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

Introduction

Because of rapid environmental changes, managers must often make strategic decisions under time constraints 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 representations (i.e., simplified models of reality) of strategic problems and of predicting the outcomes of alternative options. The ability to predict strategic outcomes, known as strategic foresight (Gavetti and Menon 2016, Csaszar and Laureiro-Martínez 2018), depends on how effectively decision-makers utilize environmental 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 closure (Tetlock et al. 2014, Tetlock 2017) and flexible thinking styles (Tetlock and Gardner 2015) make superior forecasts about uncertain future events.

If strategic foresight is fundamentally concerned with effectively utilizing cues to develop accurate mental representations (Csaszar and Laureiro-Martínez 2018), recent advances in artificial intelligence (AI), and particularly large language models (LLMs), deserve attention. Generating strategy documents 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 (Boussioux 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 relevant cues, the cognitive processes they trigger may make it harder for decision-makers to effectively identify 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-Martínez 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.

Drawing on prior research, we employ an experimental approach in which individual decision-makers evaluate two different strategic options and decide which one they expect to be more successful (Csaszar and Laureiro-Martínez 2018, Heshmati and Csaszar 2023). Specifically, we employ a 2 (time constraints: no versus yes) × 2 (LLM use: no versus yes) between-participants design, in which we study the effects of time constraints and LLM use on mental representations and strategic foresight. We find that time constraints and LLM use significantly alter mental representations but not strategic foresight. Insights from additional analyses suggest, for example, that time constraints limit cognitive flexibility and that LLM use causes information overload and lower psychological ownership of mental representations.

Our study contributes to the literature on strategic foresight and the role of LLMs in strategic decision-making. Whereas prior work has recognized the importance of both time (Ahuja et al. 2005, Peterson and Wu 2021, Kapoor and Wilde 2023) and LLMs (Csaszar et al. 2024b, Doshi et al. 2025) in strategic foresight, we investigate whether LLMs help decision-makers under time constraints to adapt their mental representations and improve strategic foresight. Our findings suggest that LLM use may amplify the cognitive bottlenecks imposed by time constraints. We conclude that fully utilizing the potential of LLM use for strategic foresight likely requires a deeper understanding of the interplay between individual cognitive processes and the contextual factors of a strategic decision. Thus, rather as a negative verdict on the effectiveness of LLMs in strategic decision-making, our findings can be understood as a call for further research on cognition and context in LLM-augmented strategic decision-making.

Theoretical Background

The Representational Approach of Strategic Decision-Making

Following the Carnegie tradition of incorporating theories from cognitive science and psychology (see, e.g., Newell and Simon 1972, Simon 1997), the representational approach to strategic decision-making draws on Brunswik’s (1952) lens model (Csaszar and Laureiro-Martínez 2018). This model assumes that individuals have limited capacity to accurately process cues from the environment and make accurate predictions. Specifically, individuals are limited in their capacity to accurately (a) identify and (b) weight the cues relevant for predicting real values (Hammond 1955). Against this background, the representational approach sees strategic foresight as a key managerial ability. Having accurate 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-Martínez (2018) presented three key characteristics of mental representations. First, breadth captures categorical diversity of a mental representation (e.g., value to customer, industry structure). 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 Time Constraints on Mental Representations and Strategic Foresight

To clarify the role of time constraints in shaping mental representations and strategic foresight, it is useful to understand their relationship with other challenges in strategic decision-making—namely, complexity (i.e., the number of criteria and alternatives and the perceived difficulty) (Nutt 1998) and uncertainty (i.e., the unpredictability of outcomes) (Milliken 1987, Hodgkinson et al. 1999, Kapoor and Wilde 2023). Whereas complexity and uncertainty are key features (and thus inherent properties) of strategic decisions (Csaszar 2018), time constraints are externally imposed conditions that can be layered onto strategic decision-making to varying degrees. As such, they can amplify the cognitive strain associated with strategic decision-making because decision-makers must deal with complexity and uncertainty in a shorter timeframe (Eisenhardt 1989, Ordóñez and Benson 1997).

Breadth and Depth.

