizational Design

What the DeepSeek story actually points toward is a research gap in how we model high-autonomy AI organizations. The question is not whether to impose KPIs or remove them. The question is what coordination mechanisms remain operative when explicit ones are withdrawn, and whether those implicit mechanisms are visible to the organization's leadership. Rahman (2021) describes the structuring power of invisible constraints in platform labor contexts. The same dynamic applies internally. Organizations that remove formal performance structures without modeling the informal and external structures that replace them are not less coordinated. They are less legible to themselves. That is a different kind of organizational risk - and one that a $60 billion AI company probably cannot afford to leave unexamined.

Gentner, D. (1983). Structure-mapping: A theoretical framework for analogy. Cognitive Science, 7(2), 155-170. Hatano, G., and Inagaki, K. (1986). Two courses of expertise. Research and Clinical Center for Child Development, 11, 27-36. Kellogg, K. C., Valentine, M. A., and Christin, A. (2020). Algorithms at work: The new contested terrain of control. Academy of Management Annals, 14(1), 366-410. Rahman, K. S. (2021). The invisible cage: Workers' reactivity to opaque algorithmic evaluations. Administrative Science Quarterly, 66(4), 945-988.

The Specific Event

On a recent weekday, Google launched an AI-powered image generation feature inside Google Earth. Within hours, the company pulled it. The reason: users immediately began generating manipulated aerial views depicting fake bombings, riots, and large-scale destruction. The Washington Post documented how the feature was exploited to produce what appeared to be photorealistic scenes of violence and crisis layered onto real geographic coordinates. Google's response was swift, but the damage to the question of AI deployment governance was already visible. The gap between what the feature was designed to do and what users actually did with it was not a narrow gap. It was structural.

This Is Not a Moderation Story

The instinct when reading this story is to frame it as a content moderation failure. That framing misses something more important. Content moderation is a downstream response to a governance design problem. What Google Earth's rapid rollback actually reveals is an organizational failure to anticipate the schema mismatch between how engineers and product designers understood the feature and how users encountered it. The engineers likely held an accurate structural model of what the tool could produce. The users who immediately exploited it held a different, equally accurate structural model of the same tool. Both groups understood the topology. Only one group had institutional authority to act on it in advance.

Awareness Without Governance Structure

This distinction maps directly onto what algorithmic literacy research describes as the awareness-capability gap (Kellogg, Valentine, and Christin, 2020). In the platform labor literature, this gap refers to workers who know algorithms govern their outcomes but cannot translate that awareness into improved performance. The governance analog is organizations that know AI systems can be misused but cannot translate that knowledge into pre-deployment structural constraints. Awareness of misuse potential is not equivalent to institutional capacity to prevent it. Google's internal teams almost certainly understood that generative AI layered onto satellite imagery of real-world locations carried misuse risk. The rollback happened in hours, which suggests the harm was not surprising in retrospect. The surprise was that no structural gate existed to catch it before release.

Folk Theories at the Organizational Level

Gentner's (1983) structure-mapping theory distinguishes between surface-level feature matching and deep relational structure. When organizations reason about AI risk, they often operate from what I would call organizational folk theories: intuitive, surface-level impressions of how a feature will be used rather than structurally grounded models of the relational space of possible uses. A structural schema for this Google Earth feature would have mapped the relational logic connecting generative AI, geocoded real-world imagery, and the informational authority that satellite images carry as apparent evidence. That combination produces a specific misuse topology that is distinct from, say, generative AI on a blank canvas. The misuse that occurred was not random. It was predictable from the structural relationships between the tool's components, not from surface observation of its intended function.

The Organizational Competence Inversion Problem

Hatano and Inagaki (1986) draw a line between routine expertise, which involves executing established procedures, and adaptive expertise, which involves constructing new responses when established procedures do not apply. AI deployment governance currently operates mostly through routine expertise. Organizations apply existing trust and safety checklists, red-team exercises borrowed from prior product categories, and post-hoc moderation pipelines. These are procedural tools built for known risk topographies. Generative AI layered onto authoritative data sources like satellite imagery represents a novel risk topology. Routine expertise fails here precisely because the structural relationships that generate harm are new, not because the organization lacks experience with AI risk in general. Rahman (2021) notes that algorithmic systems create invisible constraints that workers and users navigate without full visibility into the governing logic. The governance challenge is the inverse: organizations must develop visibility into constraint structures before deployment, not after harm occurs.

What This Means for Organizational Theory

The Google Earth case is a useful data point for organizational theorists working on AI governance not because the rollback was embarrassing but because it was fast. The speed of the rollback suggests that institutional recognition of the problem occurred quickly once harm was observable. The gap was not in organizational response capacity. It was in pre-deployment schema induction: the organizational process of building accurate structural models of novel AI capabilities before they reach users. Hancock, Naaman, and Levy (2020) argue that AI-mediated communication requires new frameworks for understanding agency and accountability. I would extend that claim to organizational governance: AI-mediated deployment requires new frameworks for structural anticipation, not just reactive moderation. Until organizations treat pre-deployment schema construction as a formal competency rather than an informal checklist exercise, rapid rollbacks will remain the primary evidence that governance happened at all.

References

Gentner, D. (1983). Structure-mapping: A theoretical framework for analogy. Cognitive Science, 7(2), 155-170.

Hancock, J. T., Naaman, M., and Levy, K. (2020). AI-mediated communication: Definition, research agenda, and ethical considerations. Journal of Computer-Mediated Communication, 25(1), 89-100.

Hatano, G., and Inagaki, K. (1986). Two courses of expertise. In H. Stevenson, H. Azuma, and K. Hakuta (Eds.), Child development and education in Japan (pp. 262-272). Freeman.

Kellogg, K. C., Valentine, M. A., and Christin, A. (2020). Algorithms at work: The new contested terrain of control. Academy of Management Annals, 14(1), 366-410.

Rahman, H. A. (2021). The invisible cage: Workers' reactivity to opaque algorithmic evaluations. Administrative Science Quarterly, 66(4), 945-988.

The Specific Event

EY recently disclosed that it has deployed what it calls an "invisible" AI router positioned behind its suite of AI tools. The router's function is to direct queries to the most cost-efficient language model capable of handling a given task, rather than defaulting every request to the most powerful and expensive option. The reported result is a reduction in token consumption of up to 60%. This is not a marginal operational tweak. For a Big Four firm running AI queries at scale across tens of thousands of professionals, a 60% reduction in token spend represents a structural intervention in how AI capacity is allocated inside a large organization.

The announcement is worth pausing on, because what EY is describing is not an AI literacy initiative. It is an infrastructural governance decision made above the level of individual workers. The router is invisible by design. Users do not choose which model handles their request. The organization has removed that decision from the worker entirely.

Governance Above the Awareness Layer

This design choice illuminates a distinction that algorithmic literacy research tends to underweight. Kellogg, Valentine, and Christin (2020) documented how algorithmic systems at work tend to create asymmetries between workers and the organizations deploying those systems. Workers develop awareness of the systems around them, but that awareness rarely translates into the capacity to influence how those systems are configured. EY's router makes this asymmetry structurally explicit. The worker does not need to know which model is being used, because the governance layer has already made that determination.

From the perspective of my own framework on Application Layer Communication, this is a meaningful case. The ALC framework is concerned with how competencies develop endogenously through participation in algorithmically-mediated environments. EY's router suggests that some organizations are now intervening at the infrastructure layer to prevent certain competence-development pathways from forming at all. If workers never encounter the decision of which model to use, they cannot develop schema-level understanding of why those differences matter. The awareness-capability gap does not close. It is simply made irrelevant by fiat.

The Routine Expertise Trap at Organizational Scale

Hatano and Inagaki (1986) drew a foundational distinction between routine expertise and adaptive expertise. Routine expertise enables reliable performance within stable task structures. Adaptive expertise enables reconfiguration when those structures change. EY's routing infrastructure optimizes for routine performance across its current AI stack, but it does so by abstracting away the structural features of the stack itself.

This is organizationally rational in the short term. Token costs are real, and 60% reduction is a significant efficiency gain. But the abstraction carries a longer-term risk that governance discussions about AI in organizations rarely surface directly. If the router is invisible, and if workers are never required to reason about model selection, then the organization's adaptive capacity becomes concentrated in whoever designed and maintains the router. Everyone else develops only procedural familiarity with the output layer. When the infrastructure changes, as it will, the organization's distributed AI competence will not transfer.

Gentner's (1983) structure-mapping theory is instructive here. Transfer across contexts depends on learners having access to structural, relational features of a domain, not just surface-level procedures. EY's architecture, however rational, systematically denies most workers access to those structural features. The organization trains procedural fluency. It does not train transferable schema.

What This Means for AI Governance Frameworks

The luxury board governance discussion circulating in parallel business press this week frames AI governance primarily as a risk management problem: boards need to understand AI well enough to ask the right questions. EY's router case reframes the problem. Governance is not only about board-level oversight. It is also about where in the organizational stack decisions get made, and what gets obscured when they are made at the infrastructure level rather than the worker level.

Sundar (2020) described the rise of machine agency as a process in which AI systems increasingly make decisions that humans formerly made, shifting accountability in ways that are not always visible to participants. The invisible router is a clean example. The efficiency gains are visible. The governance implications of removing model-selection from worker cognition are not.

The practical question for organizational theorists is whether large firms deploying AI routers, filters, and abstraction layers are building organizational AI competence or merely AI-adjacent workflow efficiency. Those are different assets with different durability. The 60% token reduction will show up in EY's cost reports. The competence deficit, if it materializes, will show up much later and be harder to attribute.

References

Gentner, D. (1983). Structure-mapping: A theoretical framework for analogy. Cognitive Science, 7(2), 155-170.

Hatano, G., & Inagaki, K. (1986). Two courses of expertise. Research and Clinical Center for Child Development Annual Report, 8, 27-36.

Kellogg, K. C., Valentine, M. A., & Christin, A. (2020). Algorithms at work: The new contested terrain of control. Academy of Management Annals, 14(1), 366-410.

Sundar, S. S. (2020). Rise of machine agency: A framework for studying the psychology of human-AI interaction. Journal of Computer-Mediated Communication, 25(1), 74-88.

The Specific Allegation

A healthcare workers' union has alleged that Kaiser Permanente deployed an algorithmic e-visit tool to screen mental health patients without real-time clinician review, potentially violating California state law. Kaiser has now skipped a public hearing on these allegations for the second consecutive year. This is not a story about AI adoption gone wrong in some abstract sense. It is a story about what happens when an organization uses an algorithm to perform a coordination function that participants - patients seeking mental health care - have no framework to recognize, evaluate, or contest.

The structural detail that matters most here is not the algorithm itself. It is the absence of any visible signal that an algorithm is doing the triage at all. Patients entering a mental health e-visit have no reliable way to know whether a clinician or a decision-tree is determining their care pathway. That asymmetry is the central problem, and it extends well beyond healthcare.

Algorithmic Triage as Coordination Failure

The ALC framework I am developing treats platforms as coordination mechanisms where competency develops endogenously through participation. Classical coordination - markets, hierarchies, professional networks - assumes some baseline of ex-ante competence among participants. A patient entering a physician's office has schema for what that interaction involves. They understand, however imperfectly, the norms, the roles, and the feedback loops. Algorithmic triage removes those schemas without replacing them.

Rahman (2021) describes this dynamic as an "invisible cage" in which workers - and by extension, platform participants generally - are constrained by algorithmic rules they cannot see, audit, or appeal. The Kaiser case materializes this metaphor in a clinical setting. A mental health patient routed away from immediate care by an algorithmic screen has no recourse mechanism because they have no awareness that an algorithmic screen existed. Kellogg, Valentine, and Christin (2020) identify this as a defining feature of algorithmic management: the work of coordination becomes opaque precisely at the moment when transparency would be most consequential.

The Awareness-Capability Gap in High-Stakes Contexts

Research on algorithmic literacy consistently finds that awareness of algorithms does not translate to improved navigation of algorithmic systems (Gagrain, Naab, and Grub, 2024). In low-stakes platform contexts - content recommendation, search ranking, gig work dispatch - this gap produces unequal outcomes across workers with otherwise identical access. In a clinical mental health context, the same gap produces something categorically different: patients who cannot advocate for themselves because they do not know what they are being evaluated by.

This distinction matters for organizational theory. The variance puzzle I examine in my dissertation - why platform participants with identical access show dramatically different outcomes - typically manifests as a performance distribution problem. Some workers capture more value than others. In the Kaiser case, the distribution problem becomes a triage problem. The patients least equipped to articulate their needs through a structured algorithmic interface are, plausibly, the patients with the most acute need for clinical judgment rather than algorithmic routing. The algorithm amplifies initial differences in communicative competence at exactly the moment when those differences should be professionally compensated for.

Governance Structures and the Procedural Substitution Problem

Kaiser's decision to skip the public hearing for a second consecutive year is an organizational choice that deserves analytical attention separate from the algorithm itself. Organizations deploying algorithmic coordination tools frequently treat the algorithm as a procedural solution to a resource constraint - here, a shortage of clinicians available for real-time triage. What the ALC framework would predict, consistent with Hatano and Inagaki's (1986) distinction between routine and adaptive expertise, is that procedural substitution fails specifically in novel or high-variance cases. Algorithmic triage built from historical patient data encodes past distributions of case severity. The patient presenting in ways that deviate from those historical patterns - which is structurally more likely in mental health than in, say, dermatology - is precisely the patient the algorithm is least equipped to handle correctly.

