Algorithmic Determinism vs. Executive Agency: A Theoretical Examination of Strategic Management in the Age of Artificial Intelligence

Author: Chen Wei

Affiliation: Zhengzhou University, Zhengzhou, China

ORCID ID: 0009-0001-5070-8256

Doi: https://doi.org/10.65326/u7y10032

Submitted 15 August 2026; Revised 16 September 2026; Revised 1 October 2026; Accepted 7 October 2026; Available online 10 October 2026; Version of Record 10 October 2026.

Volume 3, December 2026, (10032)

10032.pdf

Download PDF • 827KB

Abstract

Strategic management settled its dispute between environmental determinism and managerial choice by treating constraint as arriving from outside the firm. Artificial intelligence unsettles that arrangement, because the algorithms that filter markets, rank acquisition targets and evaluate strategic alternatives are installed by executives themselves. This conceptual paper asks where executive agency goes when strategic decisions pass through algorithmic systems. Drawing on upper echelons and managerial discretion research, sociomaterial accounts of technology and evidence on how decision-makers respond to algorithmic outputs, it develops a framework of configured discretion. Algorithmic determinism is theorized as an endogenous, cumulative constraint operating through attention, evaluation and capability. Executive agency is relocated into three practices: framing objectives, contesting outputs and reconfiguring decision architecture. The central argument is that delegation is self-binding. Each act of delegation can erode the capabilities executives need to contest the system later, so formal authority persists while real latitude shrinks. Seven propositions follow, with a state-transition model of four adaptation states (algorithmic capture, governed hybridity, executive voluntarism and fragmented drift) that recasts the two-continua model of organizational adaptation for firms whose constraints are partly self-made. The framework reconciles conflicting findings on algorithm aversion and appreciation and on the value of human override. Governance regimes moderate the loop: fiduciary duties of care and oversight, audit-committee benchmarks against core models and party-state oversight each shape how far effective discretion can fall below formal authority.

Keywords: algorithmic decision-making, executive agency, managerial discretion, strategic choice, upper echelons theory, artificial intelligence, human oversight

1.     Introduction

Kim et al. (2024) studied a public inspections department that piloted two algorithms for prioritizing its inspections. Both delivered substantial gains in prediction. Decisions did not improve. Inspectors retained their authority and used it to override the recommendations, partly to pursue organizational objectives the models did not encode, and the overrides did not produce better outcomes. The episode is small, but it holds the strategic problem of the present moment in miniature. If the algorithm is followed, the organization’s choices are fixed by a model that nobody in the room built or can fully read. If it is overridden, human judgement reasserts itself, and on this evidence it often reasserts itself badly.

Organization theory has argued about this before, in other clothes. After Child (1972) proposed that a dominant coalition chooses structure rather than receiving it from the environment, the field divided over whether managers matter at all. Population ecologists described inert organizations sorted by environmental selection (Hannan & Freeman, 1977), while strategic choice theorists defended the executive as the author of outcomes (Bourgeois, 1984). Hrebiniak and Joyce (1985) largely ended the quarrel by showing that choice and determinism are separate dimensions, so that an organization can face heavy constraint and still exercise real choice within it. Managerial discretion, the latitude of action available to executives, became the variable that governs how far upper echelons shape what their firms do (Hambrick, 2007; Wangrow et al., 2015).

Artificial intelligence reopens that settlement in a form its authors did not anticipate. The determinism they had in mind came from outside: industry structure, scarce resources, institutional pressure. Algorithmic systems produce constraint from inside. Firms buy, train and install the models that then forecast their demand, screen their markets and, with large language models, generate and evaluate strategic alternatives at a level comparable to experienced entrepreneurs and investors (Csaszar et al., 2024; Doshi et al., 2025). Those capabilities change with each model generation, so any statement of their level is dated by the time it appears in print; the argument developed here depends on the direction of that change, not on its present level. The resulting constraint is real. Executives manufactured it.

Some things about this situation are reasonably well established. Machine learning lowers the cost of prediction and raises the relative value of the human judgement needed to weigh outcomes against one another (Agrawal et al., 2019). Automating and augmenting managerial work cannot be cleanly separated, because a task augmented today generates the data that automates it tomorrow (Raisch & Krakowski, 2021). Human and material agency are entangled in practice rather than located on one side of the human–machine boundary (Leonardi, 2011; Orlikowski, 2007), and Murray et al. (2021) have shown how this entanglement changes once technologies can select actions within organizational routines.

The literature stops at an identifiable point. Work on algorithmic management examines how algorithms direct, evaluate and discipline workers, mostly on digital labour platforms (Kellogg et al., 2020; Möhlmann et al., 2021; Noponen et al., 2024). Where it reaches conventional employers, it asks how algorithms shift power between workers and their supervisors (Jarrahi et al., 2021; Keegan & Meijerink, 2025). It casts managers as the party exercising control, not as a party whose own latitude is being reshaped. Upper echelons research, for its part, treats discretion as supplied by the industry, the internal organization and the executive’s disposition (Hambrick, 2007), and its empirical work has concentrated on industry-level sources (Wangrow et al., 2015). The decision infrastructure a firm installs in its own strategy process does not appear among those sources. Two recent studies come close. Mehler et al. (2026) argue that AI reconfigures executive cognition, evaluation and discretion, so that executive influence shifts from making decisions to configuring and governing AI-enabled decision processes. Srinivas and Chetan (2026), drawing on interviews with 33 C-suite executives, find that top teams delegate analytical work to AI when problems are structured and predictable, and co-deliberate with it when they are not. Both treat configuration as a stable resting place for executive agency. Neither asks whether configuring an algorithm today narrows what the same executives can notice, judge or reverse later.

The gap has a cause. Strategic management closed the determinism debate with a model in which constraint was exogenous, so self-produced constraint never entered its vocabulary. Information systems research does theorize technology as both medium and outcome of human action (Orlikowski, 1992), but it has concentrated on frontline and professional work, where what is at stake is operational control rather than strategic direction. The two literatures cite past each other, and the evidence they might share points in opposite directions. People abandon algorithmic forecasters after watching them err (Dietvorst et al., 2015). Yet lay decision-makers weight algorithmic advice above human advice, and experts who discount it lose accuracy (Logg et al., 2019). Overriding algorithms erased their value in the inspections pilot (Kim et al., 2024); among radiologists, only those who interrogated opaque diagnostic tools went on to incorporate their outputs (Lebovitz et al., 2022). Without a theory of where executive agency goes when algorithms enter the strategy process, such findings remain a list of contradictions.

Boards and regulators carry the cost. The European Union’s Artificial Intelligence Act makes human oversight a legal requirement for high-risk systems (European Union, 2024), and corporate governance presumes that an identifiable executive answers for consequential choices. Both assumptions require an executive who can still contest what the system produces. If delegation erodes that capacity, oversight becomes nominal and the responsibility gap first described for learning machines (Matthias, 2004) moves into the boardroom. A second cost is competitive. Advice from a shared model pulls human judgements toward one another (Fügener et al., 2021), and firms that consult the same systems may find their strategies converging without anyone having chosen convergence.

This paper addresses that boundary with a conceptual framework. I argue that algorithmic determinism in strategy is endogenous and cumulative: it works through three channels (attention, evaluation and capability) and deepens with each act of delegation. Executive agency is not displaced. It relocates into three practices, namely framing, contesting and reconfiguring, and it persists only where contestability is deliberately maintained. Seven propositions and a state-transition model of four adaptation states recast Hrebiniak and Joyce’s (1985) model for firms whose constraints are partly of their own making. The propositions distinguish how prediction, optimization and generative systems act on each channel, extend to firms that buy rather than build their models, and treat the governance regime, from fiduciary law and audit committees to party-state oversight, as a moderator of how far effective discretion can fall below formal authority. The paper does not test them.

The argument is theory-building. Sources were identified through searches of the OpenAlex index conducted in August 2026, combining terms for strategic decision-making, upper echelons and managerial discretion, algorithmic management, automation and augmentation, and conjoined human–technology agency, followed by backward reading into the determinism debate of the 1970s and 1980s. Metadata for every retained source was confirmed against Crossref or OpenAlex records. I retained work that either theorizes the locus of agency in organizations or reports evidence on how decision-makers respond to algorithmic outputs, and I gave priority to studies whose findings conflict. Propositions were derived by holding the two-continua model of adaptation fixed and asking which of its assumptions fail once constraint is generated inside the firm. The next section revisits the determinism debate and the literature on technology and agency; the third develops the framework; the fourth discusses implications, limitations and the questions left open.

2.     Theoretical Background: Choice, Constraint and the Machine

2.1.          The Discretion Settlement and Its Exogenous Premise

Astley and Van de Ven (1983) organized the field along two axes, level of analysis and the deterministic or voluntaristic assumptions a theory makes about human action. Their map made visible how differently the perspectives explained change. For population ecology, organizations are too inert to adapt, and change at the population level occurs through the selection of fitting forms (Hannan & Freeman, 1977). For strategic choice, managers enact their environments, select domains and design structures (Child, 1972). Bourgeois (1984) defended the second view against deterministic organization theory, arguing that managerial choice remains central to explaining what organizations become.

Hrebiniak and Joyce (1985) dissolved the opposition. Treating choice and environmental determinism as independent continua, they derived four types of adaptation: natural selection where choice is low and determinism high, differentiation where both are high, strategic choice where choice is high and determinism low, and undifferentiated choice where both are low. The typology mattered because it allowed both literatures to be right about different firms, and about the same firm at different times.

Upper echelons theory supplied the agentic half of that bargain with content. Organizations, Hambrick and Mason (1984) proposed, reflect the values and cognitive bases of their top managers. Hambrick’s (2007) update made the relationship conditional on discretion and on the demands of the executive job, so that executive characteristics shape outcomes most where latitude is widest. Reviewing three decades of discretion research, Wangrow et al. (2015) observed that most studies measure discretion at the industry level and give little attention to the internal organization as a source of latitude or constraint.

That observation exposes the premise on which the settlement rests. In every version, the deterministic pole sits outside the executive’s reach: in markets, in institutions, in the inertia of structures inherited from predecessors. Mintzberg and Waters (1985) allowed that realized strategy blends deliberate intentions with emergent patterns nobody planned, but they located emergence in the actions of people throughout the organization. None of these accounts anticipated constraint that executives install deliberately, for good reasons, in the decision process itself.

2.2.          Technology as Medium and Outcome of Agency

The information systems literature has long resisted treating technology as either a neutral tool or an autonomous force. Orlikowski (1992) described technology as physically constructed by designers and socially constructed by users, a product of human action that then enables and constrains further action. Her later work on sociomaterial practice went further, arguing that the social and the material are constitutively entangled in everyday organizing (Orlikowski, 2007). Leonardi (2011) offered a more analytically tractable version: human and material agencies remain distinct but become imbricated, as people who perceive a technology as constraining change their routines, and people who perceive it as affording new possibilities change the technology.

These accounts were built around technologies that executed procedures. Learning algorithms differ in kind. They are opaque, comprehensive in the data they absorb, and performative in the sense that their outputs shape the reality they model (Faraj et al., 2018). Their opacity comes in three forms: corporate secrecy, the technical illiteracy of users, and the scale and mathematics of machine learning itself (Burrell, 2016). Berente et al. (2021) characterize managing AI as managing a moving frontier of autonomy, learning and inscrutability. Murray et al. (2021) carry this into organizational routines, distinguishing technologies that assist, arrest, augment or automate human action according to whether the human or the technology selects protocols and actions. Anthony et al. (2023) propose treating AI as a counterpart within a wider system of design, implementation and use, which moves the analysis from the individual user to the arrangements that surround the technology.