As time constraints increase, individuals cannot arbitrarily increase the speed at which they process environmental cues (Adelman et al. 2003). Accordingly, decision-makers under time constraints must employ strategies that enable them to make accurate predictions using fewer cues than they naturally would (Payne et al. 1988). In the language of the representational approach, using fewer cues implies that decision-makers can either (a) keep the breadth of utilized categories stable at the expense of depth (shallow representations), (b) prioritize depth at the expense of breadth (detailed representations), or (c) randomly cut breadth or depth, resulting in a contraction of both.

Cognitive flexibility refers to the ability to switch between different cognitive styles (Laureiro‐Martínez and Brusoni 2018) and enables decision-makers to incorporate a greater breadth of cues in their mental representations by attending to more distant, less familiar cues (Kiss et al. 2020). Broadening representations generally requires more effort compared with deepening because broadening requires individuals to inhibit prior perspectives and actively search for new ones (Diamond 2013). Under time constraints, however, individuals tend to adopt less effortful styles (Payne et al. 1988), suggesting that decision-makers will probably deepen existing categories rather than broaden into new ones. This logic is consistent with research on managerial decision-making suggesting that decision-makers often assign cues to existing cognitive categories (Gavetti 2012), and research on forecasting, which has suggested that cognitive flexibility requires leaving one’s “intellectual comfort zone” (Tetlock and Gardner 2015, p. 191; see also Scoblic and Tetlock 2020). A simple example from a startup evaluation task illustrates the challenge of extending breadth versus depth. After considering that a startup serves a “large domestic market,” it takes little extra effort to think of a “large international market” (greater depth). In contrast, thinking of a “high-industry rivalry” (greater breadth) is more effortful because it requires the decision-maker to inhibit the prior, market-related perspectives and shift attention to industry-related success factors (Heshmati and Csaszar 2023).

In summary, if increasing depth is associated with less effort than increasing breadth, then we can expect that decision-makers under time constraints will adopt the effort-saving strategy that prioritizes depth over breadth. Therefore, we hypothesize the following:

Consensus.

If our predictions from H1a and H1b hold (i.e., that decision-makers prefer depth over breadth), then we can expect not only a decline in the total number of cues utilized (Payne et al. 1988) but also that decision-makers employ narrower representations. However, the question arises as to whether those narrower representations are also more similar (and thus whether time constraints affect consensus).

One assumption for increased consensus would be that decision-makers receive a similar set of cues. An example of such a situation is a startup pitch, because startups typically employ pitches to convince more than one investor. For example, in their study of two popular TV shows on startups, Pollack et al. (2012) showed that startups that display an observable characteristic—for instance, a high level of preparedness—receive more funding. If decision-makers receive a similar set of cues, it may be likely that their representations capture the same salient cues (Reutskaja et al. 2011). For example, Heshmati and Csaszar (2023) showed that it takes significant effort—in their case, to invest time and financial resources to attend a strategy course—to look beyond the obvious characteristics of strategic problems. Specifically, the authors found that individuals were biased toward the more visible consumer-related cues when evaluating startups. Yet, after taking a strategy course, the ability to identify non-consumer cues increased. Similarly, research on the availability heuristic has shown that decision-makers under time constraints rely on cues that are easily observable to make decisions (Tversky and Kahneman 1973). These findings are consistent with cognitive flexibility theory, because if time constraints limit cognitive flexibility, then it is more likely that fewer decision-makers explore more distant cues, which may depend more on individual experience than on salient environmental cues. Against this background, we hypothesize the following:

Strategic Foresight.

Time constraints can impose stress on decision-makers, leading to “response without careful analysis and calculation” (Simon 1987, p. 62). Thus, time constraints can increase the risk of making poor decisions, especially when decision-makers lack experience with the task and when ambiguity is high (Klein 1993). Moreover, time constraints can restrict decision-makers’ cognitive flexibility and increase the risk of omitting important cues. Prior research has shown that considering a broad range of cues plays an important role in foresight tasks, for example, by using cognitive flexibility to draw on cues from diverse sources (Tetlock and Gardner 2015, Scoblic and Tetlock 2020), by attending to cues across multiple categories (Csaszar and Laureiro-Martínez 2018), and by recognizing the “breadth of interdependencies” among cues (Peterson and Wu 2021, p. 2384). Related research further suggests that cognitive flexibility increases the time that individuals spend considering a broad and distant set of cues (Kiss et al. 2020) and that the more time forecasters spend in a forecasting environment, and the lower the uncertainty, the better their strategic foresight (Kapoor and Wilde 2023).