Sundar (2020) argues that machine agency creates a distinct communicative relationship in which human participants implicitly defer to algorithmic outputs as authoritative. In a clinical setting, that deference is not merely a cognitive tendency - it is structurally enforced. The patient has no alternative channel and no visible indicator that deferral is even occurring.

What This Case Reveals

The Kaiser allegation is a concrete instance of what happens when algorithmic coordination replaces professional coordination without corresponding investment in participant schema. The problem is not that algorithms exist in healthcare. The problem is that organizations are deploying algorithmic triage in contexts where the awareness-capability gap is not just an inconvenience but a clinical risk, and where governance structures - as evidenced by two consecutive missed hearings - appear designed to reduce rather than increase accountability. That combination is what makes this case worth watching closely.

References

Gagrain, A., Naab, T. K., and Grub, J. (2024). Algorithmic media use and algorithm literacy. *New Media and Society*.

Hatano, G., and Inagaki, K. (1986). Two courses of expertise. In H. Stevenson, H. Azuma, and K. Hakuta (Eds.), *Child development and education in Japan* (pp. 262-272). Freeman.

Kellogg, K. C., Valentine, M. A., and Christin, A. (2020). Algorithms at work: The new contested terrain of control. *Academy of Management Annals, 14*(1), 366-410.

Rahman, H. A. (2021). The invisible cage: Workers' reactivity to opaque algorithmic evaluations. *Administrative Science Quarterly, 66*(4), 945-988.

Sundar, S. S. (2020). Rise of machine agency: A framework for studying the psychology of human-AI interaction. *Journal of Computer-Mediated Communication, 25*(1), 74-88.

The Statement and Why It Deserves Serious Attention

Shopify CEO Tobi Lütke recently endorsed a proposal by banker Eric Thor suggesting that voting rights should be weighted proportional to income tax paid, and that the poor and pensioners should lose their right to vote entirely. This is not a fringe comment buried in a reply thread. It comes from the chief executive of one of the most influential platform companies in the world, a company whose infrastructure coordinates hundreds of thousands of sellers, some generating up to $260,000 per month in sales. The statement deserves analytical attention, not because it reflects mainstream policy, but because it expresses a governance logic that is structurally consistent with how platform systems actually allocate participation rights.

Platforms Already Do This: Participation as a Function of Output

Lütke's proposal is not alien to the operational logic of the system he runs. Platform architectures routinely condition visibility, reach, and effective participation on demonstrated performance. A Shopify seller generating $260,000 monthly enjoys access to capital advance programs, preferential support tiers, and algorithmic promotion that a low-volume seller simply cannot access. This is not incidental. It is structural. Kellogg, Valentine, and Christin (2020) describe how algorithmic systems at work encode managerial preferences into automated evaluation loops, creating differentiated conditions of participation that appear neutral but are outcome-dependent. The logic Lütke expressed publicly about democracy is the logic his platform encodes privately about commerce.

The Competence Inversion Problem, Applied to Civic Systems

What makes this governance logic particularly worth interrogating is how it interacts with what I call the competence inversion problem in my own research. Classical coordination theory, whether in markets, hierarchies, or democratic systems, generally assumes that participants arrive with pre-existing capacity to engage. Platforms invert this assumption: competence is not assumed, it is produced endogenously through participation. But here is the structural tension Lütke's statement exposes. If you use output metrics, whether tax revenue or sales volume, to gate participation rights, you are not measuring competence. You are measuring the accumulated result of prior access conditions, including access to capital, networks, and infrastructure that are themselves unequally distributed.

Schor et al. (2020) document precisely this dynamic in the platform economy: workers enter platform systems with structurally unequal starting positions, but the system evaluates them as if those differences are performance signals rather than access artifacts. Lütke's democratic proposal applies this same inferential error at a civic scale. Low tax contribution is treated as evidence of low governance competence, when it may simply reflect prior structural exclusion from conditions that generate taxable income.

Folk Theories of Meritocracy in Platform Governance

There is a theoretical category that helps explain why this kind of reasoning is appealing to platform executives specifically. In my ALC framework, I distinguish between folk theories and structural schemas. Folk theories are plausible-sounding individual impressions about how a system works, often accurate at the surface level but missing the underlying structural logic. The folk theory at work in Lütke's endorsement is that economic contribution tracks civic value. This feels intuitive in a market context. But it confuses a topographic reading of economic outcomes, who is currently at the top, with a topological understanding of why those distributions exist in the first place. Rahman (2021) describes this as the invisible cage problem: algorithmic and institutional structures constrain worker trajectories in ways that are systematically invisible to the workers themselves and, apparently, to the executives designing those systems.

Why Corporate Governance Researchers Should Care

Lütke's comment is a data point about how platform-native executives conceptualize legitimate governance. Visa is simultaneously cutting 7% of its workforce, targeting technology and product teams, as its CEO pursues efficiency optimization. UMC is expanding semiconductor fabrication capacity in Singapore and Tainan to meet AI demand. These developments share a common thread: governance decisions at major technology companies are increasingly being made through an efficiency-first logic that treats participation rights, whether in employment or in civic life, as outputs to be earned rather than conditions to be guaranteed.

This is not simply an ethics problem. It is an organizational theory problem. If the executives coordinating some of the largest platforms in the world hold folk theories about governance that are structurally inconsistent with how access and outcome are actually related, then the systems they design will reproduce those errors at scale. Hatano and Inagaki (1986) distinguish routine expertise from adaptive expertise precisely on this basis: routine experts apply successful procedures without understanding when those procedures break down. What Lütke revealed is not malice. It is routine expertise applied to a domain where the structural conditions do not transfer.

The Structural Lesson

The point is not that Lütke's proposal will become policy. It is that the governance logic it expresses is already operating inside the platforms that coordinate labor, commerce, and increasingly information. Researchers studying algorithmic coordination need to take seriously the possibility that the folk theories held by platform architects are not just incidentally flawed but systematically so, in ways that compound inequality precisely because they mistake output for access, and performance for merit. That is a boundary condition worth naming clearly.

The Specific Event

This week, BBC reported that hundreds of private conversations with Anthropic's Claude chatbot were discovered to be publicly accessible online. The exposure was not the result of a sophisticated cyberattack. The conversations were simply findable. Separately, an Anthropic product executive, Dianne Penn, gave an interview describing how she uses Claude as part of her active management toolkit, embedding the chatbot into daily decisions about her team. These two stories appeared within the same news cycle, and the juxtaposition is worth examining carefully. One story describes an organization promoting AI as a managerial competency. The other describes that same organization's product leaking user data at scale. The tension between these two data points is not incidental. It is structurally informative.

Competence in the Wrong Direction

The framing of Penn's interview follows a pattern that organizational researchers should recognize immediately. A senior executive demonstrates fluency with a tool, describes specific workflows, and positions that fluency as a model for others. This is, in essence, a competence signal. The implicit argument is that AI adoption at the managerial level represents organizational sophistication. What the Claude data exposure reveals, however, is that competence in using a tool and competence in understanding what that tool is doing with your information are not the same thing. This distinction maps directly onto what Kellogg, Valentine, and Christin (2020) describe as the fundamental problem of algorithmic work: workers interact with systems whose internal logic remains opaque, and their operational fluency masks rather than resolves that opacity.

The Awareness-Capability Gap in Institutional Settings

My dissertation research on the Algorithmic Literacy Coordination framework focuses on a specific puzzle: workers who are aware that algorithms govern their outcomes still fail to improve those outcomes. Awareness does not produce capability. The Claude incident extends this logic into a different but structurally parallel domain. Anthropic's own internal executive was presumably aware that Claude processes and stores conversational data. That awareness did not translate into organizational protocols robust enough to prevent public exposure of user conversations. The gap here is not between knowing and not knowing. It is between surface-level operational awareness and genuine structural understanding of what the system does with information at the infrastructure level. Gagrain, Naab, and Grub (2024) make a related point about algorithmic media use: individuals develop folk theories about how platforms behave, but these folk theories systematically underestimate the complexity of backend processes.

What Organizational Theory Predicts Here

From an organizational theory standpoint, the Anthropic case is a fairly clean example of what happens when institutions adopt tools faster than they develop governance schemas for those tools. The managerial enthusiasm Penn describes is real and probably produces genuine short-term value. But Hatano and Inagaki's (1986) distinction between routine expertise and adaptive expertise is directly applicable. Routine expertise means knowing how to use Claude to summarize meeting notes or draft performance feedback. Adaptive expertise means understanding the conditions under which that usage creates institutional exposure, and being able to adjust behavior when those conditions change. The BBC story suggests Anthropic, as an organization, had not fully developed the second form of expertise with respect to its own product's data handling.

The Governance Schema Deficit

This is not primarily a story about a data breach. It is a story about schema deficits in organizational AI adoption. Institutions adopting AI tools are largely building procedural competence, the equivalent of knowing which buttons to push, without building the structural schemas that would allow them to reason about novel failure modes. Hancock, Naaman, and Levy (2020) argue that AI-mediated communication introduces accountability ambiguities that existing organizational structures are not designed to handle. The Claude exposure is one concrete instance of that ambiguity materializing. When an executive uses an AI tool for team management, who owns the conversational data generated in that process? What is the organization's theory of where that data lives and who can access it? The absence of public answers to those questions at Anthropic, of all organizations, is notable.

The Practical Implication

Organizations adopting AI tools for internal management need to distinguish between tool training and schema training. Teaching managers to use Claude effectively is a procedural intervention. Teaching them to reason about data residency, access permissions, and failure modes is a structural one. The ALC framework predicts that structural training produces better transfer across novel situations precisely because it targets principles rather than procedures. The Anthropic incident this week is a case study in what happens when only the procedural layer gets built out. The gap between Penn's confident managerial AI usage and the exposure of user conversations is not a contradiction. It is exactly what the awareness-capability gap predicts at the institutional level.

The Announcement and Its Structural Weight

This week, OpenAI CEO Sam Altman stated publicly that we are now inside the technological singularity, describing it as "the moment" when artificial intelligence surpasses human intelligence. This is not a forecast or a roadmap slide. Altman is making a present-tense empirical claim about where we are in historical time. Whether or not one accepts the framing, the organizational consequences of that claim deserve serious analysis, because the claim itself now functions as a coordination signal regardless of its accuracy.

I want to be precise about what I mean. When the head of the most visible AI organization in the world announces that the singularity is here, he is not merely describing a technical threshold. He is issuing a schema to every firm, regulator, and worker that must now orient around AI systems. The claim reconfigures what counts as competent behavior in algorithmically-mediated environments, independent of whether the underlying technical claim is true.

Folk Theory Inflation at Scale

Here is where organizational theory becomes useful. Research on algorithmic environments consistently distinguishes between folk theories and structural schemas (Gagrain, Naab, and Grub, 2024). Folk theories are working impressions individuals construct to explain why outcomes happen; schemas are accurate structural representations of the system's actual decision logic. The gap between these two things is consequential. Kellogg, Valentine, and Christin (2020) documented that workers in algorithmically-managed environments develop awareness of algorithmic control without developing the adaptive expertise needed to navigate it effectively. Awareness and capability are not the same variable.

Altman's singularity declaration risks industrializing folk theory at an organizational scale. When executives read "the singularity is here," the most common response will not be to interrogate the technical definition of recursive self-improvement or examine benchmark validity. The response will be to accelerate existing AI adoption initiatives on the assumption that competitive disadvantage is now immediate and severe. That is a folk-theory response: a surface-level heuristic triggered by a high-status signal, not a structural understanding of what AI systems actually do inside workflows.

The Competence Inversion Problem, Applied Upward

The Algorithmic Literacy Coordination framework I am developing at Bentley argues that platform coordination inverts classical organizational assumptions. Classical coordination theory, from markets to hierarchies to networks, assumes that competence exists before coordination begins (Schor et al., 2020). Platforms, by contrast, generate competence endogenously through participation. You do not arrive with expertise; you develop it or you do not, and the algorithmic environment amplifies those differences into power-law outcome distributions.

What Altman's announcement introduces is a version of this inversion applied to organizational leadership rather than platform workers. Senior decision-makers are now being asked to coordinate around a system they do not structurally understand, on a timeline defined by a claim they cannot independently verify. Rahman (2021) described this as the invisible cage problem: the control architecture is opaque, yet workers must respond to it as though it were legible. C-suite executives reading Altman's statement face an identical constraint. The cage is now organizational strategy itself.

Why the Google DeepMind Defection Matters Here

A related piece of news is worth connecting. A former Google DeepMind researcher published this week an account of leaving the organization after internal opposition to a Pentagon AI contract failed. The account describes a pattern in which AI ethics commitments dissolved under institutional pressure. This is not a peripheral story. It is direct evidence of how organizations respond when the gap between stated schema ("we have ethical AI commitments") and operational behavior widens under coordination pressure.

Hatano and Inagaki (1986) distinguished routine expertise from adaptive expertise on exactly this dimension. Routine expertise produces correct behavior under stable conditions; adaptive expertise produces correct behavior when conditions change. An organization that treats AI ethics as a stable procedural checklist will abandon it precisely when conditions become unstable, which is the moment ethics actually matters. The DeepMind account describes routine expertise hitting its structural limit.

What Coordination Theory Would Predict

If Altman is right that the inflection point is now, then the firms that will navigate it well are not the ones that respond fastest to his announcement. They are the ones whose internal schemas are accurate enough to distinguish what has actually changed from what has not. Hancock, Naaman, and Levy (2020) argued that AI-mediated communication fundamentally alters the conditions of human agency, but the alteration is structural, not rhetorical. A CEO press statement does not change the underlying topology of what these systems can and cannot do.