Two further strands sharpen the problem for strategy. Lindebaum et al. (2020) argue that algorithms act as supercarriers of formal rationality, embedding calculative logic so thoroughly in decision processes that alternative grounds for choice, including moral ones, lose standing. The attention-based view holds that firm behaviour follows from how organizational structures channel the attention of decision-makers (Ocasio, 1997). Algorithms that sort, flag and recommend are now among those structures. What an executive attends to, and so what counts as a strategic option at all, increasingly passes through them.

Agency itself needs a precise definition if its fate is to be traced. Emirbayer and Mische (1998) describe human agency as temporally embedded, with an iterational element that draws on past patterns, a projective element that imagines future trajectories, and a practical-evaluative element that judges among options in the present. The definition matters here because algorithms engage each element differently. Models trained on historical data are, in a sense, pure iteration. Executives contribute most where projection and evaluation are required, which is where Agrawal et al. (2019) locate the residual value of human judgement.

2.3.          Contradictory Evidence on Algorithms in Judgement

The empirical literature on algorithmic advice is large, and on five questions that matter for strategy it disagrees with itself. Table 1 sets out the disagreements and anticipates how the framework reconciles them.

Table 1

Contested Findings on Algorithmic Judgement and Their Reconciliation in the Configured Discretion Framework

Contested question	Evidence pointing one way	Evidence pointing the other way	Reconciliation in this framework
Do decision-makers resist or embrace algorithmic advice?	Confidence in algorithms collapses after errors are observed (Dietvorst et al., 2015)	Lay users prefer algorithmic to human advice; experts who discount it lose accuracy (Logg et al., 2019)	Response depends on error visibility; delayed strategic feedback favours deference (P2)
Does human override add or destroy value?	Inspectors’ overrides erased algorithmic prediction gains (Kim et al., 2024)	Radiologists who interrogated AI outputs were the ones who incorporated them into their judgements (Lebovitz et al., 2022)	Override adds value only after interrogation drawing on independent information, whose cognitive load rises with model opacity (P4)
Can AI evaluate strategic alternatives?	Aggregated LLM evaluations approximate expert rankings (Csaszar et al., 2024; Doshi et al., 2025)	AI can amplify problems in decision processes and displace substantive reasoning (Lindebaum et al., 2020; Trunk et al., 2020)	Point-in-time accuracy coexists with narrowing of options over time (P1, P6)
Should managerial tasks be automated or augmented?	Optimal allocation depends on between-task and within-task complementarity (Fügener et al., 2026)	Automation and augmentation are interdependent and feed reinforcing cycles (Raisch & Krakowski, 2021)	Augmentation drifts toward de facto automation as capabilities erode (P3)
Does algorithmic control remove or relocate agency?	Algorithmic management intensifies control, a digital Taylorism (Kellogg et al., 2020; Noponen et al., 2024)	Workers respond with market-like and organization-like tactics and resistance (Möhlmann et al., 2021)	Agency survives as contestation; for executives this depends on decision architecture (P5)

Note. P1–P7 refer to the propositions developed in the Framework Development section. Evidence in the second and third columns is summarized from the cited studies; the fourth column states the theoretical interpretation advanced in this paper.

The first disagreement concerns aversion and appreciation. Dietvorst et al. (2015) found that participants abandoned an algorithmic forecaster after seeing it make errors, even when it outperformed humans. Logg et al. (2019) found the opposite tendency among lay participants, who weighted algorithmic advice more heavily than identical human advice, while experienced forecasters discounted it and became less accurate. Glikson and Woolley’s (2020) review suggests that trust in AI depends heavily on how the system is represented and on its transparency and reliability, which implies that neither tendency is fixed.

The second concerns override. In Kim et al. (2024), discretion over algorithmic recommendations destroyed most of their value. Lebovitz et al. (2022), studying radiologists, found that opacity raised uncertainty, and that professionals who developed practices for interrogating AI outputs incorporated them into their diagnoses while the others set them aside. The third disagreement concerns whether AI can evaluate strategy at all. Doshi et al. (2025) found single large language model evaluations of business models inconsistent and biased, yet aggregated evaluations resembled expert rankings; Choudhary et al. (2025) identify conditions under which ensembles of human and AI judgements outperform either. Trunk et al. (2020), reviewing AI in strategic decision-making under uncertainty, conclude that AI can amplify problems inherent in the decision process and raise rather than lower the burden of human responsibility.

These disagreements are less contradictory than they look. Each study observes a single moment or a single type of decision. None observes how a top team’s capacity to evaluate, contest and redesign changes after it has relied on an algorithm for several years. That is the variable the framework places at its centre.

Most of this evidence comes from below the C-suite, and using it to theorize executive agency needs a justification, because strategic decisions are rarer, larger and more political than inspections or diagnoses. The studies are drawn on for the processes they document, not for their effect sizes. Each process depends on structural features that top teams share with the inspectors, radiologists and administrators studied: decision authority that formally rests with a person (Kim et al., 2024), outputs that the person cannot fully inspect (Lebovitz et al., 2022), and expertise that survives only through practice (Rinta-Kahila et al., 2023). Research on human supervision of automated systems treats these as general conditions rather than occupational ones. Complacency and automation bias share attentional roots, appear among experts as well as novices, and are not removed by training alone (Parasuraman & Manzey, 2010). The transfer to executives therefore rests on a claim about the conditions of the work, and it has to be checked against each way the C-suite differs.

Rarity strengthens the processes. Bainbridge (1983) observed that automation leaves human operators with the rare, abnormal cases, precisely the cases for which their skills are least practised. Strategic decisions are the abnormal cases of organizational life, made a handful of times a year, so executives have fewer occasions to calibrate their trust and fewer chances to keep analytical skill in use. Their outcomes also arrive later and resist attribution. Size cuts both ways. Large stakes raise scrutiny at the moment of decision, but scrutiny depends on the capacity to scrutinize, and that capacity is what Proposition 3 below says delegation erodes; high stakes raise the demand for contestation without raising its supply. Top teams are also small, so one member’s lost analytical skill removes a large share of the team’s capacity to contest.

Politics runs the other way. Strategic decisions involve coalitions, and contestation can stay alive for reasons that have nothing to do with analysis. Executives can also commission independent analysis, an option a radiologist reading a scan in real time does not have. Both are treated in the framework as conditions under which self-binding slows, and the analysis of time pressure, external frictions and shared authority below takes them up directly. The executive-level evidence that exists shows the precondition for the argument already in place: top teams report delegating structured analytical work to AI (Srinivas & Chetan, 2026), and AI literacy in top management predicts how far firms take AI up (Pinski et al., 2024). Whether erosion follows is what the propositions are written to test.

3.     Framework Development: Configured Discretion

The argument concerns a bounded class of systems. By algorithmic systems I mean learning-based models that a firm uses within its strategy process, in three architectures: supervised prediction and ranking, optimization over learned parameters, and generative language models. They are applied to environmental scanning, forecasting, the screening of investment, acquisition and market-entry options, and the evaluation of strategic alternatives. What unites them is the combination Berente et al. (2021) identify: some autonomy in producing outputs, learning from data, and inscrutability to those who rely on them. What divides them is how they constrain, and a later subsection takes the three architectures in turn. Three classes fall outside the argument. Rule-based automation that executes a procedure fixed in advance does not learn, so it cannot change what executives know without their noticing. Algorithmic management of workers is excluded as a phenomenon, although its findings are used as evidence. Fully automated operational decisions, such as pricing within bands set by management, are excluded unless their outputs feed the strategic agenda.

The framework rests on two definitions. Algorithmic determinism in strategy is the degree to which the set of strategic alternatives a firm considers, the evaluation of those alternatives and the selection among them are constrained by the outputs and architecture of algorithmic systems. It has a temporal face as well as a substantive one. Systems produce outputs at a speed that sets the tempo for everyone who must respond to them, and once rivals act on recommendations produced in seconds, the time available for human contest shrinks whatever the firm intends. Executive agency is the temporally embedded capacity of top managers to frame strategic problems, contest the outputs presented to them and redesign the processes that produce those outputs. Following Hrebiniak and Joyce (1985), the two are treated as separate dimensions at any single moment. The departure from their model is that the dimensions are coupled over time: agency exercised as delegation raises later determinism, and determinism, through the capability channel, lowers later agency. The framework is therefore a dynamic one. Its propositions describe the forces that move a firm through the space defined by the two dimensions, and a state-transition model at the end of the section describes where those forces lead. Figure 1 summarizes the argument.

Figure 1

The Configured Discretion Framework

Note. Straight arrows show the three channels through which delegation narrows effective strategic latitude; the attention channel narrows the range of alternatives independently of predictive accuracy (P1). The dotted arrow shows executive practices moderating the channels, and the curved solid arrow shows executives setting the scope of delegation. The dashed arrow is the self-binding loop (P6). The governance regime moderates the loop (P7), and independent analytical benchmarks held by an audit committee interrupt it.

3.1.          Algorithmic Determinism as Manufactured Constraint

The first channel runs through attention. Before anything can be chosen, it must be noticed. Algorithmic systems now perform much of the scanning on which strategic attention depends: they monitor markets, flag anomalies, rank opportunities and suppress what they score as noise. Kellogg et al. (2020) list restricting and recommending among the core functions of algorithmic control, and those functions apply to information flowing up an organization as much as to instructions flowing down. In a 31-month ethnography of predictive policing, Waardenburg et al. (2022) observed intermediaries who began by passing algorithmic outputs along and ended by curating them, until their own judgement had in effect replaced the algorithm’s. Policing operates under a logic of public accountability and legal constraint that corporate strategy does not share, so the case is used here for the brokerage process it documents, which depends on an information asymmetry between those who read the system and those who receive its outputs, rather than for its institutional setting. Filtering of this kind is never neutral. It carries the priorities of whoever designed or curated it, and it removes alternatives before any executive has the chance to reject them. Lindebaum et al.’s (2020) argument about formal rationality sharpens the point. A system that carries calculative logic passes upward what can be quantified and scored. Considerations that resist scoring, such as a partner’s trustworthiness or a community’s tolerance for a new plant, drop out before anyone has judged them unimportant.

Proposition 1: The greater the share of a firm’s environmental scanning that is routed through algorithmic filtering and ranking, the narrower the range of strategic alternatives its executives consider, independent of the predictive accuracy of the filtering.

The independence clause carries the argument. A filter can be accurate about what it measures and still exclude options that fall outside its training distribution, which is where strategic novelty tends to lie. Krakowski et al. (2023) showed, in the domain of chess, that AI adoption rendered previously valuable human capabilities obsolete while new human–machine capabilities, unrelated to the old ones, became the source of advantage. If that pattern holds in markets, a filter tuned to past sources of advantage will screen out the next ones.

The second channel runs through evaluation. Once alternatives are on the table, algorithms increasingly score them, and generative AI has extended this from structured forecasting to the evaluation of business models and strategies (Csaszar et al., 2024; Doshi et al., 2025). Whether executives defer to such evaluations depends, on the evidence in Table 1, on whether the algorithm’s errors are visible. Most strategic decisions are poorly suited to producing visible errors. They are infrequent, often irreversible, and judged on outcomes that arrive years later and are confounded by everything else that happened in the interval. The error signal that triggers aversion in laboratory settings (Dietvorst et al., 2015) rarely reaches a top team in usable form. What the team does observe is a recommendation that is fluent, quantified and available on demand.