In summary, when a task context is novel and uncertain (as in strategic decisions; Csaszar 2018), and if time constraints limit the decision-maker’s ability to comprehensively consider a broad set of environmental cues, the likelihood of considering a sufficient number of important cues should decline, and so should strategic foresight accuracy. Therefore, we hypothesize the following:

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 mechanism 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).

Because different mechanisms exist by which organizations can employ AI (Shrestha et al. 2019), it is helpful to locate human-LLM interaction in the broader literature. First, literature has proposed different ways of combining the predictions of humans and AI (Agrawal et al. 2022, Csaszar and Steinberger 2022, Choudhary et al. 2025). However, the scenario of human-LLM interaction differs from that studied in much prior research because transformer-based LLMs do not merely generate numerical output but enable natural language interaction with an “intelligent” system trained on vast textual corpora; thus, we can conceptualize LLMs as artifacts capable of providing interactive external representations (Csaszar 2018, Csaszar et al. 2024a). If LLMs provide representations rather than predictions (although LLMs are still predictive in the narrower sense that their answers are based on predictions about the most likely next words), then human-LLM interaction is about aggregating not predictions but rather representations.

Second, we lack a clear understanding of how individual decision-makers’ interactions with LLMs influence strategic decision-making in general and mental representations and strategic foresight in particular. To our knowledge, only two studies in the strategy domain have empirically investigated LLMs in decision-making contexts. Doshi et al. (2025) compared the aggregated LLM evaluations of 60 business models with those of human experts, whereas Csaszar et al. (2024b) investigated how LLMs can be used to generate and evaluate entrepreneurial strategies. Together, these studies demonstrate that specific mechanisms—such as aggregation and generation—can effectively enhance strategic decision-making. Such findings are encouraging for exploring additional mechanisms that may enhance strategic decision-making. For example, the LLM-human interaction might be explored at the individual level, rather than focusing on large sets of LLM-generated output (per prior studies).

In practice, decision-makers often use LLMs situationally on a one-to-one basis between the decision-maker and the LLM. However, it is unclear whether this can effectively improve strategic decisions because we know from prior research that individual LLMs can be unreliable and biased, making them prone to poor judgment (Doshi et al. 2025). Yet individual LLMs have also been shown to effectively improve underperforming strategies (albeit not through interactions with LLMs but through LLM-based, automatic completion) (Csaszar et al. 2024b), making it unclear whether LLMs are really a practical tool for strategic decision-making at the individual level.

Breadth and Depth.

A core idea of augmentation is that LLMs can help individuals overcome cognitive limitations (Raisch and Krakowski 2021). LLMs provide 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. 2024b). Therefore, we can expect decision-makers to generally consider LLMs as promising strategy tools. Yet a question arises as to whether decision-makers would employ LLMs to increase the breadth or depth of their representations.

Research on creativity and innovation suggests that LLMs are particularly helpful in creating more divergent ideas (Bouschery et al. 2023, He et al. 2025). In the strategy domain, Csaszar et al. (2024b, p. 325) found that an LLM was “not merely echoing the input that it was given but using additional knowledge from its training data.” Moreover, they suggested that LLMs can help to explore more abstract and out-of-the-box strategies. Similarly, other research has suggested that LLMs could be helpful in organizational decision-making because, based on their vast training data, they can explore more distant information (Raisch and Fomina 2025). These findings indicate that decision-makers would find LLMs to be more valuable for expanding breadth rather than depth. Therefore, we hypothesize the following:

Consensus.

Consensus refers to a mental representation’s similarity to the average representation of the crowd (Csaszar 2018, Csaszar and Laureiro-Martínez 2018). Following the rationale that individuals who rely on the same algorithms tend to use similar information, generate similar outputs, and make similar decisions (Kleinberg and Raghavan 2021, Padmakumar and He 2023), several studies in the field of human-computer interaction raised the concern that the use of LLMs may lead to a homogenization of human output. Thus, LLMs can create seemingly comprehensive and plausible output, which individuals adopt without considering more distant information that would diversify that output (Anderson et al. 2024). In this context, Doshi and Hauser (2024) found that LLM-augmented writing leads to more creative and enjoyable yet overall more similar texts.