The practical implication for organizational theory is this: singularity rhetoric functions as a coordination accelerant, and accelerants are dangerous when the underlying schema is weak. The variance in organizational outcomes over the next several years will not be explained primarily by who adopted AI fastest. It will be explained by who understood the structural features of what they adopted well enough to transfer that understanding when the next configuration arrives.

References

Gagrain, A., Naab, T., and Grub, J. (2024). Algorithmic media use and algorithm literacy. New Media and Society.

Hancock, J. T., Naaman, M., and Levy, K. (2020). AI-mediated communication: Definition, research agenda, and ethical considerations. Journal of Computer-Mediated Communication, 25(1), 89-100.

Hatano, G., and Inagaki, K. (1986). Two courses of expertise. In H. Stevenson, H. Azuma, and K. Hakuta (Eds.), Child development and education in Japan (pp. 262-272). Freeman.

Kellogg, K. C., Valentine, M. A., and Christin, A. (2020). Algorithms at work: The new contested terrain of control. Academy of Management Annals, 14(1), 366-410.

Rahman, K. S. (2021). The invisible cage: Workers' reactivity to opaque algorithmic evaluations. Administrative Science Quarterly, 66(4), 945-988.

Schor, J. B., Attwood-Charles, W., Cansoy, M., Ladegaard, I., and Wengronowitz, R. (2020). Dependence and precarity in the platform economy. Theory and Society, 49(5-6), 833-861.

The Bet That Cannot Be Unwound

A recent analysis framed Microsoft's total organizational commitment to AI as a potential liability, asking whether the company's "North Star" has become a noose. The framing is provocative, but it points toward something theoretically precise: what happens when an organization structures its entire coordination logic around a technology whose behavioral rules are still being written? This is not a strategic question about market timing. It is a structural question about what kind of competence Microsoft has actually built, and whether that competence transfers when the platform underneath it shifts.

Procedural Commitment at Scale

Microsoft's AI integration is not shallow. Copilot is embedded across Azure, Teams, Office, and its developer toolchain. The organizational transformation required to ship at that depth is real, and the internal retraining required to support it is substantial. But there is a distinction in expertise research that matters here: the difference between routine expertise and adaptive expertise (Hatano and Inagaki, 1986). Routine expertise is the capacity to execute well-defined procedures reliably. Adaptive expertise is the capacity to improvise correctly when those procedures no longer fit the situation. Microsoft has invested enormously in building routine expertise around specific AI system behaviors. The question the recent coverage implicitly raises is whether that investment is brittle.

The brittleness concern is not hypothetical. Anthropic's release of Claude Opus 5 this week, positioned as approaching the capability ceiling of more expensive frontier models at half the price, signals that the cost-capability frontier is moving faster than organizational adaptation cycles. When the underlying model economics shift, procedural expertise built around one pricing and capability tier does not automatically transfer to the next. This is precisely the transfer failure that Gentner's (1983) structure-mapping theory predicts: surface-level similarity between old and new contexts triggers inappropriate schema application, while the deeper structural features that would support correct transfer go unnoticed.

The Coordination Inversion Problem

What makes Microsoft's situation theoretically interesting is that it illustrates a particular failure mode in algorithmically-mediated coordination. Classical coordination theory assumes that organizations possess ex-ante competence relevant to the environment they are entering. Platform coordination inverts this: competence develops endogenously, through participation, as the platform's behavioral rules become legible over time (Kellogg, Valentine, and Christin, 2020). Microsoft is essentially a very large platform worker. It has developed deep competence in navigating the current behavioral rules of its AI infrastructure. But that competence was built on a specific topology of constraints, and the topology is changing rapidly.

This is not the same as saying Microsoft made a bad bet. It is saying that the organizational competence question and the strategic bet question are different problems, and conflating them produces misleading analysis. The "noose" framing suggests that total commitment was itself the error. But the more precise diagnosis is that total commitment to procedural integration, without corresponding investment in structural schema development, leaves the organization with competence that cannot transfer when model generations turn over.

What Schema-Level Understanding Would Look Like

Rahman's (2021) concept of the invisible cage is useful here. Platform-dependent workers often cannot see the structural rules shaping their outcomes because those rules are opaque by design. Microsoft's dependency on OpenAI's model roadmap creates an analogous constraint: the behavioral rules governing Copilot's capabilities are set upstream, and Microsoft's internal coordination logic is built around outputs it does not fully control. The organization that would navigate this well is one that understands the structural features of how large language model capability curves develop, how cost-performance frontiers shift, and how downstream integration assumptions need to be parameterized loosely enough to accommodate those shifts.

That kind of understanding is harder to acquire than deployment expertise, and it looks less productive in the short term. But it is the difference between an organization that is genuinely adaptive to AI infrastructure changes and one that has simply internalized the current moment's procedures very deeply.

The Broader Organizational Theory Implication

The Microsoft coverage, read alongside Anthropic's Opus 5 announcement, offers a natural experiment in organizational schema flexibility. The firms that treat specific model integrations as instances of a more general class of human-AI coordination problems will handle the next capability shift more efficiently than firms that treat each integration as a terminal deployment. This is not a recommendation to be vague about implementation. It is a recommendation to maintain a layer of structural understanding above the procedural layer, so that when the procedures become obsolete, the organization retains the analytical capacity to rebuild them correctly. That capacity is not acquired automatically through experience. It requires deliberate attention to the topology of the constraint environment, not just its current surface features.

The Specific Event

WPP, one of the largest advertising conglomerates in the world, announced a sweeping organizational restructure this week aimed at saving £500 million. Chief people officer Marie-Claire Barker acknowledged publicly that the harder challenge is not the cost savings - it is rebuilding trust, changing leadership behavior, and creating a unified culture across 90,000 employees spread across dozens of formerly distinct agency brands. That framing is worth examining carefully, because it reveals a conceptual error that organizations of this scale make repeatedly: treating coordination failure as a culture problem when it is, structurally, a competence distribution problem.

What WPP Is Actually Facing

WPP's decline has been well-documented. The restructure consolidates agency brands that historically operated with distinct identities, client relationships, and internal norms. The stated goal is simplification, but the actual coordination challenge is getting 90,000 people to operate effectively within a new organizational topology they did not develop their competencies inside. This is not a trust problem at its root. Trust is a symptom. The underlying issue is that WPP's employees built their professional schemas inside specific, bounded agency environments. Those schemas - the tacit models workers use to navigate resource allocation, client management, and creative production - do not transfer automatically when the organizational structure changes around them.

This maps directly onto what Kellogg, Valentine, and Christin (2020) identified in their review of algorithmic coordination: when the governing rules of a work environment change, workers who built routine expertise within the old system face an abrupt capability gap. They retain awareness that things have changed, but awareness does not produce adaptive response. WPP's leadership appears to understand this at an intuitive level - Barker's emphasis on "changing leadership behavior" suggests recognition that procedural retraining alone is insufficient - but the public framing still defaults to culture as the explanatory variable.

The Competence Distribution Error

Large organizational restructures consistently underestimate variance in worker competence at navigating structural change. WPP's 90,000 employees are not a uniform population. Some subset will have developed what Hatano and Inagaki (1986) call adaptive expertise - the capacity to derive operative principles from novel configurations rather than relying on memorized procedures. A much larger subset will have developed routine expertise calibrated to the specific agency context they worked inside for years or decades. Merging these populations under a unified brand without a mechanism for schema induction - teaching the structural features of the new coordination environment, not just the new procedures - produces predictable outcomes: early compliance, surface-level culture adoption, and persistent underperformance on tasks that require genuine structural understanding.

This is the variance puzzle applied to corporate restructuring. Identical access to new organizational resources and unified brand identity will produce dramatically different outcomes across employees, not because of differences in individual motivation or cultural buy-in, but because of differences in structural schema. Rahman (2021) demonstrated this dynamic in platform work contexts, where workers operating inside identical constraint structures showed power-law distributions in outcomes. The mechanism WPP is relying on - culture change through leadership modeling - addresses the topography of the new organization. It shows workers what the new environment looks like. It does not address topology: the underlying shape of how coordination actually works inside the restructured entity.

Why the Framing of "Trust" Is Doing Conceptual Work It Cannot Support

Barker's language about rebuilding trust is not wrong, but it is incomplete in a way that has practical consequences. Trust in organizational contexts is a relational resource that facilitates coordination under uncertainty (Schor et al., 2020). It reduces the friction of operating inside ambiguous structures. But trust is not a substitute for structural competence. A workforce that trusts its leadership but lacks the schemas to navigate a new coordination environment will produce well-intentioned, low-efficacy outputs. WPP's restructure will be evaluated on client retention and revenue recovery, not on employee sentiment scores. The gap between what culture interventions can deliver and what structural competence development requires is where most large-scale restructures fail to meet their financial projections.

What a Structurally Sound Intervention Would Look Like

The intervention WPP needs alongside its culture work is schema induction at scale. This means identifying the structural features of the new coordination environment - how decisions flow, how resources are allocated across former agency boundaries, how client relationships are now governed - and building training that teaches those structural principles explicitly, not just the procedures that follow from them. Gentner's (1983) structure-mapping theory predicts that workers who understand the relational structure of a new system will transfer that understanding to novel problems within it. Workers who learn only the new procedures will perform adequately on familiar tasks and fail on novel ones. For an advertising conglomerate competing in an environment where client briefs and platform constraints change constantly, that distinction between routine and adaptive expertise is not academic. It is the difference between a restructure that stabilizes the business and one that merely rebrands its decline.

WPP's £500 million target is achievable through consolidation accounting. The harder number - the one that determines whether this restructure produces a genuinely more capable organization - is the distribution of structural competence across its workforce after the transition. That number is not being measured, at least not publicly, and that is the problem worth watching.

The Specific Event

Microsoft announced this week that it is replacing OpenAI's image-generating models with its own proprietary technology across PowerPoint and Bing. This is not a minor product update. Microsoft is the single largest investor in OpenAI, holding a reported multi-billion dollar stake and hosting OpenAI's infrastructure on Azure. The decision to displace a partner's core product from flagship applications, while that partnership remains formally intact, is an organizational maneuver worth examining carefully. It signals something structural about how platform dependencies are being renegotiated at the application layer.

The Dependency Inversion

For the past several years, the dominant assumption in enterprise AI was that application developers depended on foundation model providers. Microsoft, Harvey, and thousands of smaller software companies built product experiences on top of OpenAI's API. This is the classic platform-complement relationship: the platform owns the infrastructure, the complement owns the user experience. What Microsoft's move demonstrates is that this dependency relationship is not fixed. When a platform integrator controls both the distribution surface (PowerPoint's 1.2 billion users, Bing's search index) and the financial resources to develop substitutable models, the leverage structure can reverse.

Rahman (2021) describes this dynamic in platform labor contexts as the "invisible cage," where the platform retains ultimate authority over the terms of participation regardless of how the dependency appears from the outside. Microsoft's position relative to OpenAI is structurally analogous. OpenAI's models were never integrated into PowerPoint because Microsoft lacked alternatives; they were integrated while Microsoft was developing alternatives. The visibility of that arrangement was asymmetric.

What This Means for the Application Layer

My dissertation research focuses on what I call Algorithmic Literacy Coordination, the idea that competencies on algorithmically-mediated platforms develop endogenously through participation rather than through prior knowledge. One of the core puzzles in this framework is why platform workers with identical access show dramatically different outcomes. The Microsoft-OpenAI case extends this puzzle upward from individual workers to organizational actors.

OpenAI built deep integrations with Microsoft's product suite. Engineers wrote to Microsoft's APIs, trained relationships with Microsoft's enterprise sales teams, and optimized model outputs for Microsoft's interface constraints. That participation-based competence is real, but it turns out to be asymmetrically valuable. The expertise OpenAI developed about Microsoft's distribution layer does not transfer to OpenAI's own distribution ambitions. Microsoft, by contrast, learned from OpenAI's model behavior, user interaction patterns, and enterprise use cases - knowledge that is now being deployed to build competing image generation infrastructure. This is a clean illustration of what Hatano and Inagaki (1986) distinguish as routine versus adaptive expertise. OpenAI's integration expertise was procedural and context-specific; Microsoft's observational learning was structural and transferable.

The Vertical Integration Signal and OpenAI's Legal Position

The timing here matters. Apple has filed a trade secrets lawsuit against OpenAI, and OpenAI is simultaneously making direct moves into legal software markets that startups like Harvey previously occupied. The company is being squeezed from multiple directions at once. Its most important distribution partner is replacing its models. Its own downstream partners now face competition from OpenAI itself. And a major consumer hardware company is litigating against it over proprietary information. These are not unrelated events.

Schor et al. (2020) argue that platform dependence creates structural precarity because participants lack exit options proportional to their investment in platform-specific skills. OpenAI's position currently reflects this dynamic at organizational scale. Years of API integrations, enterprise relationships, and model tuning for Microsoft-adjacent use cases have created a competence profile that is difficult to redeploy quickly. The irony is that OpenAI is simultaneously inflicting this same dynamic on its own complement ecosystem by entering legal markets directly.

The Schema Deficit at the Organizational Level

Kellogg, Valentine, and Christin (2020) note that workers embedded in algorithmic systems frequently develop folk theories about platform behavior rather than accurate structural schemas. At the organizational level, the equivalent mistake is treating a partnership agreement as a structural constraint rather than a contingent arrangement. OpenAI's reliance on Microsoft distribution appears to have been treated as durable when it was always conditional on Microsoft lacking better alternatives. The schema error was confusing a contractual relationship for a coordination equilibrium.