Proposition 2: In strategic decisions characterized by delayed and ambiguous outcome feedback, a top team’s deference to algorithmic evaluations increases with the duration of its own use of those evaluations, because the observable errors that produce algorithm aversion seldom occur.

The proposition has a boundary. Some decisions with strategic consequences produce fast and attributable feedback, as when an airline’s revenue-management system adjusts fares and observes bookings within hours. There, errors recur and can be measured, and deference should track measured performance rather than duration of use. Proposition 2 applies at the other end of the feedback spectrum, to entry, acquisition, divestment and portfolio decisions, which are the core of strategic choice and the decisions least able to correct deference through experience.

The third channel runs through capability, and it is the slowest. Teece (2007) disaggregates dynamic capabilities into sensing, seizing and reconfiguring, and Helfat and Peteraf (2015) trace each to managerial cognitive capabilities: perception and attention underpin sensing, problem-solving and reasoning underpin seizing, and language, communication and social cognition underpin the reconfiguring that requires others to follow. Capabilities of this kind are maintained by use. Rinta-Kahila et al. (2023), in a case study of cognitive automation, traced a vicious circle in which reliance on an automated system bred complacency at both individual and organizational levels, eroding the skills needed to notice when the system went wrong. Balasubramanian et al. (2022) argue that substituting machine learning for human decision-making can reduce the diversity of judgement on which organizational learning depends. Fügener et al. (2021) found experimentally that AI advice made individual judgements converge, raising individual accuracy while destroying the unique knowledge that made the group wiser than its members. Drawing on the Carnegie School, Haefner et al. (2021) set out how far machine learning can take over the search and problem-solving steps of innovation, the very activities through which those capabilities are exercised.

The capability channel strikes the three unevenly, and the dependency among them matters. Delegated scanning substitutes most directly for perception and attention, and delegated evaluation for problem-solving and reasoning. Communication stays in executive hands, because persuading an organization to act remains an executive task. Its fluency survives. Its substance does not, because a strategic argument is only as good as the reasoning beneath it. As reasoning erodes, communication hollows into the presentation of conclusions the speaker can no longer reconstruct, a kind of corporate theatre that persuades without explaining. The hollowing is easy to miss, and generative systems make it easier still, since they can supply the fluent rationale that the executive no longer produces. What boards and investors observe is the part of strategic capability that erodes last and shows its erosion least. Detecting the hollowing is possible, but not by reading the output, since a generated rationale can be indistinguishable on the page from one an executive reasoned through. Detection has to test the person rather than the document. Boards can ask executives to reconstruct a recommendation’s reasoning without the system to hand, to say what evidence would change their view, and to explain why the rejected alternatives were rejected. Reasoning that was worked through survives such questioning and varies sensibly with the question; reasoning that was borrowed tends to repeat its original phrasing or fall silent. Disclosure of where generative tools drafted strategy papers would help, as would asking for the analyses on which a recommendation rests rather than its summary.

Proposition 3: Sustained delegation of analytical and evaluative tasks in strategic decisions erodes the managerial cognitive capabilities required to contest algorithmic outputs in an ordered progression. (a) Perception and attention, which support sensing, erode first, as scanning is delegated. (b) Problem-solving and reasoning, which support seizing, erode next, as evaluation is delegated. (c) Communication, which supports reconfiguring, retains its fluency longest but loses its substance as reasoning erodes, which conceals the erosion from those who monitor the top team. The cumulative effect is a reduction in executives’ effective discretion even where their formal decision authority is unchanged.

The distinction between formal and effective discretion is what upper echelons research has lacked the means to see. A chief executive who signs every acquisition retains the authority to refuse one. If the team that once built the valuation models now reviews the outputs of a system it cannot reconstruct, the authority to refuse has outlived the capacity to refuse well.

3.2.          Architecture Matters: Prediction, Optimization and Generation

The three architectures in scope learn and constrain differently, and treating them as one would hide the variation that empirical tests need. Table 2 sets out how each acts on the three channels.

Table 2

How Three Algorithmic Architectures Act on the Channels of Algorithmic Determinism

Architecture	Attention channel	Evaluation channel	Capability channel	Typical strategic uses
Supervised prediction and ranking	Narrows by filtering: scores options against patterns in historical data and discards outliers outside the training distribution	Supplies precise point estimates that invite deference; errors become visible only where outcomes recur	Erodes estimation and pattern recognition in screening and forecasting	Demand forecasting, target screening, risk scoring
Optimization over learned parameters	Fixes the agenda through the objective function: options that cannot be expressed in it never appear	Collapses evaluation into a single ranked solution, hiding the trade-offs among objectives	Erodes the practice of weighing objectives against each other, which is the substance of framing	Capacity and network planning, portfolio allocation, pricing
Generative language models	Floods rather than filters: produces many fluent alternatives, including fabricated ones that never existed	Evaluates through fluent narrative; single evaluations are inconsistent, and persuasiveness can stand in for evidence	Erodes problem formulation and the working-through of arguments; communication hollows	Scenario drafting, strategy memos, generation and evaluation of business models

Note. Entries are theoretical interpretations grounded in the cited literature: opacity of learning systems (Burrell, 2016), hallucination in language generation (Ji et al., 2023), inconsistency and bias in single LLM evaluations of strategic alternatives (Doshi et al., 2025), cognitive offloading (Risko & Gilbert, 2016), and homogenization of generated content and strategic recommendations (Doshi & Hauser, 2024; Stoeber et al., 2026).

Supervised prediction is the architecture Proposition 1 describes most directly. It narrows attention by discarding what looks unlike the past. Optimization narrows earlier and more completely, because an option that cannot be written into the objective function is never generated at all, and the single ranked solution it returns hides the trade-offs that framing is supposed to settle. Generative language models work the other way. They expand the apparent option set with fluent alternatives, some of which are fabrications that sound as plausible as the real ones (Ji et al., 2023). Attention is flooded rather than starved, and the scarce resource becomes the capacity to tell a real option from an invented one. Their evaluations are inconsistent and biased taken singly, although aggregated evaluations come closer to expert rankings (Doshi et al., 2025). Their outputs also converge across users: access to generated ideas made individual work more creative but more similar (Doshi & Hauser, 2024), and in a comparison of 2,400 model recommendations with the choices of 218 managers, language models favoured acquiescence to institutional pressure and compressed the diversity of strategic responses (Stoeber et al., 2026).

Capability erosion differs accordingly. Reliance on prediction offloads estimation; reliance on optimization offloads the weighing of objectives; reliance on generation offloads the formulation of problems and the drafting of arguments. People use external tools to reduce internal cognitive demand, with consequences for what they later remember and can do unaided (Risko & Gilbert, 2016). Generative systems invite the broadest offloading, because they take over the work of writing through a problem, which is where much strategic reasoning takes place. The ordering in Proposition 3 should therefore be fastest for generative tools, and the hollowing of communication most pronounced where language models draft the executive’s rationale.

3.3.          Who Erodes: Turnover, Entrenchment and Team Composition

Propositions 2 and 3 describe a top team, yet top teams change. A new chief executive or a reconstituted team arrives without a history of deferring to the firm’s systems, so the clock in Proposition 2 restarts with them. Capability erosion does not restart so easily, because much of it is not stored in individual heads. Organizational memory is held in routines, structures and records as well as in people (Walsh & Ungson, 1991), and routines carry an ostensive aspect, the shared understanding of how a task is done, that persists across the people who perform it (Feldman & Pentland, 2003). A new team inherits the decision architecture, the data pipelines, the vendor contracts and the analysts whose skills were allowed to lapse. Turnover therefore resets deference faster than it restores capability.

The inheritance is not neutral. An architecture encodes the objectives, data choices and thresholds of the team that commissioned it, and so carries that team’s strategic paradigm forward in machine form. Hambrick and Fukutomi (1991) describe chief executives as increasingly committed to an enduring paradigm over their tenure, while longer team tenure is associated with more persistent strategy (Finkelstein & Hambrick, 1990). An architecture built late in a long tenure fixes that paradigm in the routines a successor must work through. Quigley and Hambrick (2012) found that a predecessor who stays on as board chair restricts the successor’s discretion and dampens strategic change; an inherited architecture can do the same without a predecessor present. It functions as an entrenchment of the prior paradigm.

Incoming executives can navigate the inheritance in three ways, each corresponding to a relation with the ostensive aspect of the routine. They can perform it as received, accepting the outputs as the way decisions are made here. They can perform it selectively, using the outputs while working around them. Or they can rewrite it, changing the objectives, data and points of contest. Hambrick and Fukutomi (1991) place experimentation early in a tenure, and the early season is when a rewrite is most feasible, because the new team has not yet come to depend on the system and still has the mandate to change it. A rewrite also requires a capability the inherited organization may lack, which is why incoming teams that bring analytical staff with them should be better placed than those that rely on the staff they find.

The difficulty of a rewrite follows from what the strategy literature calls momentum. Organizations tend to keep moving in the direction they are already moving, with long periods of continuity broken only occasionally by revolutionary change (Miller & Friesen, 1980), and a firm’s past changes raise the likelihood that it will repeat changes of the same kind (Kelly & Amburgey, 1991). An inherited objective function is momentum made executable. Once it is wired into daily operations, it is invoked thousands of times without anyone deciding to invoke it, and every invocation produces data that confirms the patterns it was built to find. Dismantling it is therefore less a single decision than a sequence. The incoming team first has to make the objective explicit, since an objective function that no one has written down in plain language cannot be argued with. It can then run a challenger objective in parallel, scoring the same decisions under the old and new specifications and comparing the results before either governs action. Decision rights over the objective itself, as distinct from decisions taken under it, need to move to the top team, and the old specification needs a date on which it lapses unless deliberately renewed. Each step converts momentum that runs by default into a choice that has to be remade.

Erosion is also uneven within a team. Top team heterogeneity in tenure, functional background and age shapes how quickly firms act and change strategy (Hambrick et al., 1996; Wiersema & Bantel, 1992), and it should shape the speed of collective erosion. Members with AI literacy and digital knowledge (Firk et al., 2022; Pinski et al., 2024) are best placed to interrogate models. They may also defer more readily, much as lay users in Logg et al. (2019) weighted algorithmic advice heavily. Industry veterans hold the domain knowledge needed to notice what a filter has excluded, but they may lack the literacy to interrogate the system, and experienced judges in the same study discounted algorithmic advice to their cost. Neither profile protects a team on its own. The protective configuration is one in which domain expertise and technical literacy sit in different members who are obliged to confront each other’s readings of the same output. Obligation is the operative word. Confrontation that depends on goodwill weakens precisely when a model’s output arrives looking authoritative, which is why a formal structure helps. In experiments on strategic decision-making, groups using dialectical inquiry or devil’s advocacy produced higher-quality recommendations than groups seeking consensus (Schweiger et al., 1986). A rotating role whose holder is charged with arguing against the model’s recommendation, using evidence from outside the system, applies the same logic to algorithmic consensus. Where interrogation depends on a single technically literate member, team-level contestation rests on one person, who can end up curating what everyone else sees, as the brokers in Waardenburg et al. (2022) did.