However, before concluding that mental representations will be more consensual when augmented by LLMs, we should consider three contingencies. First, whether LLM use leads to more similar mental representations can depend heavily on how users interact with LLMs. In their study of LLM-augmented solution search in a crowdfunding setting, Boussioux et al. (2024) argued that, through prompting, LLM users can diversify the system’s output, lowering the probability of replicating consensual LLM output. A second contingency lies in the nature of the LLM system itself. In their study on text writing, Padmakumar and He (2023) found that the similarity of texts depends on the similarity of the LLM output that users receive. Put differently, if we design LLMs to provide diverse output—for example, through prompting—we can expect that this diversity will be forwarded to users. Third, the effect of LLMs on consensus could depend on task characteristics. If we have a baseline stimulus that is similar for most decision-makers (Csaszar and Laureiro-Martínez 2018, Heshmati and Csaszar 2023), then LLM assistance could mean that decision-makers can go beyond the representations primed by salient cues, exploring more diverse categories and forming more diverse representations. This argument is coherent with our cognitive flexibility argument because LLMs reduce the mental costs of switching between broadening and deepening, allowing more idiosyncratic mental representations to develop. Therefore, we hypothesize the following:

Strategic Foresight.

Strategic foresight fundamentally depends on how decision-makers identify, interpret, and accurately utilize environmental cues to predict strategic outcomes (Csaszar 2018, Csaszar and Laureiro-Martínez 2018). Therefore, it seems plausible that LLMs, with their extensive information-processing capabilities, could enhance strategic foresight by helping decision-makers utilize a broader range of relevant cues (Weiser and Von Krogh 2023, Csaszar et al. 2024b).

However, researchers have questioned the value of LLMs for strategic decision-making (and thus strategic foresight) because LLMs impose new constraints rather than relaxing them. First, LLMs may not be helpful for strategic decision-making because they lack the theorizing ability required to effectively design strategies under uncertainty (Felin and Holweg 2024). Second, a recent study that measured brain activity during LLM use found that LLMs reduce cognitive effort and increase cognitive load (Kosmyna et al. 2025). These findings suggest that, even when LLMs provide decision-makers with cues that could enhance strategic foresight, individuals may struggle to utilize these cues effectively—for instance, because the increased cognitive load hinders their ability to retain or recall the information when making a decision. Third, LLMs can impair the sense of psychological ownership (Pierce et al. 1991) over the utilized information (Lee et al. 2022, Draxler et al. 2024). This could limit the effectiveness of LLMs for strategy problems, where a strong sense of ownership is key when communicating and justifying a strategy (Pierce et al. 2001). Fourth, because of the low effort required to generate strategic representations with LLMs (Csaszar et al. 2024b), using them carries the risk of producing unnecessarily extensive and complex representations. Prior research has suggested that more complex representations do not necessarily perform better than simpler ones (Bingham and Eisenhardt 2011, Csaszar and Ostler 2020). Hence, if LLMs make representations more complex, then we cannot necessarily expect higher levels of strategic foresight.

Conversely, we can portray LLMs as a means to relax some of the cognitive constraints associated with strategic decision-making. First, LLMs can help decision-makers to process more environmental cues and to perform more comprehensive strategy analyses, which can enhance “management’s ability to determine probabilities of different future outcomes” (Laamanen et al. 2025, p. 3) and reduce the risk of neglecting important cues (Csaszar and Laureiro-Martínez 2018). Second, LLMs can be employed to generate and evaluate strategies (Doshi et al. 2025). For example, Csaszar et al. (2024b) illustrated that strategy analyses such as Porter’s Five Forces (Porter 1979) can easily be performed by LLMs on a level similar to that enacted by business school students. Moreover, they argued that LLMs enable decision-makers to generate more unconventional strategies that would otherwise require substantial cognitive effort. Third, LLM and human strengths may be complementary because humans are context-aware and capable of theory-based reasoning (Felin and Holweg 2024), whereas LLMs have distant search abilities that go beyond the local human search scope (Bouschery et al. 2023, He et al. 2025). Therefore, LLMs may help decision-makers build more accurate representations and overcome myopia in strategic decision-making (Raisch and Fomina 2025). Thus, LLMs could provide decision-makers with cues that allow them to update inaccurate representations and increase the accuracy of strategic foresight (Kapoor and Wilde 2023).