Microsoft's replacement of OpenAI models in PowerPoint and Bing is worth tracking not because it damages one company, but because it makes visible the topology of the current AI application layer. Dependency in this ecosystem flows toward whoever controls the final distribution surface, not toward whoever built the model first. That is a structural feature, not a negotiating outcome.

References

Hatano, G., & Inagaki, K. (1986). Two courses of expertise. In H. Stevenson, H. Azuma, & K. Hakuta (Eds.), Child development and education in Japan (pp. 262-272). Freeman.

Kellogg, K. C., Valentine, M. A., & Christin, A. (2020). Algorithms at work: The new contested terrain of control. Academy of Management Annals, 14(1), 366-410.

Rahman, H. A. (2021). The invisible cage: Workers' reactivity to opaque algorithmic evaluations. Administrative Science Quarterly, 66(4), 945-988.

Schor, J. B., Attwood-Charles, W., Cansoy, M., Ladegaard, I., & Wengronowitz, R. (2020). Dependence and precarity in the platform economy. Theory and Society, 49(5-6), 833-861.

The Announcement and What It Actually Says

Monday.com filed a Form 6-K this week announcing a 20% reduction in its global workforce, attributing the decision to an "AI-driven growth strategy." The phrasing is worth pausing on. Companies routinely cite restructuring, market conditions, or strategic pivots when announcing layoffs. Framing a workforce reduction as a growth strategy, and specifically as an AI-driven one, is a different rhetorical move. It signals something about how organizational leadership currently understands the relationship between headcount and productivity in software firms. That relationship is being revised in real time, and the Monday.com announcement is one of the cleaner examples of what that revision looks like in practice.

Coordination Without Competence Transfer

The standard organizational response to AI adoption has followed a familiar script: deploy tools, train workers, and assume that training produces capability. What Monday.com's decision implies is that at least some firms have concluded the script does not work. If AI tooling had simply augmented existing workers uniformly, you would expect productivity gains distributed across the workforce. You would not expect a 20% reduction. The reduction suggests the firm has identified that AI tools do not produce equivalent output gains across all roles, and that the firm's coordination logic is being restructured around a smaller number of workers who interact effectively with those tools.

This maps directly onto what the Algorithmic Literacy Coordination framework identifies as the variance puzzle. Workers with identical access to the same platform or toolset produce dramatically different outcomes (Kellogg, Valentine, and Christin, 2020). The power-law distributions that emerge in platform labor contexts appear to be replicating inside organizational boundaries as firms integrate AI into core workflows. Monday.com's restructuring is not an anomaly. It is an organizational response to a distribution problem.

The Awareness-Capability Gap Inside the Firm

What makes the Monday.com case theoretically interesting is the distinction it forces between two kinds of AI competence. There is awareness of AI tools, meaning knowing they exist, having access to them, and receiving procedural training on their use. Then there is the structural competence to work adaptively with AI outputs, to recognize when the tool is failing, to modify inputs in response to output quality, and to integrate AI-generated work into broader organizational processes without degrading quality. Research on algorithmic literacy consistently shows that awareness does not transfer into improved outcomes (Gagrain, Naab, and Grub, 2024). Organizations that conflate the two make costly assumptions about how broadly productivity gains will distribute after an AI rollout.

Hatano and Inagaki (1986) draw a distinction between routine expertise, which is procedural and context-bound, and adaptive expertise, which is principle-driven and generalizes across novel situations. A worker trained to use Monday.com's AI features through a walkthrough tutorial has procedural knowledge. A worker who understands the structural logic of how AI-assisted project management tools handle dependency tracking, uncertainty, and output confidence has something more transferable. The layoffs suggest that the firm has concluded a significant portion of its workforce was operating at the procedural level, and that the tools have now absorbed enough of that procedural work to reduce the need for human execution at that tier.

What Organizational Theory Has to Say

Rahman (2021) describes how algorithmic systems function as "invisible cages" that constrain worker agency through opaque rule structures. The interesting inversion in the Monday.com case is that the AI is not constraining existing workers inside the organization. It is replacing the coordination function those workers served. The cage metaphor applies differently here: the workers being let go were not trapped inside the algorithmic system; they were performing tasks the system has now internalized.

Schor et al. (2020) emphasize that platform-mediated work creates forms of dependence that are structurally distinct from traditional employment. Monday.com's announcement compresses that dynamic into a single corporate event. The remaining workforce will be more deeply dependent on the AI tooling than their predecessors were, because the organizational structure is being redesigned around that dependency. That is not inherently problematic, but it does mean that the competence gap between workers who understand AI structurally and those who understand it procedurally becomes a direct determinant of employment security.

The Strategic Communication Problem

There is also something worth noting about the communication choice itself. Filing the announcement in a Form 6-K and labeling it an "AI-driven growth strategy" is an investor relations framing, not an organizational communication strategy. For the workers affected, the framing explains nothing about what competencies were missing or what the firm intends to build. Hancock, Naaman, and Levy (2020) note that AI-mediated communication alters how messages are produced and interpreted. In this case, the message is being produced for one audience - investors - while a second audience - employees - receives it as a termination rationale. That gap in intended and received meaning is itself a coordination failure, one that will shape how the remaining workforce interprets the firm's relationship to its own AI strategy going forward.

Last week, members of the United States Army received an internal email notifying them that they were burning through their allocated AI tokens at an unsustainable rate and needed to curtail use immediately. This is not, on the surface, a dramatic story. Bureaucracies run over budget. Usage caps get hit. But the specific mechanics of what happened here are worth examining carefully, because they reveal something important about how large institutions are deploying AI and why the deployment model itself is generating predictable failure.

The Token as a Coordination Signal Nobody Trained For

A token, in large language model infrastructure, is not a feature. It is a unit of computational consumption. When the Army hit its ceiling, it was not because soldiers were misusing the technology in any normative sense. They were using it exactly as advertised. The problem is that no organizational schema existed for what "appropriate use" meant at scale, across a distributed workforce, under a fixed resource constraint. The Army deployed a capability without deploying a corresponding framework for understanding how that capability is consumed.

This is a precise instance of what I have been calling the awareness-capability gap in my dissertation research. Algorithmic literacy literature consistently shows that individuals can become aware of a system's existence without developing any functional understanding of how to interact with it efficiently (Kellogg, Valentine, and Christin, 2020). Army personnel knew they had access to AI. They did not have a structural schema for what "using AI" costs at the infrastructure level. Awareness of access is not the same as competency in calibrated use.

Why Procedural Onboarding Fails Here

The institutional response to this kind of overage is almost always procedural: issue a memo, set individual caps, add a warning message to the interface. These are routine expertise responses to an adaptive expertise problem (Hatano and Inagaki, 1986). Telling a soldier to "use fewer tokens" without explaining the structural logic of why token consumption varies by task type, prompt length, and model version produces compliance without comprehension. The same overage will recur the next quarter under slightly different conditions, because the underlying schema deficit has not been addressed.

What would a structural intervention look like? It would involve teaching users the functional architecture of consumption, not just the rule. Why does a complex reasoning prompt cost more than a simple retrieval prompt? How does iterative back-and-forth with a model compound costs in ways that a single well-structured query does not? These are not technical questions reserved for engineers. They are the kind of schema-level knowledge that enables what Gentner (1983) calls far transfer: the ability to adapt behavior appropriately when the specific context changes, because you understand the relational structure underneath it.

The Organizational Theory Problem Underneath the Memo

There is a deeper issue here that connects to how organizations are structuring AI adoption broadly. The Army's situation reflects a pattern visible across institutional deployments: access is distributed rapidly, governance structures are retrofitted after the fact, and the competence development infrastructure is either absent or lagging by months. This is not a military-specific failure. It is the standard deployment sequence.

Rahman (2021) describes how algorithmic systems create what he terms "invisible cages," constraints that shape worker behavior without those workers having legible access to the rules governing their situation. The token limit memo is a visible cage, which is actually a step forward in transparency. But visibility without comprehension still produces the same coordination failure. Workers respond to the constraint as an external rule rather than integrating it into a revised understanding of how the tool works.

Schor et al. (2020) argue that platform dependence deepens when workers lack the structural knowledge to make autonomous decisions about resource allocation. The Army case is not a platform economy story, but the mechanism is analogous. When an institution deploys AI without investing in schema-level literacy, it creates dependence on centralized rationing rather than distributed competence. The memo is, in effect, an admission that the institution has no better governance instrument available.

What This Signals for AI Deployment at Scale

The Army's token crisis is a small but diagnostically rich event. It demonstrates that AI deployment without accompanying schema induction does not produce capable users. It produces high-consumption users who must be externally regulated. For organizations watching this unfold, the practical implication is straightforward: the governance problem and the training problem are the same problem. You cannot solve one with a memo while ignoring the other.

How institutions answer that challenge will determine whether AI deployment produces genuine capability gains or simply transfers the coordination burden upward, from distributed workers to centralized administrators counting tokens.

References

Gentner, D. (1983). Structure-mapping: A theoretical framework for analogy. *Cognitive Science, 7*(2), 155-170.

Hatano, G., and Inagaki, K. (1986). Two courses of expertise. In H. Stevenson, H. Azuma, and K. Hakuta (Eds.), *Child development and education in Japan* (pp. 262-272). Freeman.

Kellogg, K. C., Valentine, M. A., and Christin, A. (2020). Algorithms at work: The new contested terrain of control. *Academy of Management Annals, 14*(1), 366-410.

Rahman, H. A. (2021). The invisible cage: Workers' reactivity to opaque algorithmic evaluations. *Administrative Science Quarterly, 66*(4), 945-988.

Schor, J. B., Attwood-Charles, W., Cansoy, M., Ladegaard, I., and Wengronowitz, R. (2020). Dependence and precarity in the platform economy. *Theory and Society, 49*(5), 833-861.

A Bureaucracy Hits a Resource Ceiling It Did Not Design For

This week, members of the U.S. Army received an internal email warning that they were rapidly depleting their allocated AI tokens and needed to limit use. The message was framed as a consumption problem: people were using too much, and the supply was running out. That framing is almost certainly wrong, and the misdiagnosis reveals something important about how large organizations are failing to theorize their own relationship with AI-mediated work.

Token limits exist because AI inference is computationally expensive. Organizations purchase access in bulk, distribute it across personnel, and then discover that demand is uneven and often exceeds projections. The Army's situation is not unique. It is, however, unusually visible, and it surfaces a structural question that most organizations are quietly avoiding: when you introduce AI capability into a hierarchical institution without a coordination framework, what actually happens to the distribution of that capability?

The Variance Problem Inside Institutions

Research on platform coordination consistently finds that identical access produces dramatically different outcomes across individuals (Kellogg, Valentine, & Christin, 2020). This variance puzzle is usually discussed in the context of gig workers or content creators, but it applies with equal force inside formal organizations. Some Army personnel presumably consumed far more tokens than others. The internal email treats this as a problem of overuse. A more precise reading is that it is a problem of unequal competence in using AI tools effectively, compounded by an absence of any institutional framework for thinking about that inequality.

The distinction matters. If you respond to a consumption spike by rationing tokens, you are treating a coordination failure as a supply problem. The personnel who consumed the most tokens may have been doing genuinely productive work, or they may have been generating AI output they did not know how to evaluate or use. The token count does not distinguish between these cases. The institution has no visibility into which is true, and its response - limit use - does not help it find out.

Folk Theories at Institutional Scale

What the Army email reveals is an organization operating on folk theories about AI rather than structural schemas. Algorithmic literacy research distinguishes between folk theories, which are individual impressions about how a system works, and schemas, which are accurate structural understandings of the system's logic (Gagrain, Naab, & Grub, 2024). At the individual level, this gap predicts why awareness of algorithms does not translate to better outcomes. At the institutional level, it predicts something more troubling: organizations can deploy AI infrastructure and simultaneously have no coherent model of how that infrastructure interacts with their existing coordination mechanisms.

The Army is a hierarchy. Hierarchies coordinate through authority and rules (Kellogg et al., 2020). When a hierarchy encounters a resource it cannot allocate through authority - because it cannot observe the quality of AI use, only the quantity - it defaults to rationing. This is a legible response within hierarchical logic. It is not, however, a solution to the underlying problem, which is that the institution lacks the schema to distinguish productive AI use from unproductive AI use at scale.

Routine Expertise in an Adaptive Environment

Hatano and Inagaki (1986) distinguish between routine expertise, which is optimized for known procedures in stable environments, and adaptive expertise, which transfers to novel problems because it is grounded in structural principles rather than memorized steps. Bureaucratic organizations are engineered for routine expertise. That is not a criticism; it is the point. Standardized procedures reduce variance and enable coordination across large numbers of people with different backgrounds and training levels.

AI tools invert this logic. Their value is highest precisely in situations where standard procedures are insufficient, where synthesis, judgment, and novel problem formulation are required. This means that the personnel best positioned to use AI productively inside a bureaucracy are those capable of adaptive expertise, which is exactly the competence that bureaucratic training tends not to develop or reward. The token shortage may therefore be a symptom of a deeper mismatch: the institution imported a tool suited to adaptive work into an environment structured around routine performance.

What the Diagnosis Implies

The Army's response to its token shortage will likely be administrative: tighter allocation, usage monitoring, possibly tiered access by rank or role. These are rational responses within a hierarchical coordination framework. They will not, however, produce an institution that uses AI more effectively. For that, the organization would need what Rahman (2021) describes as structural transparency - a clear model of how the AI system operates and how individual behavior within it produces aggregate outcomes. Rationing tokens preserves the budget. It does not build the schema that would make the budget worth spending.

