Can AI Do Strategy?

Source: Strategy Science (Csaszar, Lee, Zemsky, Zenger) · 2026-03

Abstract. 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 dele­ gation 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 invi­ tation: 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 ulti­ mately a method for rediscovering what strategy is.

Introduction. The field of strategic management has long navigated waves of transformation, but the emergence of advanced artificial intelligence (AI) appears to mark a profound shift in how we will understand and practice strategy moving forward. Consider the remarkable trajectory of AI capabilities: just a few years ago, gen­ erating coherent narratives from unstructured data or simulating multifaceted decision scenarios seemed firmly beyond the reach of machines. Today, these feats are not only possible but increasingly sophisti­ cated, powered by models that process vast informa­ tion landscapes with speed and precision unattainable by human minds alone. This evolution compels us as a field to confront a foundational question: can AI do strategy? Or, perhaps, can humans and AI, by skill­ fully partitioning relevant tasks and decision rights, productively collaborate to do strategy?

At its essence, strategy involves sensing opportuni­ ties, crafting pathways to valuable future states, and adapting amid uncertainty—all tasks that demand foresight, synthesis, and bold choices. The rise of AI prompts two intertwined inquiries that lie at the heart of this special issue. First, how might AI trans­ form the very processes of strategic reasoning from initial formulation to ongoing adaptation? Second, in what ways could AI redefine the building blocks of achieving valued strategic outcomes, altering how The stakes of this inquiry extend far beyond theoret­ ical curiosity; they may touch the core survival of stra­ tegic management as a field. As AI permeates strategic decision making in an organization, we argue that our field must lead the charge in exploring its implications lest we relinquish this decision-making terrain to adja­ cent domains such as computer science. Engaging deeply with AI offers a unique opportunity for selfdiscovery: by embedding strategic principles into AI systems—essentially teaching machines to “do” strat­ egy or strategy tasks—we stand to gain profound insights into the nature of strategy itself. This recipro­ cal process could illuminate hidden patterns in how effective strategies emerge, refine our understanding of cognitive bottlenecks in human reasoning, and reveal novel ways to integrate judgment and reasoning with computation. Through these efforts, strategy scholars are ideally positioned to shape AI’s trajectory, ensuring that it serves as a force for ethical innovation and sus­ tained value creation in business and society.

To navigate this terrain, it is helpful to draw an analogy from another domain in which machine intel­ ligence and autonomy have progressed in measured stages: autonomous driving. Frameworks such as the Society of Automotive Engineers (SAE) levels delin­ eate a spectrum from basic driver assistance to full Of course, framing the “can AI do strategy” question as binary overlooks the essential facts that, though AI has significant limitations, such as causal reasoning (“Mean Articulation Machines” by McBride 2026, Strat­ egy Science (this issue), Felin and Holweg 2024, Song et al. 2026, Yang et al. 2026), along many dimensions, it is rapidly evolving (Dell’Acqua et al. 2023). Thus, a more productive approach contemplates answers along three interrelated dimensions: the present (“can”), the prospective (“could”), and the prescriptive (“should”). On the “can” front, early indicators suggest that AI is already making substantial inroads; for instance, con­ temporary models can generate and assess strategic proposals (Doshi et al. 2025) with quality rivaling that of seasoned professionals in generating plausible strate­ gic narratives and options, especially in entrepreneurial contexts (Csaszar et al. 2024). The “could” dimension probes deeper and depends on how well AI progresses in its capacity to augment or automate strategic deci­ sion making as detailed in our framework. Finally, the “should” lens introduces vital considerations of gover­ nance: even as AI’s potential expands, human elements such as ethical discernment, accountability for out­ comes, and alignment with societal values remain indispensable. This human–machine interaction under­ scores that strategy’s future lies not in automation alone but in thoughtful partitioning, integration, and gover­ nance, in which AI augments the effectiveness of strat­ egy without supplanting human responsibility.

A central goal of this special issue is to foster a shared language and a set of concepts that enable strategy researchers to build cumulatively on one another’s work, accelerating progress in this nascent area. Large language models (LLMs), in particular, represent a highly malleable technology, one that requires deliberate exploration to unlock its full value in strategic contexts. Whereas they may harbor the potential to facilitate or accelerate high-quality strate­ gic decisions, we lack clear methods for extracting that value effectively.

Related work. (Rumelt 2011). As humans practice and theorize about it, strategic decision making is, therefore, a causal enterprise, one that presupposes beliefs—often tacit beliefs—about how particular initiatives, investments, or policies generate desired outcomes. Even when such beliefs are incomplete or uncertain, they provide the basis for comparing alternative courses of action and making choices that subsequently guide all others (Van den Steen 2018).