The possibility that LLMs may either impair or relax cognitive bottlenecks raises the question of under which circumstances LLMs will truly be of help for improving strategic foresight. The decisive contingency for this question may be the level of environmental uncertainty. As Felin and Holweg (2024) argued, LLMs may be unhelpful when uncertainty is high, because under such conditions, the LLM training data, which inherently draw from the past, cannot accurately support predictions when cues that have been diagnostic in the past are no longer so and will not be in future. If, however, uncertainty is moderate, we may expect that LLMs could help utilize at least some relevant cues to improve the accuracy of strategic foresight. For example, third-party endorsement and founder characteristics are relatively stable predictors of success in crowdfunding settings (Bafera and Kleinert 2023). Given that LLMs may be helpful under settings of moderate uncertainty, we hypothesize the following:

Methods

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-Martínez 2018, Heshmati 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 participants with an LLM chatbot. For our strategic foresight task, we selected two startup pitch videos from crowdfunding campaigns on Kickstarter.com with opposite performances in terms of raising the desired funding, delivering the product on time, and commercializing the product (Csaszar and Laureiro-Martínez 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 different selection criteria for the startups that we used. First, we made sure that the startups belonged to the same Kickstarter category that Csaszar and Laureiro-Martínez (2018) used (i.e., technology). This requirement allowed us to hold industry-related success factors constant. Second, we made sure that the startups were launched within the same period. Thereby, we minimized confounds due to temporal trends in technology, consumer preferences, or market conditions. Third, we made sure the videos and accompanying texts were of similar length. This reduced the risk of different fatigue or information overload effects between the experimental conditions as well as the risk of unequal exposure to diagnostic cues. Fourth, we excluded startups that were already known to the LLM underlying the chatbot used in our study. This exclusion prevented the manipulation from being contaminated by prior model knowledge. Fifth, we discussed each startup pair within our research team in order to ensure that it was not trivially obvious which one would succeed. By increasing the likelihood that participants needed to engage in foresight rather than guessing, we intended to strengthen task validity.

By applying these criteria, we identified three potential startup pairs for our main study. From these pairs, we excluded one pair based on a pretest because one startup in the pair offered a product that was based on AI technology, and we found indications that having AI in a startup’s product might prime participants who use the LLM. We also excluded a second pair because further investigations revealed that the successful startup within this pair was a copy of a patent-protected product, raising concern regarding the startup’s future success. Thus, finally, we identified one startup pair. The first startup, a portable water filter bottle named Vitaloop, successfully raised the desired funding, delivered its product on time, and commercialized its product (meaning that it was available not only on Kickstarter.com but also on other distribution channels such as Amazon.com and Globetrotter.de, a large European outdoor gear reseller). The second startup, Tidee,¹ a charging and storage solution for power sockets, failed in all three of these aspects. Both Kickstarter campaigns were launched and closed in 2023. In our pretest (N = 165), we observed correct strategic foresight among 63.03% of our participants.

We instructed the participants to evaluate both startups and select the one they thought would be more successful based on the pitch videos. Specifically, 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 completed, (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 successful) 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 constraints: 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 influences mental representations and strategic foresight under different time conditions.

In the no time constraints condition, participants could take unlimited time to complete the evaluations and make their decisions. By contrast, in the time constraints condition, we set limits on the time participants had for (a) the evaluation (including watching the videos); (b) the decision, in terms of which startup they thought would be more successful; and (c) the estimation of the likelihood of success for both startups. For the main task—that is, evaluation of the startups—we followed recommendations from cognitive psychology for inducing time constraints (Benson and Beach 1996, Ordóñez and Benson 1997) by subtracting one standard deviation from the baseline duration. Accordingly, based on a pretest with 35 participants, we set the time available for each evaluation to three minutes. Concerning the decision on which startup participants thought would be more successful, we set the available time to 20 seconds (to ensure that the participants could make a decision on time). Concerning the estimation of the likelihood of success for both startups, we set the available time to 40 seconds (because instructions were approximately twice as long, and the estimation involved two rating activities).