This is the pattern worth watching. The Army's email is a small, specific event. But it illustrates a failure mode that will appear repeatedly as large institutions absorb AI tools without absorbing any theory of how those tools change coordination. The problem is not that people used too many tokens. The problem is that the institution has no way to tell whether that use was productive, and no framework for developing one.

The Specific Warning

MIT economist Andrew McAfee recently issued a pointed warning to executives: automating Gen Z entry-level positions to cut costs may produce short-term savings while destroying the pipeline through which organizations develop future senior talent. The warning comes as IBM and Salesforce have simultaneously doubled down on Gen Z recruitment, creating a visible split in corporate strategy. McAfee's concern is not primarily about fairness or youth unemployment. It is about organizational competence reproduction. If entry-level roles are the environments where people learn how organizations actually function, eliminating those roles eliminates the learning mechanism itself.

This is a sharper claim than it first appears. Most executive commentary on AI automation treats headcount reduction as a resource allocation question. McAfee is making a structural argument: the entry-level job is not just labor, it is a training context. Remove the context, and you remove what the context produces. That reframing has significant implications, and I think it maps onto something my own research has been circling around for some time.

Competence Does Not Transfer Without a Medium

The ALC framework I am developing distinguishes between two ways that expertise develops. Routine expertise - the procedural kind - is acquired by repeating tasks until they become automatic. Adaptive expertise, by contrast, develops when a learner encounters variation, encounters failure, and is forced to reason about underlying principles rather than surface procedures (Hatano and Inagaki, 1986). Entry-level roles, at their best, produce adaptive expertise precisely because they expose workers to the full, messy complexity of organizational life before those workers have the authority to avoid that complexity.

When organizations automate those roles, they are not just removing labor. They are removing the conditions under which adaptive expertise forms. The AI system that replaces an entry-level analyst can execute the procedure. It cannot develop a structural understanding of why the procedure exists, when it fails, and how to respond when the context shifts. That structural understanding is what McAfee is worried organizations will stop producing.

This connects directly to what Kellogg, Valentine, and Christin (2020) describe as algorithmic work: the irony is that as platforms and AI systems absorb routine tasks, the remaining human work becomes more cognitively demanding, not less. Organizations that hollow out entry-level pipelines are simultaneously increasing the cognitive demands on their senior workforce while eliminating the developmental pathway that produces people capable of meeting those demands.

The Awareness-Capability Gap at the Organizational Level

McAfee's warning also implies something the automation debate rarely addresses directly: executives understand that entry-level jobs are disappearing, but that awareness does not translate into effective organizational response. This mirrors what algorithmic literacy research consistently finds at the individual level - workers develop awareness of algorithmic constraints without developing the capability to respond to them effectively (Gagrain, Naab, and Grub, 2024). The same gap appears to operate at the organizational level when leadership recognizes a structural risk but continues optimizing for the metric that produces it.

IBM and Salesforce's contrarian move toward Gen Z investment suggests at least some firms are reasoning at the structural level rather than the procedural one. They are not asking "how do we reduce headcount costs this quarter?" They are asking "what does our talent pipeline look like in ten years if we make this decision now?" That is the difference between topographic optimization - finding the best path given current terrain - and topological reasoning, which asks how the terrain itself is being reshaped by the decisions being made.

What This Means for Organizational Theory

Rahman (2021) argues that algorithmic systems create invisible cages: constraint structures that workers navigate without fully perceiving. McAfee's warning suggests organizations are building a different kind of invisible cage for themselves. By automating the contexts that produce adaptive expertise, firms create a structural dependency on AI systems to perform functions that humans can no longer perform, not because humans lack the capacity, but because the organization eliminated the developmental environment that would have produced that capacity.

This is not an argument against automation. It is an argument for treating organizational learning environments as infrastructure rather than overhead. Schor et al. (2020) note that platform economies create new forms of worker precarity through structural dependence. The irony McAfee is identifying is that firms pursuing cost efficiency through automation may be engineering an equivalent precarity for themselves: structural dependence on AI systems for capabilities they no longer know how to develop internally. That is a coordination failure with a long lag, which is precisely why it is so easy to miss until it is expensive to reverse.

The Patent as Organizational Artifact

On July 2, Meta received a patent for an AI system that continuously records a user's voice throughout the day, transcribes the audio, and feeds it through a machine learning model to detect emotional state in real time. This is not a feature announcement. It is a granted patent describing a device architecture where ambient affective data becomes a persistent input stream. The system does not wait for a query. It listens, classifies, and infers mood continuously. That distinction - between reactive and ambient surveillance - matters enormously for how we think about algorithmic coordination and organizational control.

Ambient Inference Is a Different Category of Constraint

Most algorithmic literacy research focuses on systems where the user initiates interaction: a search query, a post, a gig acceptance. Rahman's (2021) concept of the invisible cage describes how platform algorithms construct behavioral constraints that workers experience as external and fixed, even when those constraints are encoded preferences of the platform operator. The Meta patent extends this architecture into a domain where the user does not initiate anything. The inference engine operates on the ambient residue of daily life - speech patterns, tonal variation, conversational fragments. If Rahman's invisible cage is built from behavioral data the worker voluntarily generates, this system proposes a cage built from data the user does not know they are producing in any actionable sense.

This matters for organizational theory because it changes the unit of surveillance from behavior to state. Kellogg, Valentine, and Christin (2020) document how algorithmic management systems capture and classify worker behavior to produce control at scale. But behavioral classification assumes a separable act: the worker does something, the system records it. Affective classification through ambient voice monitoring collapses that separation. The system infers an internal state from continuous acoustic output, most of which the user would not describe as purposive communication with any platform.

The Awareness-Capability Gap Gets Worse Under Ambient Conditions

The awareness-capability gap in my ALC framework describes a well-documented finding: people who know algorithms exist, and even understand their general purpose, still cannot translate that awareness into improved outcomes (Gagrain, Naab, and Grub, 2024). The gap exists because structural knowledge of an algorithm's existence does not provide a schema for responding to it effectively. Now consider what ambient affective inference does to that gap. Under conventional platform interaction, a user at least has a bounded interaction event to reason about. They post something, they notice an outcome, they form a folk theory. The folk theory may be inaccurate, but there is a signal-response loop to observe.

Ambient mood tracking eliminates that loop. The input is not a discrete action but a continuous acoustic state. There is no identifiable moment where the user can ask: what did I do, and what did the system do in response? Sundar (2020) argues that machine agency becomes opaque when the system's decision processes are not legible to users. Ambient inference is the limiting case of that opacity: the input is not legible as input at all, because it is just the auditory texture of ordinary life. The awareness-capability gap, already difficult to close through training, becomes structurally unclosable when the object of inference is something the user cannot observe themselves generating.

What This Means for Organizational Governance

Hancock, Naaman, and Levy (2020) describe AI-mediated communication as a condition where AI systems shape, augment, or generate communicative acts. The Meta patent describes something downstream of that: AI systems that classify the communicator's internal state as a byproduct of communication the person was having with someone else entirely. The governance question this raises is not primarily about privacy in the legal sense. It is about organizational power. If an employer, an advertiser, or a platform operator gains access to a continuous affective signal derived from a worker's or user's daily speech, the information asymmetry that Rahman (2021) describes as constitutive of platform control expands into a domain that has historically been considered inaccessible to institutional actors.

Schor et al. (2020) identify dependence and precarity as structural features of platform participation, not individual failures. Ambient affective surveillance does not change that structural relationship, but it deepens it by extending the platform's information reach into the worker's or user's non-platform time. The patent describes a device architecture, not a deployed product. But patents are organizational commitments. They represent capabilities that firms have decided are worth defending. The fact that this architecture was worth patenting is itself a data point about where platform coordination is heading, and what kinds of schema workers and researchers will need to navigate it.

The Event That Changes the Frame

This week, workers at Hyundai staged what is being reported as the first humanoid robot strike in history. The dispute centers on Hyundai's plans to deploy Boston Dynamics Atlas robots on car-manufacturing lines. The union's response was not simply a protest against job loss. It was a formal labor action targeting the introduction of a specific class of embodied AI agent into a physical production environment. That specificity matters enormously, and most of the coverage has missed why.

The standard framing treats this as another chapter in the automation-versus-labor story. That framing is analytically lazy. What the Hyundai strike actually illustrates is a coordination failure of a particular kind: workers and management do not share a schema for what humanoid robots are, what they do to task structure, and therefore what the legitimate domain of collective bargaining even covers. Before the robots arrive, there is no shared competence base from which to negotiate. That is not a political problem. It is an organizational theory problem.

Coordination Without Pre-Existing Competence

Classical organizational coordination assumes that the parties involved understand, at least roughly, the nature of the work being coordinated. Markets price known goods. Hierarchies issue instructions for understood tasks. Even networks depend on participants who can evaluate what their peers are offering. The introduction of humanoid robots into manufacturing disrupts this assumption at its root. Neither workers nor managers have stable, accurate schemas for what a general-purpose embodied AI agent actually does to job task structure over time. They are negotiating a contract for a production system that neither side yet understands operationally.

This is precisely the inversion that the Algorithmic Literacy Coordination framework is designed to describe. Kellogg, Valentine, and Christin (2020) documented how algorithmic systems at work create systematic asymmetries in worker knowledge, but their analysis focused primarily on software-mediated task allocation. The Hyundai case extends that logic into physical space, where the algorithmic agent is not a dispatcher or a recommender but an actor sharing a factory floor. The coordination problem is not just epistemic. It is spatial, temporal, and contractual simultaneously.

Folk Theories on the Shop Floor

What workers and union representatives currently hold are not structural schemas for humanoid robotics. They hold folk theories: informal, individually constructed impressions of what Atlas does, derived from Boston Dynamics promotional videos, media coverage, and management presentations optimized for persuasion rather than accuracy. Gagrain, Naab, and Grub (2024) make the important distinction between algorithmic awareness and algorithmic literacy. Awareness - knowing that an automated system is operating - is nearly universal. Literacy - understanding the structural logic of how that system transforms task environments - is rare and difficult to acquire without deliberate schema induction.

The strike is, in part, a rational response to that literacy deficit. When workers cannot accurately model what Atlas will do to their roles, the precautionary move is to stop deployment entirely. Schor et al. (2020) describe how platform-mediated dependence produces precarity not just through wage effects but through opacity. Workers become dependent on systems they cannot read, which means they cannot adapt, bargain, or exit strategically. The factory floor is becoming a new instance of that dynamic, with the additional complication that the opacity is not algorithmic but embodied and physically co-present.

Why Procedural Training Will Fail Here

If Hyundai's response is to offer workers specific procedural training on how to operate alongside Atlas robots in their current configuration, the research literature predicts that training will not generalize. Hatano and Inagaki (1986) drew the line between routine expertise, which is procedure-bound and context-specific, and adaptive expertise, which is principle-based and transfers across novel configurations. A humanoid robot capable of general-purpose manipulation will not stay in one configuration. Its task assignments will evolve. Workers trained to operate alongside Atlas doing task A will lack the conceptual resources to adapt when Atlas is reassigned to task B, or when a successor model arrives with different physical capabilities.

What the Hyundai situation actually requires is schema-level understanding of how embodied AI agents transform task interdependence structures. That is a harder educational investment and a less satisfying answer for a union trying to draft contract language next quarter. But the alternative is a workforce that wins today's negotiation and loses every subsequent one, because the terms of the production environment will keep shifting faster than procedural knowledge can track.

What the Strike Actually Signals

The first humanoid robot strike is not a Luddite moment. It is a signal that coordination theory has a boundary condition it has not yet formalized. When the agent being coordinated around is itself adaptive, physically embodied, and algorithmically governed, neither classical labor relations frameworks nor current platform coordination theories offer adequate tools. The workers at Hyundai are not wrong to be cautious. They are operating rationally under genuine structural uncertainty. The theoretical challenge is to build frameworks that can serve them better than the ones currently available - before the robots are already on the floor.

The Lawsuit That Exposes a Structural Failure

On July 13, 2025, twenty-six Meta employees filed a federal lawsuit in Oakland, California, alleging that the company used AI-assisted systems to target workers with disabilities and medical conditions for layoffs. The complaint is specific: not that Meta laid people off, but that an algorithmic intermediary selected which people to lay off, and did so in ways that correlate systematically with protected medical status. This is a different kind of legal claim than a standard wrongful termination suit. It is, at its core, a claim about what happens when consequential decisions are delegated to systems that no individual manager fully controls or fully understands.

The lawsuit is still in early stages, and Meta has not yet responded on the merits. But the structural problem the complaint describes is worth analyzing independently of the legal outcome, because it surfaces a dynamic that organizational theory has been slow to formalize.

Rahman's Invisible Cage, Applied to Employment Decisions

Rahman (2021) introduced the concept of the "invisible cage" to describe how algorithmic systems constrain worker behavior without workers being able to identify, contest, or appeal those constraints. The analysis in that work focused primarily on gig economy platforms: Uber drivers, TaskRabbit workers, people who could theoretically "leave" but faced structural lock-in. What the Meta lawsuit suggests is that the invisible cage logic applies equally, and perhaps more severely, inside the firm. When algorithmic scoring systems inform layoff decisions, employees cannot see the criteria, cannot interrogate the weights, and cannot challenge outputs through normal organizational channels. The cage is internal to the employment relationship itself.

This matters for organizational theory because it disrupts a foundational assumption in principal-agent models. Standard accounts of employment governance assume that managers make decisions and bear accountability for those decisions. When an AI system generates a ranked list of employees to be terminated, accountability becomes diffuse. The manager who approves the list can claim they were following a process; the team that built the algorithm can claim they were not responsible for its application; the firm can claim the system was a decision support tool, not a decision maker. This diffusion is not incidental - it is structurally produced by the architecture of AI-mediated decision systems.

