Innovation-sourcing acquisitions provide firms with rapid access to new knowledge, but their success depends on how effectively the acquiring firm integrates this knowledge. Prior research has focused on firm-level absorptive capacity, yet little is known about how such acquisitions affect inventors within acquiring firms who embody much of this capacity. We theorize that acquiring firm inventors' postacquisition performance will be shaped by the interaction between their own knowledge and the firms' characteristics, namely (a) the technological distance between the acquiring and target firms and (b) the alignment of the inventors' knowledge with the firms' postacquisition innovation trajectory. We argue that, although acquiring firm inventors generally experience a decline in postacquisition innovation performance, generalists face a smaller decline than specialists in distant acquisitions, and specialists experience a smaller decline than generalists in close acquisitions. Our predictions are conditional on whether the inventors continue to patent after the acquisition: inventors will be more likely to continue patenting when their knowledge base aligns with the firm's postacquisition innovation trajectory. Using panel data on 334 pharmaceutical acquisitions between 1990 and 2007, we find support for our hypotheses.
We show that demand surges can lead to disintermediation through an analysis of the effects of the 2011 Japanese tsunami on the U.S. used car market. The tsunami devastated Japanese new car production, which led to an increase in demand for used cars. Wholesale and retail prices for used Japanese cars rose, but dealer margins declined by 2.3 percentage points. This decline is consistent with surges in demand leading to buyers and sellers bypassing intermediaries. The decline in dealer margins was most pronounced for popular cars trading in thick markets for which dealers were relatively more dispensable for trade and for younger used vehicles. Consistent with the adjustment costs literature, Japanese dealers, unlike their non-Japanese counterparts, were unable to exit markets for generally younger used Japanese models.
Receiving venture capital can dramatically shape the trajectory and ultimate success of startup firms. Because these investments occur under significant uncertainty, investors rely heavily on subjective assessments, often backing entrepreneurs who are similar to themselves. Yet, research on this tendency has largely relied on coarse demographic categories, potentially obscuring more granular visually assessed forms of similarity that influence investment decisions. This paper introduces "face distance," a novel measure of visual similarity derived from facial recognition models, to explore this subtle channel. Using a mixedmethods approach that combines observational data from a top startup accelerator with an online controlled experiment, I find that facial similarity between an entrepreneur and an investor is a powerful predictor of investment. This relationship is more pronounced when there is relatively higher uncertainty, suggesting that facial similarity is a heuristic investors tend to rely on when concrete "hard" information is scarce. In addition, investments between visually similar entrepreneurs and investors underperform in terms of successful exits, which is consistent with a costly distortion. These findings highlight that homophily operates at a highly granular, visual level in early-stage investment and that this is not only inequitable but also a potentially inefficient heuristic.
Models are playing an increasingly important role in the development of management theory. From the journal's inception, Strategy Science has welcomed formal models, including analytic models, computational models, and simulation studies. Unfortunately, many modeling papers face first round rejection. They usually suffer from a relatively small set of issues. In this editorial, we provide specific guidance on how to address five of the most common issues in modeling papers. The goal is to provide a short practical guide for authors to enhance their chances of publication and subsequent impact.
The performance of large language models (LLMs), both good and bad, derives from their core architecture as text pattern detection and generation machines that are sensitive to the frequency of the data upon which they are trained. They are amazing "mean articulation machines" in this sense. Using conceptual analysis and recent benchmark data, the paper identifies those strategic tasks that fall within the reliable competence of LLMs and those that remain fundamentally misaligned with LLM's associationistic architecture. The result is a practical continuum identifying where LLMs offer genuine leverage and where human cognition remains indispensable. The most challenging tasks-novel scientific and strategic breakthroughs-are currently out of reach for LLMs because of inherent limitations in their architecture. Because breakthroughs are described with text does not imply that we can simply mine text for the next novel breakthrough. In clarifying the boundary of current LLM capabilities, the paper aims to help strategic decision makers deploy these tools more effectively as powerful assistants for the majority of tasks that lie on the tractable side of the continuum.