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 participants with information concerning the startups’ performance. 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 covered three main elements: (a) To make the LLM output meaningful for the task, we prompted the LLM to provide weighted pros and cons along the same strategic evaluation criteria used in previous research (Csaszar and Laureiro-Martínez 2018, Heshmati and Csaszar 2023); (b) we included transcripts of the spoken texts from the pitch videos in the prompt to provide the LLM with context information about the startups; and (c) we instructed the LLM to only provide information from the videos to avoid false information (resulting from hallucinations), which could have negatively influenced participant performance. To assess whether the LLM answered consistently with the system prompt, we investigated a subset of 347 prompts from across all participants and both startups. We distinguished between two typical types of prompts: (1) breadth: question for pros/cons (N = 159); and (2) depth: questions about facts such as pricing or weight (N = 187). With regard to breadth, the LLM answered according to its instruction in 94.97% of cases and with regard to depth in 95.21% of the cases analyzed. This finding suggests that although responses were not identical, they were sufficiently reliable and consistent with the prompts to minimize heterogeneity.

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 responsibilities 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. Subsequently, we informed participants about the study procedure. 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. During 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 condition could take as much time as they wanted. Following the main task, we asked participants to complete a brief questionnaire, which included questions about participants’ demographic information and experiences during the study and questions that served as a manipulation check (see Appendix B).

Measures

Breadth.

We measured breadth as the number of distinct categories in participants’ mental representations. To do so, we analyzed participants’ pros and cons list for both startups. In particular, we assigned all pros and cons to one of the 10 strategic categories proposed by Csaszar and Laureiro-Martínez (2018) and Heshmati and Csaszar (2023)—namely, industry structure, market size, imitability and time to market, costs, operations, value to customer, nonmarket, marketing, business model, and funding. For example, if a participant’s list of pros and cons included items categorized as value to customer and funding, we assigned a breadth of two.

To assign the pros and cons to the respective categories, we used three LLMs (gpt-4.1, claude-opus-4-1-20250805, and mistral-large-latest) (consistency of LLM classifications: α = 0.93; average interrater correlation: r = 0.93). We implemented a Python script that ensured identical prompting across models and submitted the items in small batches to avoid context-length limitations. In cases where pros or cons were not assigned to the same category by all LLMs, we chose the category assigned by at least two LLMs. In the 45 cases (1.44%) where all three LLMs suggested different classifications, we manually assigned a category after discussion in the author team. To validate the quality of LLM classification, we calculated Krippendorff’s alpha between LLM classification and human classification (via independent coding conducted by the first two authors, followed by a thorough discussion in cases of disagreement) across a subset of 220 pros and cons. Krippendorff’s alpha was 0.89, suggesting acceptable agreement (Marzi et al. 2024).

Depth.

We measured depth through the average number of items per category by dividing the total number of pros and cons by breadth (Csaszar 2018). Thus, depth indicated how much detail a participant provided within each category. For instance, if a participant listed two customer items and one industry structure item, depth was 1.5 (i.e., 3 items across 2 categories).

Consensus.

We calculated consensus (i.e., the degree of similarity between individuals’ mental representations and average mental representations) to assess the proximity of individuals’ mental representations to those of the crowd. We measured consensus as the Manhattan distance of the individuals’ 20-dimensional category vector (representation of 10 categories per startup) to the centroid, that is, the average vector of all participants according to previous research (Csaszar 2018, Csaszar and Laureiro-Martínez 2018). To ensure that higher values reflected greater consensus, we multiplied the Manhattan distances by −1. As a result, a value of, for example, −1.3 indicated closer agreement with the crowd than a value of −2.0.

Strategic Foresight.