The Awareness-Capability Gap in Reverse

My dissertation research on Algorithmic Literacy Coordination focuses primarily on a particular gap: workers who know that algorithms govern their outcomes but cannot translate that awareness into effective adaptive behavior (Kellogg, Valentine, and Christin, 2020). The Meta case presents the inverse problem. Here, the workers allegedly harmed had no meaningful awareness that an AI system was influencing their employment status at all. The awareness-capability gap assumes awareness exists but does not convert to capability. The Meta situation describes a condition where awareness is precluded by design.

This is an important boundary condition for algorithmic literacy frameworks. Literacy-based interventions assume that making workers aware of algorithmic systems enables them to respond more effectively. But that assumption breaks down entirely when the algorithmic system in question operates within the firm's internal governance processes rather than in a visible, interactive platform environment. A content creator on YouTube can observe engagement patterns and adjust. An employee who has been algorithmically scored for termination has no equivalent feedback loop, no interface, and no opportunity to adapt. The coordination problem here is not one of schema development - it is one of structural opacity by design.

What This Means for Organizational Governance

The harder theoretical question is what kind of governance apparatus is adequate to this problem. Schor et al. (2020) argued that platform dependence produces a specific form of precarity rooted in informational asymmetry between workers and platform operators. The Meta allegations extend that argument into traditional employment, suggesting that precarity is no longer a feature exclusive to gig work. Full-time employees at one of the world's largest technology firms may face consequential algorithmic judgments they have no mechanism to contest.

Hancock, Naaman, and Levy (2020) identified a core tension in AI-mediated communication: the systems that mediate interaction between people also shape what participants believe about each other and themselves. In a hiring and firing context, the mediated signal is not a message but a person's continued employment. The stakes of that mediation are categorically different from the stakes of a content recommendation.

The Meta lawsuit will likely produce a legal ruling about disparate impact under existing disability discrimination frameworks. That ruling, whatever it says, will not resolve the deeper organizational design question: whether firms that delegate consequential employment decisions to algorithmic systems bear full accountability for the distributional outcomes those systems produce. That question requires a theoretical framework that neither employment law nor current platform theory has fully built. The Meta case is a prompt to start building it.

References

Hancock, J. T., Naaman, M., and Levy, K. (2020). AI-mediated communication: Definition, research agenda, and ethical considerations. Journal of Computer-Mediated Communication, 25(1), 89-100.

Kellogg, K. C., Valentine, M. A., and Christin, A. (2020). Algorithms at work: The new contested terrain of control. Academy of Management Annals, 14(1), 366-410.

Rahman, K. S. (2021). The invisible cage: Workers' reactivity to opaque algorithmic evaluations. Administrative Science Quarterly, 66(4), 945-988.

Schor, J. B., Attwood-Charles, W., Cansoy, M., Ladegaard, I., and Wengronowitz, R. (2020). Dependence and precarity in the platform economy. Theory and Society, 49(5-6), 833-861.

The Specific Event

Reporting this week confirms that Amazon is actively pushing to enforce stricter compliance with its automated staffing recommendation systems in warehouse operations. The problem: some warehouse managers are simply overriding the system. Amazon's position is that this override behavior is degrading decision quality and that stricter enforcement is necessary to bring managers back into alignment with algorithmic recommendations. This is not a story about technology failing. It is a story about what happens when an organization deploys algorithmic coordination without first resolving the competence structure that surrounds it.

The Inversion Problem at Amazon's Warehouse Level

Classical coordination theory, whether market-based, hierarchical, or relational, assumes that participants arrive with pre-existing competence relevant to the coordination mechanism being deployed. A manager in a traditional hierarchy is assumed to already know how to interpret staffing signals and make personnel decisions. Amazon's automated staffing system violates this assumption by inverting the competence relationship entirely. The algorithm now holds the relevant structural knowledge about demand patterns, labor efficiency curves, and shift optimization. The human manager has been repositioned as an implementer, not a decision-maker. The resistance Amazon is encountering is a predictable consequence of deploying a coordination mechanism that the workforce was not prepared to inhabit.

This maps directly onto what Kellogg, Valentine, and Christin (2020) identify as one of the central tensions in algorithmic workplace systems: the gap between algorithmic authority and human interpretive capacity. When workers lack the structural schema to understand why an algorithm produces a given recommendation, they do not defer to the algorithm. They override it. This is not irrationality. It is the rational response of agents operating with folk theories rather than accurate structural understanding of the system they are embedded in.

Why Override Behavior Is Diagnostic, Not Disciplinary

Amazon's framing treats override behavior as a compliance failure to be corrected through enforcement. This framing misidentifies the problem. Override behavior is actually a diagnostic signal indicating that the awareness-capability gap is structurally present in the organization. Managers are aware, presumably, that the staffing system is producing recommendations. What they lack is the schema-level understanding of how those recommendations are generated, what variables they optimize for, and when deferring to them produces better outcomes than their own intuitions. Enforcement closes the override loop but does not close the competence gap. The underlying literacy deficit persists.

Hatano and Inagaki (1986) distinguish between routine expertise and adaptive expertise in ways that are directly applicable here. A manager trained procedurally to "accept the system's recommendation unless X condition holds" has developed routine expertise. When novel staffing conditions arise, that routine breaks down and the override instinct reasserts itself. Adaptive expertise, by contrast, requires understanding the structural principles behind the recommendation so the manager can assess when deference is warranted and when genuine exception conditions exist. Amazon appears to be pursuing the enforcement of routine expertise when the situation calls for the development of adaptive expertise.

The Organizational Theory Frame

Rahman (2021) describes algorithmic management as producing an "invisible cage," where workers are subject to constraints they cannot see, interpret, or contest. Amazon's warehouse situation illustrates a specific organizational variant of this: the cage is visible enough that managers know it exists, but opaque enough that they cannot reason about its structure. The result is not compliance and not productive resistance. It is arbitrary override, which is the worst of both possibilities from an organizational efficiency standpoint. The manager neither defers intelligently nor exercises informed judgment. They simply substitute their own heuristic for the algorithm's recommendation without a principled basis for doing so.

Schor et al. (2020) emphasize that dependence and precarity in platform-mediated work environments are partly a function of information asymmetry between the platform and the worker. The Amazon warehouse case extends this insight into an intra-organizational register. The information asymmetry is not between Amazon and an external gig worker. It is between Amazon's algorithmic infrastructure and its own mid-level managers. The organization has created internal information precarity.

What This Means for Coordination Theory

The practical implication here is narrow and specific: enforcement without schema induction will not resolve Amazon's override problem. It will suppress visible override behavior while leaving the underlying competence gap intact, likely producing subtler forms of workaround behavior that are harder to detect and measure. The theoretically interesting implication is broader. Amazon's situation provides a real-world case for the claim, central to the ALC framework, that algorithmic coordination systems do not assume pre-existing competence and therefore generate competence deficits endogenously. The variance in manager behavior across Amazon warehouses is not explained by natural ability differences alone. It is explained by differential schema development in an environment that has not systematically invested in producing it.

References

Hatano, G., and Inagaki, K. (1986). Two courses of expertise. In H. Stevenson, H. Azuma, and K. Hakuta (Eds.), Child development and education in Japan (pp. 262-272). Freeman.

Kellogg, K. C., Valentine, M. A., and Christin, A. (2020). Algorithms at work: The new contested terrain of control. Academy of Management Annals, 14(1), 366-410.

Rahman, H. A. (2021). The invisible cage: Workers' reactivity to opaque algorithmic evaluations. Administrative Science Quarterly, 66(4), 945-988.

Schor, J. B., Attwood-Charles, W., Cansoy, M., Ladegaard, I., and Wengronowitz, R. (2020). Dependence and precarity in the platform economy. Theory and Society, 49(5), 833-861.

The Specific Event

Twenty-six former Meta employees have filed a lawsuit alleging that the company used AI tools to disproportionately target workers on medical leave for inclusion in its recent mass layoffs. As reported by Reuters and subsequently covered across multiple outlets, the plaintiffs claim Meta's algorithmic systems identified which workers to dismiss based partly on leave status, a protected category under federal and state employment law. Meta has not confirmed the specific mechanics of its internal workforce management tools. But the lawsuit itself, regardless of its eventual legal outcome, surfaces a structural problem that organizational theory has not yet adequately addressed: what happens when algorithmic coordination systems are turned inward, from managing platform workers to managing employees?

The Invisible Cage Moves Inside the Firm

Rahman's (2021) concept of the invisible cage describes how platform firms constrain worker behavior through algorithmic monitoring without workers having meaningful access to the logic governing their evaluations. The canonical cases involve gig workers on platforms like Uber or TaskRabbit, where output metrics drive visibility and assignment. The Meta lawsuit suggests that the same structural dynamic now operates inside formally employed, salaried workforces. The employees in question were not independent contractors. They had HR protections, written leave policies, and legal rights. None of that appears to have interrupted the algorithmic signal that marked them as candidates for removal. The cage is no longer metaphorical and it is no longer limited to the platform economy.

Awareness Without Navigability

A central claim of my dissertation research is that algorithmic awareness does not translate to improved outcomes - what I call the awareness-capability gap. This case illustrates a more troubling corollary: in internal algorithmic governance, workers may have zero awareness at all. Kellogg, Valentine, and Christin (2020) catalogued the ways algorithmic management systems at work remain opaque even to the workers they govern. But their analysis focused on behavioral monitoring in frontline and gig settings. The Meta case involves knowledge workers, managers, and professionals who almost certainly had no way to observe that a workforce management algorithm was encoding their leave status as a performance or retention signal. You cannot navigate a constraint you cannot see, and you cannot contest a decision when the decisional logic is proprietary.

Topology of Corporate Algorithmic Power

I have written in previous posts about the distinction between topology and topography in algorithmic environments. Topography is the surface map of a platform's rules and features. Topology is the underlying shape of constraints that persists across surface changes. In the Meta case, the topological structure is a power asymmetry: the firm has complete visibility into worker behavior, status, and signals, while workers have none into how those inputs are weighted. Schor et al. (2020) describe a similar asymmetry in platform labor markets and link it directly to precarity. What this lawsuit reveals is that precarity is no longer a condition unique to gig workers. Formal employment status does not dissolve the topology of algorithmic power; it simply disguises it beneath the vocabulary of HR policy.

The Governance Deficit That Legal Action Cannot Fix

The instinct to litigate is understandable, but lawsuits address specific harms after the fact. They do not produce schema-level understanding of how AI systems are being used in workforce decisions. Sundar (2020) argues that as machine agency increases in communication systems, human accountability structures lag considerably. The Meta case is an instance of that lag becoming legally consequential. The employees who were laid off could not have developed adaptive expertise around a system they did not know existed. There was no folk theory to refine into a structural schema, because there was no available information at all. Gentner's (1983) structure-mapping framework requires that learners have access to at least one analogous case to induce relational structure. Fully opaque systems deny that prerequisite entirely.

What This Means for Organizational Theory

Organizational theory has a well-developed literature on procedural justice and its role in sustaining employee trust and compliance. That literature assumes humans make the consequential decisions, even when those decisions are poorly explained. The Meta lawsuit introduces a different situation: a case where the consequential decision may have been made, or materially shaped, by an algorithmic system that no individual employee designed for that specific purpose and that no individual manager explicitly directed to produce that outcome. Rahman (2021) called this diffusion of accountability a feature, not a bug, of algorithmic control. For researchers studying coordination, the practical question is not only whether AI tools can discriminate, but whether existing organizational governance frameworks have any traction on systems that operate below the threshold of deliberate human choice.

References

Gentner, D. (1983). Structure-mapping: A theoretical framework for analogy. Cognitive Science, 7(2), 155-170.

Kellogg, K. C., Valentine, M. A., and Christin, A. (2020). Algorithms at work: The new contested terrain of control. Academy of Management Annals, 14(1), 366-410.

Rahman, H. A. (2021). The invisible cage: Workers' reactivity to opaque algorithmic evaluations. Administrative Science Quarterly, 66(4), 945-988.

Schor, J. B., Attwood-Charles, W., Cansoy, M., Ladegaard, I., and Wengronowitz, R. (2020). Dependence and precarity in the platform economy. Theory and Society, 49(5-6), 833-861.

Sundar, S. S. (2020). Rise of machine agency: A framework for studying the psychology of human-AI interaction. Journal of Computer-Mediated Communication, 25(1), 74-88.

A Leadership Bet With Structural Stakes

Citigroup is two years into a significant organizational gamble. Tim Ryan, described in recent reporting as a "mild-mannered Bostonian," was tasked with modernizing one of the world's largest banks' technology infrastructure, and by most accounts, the job is nowhere near complete. The framing in the coverage is telling: Ryan's challenge is described not as a technical problem but as an organizational one. Citi is not missing the right software. It is missing the right coordination structure to make its technology investments produce consistent, scalable outcomes.

That distinction matters more than most corporate technology reporting acknowledges. The question is not what tools Citi deploys. The question is whether the organization has developed the internal competence architecture to use those tools effectively and to transfer that competence across divisions, roles, and system transitions.

When Expertise Does Not Travel

Hatano and Inagaki (1986) drew a distinction between routine expertise and adaptive expertise that applies directly here. Routine expertise is the ability to execute known procedures in stable contexts. Adaptive expertise is the ability to recognize when procedures no longer apply and to construct new responses from structural principles. Large legacy financial institutions like Citi are, by design, optimized for routine expertise. Their compliance requirements, regulatory frameworks, and internal auditing processes all reward procedural consistency over adaptive flexibility.

