AI-Augmented Strategic Decision-Making Under Time Constraints: An Experimental Study on Mental Representations and Strategic Foresight

Source: Strategy Science (Kanis, Mann, Stumpf-Wollersheim) · 2026-02-04

Abstract. Strategic foresight—that is, the ability to predict strategic outcomes—depends on how decision-makers represent strategic problems. Time constraints and large language mod­ els (LLMs) are increasingly salient factors shaping this process. We study how both jointly affect mental representations and strategic foresight in a startup evaluation task (Nˆ 348). Using a 2 × 2 experimental design, we show that both time constraints and LLM use signifi­ cantly 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 ave­ nues for future research on LLM use and strategic foresight, particularly regarding the inter­ play between individual cognitive processes and the contextual factors of strategic decisions.

Introduction. Because of rapid environmental changes, managers must often make strategic decisions under time con­ straints in order to seize opportunities on time and increase firm performance (Eisenhardt 1989, Bettis and Hitt 1995, Baum and Wally 2003, Csaszar and Levinthal 2016). Time constraints present managers with the challenge of rapidly developing mental repre­ sentations (i.e., simplified models of reality) of strate­ gic problems and of predicting the outcomes of alternative options. The ability to predict strategic out­ comes, known as strategic foresight (Gavetti and Menon 2016, Csaszar and Laureiro-Mart ́ınez 2018), depends on how effectively decision-makers utilize environ­ mental cues to update their mental representations (Kapoor and Wilde 2023). Prior research indicates that “intimate knowledge” about firm innovations (Ahuja et al. 2005, p. 795), strategy courses (Heshmati and Csaszar 2023), and anticipating “things that can go wrong” (Peterson and Wu 2021, p. 2358) can improve strategic foresight. Similarly, research on forecasting has shown that individuals with a lower need for clo­ sure (Tetlock et al. 2014, Tetlock 2017) and flexible thinking styles (Tetlock and Gardner 2015) make supe­ rior forecasts about uncertain future events.

If strategic foresight is fundamentally concerned with effectively utilizing cues to develop accurate mental representations (Csaszar and Laureiro- Mart ́ınez 2018), recent advances in artificial intelli­ gence (AI), and particularly large language models (LLMs), deserve attention. Generating strategy docu­ ments and slide decks with LLMs requires little effort in terms of temporal, cognitive, or monetary resources (Csaszar et al. 2024b), making it an attractive solution to overcome the information-processing constraints imposed by time constraints (Simon 1997). Moreover, prior research has found that LLMs can generate and evaluate strategies (Doshi et al. 2025) and support humans in creating strategically viable ideas (Bous­ sioux et al. 2024).

Because using LLMs to generate abundant amounts of “polished” strategic presentations requires little effort (Csaszar et al. 2024b), the risk increases that LLMs impose new cognitive bottlenecks and amplify existing ones. Prior work suggested that LLMs can reduce psychological ownership (Draxler et al. 2024), cognitive engagement, and memory recall while also leading to information overload (Kosmyna et al. 2025). Even when LLMs provide decision-makers with rele­ vant cues, the cognitive processes they trigger may make it harder for decision-makers to effectively iden­ tify and utilize them, especially under conditions where time is constrained (Simon 1997).

Against this background, we investigate how time constraints and LLM use affect mental representations and strategic foresight. Specifically, we study how time constraints and LLM use alter the breadth, depth, and consensus of mental representations and strategic foresight (Csaszar and Laureiro-Mart ́ınez 2018). We hypothesize that time constraints “truncate” mental representations, make them more consensual, and impair strategic foresight. In contrast, we hypothesize that LLM use exerts largely opposite effects on mental representations.

Related work. Theoretical Background The Representational Approach of Strategic Decision-Making Following the Carnegie tradition of incorporating the­ ories from cognitive science and psychology (see, e.g., Newell and Simon 1972, Simon 1997), the representa­ tional approach to strategic decision-making draws on Brunswik’s (1952) lens model (Csaszar and Laureiro- Mart ́ınez 2018). This model assumes that individuals have limited capacity to accurately process cues from the environment and make accurate predictions. Speci­ fically, individuals are limited in their capacity to accu­ rately (a) identify and (b) weight the cues relevant for predicting real values (Hammond 1955). Against this background, the representational approach sees strate­ gic foresight as a key managerial ability. Having accu­ rate mental representations means being able to more accurately predict strategic outcomes, making better decisions, and potentially foster firm performance (Gary and Wood 2011, Csaszar 2018).