First, we measured strategic foresight as a binary variable by asking which of the two startups participants expected to be more successful. We assigned 1 to participants who selected Vitaloop as the more successful startup and 0 to participants who selected Tidee. Second, we introduced a continuous measure for strategic foresight ranging from −1 (incorrect foresight) to 1 (correct foresight) by asking participants to rate the likelihood of success for both startups on a scale from 0 (not successful) to 100 (highly successful). We calculated the continuous measure of strategic foresight using the following formula adapted from Csaszar and Laureiro-Martínez (2018): (LikelihoodVitaloop − LikelihoodTidee)/(LikelihoodVitaloop + LikelihoodTidee). This continuous measure allowed us to capture nuance about how strongly participants contrast the success prospects of the two startups and served as a robustness check for our main binary variable.

Participants

Prior to data collection, we performed G*Power analyses (assuming a medium effect size, d = 0.50, and power = 0.80), resulting in a minimum sample size of N = 256 participants (64 per condition). We recruited more participants to be able to reliably detect effect sizes somewhat smaller than the assumed medium effect and to have a comparable sample size as compared with Csaszar and Laureiro-Martínez (2018). We eventually recruited N = 410 participants from Prolific, who received compensation of £3.50 ($4.66) for their participation. We required participants to have achieved a 98% approval rate in previous studies and to have completed at least 200 tasks on Prolific to ensure that only participants with a track record of high-quality submissions participated. Moreover, to maintain cultural and contextual comparability, we required participants to be fluent in English and to reside in Australia, Canada, Ireland, New Zealand, the United Kingdom, or the United States because perceptions of startups can be influenced by regional market familiarity and economic conditions (House et al. 2004). Finally, we required participants to have previously held a position with decision-making responsibilities in business strategy.

Prior to conducting our analyses, we excluded 62 participants (15.21%) for the following reasons (some participants were excluded for multiple reasons): 13 participants did not use the LLM in the LLM condition, 40 participants failed to accomplish the task (i.e., they misunderstood the task, provided incomprehensible answers, or did not answer questions that were essential to calculate breadth, depth, or consensus), nine participants did not watch both startup videos, one participant failed to decide between startups, two participants had prior knowledge of one of the startups, five participants reported technical issues, and one participant admitted to using external sources of information (e.g., LLMs or the Internet). Thus, our final sample consisted of N = 348 participants.

Participants were on average M = 42.06 (SD = 11.92) years old; 183 were male (52.59%), 164 were female (47.13%), and one reported a different gender (0.29%); 264 (75.86%) held a college degree. In terms of current occupation, 110 participants classified their position as top-level employee, 133 as manager, 37 as team leader, 56 as employee, and 12 as other occupational status. Participants reported an average of M = 13.27 years (SD = 10.13) of professional experience in their current field. On a 7-point Likert scale from 1 (low) to 7 (high), participants rated their overall experience in making decisions related to business strategy with an average of M = 5.17 (SD = 1.19), and 109 participants (31.32%) reported previous participation in a strategy course on a business-school level. Finally, participants in the LLM condition (N = 156) rated their general competence of operating LLMs on a seven-point Likert scale from 1 (low) to 7 (high) with M = 5.53 (SD = 1.14).

Results

Manipulation Check

To assess the effectiveness of our time constraints manipulation, we included an adapted perceived time pressure scale from Denovan and Dagnall (2019) in the post-task questionnaire (α = 0.97; see Appendix C) and conducted t-tests. Participants reported a higher perceived time pressure in the time constraints conditions in both the no LLM and the LLM conditions. In the no LLM condition, time constraints increased the perceived time pressure from M = 2.05 (SD = 0.90) to M = 4.26 (SD = 1.66); t(190) = −11.12, p < 0.01. Similarly, time constraints in the LLM condition increased the perceived time pressure from M = 2.61 (SD = 1.29) to M = 5.56 (SD = 1.21); t(154) = −14.64, p < 0.01. Concerning our LLM manipulation, we checked whether participants had used the LLM in the study. As described above, we excluded 13 participants in one of the two LLM conditions who did not use the LLM as required. To ensure that participants who were not in one of the LLM conditions did not use LLMs, we instructed them not to use any external sources of information. Our LLM manipulation was successful because participants in the LLM condition sent on average M = 3.83 (SD = 2.19) prompts to the LLM across both startup evaluations.