Viewed in this way, strategy approaches vary in the depth of causal reasoning and analysis that they demand. Some approaches emphasize prediction and planning within existing structures, treating strategy as forecasting and alignment with forecasts (e.g., Ans­ off 1965). Others focus on estimating the causal effects of deliberate actions or resource commitments (e.g., Andrews 1971, Barney 1991). Still others highlight counterfactual reasoning and analogical extrapolation across contexts, perhaps framing strategy as the selec­ tion of different positions or configurations that are distinct from those occupied or composed at the present (Porter 1980, 1996; Gavetti et al. 2005). Another approach emphasizes strategy as the construction of new strategic models that redefine problems, model boundaries, and action spaces (e.g., Nickerson and Zenger 2004, Zott and Amit 2010, Felin and Zenger 2017, Teece 2018). Rather than treating these approaches as competing, we inter­ pret these variations as emphasizing and demanding different levels of causal cognition.

To organize these differences, we depict strategic decision making as a hierarchy of increasingly sophis­ ticated forms of causal cognition. We build on and extend Pearl’s (2000) ladder of causation, which distin­ guishes among reasoning based on observed associa­ tions, interventions, and counterfactuals, adapting this logic to the domain of strategy. In doing so, we distin­ guish counterfactual reasoning about new actions from reasoning about entirely new states and the causal models required to achieve them (Lee and Bettis 2023). This extension yields four levels of strategic decision making that progress in depth and scope of causal reasoning. Table 1 provides an overview of these four levels, summarizing the characteristic ques­ tions posed at each level and the corresponding object of strategic reasoning. At lower levels, strategizing is 3.1. A Dialogue and Debate “Can AI Do Strategy? A Dialogue and Debate” (Chat­ terji et al. 2026 Strategy Science (this issue)) is a compi­ lation of short essays that emerged from a two-panel dialogue during the AI–Strategy Conference, cospon­ sored by the ION Management Science Laboratory at the University of Utah and Strategy Science.

Method. The causal ladder takes strategy’s cognitive archi­ tecture seriously and asks what kinds of reasoning strategic tasks demand. The delegation ladder takes a performance-first view and asks when organizations will trust AI with strategic discretion regardless of how the AI arrives at its outputs. Whether these premises converge (AI expands its performance of strategy as it develops its capacity to causally reason) or diverge (AI expands its per­ formance of strategy through means other than causal rea­ soning) is an empirical question that the field must answer. We maintain both ladders precisely because we do not yet know which premise is correct, and the research agenda benefits from taking each seriously. anchored in observed data and projections from exist­ ing patterns; at higher levels, strategy making increas­ ingly departs from present data and requires model extensions and model building that enables projection beyond what has been directly observed. In the sec­ tions that follow, we elaborate each level in turn and provide illustrative examples, building in the process a more robust picture of AI’s capacity to contribute to strategic decision making as well as the role of humans, which we summarize in Table 2.

2.1.1. Level 1: Analytic–Predictive Strategy. Level 1 is analytic–predictive strategy, demanding diagnostic and predictive reasoning. Strategic decision making at this level asks, what should I do given that I observe X or expect to observe X? The task is to identify regulari­ ties, estimate trends, and extrapolate their implications for action. Although future-oriented, this reasoning predicts outcomes from existing data within a given model structure and action space. For example, based on historical sales patterns and macroeconomic indica­ tors, an organization may adjust capacity, inventory, or hiring. The strategist need not know the precise causal mechanism behind sales growth; simply recog­ nizing forecast growth is sufficient.

Firms improve decision making here by aggregating diverse data sources to enhance predictive accuracy. AI is particularly powerful at this level. Machine learning systems detect regularities and generate probabilistic forecasts, whereas LLMs extract signals from unstruc­ tured data, synthesize qualitative information, and translate forecasts into actionable narratives. Together, these tools expand the informational basis of predictive analysis.

Humans remain central in defining which questions to ask, identifying relevant performance dimensions, judging which forecasts matter, and determining how predictions translate into concrete responses. Deci­ sions about relevance, risk, and action continue—at least for now—to rest with human judgment.

2.1.2. Level 2: Intervention-Oriented Strategy. Level 2 focuses on establishing the causal effects of actions, choices, and interventions. The predictive capabilities of Level 1 are augmented by forming causal represen­ tations that support identification and estimation of the effects of deliberate actions still within an estab­ lished and previously observed set of strategic options. Strategic problem framing takes the form of what hap­ pens if I do X? Analytics estimate how outcomes change when familiar actions are taken under different configurations or intensities observable in data.