The relationship between divestitures and acquisitions is generally presented in three ways: to free up resources for future acquisitions, to remove redundant parts of a previously acquired firm, or due to underperformance of the combined firm. We propose an additional relationship: if an announced acquisition fails to close, the bidder may pivot to divest resources related to the target firm, particularly when the bidder lacks keystone resources- critical assets that are essential to unlock the value of other resources held by the bidder-that would have been gained through the acquisition. To test this relationship, we augment previous methodological approaches with a novel method: matching successful and unsuccessful bids using the perceived risk of deal failure by using arbitrage spreads between the announced and spot prices of the target. Consistent with this argument, we find that bidding firms are more likely to make divestitures in sectors related to the target after a failed bid, and this effect is amplified under specific conditions: when the target's resources are highly complementary to the bidder's, when the target initiates the termination, when the bidder's stranded assets have a high opportunity cost, and when the focal business is distant from the bidder's core operations.
What happens in industries where firms delegate strategy choices to private artificial intelligence (AI) agents? Would markets spiral into hypercompetition or settle into a comfortable status quo? We develop a formal model in which AI agents consider large business-model catalogs, predict performance, select, and learn from realized outcomes. Our representation accommodates existing AI paradigms, allowing for substantial increases in scale and computational capacity. We show that market dynamics converge to a self-confirming equilibrium; along the realized path, AI agents become well calibrated, and their choices become subjectively optimal-even though objectively superior business models may remain unexplored. This convergence can indeed sustain high profits. However, it also produces strategic lock-in; novel business-model implementations become rare long before catalogs are exhausted. This creates a distinct role for humans. A single episode of human-driven frame expansion-introducing a genuinely new business model to a catalog-can disrupt the AI-induced equilibrium and initiate strategic renewal. Yet, the ability to do so does not imply that it will be done. When the prevailing equilibrium is sufficiently lucrative, managers rationally refrain from triggering renewed learning. Our results clarify where humans still matter in AI-enabled strategy: deciding when to change the frame and not merely optimizing within it.
Structural holes' vast empirical literature has ostensibly advanced our knowledge of strategic behavior, outcomes, and dynamics. But this empirical research predominantly relies on Burt's opaque and misunderstood constraint index. This paper shows that empirical research using constraint does not reliably support structural holes theory but does support simpler insights related to available alternatives, exchange partner diversity, and concentration. Using formal, computational, and empirical methods, this paper reveals that network constraint almost exclusively operationalizes dyadic constructs and not the broader structural network constructs with which constraint is associated. The paper provides a path for reinterpreting extant constraint-based research and for guiding future network research.
The architecture of innovation ecosystems-the distribution of productive activities and the structure of exchanges that integrate outputs-varies widely, and it has major implications for how ecosystems create value and which participants capture value. Existing strategy research lacks a framework for explaining why different ecosystem architectures arise in different contexts. In this paper, I argue that ecosystem architectures can align with market contexts. I show that architectures differ in how well they handle environmental dynamism, systemic uncertainty, and demand heterogeneity. By integrating Williamson's foundational work on hybrid governance arrangements with the game-theoretic approach to technical coordination, I uncover a coordination trade-off, between speed and scope, and an architectural trilemma in which a given architecture can provide at most two out of three adaptation attributes. The discriminating alignment framework I develop produces a map that predicts which architectures are aligned with which configurations of environmental parameters. The paper contributes to strategic management by situating ecosystems in the markets-and-hierarchies framework of institutional economics with greater granularity than prior work and by developing a novel approach for analyzing how governance structures and architectures facilitate coordination among ecosystem participants.
Existing work on AI and strategy examines the effects of general-purpose models, implicitly assuming that strategists will use whatever tools are available and that human-AI interaction is largely a matter of prompting, complacency, or aversion. Inspired by Herbert Simon's work on system architecture, we instead adopt a design view in which purposeful AI system design and use-not mere access to generic models-can itself be a dynamic capability and a source of competitive advantage. Building on this perspective, we develop Aristotle, an agentic multiagent AI system for theory-based strategic decision making, and study how it shapes strategists' reasoning, beliefs, and strategies compared with general AI assistance or no AI assistance. We document the design journey of Aristotle, highlighting design choice trade-offs in strategy framework, human-AI integration, and cost, and then implement a streamlined three-agent version suitable for experimental testing. In a randomized experiment with 976 managers comparing this agentic AI system, general AI (GPT-4o), and a human-only condition, we find that experienced managers achieve quality improvements without confidence inflation, whereas highly educated managers exhibit confidence gains without corresponding quality improvements. We establish user-system-problem fit as a core design dimension requiring alignment between architectural complexity and practitioner expertise. We abductively derive a five-dimensional taxonomy that maps the design space for agentic AI systems and a methodological roadmap that enable researchers and practitioners to experiment with, evaluate, and iteratively improve their AI system design choices.