3.4.          Executive Agency Relocated: Framing, Contesting and Reconfiguring

If agency does not disappear, where does it go? The framework identifies three practices, each corresponding to a temporal element of agency as defined by Emirbayer and Mische (1998). A caution against dualism comes first. The sociomaterial argument with which this paper began (Orlikowski, 2007) implies that executives never step outside the system to act upon it. They frame through dashboards and model documentation, contest through the explanations the system provides, and reconfigure within the options a platform allows. The practices are themselves technologically mediated, and the AI shapes the capacity to act on it. What varies is the degree of mediation: how far each practice draws on resources independent of the system it addresses. Contesting a model through its own explanations is heavily mediated; contesting it with an analysis produced separately is less so. This degree of independence, rather than a clean separation of human from machine, is what the framework treats as the condition for agency. Independence has a price that rises over time. A system designed to absorb analytical work removes the occasions on which that work is practised, so maintaining a parallel capacity means deliberately doing by hand, at higher cognitive cost, what the system would do instantly (Risko & Gilbert, 2016; Sweller, 1988). The effort is an investment against the capability channel, and like most such investments it is easiest to cut when the system appears to be working well, which is exactly when its absence will not be noticed.

Framing is projective. It consists of specifying what a strategic decision is for: the objectives, the trade-offs, the time horizon, the constraints the firm will not cross. Algorithms optimize objectives; they do not choose them. Agrawal et al. (2019) locate the enduring value of human judgement in determining the payoffs attached to predictions, and the inspections pilot illustrates what happens when that work is left implicit. Inspectors overrode the models partly because the models ignored organizational objectives the inspectors cared about (Kim et al., 2024). Framing would have written those objectives into the system before it ran.

Contesting is practical-evaluative. It covers interrogation of a model’s inputs, assumptions and failure modes, and the decision to override, adjust or accept. The contradiction in Table 1 between Kim et al. (2024) and Lebovitz et al. (2022) dissolves once override is distinguished from contestation. Overriding on unexamined intuition replaces one uninspected judgement with another. Overriding after interrogation brings information the model lacks into the decision. Weick et al. (2005) describe sensemaking as retrospective and action-oriented, a matter of noticing and labelling cues before deciding what they mean, and contestation is sensemaking applied to machine outputs. Interrogation has a cognitive cost. Problem-solving under high cognitive load leaves little capacity for learning (Sweller, 1988), and opaque neural models impose heavy load on anyone who tries to reconstruct their reasoning. Explanations can reduce the load, but human explanation is contrastive and selective (Miller, 2019), and explanations supplied by the system change how users weigh information, sometimes inducing confirmation bias (Bauer et al., 2023). The more an executive’s interrogation relies on the system’s own explanations, the more mediated it is in the sense defined above.

Proposition 4: Executive overrides of algorithmic recommendations improve strategic decisions when they are preceded by interrogation that draws on information independent of the system’s own explanations, and degrade them when they rest on unexamined intuition.

Interrogation also takes time, and strategic agility is valuable. The trade-off is real but narrower than it appears. Eisenhardt (1989) found that fast strategic decision-makers used more real-time information and considered more alternatives, not fewer, and sought advice from experienced counsellors, so speed and thoroughness were not opposed in her firms. The cost of contestation falls if most of it is done before decisions are needed: standing interrogation of a model’s assumptions, failure modes and validation record, performed between decisions, leaves only a short check at the moment of choice. The remaining time cost can be rationed by reversibility, with reversible decisions proceeding on lightly contested recommendations and irreversible ones, such as acquisitions or exits, receiving the full procedure. Bainbridge’s (1983) irony applies here in strategic form: the decisions that arrive under the greatest time pressure are often the ones for which an unpractised team is least able to take over from the system.

Contestation at the top has a resource that frontline contestation lacks. Firms are coalitions whose goals are negotiated among members with different interests (Cyert & March, 1963), and strategic decisions in top teams are political as well as analytical (Eisenhardt & Bourgeois, 1988). Coalition-building buffers algorithmic determinism, because a recommendation must be sold to the people whose support it needs, and those people bring objectives and information the model does not encode. A division head who will lose resources under the recommended allocation will contest it whatever the model says. The buffer has two limits. Political contest can be as uninterrogated as intuitive override, protecting interests rather than improving decisions. And a recommendation that arrives with the authority of a quantified model shifts the burden of proof onto those who object, which weakens coalitions that cannot produce counter-analyses of their own. Politics slows determinism most where coalition members retain independent analytical capacity, which returns the argument to the degree of mediation.

Reconfiguring needs a fuller statement. In Emirbayer and Mische’s (1998) account, the iterational element of agency is the selective reactivation of past patterns of thought and action, the habits that let actors proceed without deliberating anew. Algorithms trained on historical data perform a version of this reactivation at scale, and once embedded in a decision routine they fix which past patterns are carried forward. Routines have an ostensive aspect, the shared understanding of how the task is done, and a performative aspect, the specific performances that enact it (Feldman & Pentland, 2003). Reconfiguring is the deliberate change of the ostensive aspect of a decision routine: who sees an output, at which stage, with what data, and with what right to stop it. It is projective agency aimed at the iterational element. By changing the pattern that will later be performed without reflection, executives re-anchor the habits of their successors and of their own future selves.

Information systems research gives the design choices concrete form. Shrestha et al. (2019) distinguish full delegation to AI, sequential hybrid structures and aggregated structures, and Puranam (2021) treats the division of labour between humans and algorithms as an organization design problem. Murray et al. (2021) distinguish augmenting technologies from arresting ones, which halt a routine until a human acts. A useful further distinction separates a human in the loop, whose action is required before a decision proceeds, from a human on the loop, who supervises a process that proceeds unless stopped. Grønsund and Aanestad (2020) found that AI adoption in one organization produced human-in-the-loop configurations in which people audited and altered the algorithm rather than being replaced by it. Points of human contest are in-the-loop positions placed where objectives are set and alternatives are eliminated. Design choices of this kind are often made by technologists rather than by executives, and they are shaped by internal politics: in an airline studied by Giannitsas et al. (2026), some parts of the organization pushed AI toward automation and others toward augmentation, depending on which institutional logic shaped how actors configured their trust in the systems. Pachidi et al. (2021) showed how symbolic conformity by incumbents and symbolic advocacy by technologists reinforced one another until an organization’s way of knowing its customers had been replaced, without anyone having decided to replace it.

Proposition 5: The effect of algorithmic adoption on executive agency depends less on the sophistication of the algorithm than on whether the decision architecture preserves in-the-loop points of human contest at the stages where strategic objectives are set and alternatives are eliminated.

Kim et al. (2024) report a finding that supports the first half of this proposition: simple data-driven heuristics captured most of the available gains, and additional sophistication added little. What determined the value captured was how decision authority was managed.

3.5.          The Self-Binding Loop: From Augmentation to De Facto Automation

The three channels and the three practices are linked over time, and the link is what distinguishes algorithmic from environmental determinism. Environmental constraint does not depend on how executives responded to it last year. Algorithmic constraint does. Each delegation shifts some analytical work from people to systems; each shift reduces the practice that maintains the capabilities needed for contestation; reduced contestation makes further delegation easier to justify and harder to reverse. Raisch and Krakowski (2021) warned that overemphasizing either automation or augmentation sets off reinforcing cycles with harmful outcomes. The loop specifies one such cycle at the top of the firm.

It also specifies how a firm passes from augmentation to automation without deciding to. The passage has three stages. In the first, the system augments: the human is in the loop, the system informs, and the executive selects among alternatives with reasons of her own. In the second, the human remains formally in the loop but acts on the loop, ratifying recommendations that she no longer reconstructs, because Proposition 2’s deference has accumulated and Proposition 3’s erosion has begun. In the third, the arresting checkpoints remain in the process diagram but stop arresting anything, because nothing reaches them that the executive can evaluate independently. In Murray et al.’s (2021) terms, an augmenting technology has become an automating one while the organization chart is unchanged. The transition is invisible to governance because each stage preserves the signatures of the one before.

The cycle is self-binding in a precise sense. Executives choose, at time one, an arrangement that limits what they can competently choose at time two. Unlike a contractual precommitment, the binding is not visible in any document. Its consequences extend beyond the individual firm when several firms depend on similar models. Advice from a common system pulls judgements together (Fügener et al., 2021), and Hudson and Morgan (2026) find that industry exposure to AI and heterogeneity in board networks each lowered firms’ idiosyncratic risk but raised it in combination, which indicates that human governance structures and AI exposure interact rather than operating separately.

Strategic distinctiveness refers here to the distance between a firm’s strategic positions and those of the rivals with which it competes in the same market domain, whether that domain is an industry or a niche within one. Kleinberg and Raghavan (2021) show formally that when many decision-makers adopt the same algorithm, overall decision quality can fall even if the algorithm is more accurate than each decision-maker alone, because their errors become correlated. The resulting convergence is a new route to a familiar outcome. DiMaggio and Powell (1983) explained why organizations in a field grow alike through coercive, mimetic and normative pressures, with mimetic isomorphism arising from uniform responses to uncertainty. Algorithmic monoculture is mimetic isomorphism performed by a shared model: firms need not observe and copy each other, because they consult the same system and receive similar answers. Stoeber et al.’s (2026) finding that language models amplify institutional pressure suggests that generative systems carry normative pressures as well. Fine-tuning an open or licensed foundation model on proprietary data restores some distinctiveness, but only in what the firm adds; the shared base still carries shared blind spots.

Proposition 6: Firms that delegate strategic analysis to algorithmic systems without investing in contestability will exhibit declining strategic distinctiveness relative to the rivals in their market domain. The decline will be steepest where those rivals rely on the same model providers or training data, and it will be moderated, though not removed, where firms fine-tune shared foundation models on proprietary data.

3.6.          When the Model Is Bought: The Vendor as Shadow Agent

The argument so far assumes a firm that builds or trains its own models. Most firms buy them, as analytical platforms, forecasting services or general-purpose language models accessed under licence. Buying and making are rarely exclusive: in a study of 3,143 European firms, buying ready-made AI software was increasingly common, and firms that bought often also built, with the two forms of sourcing acting as complements (Hoffreumon et al., 2024). Supply is concentrated. A small number of large technology firms dominate AI provision and control the cloud computing on which it runs, which leaves AI users dependent on them (Jacobides et al., 2021). Dependence produces lock-in through switching costs (Farrell & Klemperer, 2007), and in cloud computing, firms report that lock-in deepens as more of their computing migrates to a provider, with contracts and open standards as the principal defences (Opara-Martins et al., 2016).

Agency theory clarifies what changes. Jensen and Meckling (1976) define agency costs as the losses that arise when a principal delegates to an agent whose interests diverge, together with the costs of monitoring and bonding meant to limit them. A vendor whose model shapes what a top team notices and how it ranks its options is an agent of this kind, though no contract names it as one. Its interests diverge in identifiable ways: it designs for its whole customer base rather than for the firm, it benefits from the switching costs that lock the firm in, and it controls the information needed to monitor it. In the terms of upper echelons theory, the vendor becomes a shadow member of the top team, one that shapes strategic attention and evaluation without sitting on the team, answering to its board or bearing its accountability. Buying moves part of the reconfiguring practice to this shadow member, since training data, objective functions and the timing of updates are set by the provider.