Csaszar and Laureiro-Mart ́ınez (2018) presented three key characteristics of mental representations. First, breadth captures categorical diversity of a mental representation (e.g., value to customer, industry struc­ ture). Second, depth reflects the within-category detail (e.g., large domestic market, large international market). Third, consensus captures the similarity to the crowd’s representation. A high level of consensus indicates a high alignment with the wisdom of the crowd or close agreement with a typical member of the audience that is decisive for success (e.g., investors).

The Effect of Large Language Models on Mental Representations and Strategic Foresight We can conceptualize LLMs as tools that are—in many cases inexpensively—capable of dealing with a wide range of tasks, including professional writing (Noy and Zhang 2023), creativity (Jia et al. 2024), and customer support (Brynjolfsson et al. 2025). Yet we can also conceptualize LLMs not merely as a (cheap) tool for completing text-based tasks but rather as a mecha­ nism that can affect an organization’s intelligence as a whole by changing how it represents, searches for, and aggregates information (Csaszar and Steinberger 2022, Csaszar et al. 2024b, Raisch and Fomina 2025). This perspective is valuable because it puts forth the question of how organizations can employ LLMs to improve their performance and establish competitive advantages (Raisch and Krakowski 2021).

Breadth and Depth. A core idea of augmentation is that LLMs can help individuals overcome cognitive limitations (Raisch and Krakowski 2021). LLMs pro­ vide instant access to an abundance of strategy-related information; that is, LLMs have learned from countless instances of strategies, outcomes of strategies, and strategy analyses (Csaszar et al.

Method. Task We adapted an established task in the field of strategic foresight—namely, evaluation of two startups based on pitch videos and selection of the more successful startup (Csaszar and Laureiro-Mart ́ınez 2018, Hesh­ mati and Csaszar 2023). In particular, we adapted this task in two ways: (1) We introduced time constraints to some participants, and (2) we provided some parti­ cipants with an LLM chatbot. For our strategic fore­ sight task, we selected two startup pitch videos from crowdfunding campaigns on Kickstarter.com with oppo­ site performances in terms of raising the desired funding, delivering the product on time, and commercializing the product (Csaszar and Laureiro-Mart ́ınez 2018). Initially, we collected 58 potential startup videos. To improve comparability between the startup pairs, and to ensure that the selection task was not trivial, we applied differ­ ent selection criteria for the startups that we used. First, We instructed the participants to evaluate both startups and select the one they thought would be more successful based on the pitch videos. Specifi­ cally, participants (a) watched the video of the first startup, followed by the collection of pros and cons from the perspective of a potential investor, and assigned weights between 0 and 7 to each pro and con indicating importance; (b) watched the video of the second startup, followed by the collection of pros and cons, and assigned weights between 0 and 7 to each pro and con; and (c) after both evaluations were com­ pleted, (i) indicated which startup they thought would be more successful and (ii) estimated the likelihood of success for each startup on a scale from 0 (not success­ ful) to 100 (highly successful). We randomized the sequence of the startup videos between participants to avoid order effects.

Design and Manipulation To test our hypotheses, we implemented a 2 (time con­ straints: no versus yes) × 2 (LLM use: no versus yes) between-participants design and randomly assigned our participants to the experimental conditions. This design allowed us to investigate how LLM use influ­ ences mental representations and strategic foresight under different time conditions.

In the LLM condition, we provided participants with an LLM chatbot (based on the LLM gpt-4o-2024-11-20) that we integrated into the study interface. We tested whether the LLM had any prior knowledge of the startups to ensure that it could not provide partici­ pants with information concerning the startups’ perfor­ mance. Additionally, we designed a system prompt for each startup evaluation to instruct the LLM about the study context and to provide participants with helpful output (see Appendix A). Both system prompts cov­ ered three main elements: (a) To make the LLM output meaningful for the task, we prompted the LLM to pro­ vide weighted pros and cons along the same strategic evaluation criteria used in previous research (Csaszar and Laureiro-Mart ́ınez 2018, Heshmati and Csaszar 2023); (b) we included transcripts of the spoken texts from the pitch videos in the prompt to provide the Procedure At the beginning of the study, we asked participants (a) whether they were familiar with any of the startups, (b) whether they held any decision-making responsibil­ ities in business strategy, (c) not to look up any external information (to ensure that they did not search the Internet or use external AI to find out which startup had been successful), and (d) to ensure their devices had activated audio through an audio check. Subse­ quently, we informed participants about the study pro­ cedure. We asked participants in the LLM condition to use the LLM provided during startup evaluation and informed participants in the time constraints condition that they had three minutes to complete the evaluation for each startup, 20 seconds to decide which startup they thought would be more successful, and 40 seconds to rate the likelihood of success for both startups. Dur­ ing the task, we displayed the remaining time on the respective study pages in red boxes in the upper-right corners. Participants in the no time constraints condi­ tion could take as much time as they wanted.