Can artificial intelligence (AI) do strategy? This question is both urgent and foundational: urgent because AI is already reshaping strategic practice and foundational because answering it forces us to articulate what strategy actually is. In this introductory essay to the Strategy Science Special Issue on AI and Strategy, we propose a dual-ladder framework: a causal ladder that maps the cognitive hierarchy of strategic tasks and a delegation ladder that specifies when organizations will grant AI autonomy over those tasks. A core insight emerges: AI will not enter strategy where required reasoning is deepest but where its performance is most measurable. We organize the Special Issue contributions around what AI can do today, could do as capabilities develop, and should do given the imperatives of accountability and human judgment. We close with a challenge and an invitation: if strategy scholars do not define good strategizing precisely enough to be encoded, tested, and refined, other disciplines will, embedding thinner conceptions of strategy into the tools managers use. Teaching machines to strategize and support strategizing is ultia method for what is.
Benchmarks have helped fuel rapid progress in large language models (LLMs) across a variety of domains including math, science, dialogue, and coding. Yet no existing benchmark adequately captures the defining elements of strategic decision making: uncertainty, complexity, irreversible multiperiod moves, and delayed or noisy feedback. This gap limits our ability to assess and guide LLMs’ capabilities in strategy. We propose that established strategy teaching simulations provide an ideal benchmarking approach because (1) they approximate the essential features of real-world strategy, and (2) they do so in a controlled, replicable environment suitable for evaluation. To demonstrate this, we assess the performance of 21 proprietary and 13 open-source LLMs on the Back Bay Battery (BBB) simulation, a widely used exercise in strategy and innovation courses. The simulation requires balancing short-term profitability against long-term competitive positioning while integrating complex information about customer preferences and technological change. We built an interface enabling LLMs to interact with the simulation as though encountering it for the first time, masking identifiers to reduce contamination from prior training data. Our results show clear progress in composite BBB performance: Later models generally outperform earlier versions, and reasoning-focused models from late 2024–early 2025 (e.g., o4-mini, Claude Sonnet 4, Gemini 2.0 Flash) exceed even the average scores of historical MBA student cohorts. However, frontier models from mid-to-late 2025 (e.g., GPT-5, Claude Opus 4.5, Gemini 3) have declined, underperforming both earlier LLMs and MBA students. This decline is partially explained by a systematic bias toward exploiting the core business at the expense of investing in future growth. Overall, these findings highlight impressive advances in LLMs’ strategic abilities since their inception. At the same time, we document current frontier models’ surprising weakness in managing strategic uncertainty. This paper pioneers and provides guidance for using simulation-based benchmarking as a productive framework for strategy researchers to track progress, identify blind spots, and shape the trajectory of strategy-specific LLM capabilities. History: Accepted for the Special Issue: Can AI Do Strategy? Supplemental Material: The online appendix is available at https://doi.org/10.1287/stsc.2025.0444 .
We examine the role of predictions in acquisition decision making using stock market reactions as a context to formally highlight the foundations and implications of artificial intelligence (AI)-driven foresight. Drawing on behavioral perspectives, we propose that predictions related to market reactions can provide valuable foresight by capturing the wisdom of crowds of market participants and their assessments of value creation. As a result, these predictions, whereas probabilistic in nature, can enhance acquisition decision making in areas such as deal selection and target identification. Furthermore, we argue that predictions and the foresight they provide shape managerial expectations, and when actual market reactions deviate from predictions, they stimulate additional information gathering, which is reflected in processes such as deal completion. We provide evidence supporting these arguments by developing a novel measure of predicted market reactions that extrapolates prior reactions using machine learning models. Our findings highlight the informational value that predictions confer in acquisition decision making and provide formal support for investing in predictive capabilities and AI in such contexts. More broadly, we contribute to a richer understanding of the role of predictions and AI-driven foresight in strategic decision making by demonstrating not just their ex ante value in guiding managerial choices but also their ex post effects in terms of stimulating learning and subsequent information gathering.
Strategic foresight-that is, the ability to predict strategic outcomes-depends on how decision-makers represent strategic problems. Time constraints and large language models (LLMs) are increasingly salient factors shaping this process. We study how both jointly affect mental representations and strategic foresight in a startup evaluation task (N = 348). Using a 2 x 2 experimental design, we show that both time constraints and LLM use significantly alter the characteristics of mental representations. Despite these representational shifts, neither time constraints nor LLM use are found to significantly change strategic foresight. Additional analyses indicate, for instance, that LLM use increases information overload and reduces psychological ownership. Our findings can be viewed as a cautionary case for the effectiveness of LLM use in strategic decision-making. Thus, our findings suggest several avenues for future research on LLM use and strategic foresight, particularly regarding the interplay between individual cognitive processes and the contextual factors of strategic decisions.