Two implications follow. The first is that procurement becomes a form of framing. Framing specifies what a decision is for. When a model is bought, those specifications are fixed in the procurement process: in the objectives against which a vendor’s model is selected, the evaluation metrics written into service agreements, the data the firm agrees to supply and the terms under which configurations can be changed. Whatever the firm does not specify is set by the provider’s defaults. Because switching costs rise after adoption, the framing done at procurement persists long after the people who did it have moved on. The second implication concerns contesting, which depends on what the firm retains: access to model documentation and validation results, contractual rights to audit, and an in-house analytical capacity able to produce an independent view. These are the monitoring mechanisms agency theory would prescribe. The arrangement also alters the economics of monitoring. In Jensen and Meckling’s (1976) account, principals spend on monitoring and agents spend on bonding to reassure them. Here, the information needed to monitor sits with the agent, model opacity raises the cost of every check, and the vendor’s bonding takes the form of certifications and assurance reports that it designs itself. Monitoring costs rise while their effectiveness falls, which is why the retained in-house capacity matters more than any contractual promise of transparency. The complementarity Hoffreumon et al. (2024) observe suggests that retaining such capacity is common practice. This extends Proposition 5 to purchased systems: the decision architecture that preserves points of human contest includes the contract and the retained analytical function as well as the internal workflow.

3.7.          Governance Regime as a Moderator of Effective Discretion

The framework so far has held formal authority constant and let effective discretion vary. Governance regimes determine how far the two can drift apart. They do so through three levers: who holds formal authority over the decision to delegate, what forms of contest are legitimate, and what penalty attaches to exercising formal authority without the information that effective discretion requires. Two regimes illustrate the range, the Anglo-American model of dispersed shareholding with board oversight, and the Party-dominated governance of Chinese state-owned enterprises.

In the Anglo-American model, corporate law is the main lever, and it acts precisely on the gap between formal and effective discretion. Delaware doctrine, the most developed, illustrates the point; it is jurisdiction-specific and what follows is analysis rather than legal advice. Directors breach their duty of care when they approve a major transaction without adequately informing themselves, as the board in Smith v. Van Gorkom (1985) did when it approved a cash-out merger without examining the company’s intrinsic value. Reliance on a model output that no one in the boardroom or management has examined resembles that fact pattern, since formal authority is exercised without an informed basis. Oversight liability is narrower. Under In re Caremark Int’l Inc. Derivative Litig. (1996), it requires a sustained or systematic failure to ensure that reporting systems exist, and Marchand v. Barnhill (2019) applied the doctrine to a board that made no good-faith effort to oversee a mission-critical compliance risk. Where algorithmic systems become central to how a firm makes its strategic choices, a board with no system for monitoring how they perform and fail is exposed to the same reasoning. The law thus penalizes the divergence the self-binding loop produces. Enriques and Zetzsche (2020) warn against the belief that corporate technologies can substitute for boards’ judgement, and the framework gives the warning a specific form: the risk is not that boards hand authority to algorithms but that they keep authority while losing the information needed to exercise it.

Audit committees supply the institutional mechanism that can interrupt the loop rather than merely penalize it. Internal audit, reporting to the audit committee as the third line of defence, can surface ineffective risk management around AI that management would not report on its own, although it can be captured by the management it audits (Schuett, 2025). Raji et al. (2020) set out how deploying organizations can audit their own algorithmic systems end to end, and Mökander et al. (2021) describe what ethics-based auditing can and cannot achieve. For strategic systems, the audit product that matters is an independent analytical benchmark: periodic human-produced or separately modelled analyses of decisions the core system also analyses, with divergences reported to the audit committee rather than to the managers who rely on the system. In the terms of Figure 1, the benchmark cuts the self-binding loop at the point where capability erosion would otherwise weaken contestation, because it keeps independent analysis in use whatever management does.

Regulation adds a lever whose reach should not be overstated. The European Union’s Artificial Intelligence Act imposes human-oversight requirements on high-risk systems and obliges deployers to assign that oversight to people with the necessary competence and authority (European Union, 2024). Its high-risk categories concern uses such as biometrics, critical infrastructure, employment, access to essential services including creditworthiness assessment, law enforcement and the administration of justice. Most strategy-support systems fall outside them. The Act matters here as a template that boards may adopt voluntarily and as a sign of how regulators conceive of oversight, and its reliance on the competence of overseers runs directly into the capability channel.

The party-state regime changes the levers. Jin et al. (2022) describe the governance of China’s state-owned enterprises as Party-dominated, combining Party leadership, state ownership, cadre management of executives, Party participation in corporate decision-making and intra-Party supervision. Jiang and Kim (2020) caution that the dominant governance problem in Chinese listed firms lies between controlling and minority shareholders, so findings derived from dispersed ownership cannot be transferred automatically, and concentrated control in private firms also places the decision to delegate with a controlling party rather than with professional executives. Under Party-dominated governance, formal authority is shared before any algorithm enters, because major decisions pass through a political layer. That layer is a structural point of contest, which should slow the loop. Its criteria are political, however, so contestation there may resemble uninterrogated override more than the independent interrogation Proposition 4 requires. Where adoption of AI is encouraged by state policy, delegation is partly mandated rather than chosen, and manufactured constraint moves back toward the exogenous determinism of the original debate. Accountability runs through intra-Party supervision as well as corporate law, so the penalty lever operates through political rather than fiduciary channels. Whether political contestation substitutes for executive agency or merely replaces one determinism with another depends on what the political layer brings to the decision. Where it applies criteria fixed in advance by policy, without interrogating the model or the alternatives, it substitutes political determinism for algorithmic determinism, and the executives’ effective discretion shrinks from both sides. Where it requires management to justify recommendations analytically and can call on independent analysis of its own, it adds contest of the kind Proposition 4 describes. The framework predicts the first pattern where political review is formalistic and the second where it is substantive, and the difference is an empirical question for research on state-owned enterprises.

Proposition 7: The governance regime moderates the self-binding loop. Regimes that penalize the uninformed exercise of formal authority, or that give an independent body such as an audit committee an analytical benchmark against the core models, slow the divergence of effective from formal discretion. Regimes in which formal authority is shared with a political layer also slow the divergence, but because contestation there follows political rather than analytical criteria, they preserve discretion without necessarily improving decisions.

3.8.          A State-Transition Model of Strategic Adaptation

The two dimensions define a space, and the propositions describe the forces that move firms through it. Hrebiniak and Joyce (1985) presented their four types as positions in a static matrix, which suited a model in which choice and determinism are independent. Here they are coupled over time, so the four types are better treated as states in a dynamic model, each with characteristic paths of entry and exit. Figure 2 shows the states and the transitions among them, and Table 3 describes each.

Figure 2

A State-Transition Model of Strategic Adaptation Under Algorithmic Constraint

Note. States are positioned by their levels of algorithmic determinism and effective executive agency at a given time. Arrow weight indicates the relative cost of the transition: thin dashed arrows are cheap, gradual transitions that occur by default; heavy arrows are costly transitions that require deliberate investment or an external shock.

Table 3

States and Transitions in Strategic Adaptation Under Algorithmic Constraint

State	Determinism and agency	Hrebiniak and Joyce (1985) analogue	Typical entry path	Exit path and its cost	Practices that hold or move the state
Algorithmic capture	High determinism, low agency	Natural selection	Gradual erosion from governed hybridity (P2, P3, P6); consolidation of fragmented tools on one platform	Costly and lumpy: rebuild analytical teams, rewrite objectives, renegotiate contracts; usually needs turnover, activist pressure or audit findings	Independent benchmarks; restored in-the-loop checkpoints where options are eliminated
Governed hybridity	High determinism, high agency	Differentiation	Adoption with framing from executive voluntarism; costly exit from capture	Cheap and gradual decline into capture if contestability is not funded	Rotation of analysis back to human teams; documented interrogation before high-stakes acceptance; audit of what filters exclude
Executive voluntarism	Low determinism, high agency	Strategic choice	Deliberate limitation of algorithmic use by a confident top team	Moderate: adoption with retained framing leads to hybridity	Pooling of human and AI judgements; algorithms used for prediction while framing stays in the team
Fragmented drift	Low determinism at the strategic level, low agency	Undifferentiated choice	Uncoordinated adoption of tools by units and individuals without framing from the top	Toward capture when one platform standardizes use; toward voluntarism or hybridity only through framing	Framing first: articulating objectives before deciding which analyses to delegate and to which systems

Note. States are ideal types. Prediction gains associated with algorithms are documented by Kim et al. (2024) and Doshi et al. (2025); convergence and capability erosion by Fügener et al. (2021) and Rinta-Kahila et al. (2023); pooling of human and AI judgements by Choudhary et al. (2025); unsanctioned use of generative AI by Dittmar and Sposato (2026).

Algorithmic capture resembles natural selection, except that the selecting is done by the firm’s own systems. Governed hybridity resembles differentiation: the firm accepts substantial algorithmic constraint and exercises choice within and against it. Executive voluntarism is classical strategic choice and should not be romanticized. Algorithms predict better than unaided experts in many settings (Kim et al., 2024; Logg et al., 2019), and a top team that refuses them may simply be choosing worse.

The fourth state needs to be distinguished from organizational drift as the strategy literature has known it. Johnson’s (1988) account of incremental strategic change locates its source in the shared assumptions through which managers interpret their situation, so that drift follows from a coherent but outdated paradigm. Fragmented drift has the opposite structure. It arises where algorithmic tools have entered the firm through units and individuals rather than through the top team, including the informal, unsanctioned use of generative AI that Dittmar and Sposato (2026) describe as reducing an organization’s visibility into how work is done. Algorithmic determinism at the strategic level is low, because no single system dominates the strategy process, but many local decisions are shaped by different tools pulling in different directions. Executive agency is low, because no one has framed what the tools are for. The result is not inertia but incoherence: the firm changes, but not in any direction its top team chose. Fragmented drift is unstable in a particular way. When a single platform is adopted to bring order to scattered use, the firm can pass directly into capture without passing through hybridity.

The model also clarifies the switching costs between states, which are asymmetric. The path from hybridity into capture is incremental: each act of delegation is small, locally justified and cheap, and its costs, in eroded skill and narrowed attention, are paid gradually and out of sight. Organizational path dependence takes this form, with self-reinforcing processes narrowing the range of options until the organization is locked in (Sydow et al., 2009). The path out is lumpy and visible. Exit requires rebuilding analytical teams, re-specifying objectives, renegotiating or replacing contracts and accepting a period of slower decisions, and switching costs of this kind are what sustain lock-in in markets generally (Farrell & Klemperer, 2007). Drifting into capture is therefore much cheaper than leaving it, and the asymmetry grows with time spent in the captured state. The transitions also run on different clocks, though the estimates that follow are conjectures for empirical work rather than findings. Deference in Proposition 2 should build over a few decision cycles, which for most strategic decisions means quarters. Capability erosion in Proposition 3 should take longer, on the order of several planning cycles, because skills decay only as the occasions to use them disappear. Exit from capture should take longest of all, since rebuilding analytical teams and renegotiating contracts proceeds at the pace of hiring and contract renewal, typically years. Sydow et al. (2009) note that breakouts from lock-in usually require deliberate intervention or an external shock.

External parties supply some of those shocks. Activist investors win concessions in most of their campaigns and raise turnover among chief executives of target firms (Brav et al., 2008). The framework suggests where an activist’s leverage is greatest: in the blind spots that a captured firm shares with rivals using the same models. An activist who commissions an independent analysis, built on data or considerations the firm’s systems exclude, is producing the counter-analysis the captured top team can no longer produce. Where Proposition 6’s convergence holds, the same blind spot recurs across an industry, which makes it a natural target for a contrarian thesis. Unions and worker representatives have begun to treat algorithmic decision-making as a subject of collective bargaining (De Stefano, 2019), and auditors and regulators add further outside contest. These frictions do not guarantee good contestation, since an activist’s objection may be as uninterrogated as any intuitive override, but they make complete capture unstable in firms exposed to them. Capture should be most complete where such outside parties are weak or absent.