The problem is that technology modernization is not a stable context. The environment changes during the transformation itself. New integration layers introduce dependencies that did not exist when the initial procedures were designed. Workers who are highly competent in the legacy system are often the least equipped to navigate the transitional architecture, not because they lack intelligence, but because their expertise is bound to a topology that is being actively dismantled. This is precisely the transfer failure that Gentner's (1983) structure-mapping theory predicts: surface-level procedural knowledge does not transfer when the relational structure of the new environment differs from the old one.

The Coordination Problem That Precedes the Technical One

The Algorithmic Literacy Coordination framework I am developing treats platforms and algorithmically-mediated systems as environments where competence cannot be assumed to pre-exist. Classical coordination theory, whether through markets, hierarchies, or networks, assumes that actors arrive with sufficient understanding to participate meaningfully. The evidence from platform labor research, particularly Kellogg, Valentine, and Christin (2020), shows that this assumption fails systematically when the coordination mechanism is algorithmic. Workers with identical access show dramatically different outcomes, and the difference cannot be attributed to natural ability alone.

Citi's technology modernization surfaces the same structural problem in a traditional organizational setting. Ryan's engineers, analysts, and operations staff have identical nominal access to the new systems. But access is not competence, and competence is not coordination. What Citi's leadership appears to be navigating is not a technology deficit but a schema deficit: the organization lacks a shared structural understanding of how its new systems relate to one another, which makes it impossible to coordinate responses when those systems behave unexpectedly.

Why General Training Outperforms Specific Training in Transitions

The counterintuitive prediction from my dissertation research is that general schema-induction training, teaching people the structural logic of a system rather than its specific procedures, produces better transfer outcomes than platform-specific procedural training, even when the procedural training produces faster initial performance. That prediction has direct relevance to what Citi is attempting.

If Ryan's team trains workers on the specific workflows of the new infrastructure, those workers will perform well until the next round of changes. If instead the training develops a structural understanding of how data flows, how dependencies form, and how failure propagates across integrated systems, those workers will retain adaptive capacity through subsequent transitions. Rahman's (2021) concept of the invisible cage is useful here: organizations often build procedural training that inadvertently constrains the very adaptability the organization needs. The cage is constructed out of well-intentioned specificity.

What Competence Architecture Actually Requires

Citi's bet on Tim Ryan is ultimately a bet on whether one leader can shift an organization's internal competence architecture without a corresponding shift in how the organization develops and transfers expertise. Based on what has been reported, the approach remains heavily top-down and personality-dependent. That is a fragile structure for a problem that is fundamentally distributed. The variance in outcomes across Citi's technology divisions will continue to look like a personnel problem until the organization recognizes it as a coordination problem, one that requires schema-level investment, not just procedural retraining or executive leadership.

Two years in, the reporting suggests Ryan is competent and the progress is real but incomplete. That trajectory is consistent with what coordination theory would predict when the underlying competence architecture has not yet been restructured to match the new environment.

HDFC Bank's annual report, released Saturday, disclosed that its total workforce fell by more than 3,300 employees over the fiscal year ending March 31, dropping to 211,178 staff as new hires declined by 3,811. The bank attributed the reduction directly to operational automation. This is not a restructuring announcement dressed up as efficiency. It is a clean, documented case of algorithmic systems absorbing tasks that previously required human coordination, and it surfaces a theoretical problem that organizational research has not adequately resolved: when automation reduces headcount, what exactly has been displaced?

The Displacement Is Not About Tasks - It Is About Coordination

The conventional framing treats workforce automation as task substitution. Algorithms perform discrete functions previously performed by workers. That framing is technically accurate but theoretically shallow. What HDFC's numbers actually represent is a shift in the coordination mechanism itself. The bank is not simply replacing workers with software. It is replacing a human-mediated coordination layer with an algorithmic one. This distinction matters because coordination mechanisms carry different competence requirements. Classical coordination theory, whether grounded in markets, hierarchies, or networks, assumes that participants arrive with pre-existing competence relevant to the coordination task (Kellogg et al., 2020). Algorithmic coordination inverts this assumption. The platform determines what competence looks like in real time, and workers who cannot adapt to that determination are not reassigned. They are removed.

What the Awareness-Capability Gap Predicts Here

Research on algorithmic literacy consistently shows that awareness of automation does not translate into improved outcomes for workers subject to it (Gagrain et al., 2024). HDFC's employees presumably knew that automation was expanding within the bank's operations. Indian banking has been a documented site of digital transformation investment for several years. That awareness did not function as protection. The workers who exited the bank's payroll were not ignorant of the trend. They lacked the structural schema - the accurate, transferable understanding of how the automated system was reorganizing the competence requirements around them - that would have allowed them to reposition within the new coordination logic. Folk theory about automation, meaning the general impression that "automation is replacing jobs," is not the same as a structural understanding of which coordination roles become more or less valuable as algorithmic systems absorb routine operations (Kellogg et al., 2020).

Power-Law Outcomes Within a Shrinking Distribution

The variance puzzle that motivates my dissertation research applies here in a compressed form. When a platform or algorithm-dependent organization like a large retail bank automates at scale, it does not produce uniform outcomes across its workforce. It produces a sharper version of the power-law distribution that already characterized performance within the institution. Workers whose competencies align with the new coordination structure - those who can interact productively with algorithmic systems, interpret their outputs, and act on them without procedural scaffolding - retain and in some cases expand their organizational relevance. Workers whose competencies were embedded in the routines that automation absorbed face exit. Rahman (2021) describes this dynamic as the invisible cage: the algorithmic system sets the terms of participation, and those terms are not transparent to the workers subject to them. HDFC's 3,300-person reduction is the observable output of that cage closing around a specific competence distribution.

The Organizational Theory Problem Automation Creates

Standard organizational theory, including the competency frameworks that dominate human resource management research, is built around a relatively stable definition of what a competent employee looks like within a given role. Automation in banking and similar sectors destabilizes that definition faster than organizations can revise their competency frameworks. The result is that organizations are making workforce decisions - who to hire, who to retain, who to let go - using criteria that were calibrated to a coordination environment that no longer exists. Hatano and Inagaki (1986) drew a foundational distinction between routine expertise, which is procedure-bound and context-specific, and adaptive expertise, which is principle-based and transferable. HDFC's reduction is, among other things, evidence that routine expertise becomes a liability in algorithmically reorganized environments. The competence that allowed a bank employee to perform well in a human-mediated coordination structure does not automatically transfer to an algorithmic one.

What This Means Beyond Banking

HDFC is not an outlier. It is an early data point in a pattern that will repeat across industries where algorithmic coordination is expanding. The theoretical implication worth tracking is not whether automation reduces employment - that question is empirically settled in the affirmative for specific task categories. The more consequential question is whether organizations can identify, in advance, which competencies survive the transition and design deliberate pathways for workers to develop them. The evidence from algorithmic literacy research suggests that procedural training targeted at specific automated systems produces brittle competence (Gentner, 1983). Schema-based training, focused on the structural logic of how algorithmic coordination differs from human-mediated coordination, is more likely to produce the adaptive expertise that survives platform shifts. HDFC's annual report will not say any of this. But the 3,300-person gap in its workforce is the data that makes the question urgent.

References

Gagrain, A., Naab, T. K., & Grub, J. (2024). Algorithmic media use and algorithm literacy. New Media & Society.

Gentner, D. (1983). Structure-mapping: A theoretical framework for analogy. Cognitive Science, 7(2), 155-170.

Hatano, G., & Inagaki, K. (1986). Two courses of expertise. In H. Stevenson, H. Azuma, & K. Hakuta (Eds.), Child development and education in Japan (pp. 262-272). Freeman.

Kellogg, K. C., Valentine, M. A., & Christin, A. (2020). Algorithms at work: The new contested terrain of control. Academy of Management Annals, 14(1), 366-410.

Rahman, H. A. (2021). The invisible cage: Workers' reactivity to opaque algorithmic evaluations. Administrative Science Quarterly, 66(4), 945-988.

The Numbers Behind the Announcement

HDFC Bank's annual report, released this past Saturday, contains a detail that deserves more analytical attention than it has received: the bank's total workforce fell by over 3,300 employees to 211,178, with new hires down by 3,811 year-over-year. The official framing attributes this to automation of operations. This is not a layoff announcement, and that distinction matters. The bank did not eliminate jobs directly. It simply stopped replacing them. The headcount declined through attrition accelerated by algorithmic substitution, which is a structurally different process than a discrete reduction-in-force event, and one that tends to escape the scrutiny that a formal layoff would attract.

Attrition as Algorithmic Strategy

What HDFC Bank is executing is a form of displacement that Kellogg, Valentine, and Christin (2020) would recognize as characteristic of algorithmic management: the gradual enclosure of human judgment within automated decision systems, accomplished incrementally rather than through visible rupture. The strategy is organizationally elegant. It produces no single moment of accountability. There is no severance event, no press release about workforce reduction, and no union grievance filed against a specific decision. Instead, the system simply absorbs the functions that departing workers performed, and the headcount number drifts downward in an annual report footnote.

This matters for organizational theory because it complicates how we account for algorithmic labor displacement. The standard framing in the platform economy literature treats displacement as a substitution event: a human task is identified, automated, and the human worker is removed (Schor et al., 2020). HDFC's approach is subtler. The automation precedes the vacancy, rendering the replacement invisible at the individual level. No specific worker can point to a machine that took their role, because the machine absorbed the role category before the next hire could fill it.

The Competence Distribution Problem

From the perspective of my dissertation research on the Algorithmic Literacy Coordination framework, HDFC's situation raises a more specific question: what happens to the remaining 211,178 employees as the task distribution within the bank shifts? Automation of operations does not affect all roles uniformly. It compresses the lower end of the task complexity distribution, eliminating procedural and rule-based work, while leaving or expanding roles that require adaptive judgment, exception handling, and client-facing discretion. This is precisely the distinction Hatano and Inagaki (1986) draw between routine expertise and adaptive expertise.

The workforce that remains after this kind of restructuring is not simply a smaller version of what existed before. It is a workforce whose average task demands have shifted upward in complexity, but whose training history was built for a different distribution of work. The employees who remain were hired and developed under conditions where procedural competencies had value. Now the procedural layer has been automated away, and the residual workforce faces a coordination environment that requires structural understanding rather than procedural recall. Whether organizations like HDFC invest in the schema-level retraining that would support this transition, or whether they simply accept degraded adaptive performance from workers holding competencies that no longer match the task environment, is an empirical question with significant organizational consequences.

The Awareness-Capability Gap at the Institutional Level

There is a version of the awareness-capability gap that operates not at the individual level but at the institutional one. Algorithmic literacy research documents how individual workers can become aware that algorithms govern their outcomes without gaining the structural understanding needed to perform effectively within those constraints (Gagrain, Naab, and Grub, 2024). HDFC's automation announcement suggests something analogous can occur organizationally. The bank clearly has awareness that automation is reshaping its operations - the annual report documents this explicitly. What is less clear is whether institutional decision-making reflects structural understanding of what the post-automation workforce actually needs in order to coordinate effectively, or whether the organization is operating on a folk theory that smaller headcount plus automation equals equivalent output.

Rahman (2021) describes the invisible cage as the condition in which algorithmic systems constrain worker behavior without workers being able to perceive or articulate the source of those constraints. For the remaining HDFC workforce, the cage is not invisible in the sense Rahman describes - most employees will be aware that automation has reshaped their environment. The problem is the gap between that awareness and the structural competency needed to operate effectively within it. Closing that gap requires institutional investment in what the ALC framework calls schema induction: teaching workers the structural logic of the algorithmic systems they now inhabit, not just the procedural steps those systems require. Whether HDFC's headcount reduction was accompanied by that kind of investment is not visible in the annual report number. That absence of visibility is itself diagnostic.

References

Gagrain, A., Naab, T. K., and Grub, J. (2024). Algorithmic media use and algorithm literacy. New Media and Society.

Hatano, G., and Inagaki, K. (1986). Two courses of expertise. In H. Stevenson, H. Azuma, and K. Hakuta (Eds.), Child development and education in Japan. Freeman.

Kellogg, K. C., Valentine, M. A., and Christin, A. (2020). Algorithms at work: The new contested terrain of control. Academy of Management Annals, 14(1), 366-410.

Rahman, H. A. (2021). The invisible cage: Workers' reactivity to opaque algorithmic evaluations. Administrative Science Quarterly, 66(4), 945-988.

Schor, J. B., Attwood-Charles, W., Cansoy, M., Ladegaard, I., and Wengronowitz, R. (2020). Dependence and precarity in the platform economy. Theory and Society, 49(5-6), 833-861.

The Lawsuit as Organizational Signal

Apple filed suit against OpenAI on Friday, alleging that OpenAI's nascent hardware business is, in the company's own framing, "rotten to its core." The specific claim involves theft of trade secrets carried by former Apple employees who moved to OpenAI's hardware division. This is not a generic intellectual property dispute. It is a boundary dispute about where one organization ends and another begins, and what knowledge travels with the people who cross that boundary. That question sits at the intersection of organizational theory, platform coordination, and the increasingly unstable labor markets surrounding AI development.

Knowledge Portability and the Competence Problem

The core tension in Apple's lawsuit is about what kind of knowledge former employees took with them. Trade secret law distinguishes between general competence, which workers are legally permitted to carry, and proprietary information, which they are not. But this distinction is harder to operationalize than it appears. When an engineer spends years developing expertise inside a specific organizational system, their competence becomes entangled with the structural features of that system. The question of what counts as "their" knowledge versus "Apple's" knowledge is not merely legal; it is an organizational and cognitive problem.