Whereas prediction remains the goal at Level 2, the effort is no longer forecasting trends but forecasting causal effects. The aim is to assess which actions by the firm or close analogs operating under comparable Because moving from Level 1 to Level 2 requires iden­ tifying the causal links connecting actions to outcomes, it demands more sophisticated tools and estimation methods than Level 1. For instance, DoorDash’s analyt­ ics group may observe a strong correlation between dis­ count coupons and purchase likelihood and conclude that aggressive couponing is merited. However deeper Level 2 analysis reveals a substantial selection bias. Coupons only cause a specific subset of consumers to purchase, and therefore, a very targeted couponing strategy is sufficient.

Discussion. 3.2. Assessing AI’s Current Strategic Capabilities A natural starting point for delineating new research streams is to clarify the cognitive tasks involved in strategic decision making and to assess the extent to which AI can currently perform them. The paper “Mean Articulation Machines” (McBride 2026, Strategy Science (this issue)) offers precisely this conceptual groundwork. It highlights three foundational modes of intelligence: associative, causal, and interventional. Associative intelli­ gence excels at identifying patterns and regularities across large information sets; causal intelligence models mechanisms and evaluates how interventions propagate through a system; interventional intelligence goes further still, constructing and comparing counterfactual futures. McBride’s “associative intelligence” corresponds closely to Level 1 of the causal ladder and “causal intelligence” maps onto Level 2, whereas “interventional intelligence” roughly maps to our Levels 3 and 4: the ability to reason about the effects of actions in untested contexts and, at the highest reaches, to construct new causal models alto­ gether. These distinctions explain why strategic reason­ ing comprises multiple heterogeneous tasks—prediction, analogy, causal explanation, foresight, and intervention design—each with its own cognitive demands.

Within this framework, contemporary AI systems appear as highly capable associative reasoners with strong articulation abilities yet limited causal discrimi­ nation and minimal interventional capacity. Humans, by contrast, possess richer causal reasoning and contex­ tual judgment although they lack the breadth, speed, and consistency of machine-based pattern extraction. These contrasts provide a principled basis for identify­ ing when AI can augment strategists and when human cognition remains indispensable.

Two empirical papers place these distinctions into practice. “The Strategic Value of Predictions in Acqui­ sition Decision Making” (Kumar et al. 2026, Strategy Science (this issue)) evaluates AI performance in pre­ diction tasks using stock market reactions to acquisi­ tions as a setting. The study shows that machine learning models, trained on extensive historical data, can outperform human decision makers in domains in Across these three opening papers, a coherent pic­ ture of what AI can do today emerges. Strategic tasks vary in the mix of associative and causal reasoning they require, and AI’s current strengths and limitations map systematically onto this spectrum. Human–AI complementarity, thus, becomes a natural organizing principle: machines contribute breadth, speed, and associative capability; humans contribute contextual interpretation, causal discrimination, and theoretical reframing. At the same time, the evidence cautions against naïve integration of AI into strategic work as misalignment between task demands and AI capabili­ ties can distort rather than enhance decision quality.

3.3. Benchmarking AI’s Strategic Performance:

Designs, Tests, and Simulations To move beyond mapping current capabilities and explore what AI could do, the field requires rigorous methods for tracking progress. Benchmarks have aided and accelerated progress in LLMs across a variety of domains, including math, science, dialogue, coding, and professional certification exams. AI has now sur­ passed various human benchmarks in coding, reading comprehension, multimodal reasoning, and even PhD- level science questions. In other areas, LLM bench­ marks provide an objective yardstick for tracking vari­ ous dimensions of LLM performance to inform the limits of their use in specific domains (e.g., UC Berkeley SkyLab 2025).

Conclusion. As AI’s role in strategy expands, the human role will look increasingly like governance: specifying objectives and constraints, adjudicating among competing causal models, building commitment and legitimacy, and bear­ ing accountability for outcomes that cannot be dele­ gated to algorithms. Strategy research must explore not only AI-assisted tools but governance designs: override rights, escalation rules, accountability regimes that pre­ serve judgment when recommendations emerge from opaque systems. Moreover, if, as Neshenko and Ryall suggest, AI optimizes so effectively within existing models that humans lose incentive to innovate, the result may be stability without progress: an equilibrium in which AI constrains rather than enhances long-run value creation. In any AI-augmented strategy future, the human role is not residual. It is the locus of judg­ ment and responsibility that makes strategic decision making meaningful.

The agenda we propose carries implications for train­ ing scholars in strategy. Strategy doctoral programs need not produce engineers, but they should produce scholars capable of collaborating with technical collea­ gues, posing strategy-grounded design requirements for the AI-assisted tools, and evaluating sociotechnical sys­ tems credibly. Departments and journals must develop criteria recognizing benchmark creation, artifact build­ ing, and rigorous evaluation as legitimate academic out­ puts that can advance our understanding of effective strategic decision making.

The development of AI highlights an opportunity that the field has long ignored. Strategy has occupied an uncomfortable position: too applied for pure social science, too conceptual for engineering, too particular for economics, too focused on performance for organi­ zational behavior.