4.     Discussion

The gap identified in the Introduction concerned the absence of any account of where executive agency goes when algorithmic systems enter strategic decisions, and of whether configuring those systems narrows later latitude. The framework answers the first question by relocating agency into framing, contesting and reconfiguring, practices that remain technologically mediated and depend on how far they draw on resources independent of the system. It answers the second with a qualified yes: configuration is not a stable resting place, because delegation erodes the capabilities on which contestation depends, and the governance regime determines how quickly. Neither the measurement of these constructs nor the reversibility of erosion is settled here.

4.1.          Theoretical Implications

For upper echelons theory, the framework adds an endogenous source of discretion. Wangrow et al. (2015) noted that the internal organization had been neglected as a source of executive latitude. The decision infrastructure a firm installs is part of that internal organization, and unlike industry structure it is chosen by the executives whose discretion it shapes. The distinction between formal and effective discretion implies that measures based on authority or structural power will overstate executive latitude in firms that rely heavily on algorithmic analysis. The argument about inherited architecture adds a mechanism by which a predecessor’s paradigm survives succession. Mehler et al. (2026) reach a related conclusion when they describe a hybrid upper echelons in which human and algorithmic actors jointly shape strategy. The present framework differs in treating that hybridity as unstable, with a default trajectory toward capture, and in placing the vendor inside the upper echelon as a shadow member.

For the determinism debate, the contribution is a qualification of the two-continua model. Hrebiniak and Joyce (1985) treated choice and determinism as independent. When constraint is manufactured, they are independent at a single moment and coupled over time, because agency exercised as delegation raises future determinism. The appropriate representation is therefore a state-transition model rather than a static matrix, with asymmetric costs of moving between states. This links the old debate to the sociomaterial tradition, in which technology is both outcome and medium of action (Orlikowski, 1992), and the notion of degrees of mediation offers a way to keep that tradition’s insight without abandoning the claim that some practices give executives more independent purchase on their systems than others.

For research on AI in management, the framework turns the conflicting findings in Table 1 into conditional claims and differentiates them by architecture. Aversion and appreciation are responses to error visibility. Override helps or harms depending on whether interrogation draws on independent information. Accurate evaluation at one moment can coexist with narrowing of options over several years, and the form of narrowing differs between prediction, optimization and generation. Reviews of the field have tended to organize findings by application or by level of analysis (Bankins et al., 2024; Keding, 2021); organizing them by the channel and temporal element of agency they engage offers a different way to read them together.

4.2.          Implications for Executives, Boards and Regulators

For executives, contestability should be managed as a capability with a budget. A top team can keep human analytical skill in use by periodically producing its own analyses of decisions the system also analyses, can make interrogation a documented step for irreversible decisions, and can write its objectives into procurement rather than leaving them to a vendor’s defaults. Leyer and Schneider (2021) argue that whether AI threatens or empowers managers depends on how decisions are divided between them, which places the division itself within the executive remit. Incoming teams should treat an inherited architecture as a strategic commitment of their predecessors, open to review in the early season of their tenure.

For boards, oversight arrangements that confirm a human signed off on a decision will not detect self-binding, since formal authority is precisely what persists. Boards could instead ask who built the models, who can explain their failure modes, when management last rejected a recommendation and on what independent basis, and whether the audit committee holds a benchmark against the core systems. Hudson and Morgan (2026) show that board composition and industry AI exposure interact in shaping firm risk, and not in the direction one might have assumed, so boards would do well to treat their own capacity for contestation as something to assess rather than presume.

For regulators, the framework implies that meaningful human control has a capability precondition that a legal requirement cannot create on its own. Santoni de Sio and Mecacci (2021) argue that responsibility gaps around AI take several forms and that closing them requires meaningful rather than nominal human control. Newell and Marabelli (2015) called a decade ago for attention to the long-term effects of algorithmic decision-making beyond the firm; erosion of the human capacity to contest is one such effect that unfolds inside it.

4.3.          Limitations

The framework is conceptual and its propositions are untested. Several of the studies on which it draws were conducted outside strategic decision-making, in public inspections, radiology, policing, routine administrative automation, chess and platform work (Kim et al., 2024; Krakowski et al., 2023; Lebovitz et al., 2022; Rinta-Kahila et al., 2023; Waardenburg et al., 2022). The transfer to top management teams rests on structural features those settings share with the C-suite, but it may fail where coalition politics or the scale of strategic decisions change the processes involved. The closest evidence on executives themselves comes from interview studies and a conference paper (Mehler et al., 2026; Srinivas & Chetan, 2026), which are recent and not yet replicated.

The literature search was conducted by a single author, in English, through one bibliographic index with metadata verification in a second, and without a formal screening protocol. Relevant work in other languages, including Chinese-language research on digital transformation in state-owned and private enterprises, was not covered, and the treatment of party-state governance is conceptual. The legal analysis is confined to Delaware doctrine as an illustration. The analysis of purchased models combines firm-level sourcing evidence with agency theory and reasoning about contracts; it does not draw on studies of how top teams actually negotiate with AI vendors, which remain scarce. Generative AI is changing quickly, and the balance between the three architectures, and between the channels they act on, may shift as models improve.

4.4.          Directions for Future Research

Each proposition implies a specific empirical question, and several suit the phenomenon-based theorizing von Krogh (2018) recommended for AI in organizations. Table 4 sets out the central constructs and indicative measures for each proposition.

Table 4

Constructs and Indicative Measures for Testing the Propositions

No.	Independent construct and indicative measure	Dependent construct and indicative measure	Key controls or boundary conditions
P1	Share of environmental scanning routed through algorithmic filtering, including retrieval and summarization of unstructured sources by language models (inventory of scanning sources, tool logs, prompt and source logs; where logs are unavailable, share of briefing documents built on AI-generated summaries)	Breadth of strategic alternatives considered (distinct options recorded in board papers and strategy documents)	Predictive accuracy of filters; industry dynamism; architecture (Table 2)
P2	Duration of the top team’s use of algorithmic evaluations	Deference: share of recommendations adopted without modification in decision logs or board minutes	Feedback delay (decision-to-outcome interval); ambiguity (concurrent confounding events, dispersion of analyst explanations); team turnover
P3	Intensity of delegation of scanning and evaluation tasks	Executives’ ability to reconstruct or critique analyses unaided, measured in sequence (attention, then reasoning, then substance of communication)	Executive age and tenure; heterogeneity of the team; architecture used
P4	Override preceded by interrogation drawing on independent information versus override without it	Quality of decision outcomes relative to the algorithm’s recommendation	Opacity and cognitive load of the model; reversibility of the decision
P5	Presence and placement of in-the-loop points of human contest	Effective executive discretion; share of decisions altered at contest points	Algorithm sophistication; build versus buy; contract terms
P6	Delegation without contestability investment; share of rivals using the same model providers	Strategic distinctiveness: distance of a firm’s positions from rivals in its market domain over time	Fine-tuning on proprietary data; industry concentration of providers
P7	Governance regime: fiduciary exposure, audit-committee benchmarks, party-state oversight	Rate of divergence between formal and effective discretion	Ownership concentration; regulatory coverage of the systems involved

Note. Measures are indicative and drawn from the operational definitions given in the text. Delay and ambiguity measures for P2 and age and tenure controls for P3 follow the arguments in the Framework Development section and the evidence on tenure effects in Finkelstein and Hambrick (1990) and Hambrick and Fukutomi (1991).

Two design issues deserve emphasis. Tests of Proposition 3 must control for executive age and tenure, since cognitive change with age and the shifts in attention and commitment that accompany long tenure (Finkelstein & Hambrick, 1990; Hambrick & Fukutomi, 1991) would otherwise be mistaken for erosion caused by delegation. A difference-in-differences comparison of executives of similar age and tenure in firms that adopt algorithmic strategy tools at different times would separate the two. Tests of Proposition 6 require industry-level data on shared model providers and on the dispersion of strategic positions over time, and a changeover in a dominant model provider could serve as a natural experiment. Proposition 4 can be tested experimentally by manipulating whether override is preceded by structured interrogation, extending the work of Choudhary et al. (2025) on human–AI ensembles to strategic tasks. Comparative work across governance regimes, including Chinese enterprises with different ownership structures, would test Proposition 7 and the transitions in Figure 2. Research on succession could ask whether incoming teams restore capability or inherit its loss through routines.

5.     Conclusion

Strategic management once resolved the tension between environmental determinism and managerial choice by showing that the two could coexist. That resolution assumed constraint arrived from outside. Algorithmic systems change the assumption, because executives install the constraints that then shape what they notice, how they judge and, over time, what they are capable of judging. This paper has argued that executive agency survives the arrival of these systems by relocating into framing, contesting and reconfiguring, and that it survives only as long as the capabilities those practices require are kept in use. Delegation is self-binding: formal authority can remain intact while effective discretion contracts. The seven propositions and the state-transition model specify when this happens, how it differs across prediction, optimization and generation, and what interrupts it, from audit-committee benchmarks to the governance regime itself. Whether firms drift toward algorithmic capture or sustain governed hybridity is not decided by the technology. It is decided by executives who understand that keeping the ability to disagree with their own systems costs something, and who choose to pay for it.

Declarations

Funding. This research received no external funding.

Conflicts of Interest. The author declares no conflicts of interest.

Ethics Statement. This study is a conceptual analysis of published literature. It involved no human participants, personal data or animal subjects, and ethical approval was not required.

Data Availability. No new data were created or analysed in this study. All sources discussed are cited in the reference list.

Use of AI and AI-assisted technologies. Generative artificial intelligence tools were used solely to improve the readability and language of the manuscript. The author conceived the research question, developed the analytical framework, appraised and interpreted all sources, and drew all conclusions. The author reviewed and edited all content and take full responsibility for the accuracy and integrity of the published work.

References

• Agrawal, A., Gans, J. S., & Goldfarb, A. (2019). Exploring the impact of artificial intelligence: Prediction versus judgment. Information Economics and Policy, 47, 1–6. https://doi.org/10.1016/j.infoecopol.2019.05.001

• Anthony, C., Bechky, B. A., & Fayard, A.-L. (2023). “Collaborating” with AI: Taking a system view to explore the future of work. Organization Science, 34(5), 1672–1694. https://doi.org/10.1287/orsc.2022.1651

• Astley, W. G., & Van de Ven, A. H. (1983). Central perspectives and debates in organization theory. Administrative Science Quarterly, 28(2), 245–273. https://doi.org/10.2307/2392620

• Bainbridge, L. (1983). Ironies of automation. Automatica, 19(6), 775–779. https://doi.org/10.1016/0005-1098(83)90046-8

• Balasubramanian, N., Ye, Y., & Xu, M. (2022). Substituting human decision-making with machine learning: Implications for organizational learning. Academy of Management Review, 47(3), 448–465. https://doi.org/10.5465/amr.2019.0470

• Bankins, S., Ocampo, A. C., Marrone, M., Restubog, S. L. D., & Woo, S. E. (2024). A multilevel review of artificial intelligence in organizations: Implications for organizational behavior research and practice. Journal of Organizational Behavior, 45(2), 159–182. https://doi.org/10.1002/job.2735

• Bauer, K., von Zahn, M., & Hinz, O. (2023). Expl(AI)ned: The impact of explainable artificial intelligence on users' information processing. Information Systems Research, 34(4), 1582–1602. https://doi.org/10.1287/isre.2023.1199

• Berente, N., Gu, B., Recker, J., & Santhanam, R. (2021). Managing artificial intelligence. MIS Quarterly, 45(3), 1433–1450. https://doi.org/10.25300/misq/2021/16274

• Bourgeois, L. J. (1984). Strategic management and determinism. Academy of Management Review, 9(4), 586–596. https://doi.org/10.2307/258482

• Brav, A., Jiang, W., Partnoy, F., & Thomas, R. (2008). Hedge fund activism, corporate governance, and firm performance. The Journal of Finance, 63(4), 1729–1775. https://doi.org/10.1111/j.1540-6261.2008.01373.x

• Burrell, J. (2016). How the machine ‘thinks’: Understanding opacity in machine learning algorithms. Big Data & Society, 3(1), Article 2053951715622512. https://doi.org/10.1177/2053951715622512

• Child, J. (1972). Organizational structure, environment and performance: The role of strategic choice. Sociology, 6(1), 1–22. https://doi.org/10.1177/003803857200600101

• Choudhary, V., Marchetti, A., Shrestha, Y. R., & Puranam, P. (2025). Human-AI ensembles: When can they work? Journal of Management, 51(2), 536–569. https://doi.org/10.1177/01492063231194968

• Csaszar, F. A., Ketkar, H., & Kim, H. (2024). Artificial intelligence and strategic decision-making: Evidence from entrepreneurs and investors. Strategy Science, 9(4), 322–345. https://doi.org/10.1287/stsc.2024.0190

• Cyert, R. M., & March, J. G. (1963). A behavioral theory of the firm. Prentice-Hall.