The long-standing unitary-actor assumption in strategy research-treating firms as monolithic entities with coherent preferences-misses that organizations are coalitions of individuals with diverse and often conflicting goals. Although behavioral perspectives have challenged this assumption, the field lacks an operational method for deriving an organizational utility function from the disparate preferences of its members and the specific structures used to aggregate them. We develop a mathematical framework that (i) maps individual utility functions into choice probabilities via a random-utility model, (ii) combines those probabilities using an explicit aggregation structure (e.g., unanimity or polyarchy), and (iii) recovers an organizational utility function that rationalizes the collective behavior. This establishes organizational utility functions as operationally meaningful: they summarize and predict organizational choice, yet are generally not simple averages of members' utilities. Instead, aggregation structures systematically reshape preferences- unanimity approximates the pointwise minima of underlying utility functions, amplifying risk aversion; polyarchy approximates the pointwise maxima, promoting risk-seeking. We illustrate strategic implications in Cournot competition and principal-agent settings, showing how internal aggregation structures shift competitive and collaborative outcomes. Overall, the framework provides a parsimonious way to retrofit unitary-actor models with behaviorally grounded organizational preferences, reconciling the coalition view of the firm with rigorous and tractable strategic analysis.
This article may be used only for the purposes of research, teaching, and/or private study. Commercial use or systematic downloading (by robots or other automatic processes) is prohibited without explicit Publisher approval, unless otherwise noted. For more information, contact permissions@informs.org. The Publisher does not warrant or guarantee the article's accuracy, completeness, merchantability, fitness inclusion of an advertisement in this article, neither constitutes nor implies a guarantee, endorsement, or support
Analogical reasoning is central to strategy because it offers a basis for decision making in uncertain and data sparse contexts. Its effectiveness as a process depends not only on retrieving candidate analogies but on correctly matching them to the focal problem because a poorly chosen analogy can mislead decision makers and produce costly errors of commission. We investigate how humans and large language models (LLMs) perform at analogical reasoning through an exploratory study that extends classic analogical transfer designs by introducing multiple source analogs and target problems. Our results reveal a tradeoff: Humans in our sample frequently overlooked valid analogies (low recall) but rarely misapplied them (high precision); LLMs, in contrast, did not miss valid analogies (high recall) but often surfaced spurious, even if internally coherent matches (low precision). These findings suggest a complementary division of labor: LLMs might serve as expansive retrieval engines, generating a broad set of candidate analogies, whereas humans adjudicate their contextual fit through superior causal matching. This highlights a possible pathway for artificial intelligence (AI)-human collaboration in strategy making while underscoring the risks of over-reliance on AI-generated analogies until these models can improve their performance at matching analogies to problems.
We examine "reputational herding," where decision makers follow the herd to maintain or improve their reputation, rather than to make optimal decisions for the firm. We propose this as an additional mechanism for firms' mimetic behaviors. We model the mechanism and demonstrate that decision makers are less likely to herd when they have a high reputation, as it strengthens their confidence in private signals and buffers them against reputational penalties in the event of a bad outcome. Conversely, herding is more likely when experts have highly correlated information (high "signal correlation"). The model's prediction on the effect of reputation differs from previous research; the model's secondary predictions on the effect of signal correlation distinguish reputational herding from informational herding. We test the theory in the context of sell-side stock analysts and employ a difference-in-differences design to compare award-winning analysts and runners-up with similar ability to assess the impact of a reputation shock on herding behavior. The overall results are consistent with the reputational herding mechanism. We discuss several managerial implications of the model and provide recommendations regarding how to mitigate the moral hazard problem of reputational herding.
This research examines how information access affects competition in offline markets. I study the relationship between broadband internet availability and brick-andmortar retailer survival in the United States from 1999 through 2008, leveraging a period when competition from e-commerce channels was minimal and broadband's primary effect on retailers stemmed from reduced search costs. Using an instrumental variables estimation strategy that links broadband availability to local slope terrain, I find that broadband availability decreased the likelihood of retailer exit. These effects were more pronounced for retailers that faced higher discovery-related search costs ex ante, specifically young and independent establishments and those in dense or urban markets. The findings suggest that online information access can reshape offline competitive dynamics by altering consumer search and discovery.