This maps directly onto a distinction I work with in my dissertation research. Hatano and Inagaki (1986) differentiated between routine expertise, which is procedural and context-specific, and adaptive expertise, which involves understanding structural principles that transfer across contexts. The irony of trade secret litigation is that it implicitly assumes all high-value expertise is proprietary and context-specific. But if a significant portion of what top engineers know is structural and transferable, then restricting its movement does not protect competitive advantage so much as it distorts the labor market for adaptive expertise.

The Organizational Boundary as an Algorithmic Problem

What makes this case particularly interesting from an organizational theory standpoint is that it involves OpenAI's hardware ambitions, not its software products. OpenAI is attempting to expand its operational scope into physical device manufacturing, a domain that Apple has spent decades optimizing through proprietary supply chains, design processes, and embedded organizational knowledge. The lawsuit signals that Apple views this expansion as a direct threat enabled by illegitimate knowledge transfer.

Rahman (2021) described how platform firms construct "invisible cages" around workers by controlling the informational environment in which work occurs. Apple's lawsuit reverses this dynamic in an interesting way. Here, it is the departing worker who carries embedded system knowledge outward, and the platform firm that is attempting to reassert control over that knowledge after the fact. The organizational boundary, normally reinforced through employment agreements and access controls, is revealed as permeable at the level of human cognition.

What This Reveals About AI Firm Competition

The broader context matters here. OpenAI is not simply building a chatbot company. It is attempting to construct a vertically integrated AI hardware and software ecosystem that would compete directly with Apple's core business model. Apple's lawsuit, timed to OpenAI's hardware push, is better understood as a coordination failure between two organizations that are now in direct structural competition for the same market position. The legal mechanism is trade secret law, but the underlying dynamic is a race to establish organizational competence in a domain where the relevant expertise is scarce and highly concentrated in a small number of individuals.

Kellogg, Valentine, and Christin (2020) observed that algorithmic coordination systems create power-law distributions in outcomes, where small initial differences in competence get amplified over time. The same logic applies to organizational competition in AI hardware. If OpenAI successfully recruits engineers with embedded Apple knowledge, it compresses years of organizational learning into a short window. Apple's lawsuit is an attempt to prevent exactly that compression. Whether the courts will draw the line at trade secrets or whether they will effectively prohibit the transfer of adaptive expertise is an open question with significant implications for how AI firms can compete.

The Practical Implication for Organizational Theory

This case should prompt organizational theorists to revisit the concept of firm boundaries in knowledge-intensive industries. Classical boundary theory, derived from transaction cost economics, treats the firm as a governance structure that internalizes transactions when market mechanisms fail (Williamson, 1981). But in AI development, the relevant boundary is not transactional; it is epistemic. The question is not who owns the contract but who holds the schema. Apple's lawsuit reveals that firms are increasingly aware of this distinction, even if the legal system has not yet developed adequate tools to address it. That gap between organizational reality and institutional response is where the most consequential disputes in AI governance are likely to unfold over the next several years.

The Specific Event

This week, OpenAI announced that Fidji Simo, the company's CEO of AGI Deployment, is stepping down from her role following a significant medical leave. She will remain as a part-time advisor. The departure is notable not because executive turnover is unusual, but because of what the role itself represents. AGI Deployment is not a conventional business unit. It sits at the intersection of technical capability and organizational coordination, responsible for translating frontier model development into structured, governable deployment at scale. Losing the executive accountable for that translation layer is a meaningful structural event, not a personnel footnote.

Why the Role Itself Is the Story

Most coverage of Simo's departure will focus on the timing, her health, or the internal politics at OpenAI. I think those framings miss the more interesting organizational question: what does it mean to coordinate deployment of a system whose capabilities are not fully known in advance? AGI Deployment is not a product management function in the conventional sense. It requires continuous adaptation to a moving technical target. That is precisely the condition under which classical coordination mechanisms - markets, hierarchies, procedural rules - tend to break down.

This connects directly to what I have been working through in my dissertation. Kellogg, Valentine, and Christin (2020) identify algorithmic management as a distinct coordination form, one that cannot be reduced to either hierarchical control or market-based incentives. The problem at OpenAI is a version of the same puzzle at a different level of analysis: when the coordinating system itself is the object of ongoing development, the organizational roles designed to manage it cannot rely on stable procedural knowledge. Simo's position required adaptive expertise in Hatano and Inagaki's (1986) sense - an understanding of structural principles, not just deployment checklists.

The Competence Assumption Problem

OpenAI's organizational structure reveals a tension I find underexamined in the governance literature. The company has built an elaborate institutional apparatus around AGI - safety boards, external advisors, deployment frameworks - that implicitly assumes the humans managing these systems possess ex-ante competence for the task. But AGI Deployment as a function has no established professional template. There is no prior cohort of executives who have done this job at this scale, because the job did not exist five years ago.

This is analogous to what Rahman (2021) calls the invisible cage problem, where platform workers must navigate constraints that are opaque, shifting, and not fully visible even to those nominally in charge. In Rahman's analysis, the workers bear the costs of this opacity. In OpenAI's case, the executive layer bears it. The role of CEO of AGI Deployment requires coordinating across technical, regulatory, and commercial domains simultaneously, with each domain operating on a different timescale and producing different feedback signals. That is an extraordinarily high cognitive load to sustain, particularly without clear precedent for what good performance in the role looks like.

The Simultaneous IPO Filing Context

This leadership transition also arrives as OpenAI and Anthropic have both confidentially filed for IPOs. This timing matters. An IPO process imposes its own coordination demands - investor relations, regulatory disclosure, financial auditing - that sit awkwardly alongside the epistemic uncertainty baked into AGI deployment governance. Public markets require predictable, reportable performance metrics. AGI deployment governance, by contrast, involves managing systems whose behavioral boundaries are probabilistic and contested.

Schor et al. (2020) argue that platform dependence creates structural precarity not just for workers but for the organizational forms built around platforms. OpenAI moving toward public markets while simultaneously navigating a leadership gap in its core deployment function is an instance of this structural tension at the firm level. The coordination demands of being a public company and the coordination demands of deploying frontier AI systems pull in different directions. One rewards stability and predictability; the other demands continuous adaptive revision.

What This Means for Organizational Theory

The Simo departure is worth watching not as a corporate drama but as an organizational design problem that the field has not yet adequately theorized. Classical organizational theory assumes that executive roles are defined by relatively stable task environments. The ALC framework I am developing in my dissertation argues that algorithmically-mediated environments are structurally different because competence cannot be assumed at entry and must develop endogenously through participation. That principle, developed in the context of platform workers and content creators, appears to generalize upward. It applies to the executives nominally in charge of these systems as much as to anyone working within them.

The question OpenAI now faces is not who replaces Simo, but what organizational structure can sustain adaptive coordination in a domain where the ground keeps shifting. That is a theoretical problem before it is a personnel one.

References

Hatano, G., and Inagaki, K. (1986). Two courses of expertise. In H. Stevenson, H. Azuma, and K. Hakuta (Eds.), Child development and education in Japan (pp. 262-272). Freeman.

Kellogg, K. C., Valentine, M. A., and Christin, A. (2020). Algorithms at work: The new contested terrain of control. Academy of Management Annals, 14(1), 366-410.

Rahman, H. A. (2021). The invisible cage: Workers' reactivity to opaque algorithmic evaluations. Administrative Science Quarterly, 66(4), 945-988.

Schor, J. B., Attwood-Charles, W., Cansoy, M., Ladegaard, I., and Wengronowitz, R. (2020). Dependence and precarity in the platform economy. Theory and Society, 49(5-6), 833-861.

What Actually Happened

This week, Microsoft announced the elimination of 3,200 Xbox positions across two fiscal years, alongside the closure of four game development studios. Xbox CEO Asha Sharma has been direct about the reasoning: the division is being restructured around a narrower strategic footprint. Four studios gone, thousands of workers cut, and an implicit question hanging over the entire episode - if Microsoft is walking away from this much of its gaming infrastructure, who would even want what remains? The business press is framing this as an acquisition puzzle. I think that framing misses the more theoretically interesting problem.

Capability Is Not the Same as Headcount

The standard coverage of layoffs in mature technology divisions tends to treat organizational capability as proportional to personnel count. Reduce headcount, reduce capability. This is wrong in a way that organizational theory has documented for decades. What Microsoft is actually shedding is not capability in the abstract but specific coordination structures: the workflows, tacit knowledge networks, and domain-specific routines that allowed those four studios to produce games. Those structures do not transfer cleanly to an acquirer. They dissolve when the teams dissolve.

This connects to a distinction Hatano and Inagaki (1986) drew between routine expertise and adaptive expertise. Routine expertise - the procedural knowledge of how to ship a game on a specific engine, within a specific studio culture, under a specific review process - is extraordinarily difficult to acquire through purchase. You cannot buy the organizational conditions that produced it. What a hypothetical acquirer would actually be purchasing is the brand equity, the IP, and whatever institutional memory survives the transition. The operational expertise embedded in the people who are leaving is gone.

The Coordination Structure Nobody Is Pricing

Kellogg, Valentine, and Christin (2020) argue that algorithmic coordination systems create forms of worker dependence that are structurally invisible to outside observers. The Xbox situation illustrates a non-algorithmic version of the same problem. The coordination structures inside a studio - how producers communicate with engineers, how creative direction gets transmitted and revised, how failure modes get caught before shipping - are invisible to any external party trying to value the asset. They do not appear on a balance sheet. They do not survive an acquisition announcement intact.

This is why the "who would buy it" framing is the wrong question. The right question is what would actually transfer in a sale, and the answer is substantially less than the asking price implies. Rahman (2021) describes how organizational constraints become invisible cages precisely because they are embedded in relational and procedural context rather than formal rules. The inverse applies here: organizational capability is similarly invisible because it lives in the same relational and procedural context. Strip that context through mass layoffs and studio closures, and the asset becomes a shell carrying a brand name.

What This Reveals About Microsoft's Actual Strategic Bet

Microsoft is not restructuring Xbox because it failed to understand gaming. It is restructuring because the coordination costs of maintaining a large, heterogeneous portfolio of studios no longer pencil out against the returns, particularly when AI-assisted development tools are beginning to compress the labor inputs required for certain production tasks. The Cursor field CTO's observation this week - that the AI era makes the two-pizza team rule obsolete because two pizzas is now too much pizza - is not just a clever line. It is a directional claim about how production coordination is changing.

If smaller teams with AI tooling can now approach the output of larger traditional studios, then the marginal value of maintaining sprawling studio infrastructure drops. Microsoft appears to be acting on that premise. Whether that premise is correct is an empirical question the industry will answer over the next several years. But the organizational logic is coherent: you shed coordination overhead when the coordination premium disappears.

The Transfer Problem Remains

What neither Microsoft nor any potential acquirer has solved is the transfer problem. The workers being let go carry schema-level knowledge - structural understanding of how games get made - that is precisely the kind of adaptive expertise Hatano and Inagaki (1986) identified as the durable component of expertise. Procedural knowledge atrophies when separated from its organizational context. Structural knowledge can, in principle, transfer. The individuals leaving Microsoft's studios will land somewhere, and they will bring that structural knowledge with them. The organizational capability, however - the coordinated whole - will not reconstitute elsewhere. That part is lost.

Microsoft is making a calculated bet that the capability it is destroying was less valuable than the coordination costs it was incurring to maintain it. That may be correct. It is also a one-way door.

The Specific Claim and Why It Matters

David Pan, field CTO at Cursor, published an argument this week that Jeff Bezos' famous two-pizza team rule needs revision for the AI era. The original rule held that teams should be small enough to be fed by two pizzas, a heuristic for limiting coordination costs and preserving autonomy. Pan's revision is blunt: in the AI era, two pizzas is too much pizza. The claim is not merely that teams should shrink. It is that AI tooling changes the fundamental unit of productive coordination.

This is a more consequential claim than it first appears. Bezos' rule was never really about headcount. It was a proxy for cognitive overhead, the cost of shared context, synchronization, and inter-personal dependency. Pan is arguing that AI agents absorb enough of that overhead to change the optimal team size. That argument deserves serious theoretical scrutiny before anyone restructures an engineering organization around it.

What the Two-Pizza Rule Was Actually Encoding

The coordination logic behind small teams is well-established. As team size increases, communication links grow at approximately n(n-1)/2, producing superlinear increases in coordination cost (Kellogg, Valentine, and Christin, 2020). Small teams exist not because people work better in proximity, but because shared context degrades with each additional node in the communication network. The pizza heuristic was a memorable encoding of that structural reality.

Pan's argument, as reported, is that AI coding assistants like Cursor compress the execution work previously distributed across multiple engineers. One developer with strong AI tool fluency can now do what previously required three or four. If that is true at scale, then team size should shrink because the coordination problem is being partially offloaded to the human-AI dyad rather than to human-human collaboration.

The problem with this reasoning is that it conflates execution capacity with coordination capacity. Reducing headcount does not eliminate the need for shared context, decision authority, and organizational memory. It concentrates those demands on fewer people, each of whom is now managing a more complex cognitive load through their AI interface. Smaller teams with AI augmentation may have lower communication overhead but higher individual cognitive brittleness.

The Routine-Adaptive Expertise Problem Inside AI-Augmented Teams

This is where the organizational theory gets interesting. Hatano and Inagaki (1986) draw a foundational distinction between routine expertise, the a