• De Stefano, V. (2019). "Negotiating the algorithm": Automation, artificial intelligence, and labor protection. Comparative Labor Law & Policy Journal, 41(1), 15–46.

• Dietvorst, B. J., Simmons, J. P., & Massey, C. (2015). Algorithm aversion: People erroneously avoid algorithms after seeing them err. Journal of Experimental Psychology: General, 144(1), 114–126. https://doi.org/10.1037/xge0000033

• DiMaggio, P. J., & Powell, W. W. (1983). The iron cage revisited: Institutional isomorphism and collective rationality in organizational fields. American Sociological Review, 48(2), 147–160. https://doi.org/10.2307/2095101

• Dittmar, E. C., & Sposato, M. (2026). Shadow learning and knowledge redistribution: How workers appropriate generative AI outside formal organisational systems. Knowledge and Process Management. Advance online publication. https://doi.org/10.1002/kpm.70145

• Doshi, A. R., & Hauser, O. P. (2024). Generative AI enhances individual creativity but reduces the collective diversity of novel content. Science Advances, 10(28), Article eadn5290. https://doi.org/10.1126/sciadv.adn5290

• Doshi, A. R., Bell, J. J., Mirzayev, E., & Vanneste, B. S. (2025). Generative artificial intelligence and evaluating strategic decisions. Strategic Management Journal, 46(3), 583–610. https://doi.org/10.1002/smj.3677

• Eisenhardt, K. M. (1989). Making fast strategic decisions in high-velocity environments. Academy of Management Journal, 32(3), 543–576. https://doi.org/10.2307/256434

• Eisenhardt, K. M., & Bourgeois, L. J. (1988). Politics of strategic decision making in high-velocity environments: Toward a midrange theory. Academy of Management Journal, 31(4), 737–770. https://doi.org/10.2307/256337

• Emirbayer, M., & Mische, A. (1998). What is agency? American Journal of Sociology, 103(4), 962–1023. https://doi.org/10.1086/231294

• Enriques, L., & Zetzsche, D. A. (2020). Corporate technologies and the tech nirvana fallacy. Hastings Law Journal, 72(1), 55–98. https://repository.uclawsf.edu/hastings_law_journal/vol72/iss1/2

• European Union. (2024). Regulation (EU) 2024/1689 of the European Parliament and of the Council of 13 June 2024 laying down harmonised rules on artificial intelligence and amending Regulations (EC) No 300/2008, (EU) No 167/2013, (EU) No 168/2013, (EU) 2018/858, (EU) 2018/1139 and (EU) 2019/2144 and Directives 2014/90/EU, (EU) 2016/797 and (EU) 2020/1828 (Artificial Intelligence Act). Official Journal of the European Union, L series, 2024/1689 (12 July 2024). http://data.europa.eu/eli/reg/2024/1689/oj

• Faraj, S., Pachidi, S., & Sayegh, K. (2018). Working and organizing in the age of the learning algorithm. Information and Organization, 28(1), 62–70. https://doi.org/10.1016/j.infoandorg.2018.02.005

• Farrell, J., & Klemperer, P. (2007). Coordination and lock-in: Competition with switching costs and network effects. In M. Armstrong & R. Porter (Eds.), Handbook of industrial organization (Vol. 3, pp. 1967–2072). Elsevier. https://doi.org/10.1016/S1573-448X(06)03031-7

• Feldman, M. S., & Pentland, B. T. (2003). Reconceptualizing organizational routines as a source of flexibility and change. Administrative Science Quarterly, 48(1), 94–118. https://doi.org/10.2307/3556620

• Finkelstein, S., & Hambrick, D. C. (1990). Top-management-team tenure and organizational outcomes: The moderating role of managerial discretion. Administrative Science Quarterly, 35(3), 484–503. https://doi.org/10.2307/2393314

• Firk, S., Gehrke, Y., Hanelt, A., & Wolff, M. (2022). Top management team characteristics and digital innovation: Exploring digital knowledge and TMT interfaces. Long Range Planning, 55(3), Article 102166. https://doi.org/10.1016/j.lrp.2021.102166

• Fügener, A., Grahl, J., Gupta, A., & Ketter, W. (2021). Will humans-in-the-loop become borgs? Merits and pitfalls of working with AI. MIS Quarterly, 45(3), 1527–1556. https://doi.org/10.25300/misq/2021/16553

• Fügener, A., Walzner, D. D., & Gupta, A. (2026). Roles of artificial intelligence in collaboration with humans: Automation, augmentation, and the future of work. Management Science, 72(1), 538–557. https://doi.org/10.1287/mnsc.2024.05684

• Giannitsas, D., Sun, R., & Baptista, J. (2026). In artificial intelligence (AI) we (dis)trust? Navigating institutional pressures for automation and augmentation in the implementation of AI in organizations. Information and Organization, 36(1), Article 100609. https://doi.org/10.1016/j.infoandorg.2026.100609

• Glikson, E., & Woolley, A. W. (2020). Human trust in artificial intelligence: Review of empirical research. Academy of Management Annals, 14(2), 627–660. https://doi.org/10.5465/annals.2018.0057

• Grønsund, T., & Aanestad, M. (2020). Augmenting the algorithm: Emerging human-in-the-loop work configurations. The Journal of Strategic Information Systems, 29(2), Article 101614. https://doi.org/10.1016/j.jsis.2020.101614

• Haefner, N., Wincent, J., Parida, V., & Gassmann, O. (2021). Artificial intelligence and innovation management: A review, framework, and research agenda. Technological Forecasting and Social Change, 162, Article 120392. https://doi.org/10.1016/j.techfore.2020.120392

• Hambrick, D. C. (2007). Upper echelons theory: An update. Academy of Management Review, 32(2), 334–343. https://doi.org/10.5465/amr.2007.24345254

• Hambrick, D. C., & Fukutomi, G. D. S. (1991). The seasons of a CEO's tenure. Academy of Management Review, 16(4), 719–742. https://doi.org/10.2307/258978

• Hambrick, D. C., & Mason, P. A. (1984). Upper echelons: The organization as a reflection of its top managers. Academy of Management Review, 9(2), 193–206. https://doi.org/10.2307/258434

• Hambrick, D. C., Cho, T. S., & Chen, M.-J. (1996). The influence of top management team heterogeneity on firms' competitive moves. Administrative Science Quarterly, 41(4), 659–684. https://doi.org/10.2307/2393871

• Hannan, M. T., & Freeman, J. (1977). The population ecology of organizations. American Journal of Sociology, 82(5), 929–964. https://doi.org/10.1086/226424

• Helfat, C. E., & Peteraf, M. A. (2015). Managerial cognitive capabilities and the microfoundations of dynamic capabilities. Strategic Management Journal, 36(6), 831–850. https://doi.org/10.1002/smj.2247

• Hoffreumon, C., Forman, C., & van Zeebroeck, N. (2024). Make or buy your artificial intelligence? Complementarities in technology sourcing. Journal of Economics & Management Strategy, 33(2), 452–479. https://doi.org/10.1111/jems.12586

• Hrebiniak, L. G., & Joyce, W. F. (1985). Organizational adaptation: Strategic choice and environmental determinism. Administrative Science Quarterly, 30(3), 336–349. https://doi.org/10.2307/2392666

• Hudson, K., & Morgan, R. E. (2026). Industry exposure to artificial intelligence, board network heterogeneity, and firm idiosyncratic risk. Journal of Management Studies, 63(2), 596–630. https://doi.org/10.1111/joms.13127

• In re Caremark Int’l Inc. Derivative Litig., 698 A.2d 959 (Del. Ch. 1996).

• Jacobides, M. G., Brusoni, S., & Candelon, F. (2021). The evolutionary dynamics of the artificial intelligence ecosystem. Strategy Science, 6(4), 412–435. https://doi.org/10.1287/stsc.2021.0148

• Jarrahi, M. H., Newlands, G., Lee, M. K., Wolf, C. T., Kinder, E., & Sutherland, W. (2021). Algorithmic management in a work context. Big Data & Society, 8(2), Article 20539517211020332. https://doi.org/10.1177/20539517211020332

• Jensen, M. C., & Meckling, W. H. (1976). Theory of the firm: Managerial behavior, agency costs and ownership structure. Journal of Financial Economics, 3(4), 305–360. https://doi.org/10.1016/0304-405X(76)90026-X

• Ji, Z., Lee, N., Frieske, R., Yu, T., Su, D., Xu, Y., Ishii, E., Bang, Y. J., Madotto, A., & Fung, P. (2023). Survey of hallucination in natural language generation. ACM Computing Surveys, 55(12), 1–38. https://doi.org/10.1145/3571730

• Jiang, F., & Kim, K. A. (2020). Corporate governance in China: A survey. Review of Finance, 24(4), 733–772. https://doi.org/10.1093/rof/rfaa012

• Jin, X., Xu, L., Xin, Y., & Adhikari, A. (2022). Political governance in China’s state-owned enterprises. China Journal of Accounting Research, 15(2), Article 100236. https://doi.org/10.1016/j.cjar.2022.100236

• Johnson, G. (1988). Rethinking incrementalism. Strategic Management Journal, 9(1), 75–91. https://doi.org/10.1002/smj.4250090107

• Keding, C. (2021). Understanding the interplay of artificial intelligence and strategic management: Four decades of research in review. Management Review Quarterly, 71(1), 91–134. https://doi.org/10.1007/s11301-020-00181-x

• Keegan, A., & Meijerink, J. (2025). Algorithmic management in organizations? From edge case to center stage. Annual Review of Organizational Psychology and Organizational Behavior, 12, 395–422. https://doi.org/10.1146/annurev-orgpsych-110622-070928

• 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. https://doi.org/10.5465/annals.2018.0174

• Kelly, D., & Amburgey, T. L. (1991). Organizational inertia and momentum: A dynamic model of strategic change. Academy of Management Journal, 34(3), 591–612. https://doi.org/10.2307/256407

• Kim, H., Glaeser, E. L., Hillis, A., Kominers, S. D., & Luca, M. (2024). Decision authority and the returns to algorithms. Strategic Management Journal, 45(4), 619–648. https://doi.org/10.1002/smj.3569