Discussion. Time constraints impose limitations on the number of cues that decision-makers can utilize for predicting the outcomes of strategic options. Because “changing a representation could be as simple as changing the prompt of an LLM” (Csaszar et al. 2024b, p. 332), LLM technology appears promising for overcoming cogni­ tive bottlenecks imposed by time constraints. Theoreti­ cally, studying the interplay of strategic foresight, time constraints, and LLMs allows us to obtain “behaviorally plausible” (Gavetti et al. 2007, p. 525) insights in the increasingly frequent use of AI for strategic decisionmaking (Moore et al. 2025). From a practical perspec­ tive, decision-makers receive indications about what to realistically expect from employing LLMs for strategic decision-making in different conditions.

Theoretical Implications Our study contributes to the literature on strategic fore­ sight (see, e.g., Ahuja et al. 2005, Csaszar and Laureiro- Mart ́ınez 2018, Kapoor and Wilde 2023) in general and the use of LLMs in strategic decision-making in partic­ ular (Csaszar et al. 2024b, Doshi et al. 2025).

First, our findings provide empirical evidence on how context factors—specifically, time constraints— can shape the structure of mental representations underlying strategic foresight. Whereas prior work has theorized that time constraints may lead to differ­ ent mental representations (Csaszar and Levinthal 2016, Csaszar 2018), we show that they selectively reduce breadth without significantly affecting depth. This nuance refines prior assumptions and suggests that, under time constraints, decision-makers rely on narrower but still elaborated mental representations. Moreover, we observe an increase in consensus, imply­ ing that time constraints may foster representational convergence because individuals rely on similar, local, and salient cues (Tversky and Kahneman 1973).

Second, prior work has also suggested that time in general is an important factor in the context of strategic foresight—for example, when making predictions about the execution time of projects (Peterson and Wu 2021) or spending time on forecasting platforms (Kapoor and Wilde 2023). We add to this research by proposing and testing LLM use as a mechanism to help decisionmakers instantly generate relevant cues (Csaszar, et al. 2024b), particularly when time constraints limit the cog­ nitive flexibility to consider a broad set of cues (Tetlock and Gardner 2015, Laureiro-Mart ́ınez and Brusoni 2018, Scoblic and Tetlock 2020). That LLMs increased the share of non-consumer items in participants’ mental representations under time constraints indicates that they can help decision-makers identify distant cues that would otherwise remain unnoticed within the available time. This finding contributes to the prior literature showing that it can require considerable temporal effort to learn how to identify non-consumer cues (Heshmati and Csaszar 2023).

Third, we advance the literature on strategic foresight and the representational approach by demonstrating that changes in representational characteristics— whether induced by time constraints or LLM use—do not necessarily translate into changes in strategic fore­ sight. Previous research has stressed that representa­ tional characteristics are linked to strategic foresight (Gary and Wood 2011, Csaszar and Laureiro-Mart ́ınez 2018). We extend this prior work by emphasizing that whether mental representations improve strategic fore­ sight could depend on how representations are gener­ ated.

Limitations. and Future Research As with all studies, our work has limitations that open avenues for further research. First, our use of Kickstarter startup evaluations as the strategic decisionmaking context may raise questions of generalizability. However, this design has proven to be useful in various contexts, such as group tasks (Csaszar and Laureiro- Mart ́ınez 2018) and strategy courses (Heshmati and Csaszar 2023), and has high ecological validity because it offers real-world ambiguity, complexity, and uncer­ tainty. We made some adaptations to prior versions of the task (Csaszar and Laureiro-Mart ́ınez 2018, Hesh­ mati and Csaszar 2023), which may limit comparabil­ ity to extant findings. Specifically, participants in our version could assign higher weights than in the initial version, which gave them the opportunity to express more nuanced cue weights. Moreover, we chose a dif­ ferent, more recent pair of startups, which allowed us to ensure that the startups’ campaigns were not part of the training data on which our LLM was based. Still, we suggest that future research could apply the representational approach and LLM augmentation to other high-stakes contexts, such as mergers and acquisitions, new market entries, or sustainability transitions.