• Kleinberg, J., & Raghavan, M. (2021). Algorithmic monoculture and social welfare. Proceedings of the National Academy of Sciences, 118(22), Article e2018340118. https://doi.org/10.1073/pnas.2018340118

• Krakowski, S., Luger, J., & Raisch, S. (2023). Artificial intelligence and the changing sources of competitive advantage. Strategic Management Journal, 44(6), 1425–1452. https://doi.org/10.1002/smj.3387

• Lebovitz, S., Lifshitz-Assaf, H., & Levina, N. (2022). To engage or not to engage with AI for critical judgments: How professionals deal with opacity when using AI for medical diagnosis. Organization Science, 33(1), 126–148. https://doi.org/10.1287/orsc.2021.1549

• Leonardi, P. M. (2011). When flexible routines meet flexible technologies: Affordance, constraint, and the imbrication of human and material agencies. MIS Quarterly, 35(1), 147–167. https://doi.org/10.2307/23043493

• Leyer, M., & Schneider, S. (2021). Decision augmentation and automation with artificial intelligence: Threat or opportunity for managers? Business Horizons, 64(5), 711–724. https://doi.org/10.1016/j.bushor.2021.02.026

• Lindebaum, D., Vesa, M., & den Hond, F. (2020). Insights from “The Machine Stops” to better understand rational assumptions in algorithmic decision making and its implications for organizations. Academy of Management Review, 45(1), 247–263. https://doi.org/10.5465/amr.2018.0181

• Logg, J. M., Minson, J. A., & Moore, D. A. (2019). Algorithm appreciation: People prefer algorithmic to human judgment. Organizational Behavior and Human Decision Processes, 151, 90–103. https://doi.org/10.1016/j.obhdp.2018.12.005

• Marchand v. Barnhill, 212 A.3d 805 (Del. 2019).

• Matthias, A. (2004). The responsibility gap: Ascribing responsibility for the actions of learning automata. Ethics and Information Technology, 6(3), 175–183. https://doi.org/10.1007/s10676-004-3422-1

• Mehler, A., Hein, A., & Krcmar, H. (2026, June). Hybrid upper echelons: A theorizing review on AI in executive decision-making. In ECIS 2026 Proceedings (Article 21). Association for Information Systems. https://aisel.aisnet.org/ecis2026/litrev/litrev/21

• Miller, D., & Friesen, P. H. (1980). Momentum and revolution in organizational adaptation. Academy of Management Journal, 23(4), 591–614. https://doi.org/10.2307/255551

• Miller, T. (2019). Explanation in artificial intelligence: Insights from the social sciences. Artificial Intelligence, 267, 1–38. https://doi.org/10.1016/j.artint.2018.07.007

• Mintzberg, H., & Waters, J. A. (1985). Of strategies, deliberate and emergent. Strategic Management Journal, 6(3), 257–272. https://doi.org/10.1002/smj.4250060306

• Möhlmann, M., Zalmanson, L., Henfridsson, O., & Gregory, R. W. (2021). Algorithmic management of work on online labor platforms: When matching meets control. MIS Quarterly, 45(4), 1999–2022. https://doi.org/10.25300/misq/2021/15333

• Mökander, J., Morley, J., Taddeo, M., & Floridi, L. (2021). Ethics-based auditing of automated decision-making systems: Nature, scope, and limitations. Science and Engineering Ethics, 27(4), Article 44. https://doi.org/10.1007/s11948-021-00319-4

• Murray, A., Rhymer, J., & Sirmon, D. G. (2021). Humans and technology: Forms of conjoined agency in organizations. Academy of Management Review, 46(3), 552–571. https://doi.org/10.5465/amr.2019.0186

• Newell, S., & Marabelli, M. (2015). Strategic opportunities (and challenges) of algorithmic decision-making: A call for action on the long-term societal effects of ‘datification’. Journal of Strategic Information Systems, 24(1), 3–14. https://doi.org/10.1016/j.jsis.2015.02.001

• Noponen, N., Feshchenko, P., Auvinen, T., Luoma-aho, V., & Abrahamsson, P. (2024). Taylorism on steroids or enabling autonomy? A systematic review of algorithmic management. Management Review Quarterly, 74(3), 1695–1721. https://doi.org/10.1007/s11301-023-00345-5

• Ocasio, W. (1997). Towards an attention-based view of the firm. Strategic Management Journal, 18(S1), 187–206. https://doi.org/10.1002/(SICI)1097-0266(199707)18:1+<187::AID-SMJ936>3.0.CO;2-K

• Opara-Martins, J., Sahandi, R., & Tian, F. (2016). Critical analysis of vendor lock-in and its impact on cloud computing migration: A business perspective. Journal of Cloud Computing, 5(1), Article 4. https://doi.org/10.1186/s13677-016-0054-z

• Orlikowski, W. J. (1992). The duality of technology: Rethinking the concept of technology in organizations. Organization Science, 3(3), 398–427. https://doi.org/10.1287/orsc.3.3.398

• Orlikowski, W. J. (2007). Sociomaterial practices: Exploring technology at work. Organization Studies, 28(9), 1435–1448. https://doi.org/10.1177/0170840607081138

• Pachidi, S., Berends, H., Faraj, S., & Huysman, M. (2021). Make way for the algorithms: Symbolic actions and change in a regime of knowing. Organization Science, 32(1), 18–41. https://doi.org/10.1287/orsc.2020.1377

• Parasuraman, R., & Manzey, D. H. (2010). Complacency and bias in human use of automation: An attentional integration. Human Factors, 52(3), 381–410. https://doi.org/10.1177/0018720810376055

• Pinski, M., Hofmann, T., & Benlian, A. (2024). AI literacy for the top management: An upper echelons perspective on corporate AI orientation and implementation ability. Electronic Markets, 34(1), Article 24. https://doi.org/10.1007/s12525-024-00707-1

• Puranam, P. (2021). Human–AI collaborative decision-making as an organization design problem. Journal of Organization Design, 10(2), 75–80. https://doi.org/10.1007/s41469-021-00095-2

• Quigley, T. J., & Hambrick, D. C. (2012). When the former CEO stays on as board chair: Effects on successor discretion, strategic change, and performance. Strategic Management Journal, 33(7), 834–859. https://doi.org/10.1002/smj.1945

• Raisch, S., & Krakowski, S. (2021). Artificial intelligence and management: The automation–augmentation paradox. Academy of Management Review, 46(1), 192–210. https://doi.org/10.5465/amr.2018.0072

• Raji, I. D., Smart, A., White, R. N., Mitchell, M., Gebru, T., Hutchinson, B., Smith-Loud, J., Theron, D., & Barnes, P. (2020). Closing the AI accountability gap. In Proceedings of the 2020 Conference on Fairness, Accountability, and Transparency (pp. 33–44). ACM. https://doi.org/10.1145/3351095.3372873

• Rinta-Kahila, T., Penttinen, E., Salovaara, A., Soliman, W., & Ruissalo, J. (2023). The vicious circles of skill erosion: A case study of cognitive automation. Journal of the Association for Information Systems, 24(5), 1378–1412. https://doi.org/10.17705/1jais.00829

• Risko, E. F., & Gilbert, S. J. (2016). Cognitive offloading. Trends in Cognitive Sciences, 20(9), 676–688. https://doi.org/10.1016/j.tics.2016.07.002

• Santoni de Sio, F., & Mecacci, G. (2021). Four responsibility gaps with artificial intelligence: Why they matter and how to address them. Philosophy & Technology, 34(4), 1057–1084. https://doi.org/10.1007/s13347-021-00450-x

• Schuett, J. (2025). Frontier AI developers need an internal audit function. Risk Analysis, 45(6), 1332–1352. https://doi.org/10.1111/risa.17665

• Schweiger, D. M., Sandberg, W. R., & Ragan, J. W. (1986). Group approaches for improving strategic decision making: A comparative analysis of dialectical inquiry, devil’s advocacy, and consensus. Academy of Management Journal, 29(1), 51–71. https://doi.org/10.2307/255859

• Shrestha, Y. R., Ben-Menahem, S. M., & von Krogh, G. (2019). Organizational decision-making structures in the age of artificial intelligence. California Management Review, 61(4), 66–83. https://doi.org/10.1177/0008125619862257

• Smith v. Van Gorkom, 488 A.2d 858 (Del. 1985).

• Srinivas, R., & Chetan, S. S. (2026). Integrating artificial intelligence in strategic decision-making: Contexts for delegation and augmentation. Group Decision and Negotiation, 35(3), Article 59. https://doi.org/10.1007/s10726-026-10016-x

• Stoeber, T., Hammerschmidt, J., Lundervold, A., Kromidha, E., Kanbach, D. K., & Kraus, S. (2026). AI strategy under institutional pressure: Strategic conformity and decision-making in large language models. Journal of Business Research, 212, Article 116227. https://doi.org/10.1016/j.jbusres.2026.116227

• Sweller, J. (1988). Cognitive load during problem solving: Effects on learning. Cognitive Science, 12(2), 257–285. https://doi.org/10.1207/s15516709cog1202_4

• Sydow, J., Schreyögg, G., & Koch, J. (2009). Organizational path dependence: Opening the black box. Academy of Management Review, 34(4), 689–709. https://doi.org/10.5465/amr.34.4.zok689

• Teece, D. J. (2007). Explicating dynamic capabilities: The nature and microfoundations of (sustainable) enterprise performance. Strategic Management Journal, 28(13), 1319–1350. https://doi.org/10.1002/smj.640

• Trunk, A., Birkel, H., & Hartmann, E. (2020). On the current state of combining human and artificial intelligence for strategic organizational decision making. Business Research, 13(3), 875–919. https://doi.org/10.1007/s40685-020-00133-x

• von Krogh, G. (2018). Artificial intelligence in organizations: New opportunities for phenomenon-based theorizing. Academy of Management Discoveries, 4(4), 404–409. https://doi.org/10.5465/amd.2018.0084

• Waardenburg, L., Huysman, M., & Sergeeva, A. V. (2022). In the land of the blind, the one-eyed man is king: Knowledge brokerage in the age of learning algorithms. Organization Science, 33(1), 59–82. https://doi.org/10.1287/orsc.2021.1544

• Walsh, J. P., & Ungson, G. R. (1991). Organizational memory. Academy of Management Review, 16(1), 57–91. https://doi.org/10.2307/258607

• Wangrow, D. B., Schepker, D. J., & Barker, V. L. (2015). Managerial discretion. Journal of Management, 41(1), 99–135. https://doi.org/10.1177/0149206314554214

• Weick, K. E., Sutcliffe, K. M., & Obstfeld, D. (2005). Organizing and the process of sensemaking. Organization Science, 16(4), 409–421. https://doi.org/10.1287/orsc.1050.0133

• Wiersema, M. F., & Bantel, K. A. (1992). Top management team demography and corporate strategic change. Academy of Management Journal, 35(1), 91–121. https://doi.org/10.2307/256474

Hashtags:

#AlgorithmicDeterminism #ExecutiveAgency #ArtificialIntelligence #StrategicManagement #ManagerialDiscretion #UpperEchelonsTheory #StrategicChoice #CorporateGovernance #HumanOversight #AlgorithmicCapture #DecisionArchitecture #GenerativeAI #FiduciaryDuty #AuditCommittee #MachineLearning #AIStrategy #AIGovernance #Sociomateriality #BusinessLeadership #OrganizationalBehavior #U7Y #AcademicResearch