Descriptive Statistics and Correlations

In order to obtain an overview of our main findings, we display the results for strategic foresight across our experimental conditions in Table 1.²

Table 1. Experimental Design and Results for Strategic Foresight

To obtain a general understanding of the relationships among key variables of our study, we investigated the descriptive statistics and correlations across our experimental conditions (see Table 2). Similar to prior research on mental representations and strategic foresight (Csaszar and Laureiro-Martínez 2018, Heshmati and Csaszar 2023), we found that both measures of strategic foresight correlated significantly negatively with gender (female = 1) (strategic foresight binary: r = −0.15, p < 0.01; strategic foresight continuous: r = −0.18, p < 0.01). Additionally, in line with prior research (Csaszar and Laureiro-Martínez 2018, Heshmati and Csaszar 2023), we observed a negative correlation between breadth and depth (r = −0.30, p < 0.01) and a negative correlation between breadth and consensus (r = −0.80, p < 0.01). Prior research also found that individuals develop broader mental representations and achieve higher levels of strategic foresight when completing a strategy course (Heshmati and Csaszar 2023). Inspired by these findings, we investigated whether participants who had taken a strategy course exhibited higher levels of strategic foresight and greater breadth. We did not observe a significant relationship between strategic foresight and participation in a strategy course (correlation between strategic foresight binary and strategy course: r = 0.02, p = 0.76, correlation between strategic foresight continuous and strategy course: r = −0.02, p = 0.70). Similarly, we did not observe a significant correlation between participation in a strategic course and breadth (r = −0.04, p = 0.50).

Table 2. Descriptive Statistics and Correlations Among Key Variables

Time Constraints

To test our hypotheses regarding the influence of time constraints on mental representations and strategic foresight, we conducted several t-tests (see Table 3).

Table 3. Independent t-Tests Between the No Time Constraints and Time Constraints Conditions in Both the No LLM (Left) and LLM (Right) Conditions

In line with Hypothesis 1a, time constraints reduced the breadth of mental representation in both the no LLM and LLM conditions. In the no LLM condition, participants without time constraints (M = 4.10, SD = 1.48) displayed broader mental representations than participants with time constraints (M = 3.23, SD = 1.33); t(190) = 4.31, p < 0.01. We observe a similar pattern in the LLM condition, where breadth was higher without time constraints (M = 9.42, SD = 4.90) than in the presence of time constraints (M = 4.15, SD = 2.56); t(154) = 8.18, p < 0.01.

Confirming Hypothesis 1b, time constraints did not significantly affect depth in either the no LLM or LLM condition. Specifically, in the no LLM condition, depth was not statistically different between participants without time constraints (M = 2.39, SD = 0.96) and participants with time constraints (M = 2.43, SD = 1.08); t(190) = −0.25, p = 0.40. In the LLM condition, we observed a similar pattern (M = 1.59, SD = 0.59 versus M = 1.48, SD = 0.73); t(154) = 1.00, p = 0.16.

In line with Hypothesis 1c, we observed that time constraints increased consensus in both conditions. In the no LLM condition, participants without time constraints (M = −1.01, SD = 0.38) exhibited lower consensus than participants under time constraints (M = −0.87, SD = 0.31); t(190) = −2.69, p < 0.01. In the LLM condition, a similar pattern emerged (M = −2.00, SD = 0.89 versus M = −1.23, SD = 0.62); t(154) = −6.20, p < 0.01.

Contradicting Hypothesis 1d, we did not find a significant negative effect of time constraints on strategic foresight, although we observed some borderline effects in the no LLM condition. Regarding the binary measure for strategic foresight, we observed that participants in the no time constraints condition displayed slightly higher foresight on average (M = 0.62, SD = 0.49) than participants in the time constraints condition (M = 0.54, SD = 0.50); the t-test, however, was not significant; t(190) = 1.08, p = 0.14. Using the continuous measure of strategic foresight confirmed this observation: Participants in the no time constraints condition displayed slightly higher foresight on average (M = 0.16, SD = 0.38) than participants in the time constraints condition (M = 0.08, SD = 0.34); however, the t-test was not significant, t(190) = 1.61, p = 0.05. Results were clearer in the LLM condition, where the means between the no time constraints and the time constraints conditions were almost identical. Regarding the binary measure of strategic foresight, the t-test was not significant (no time constraints: M = 0.66, SD = 0.48; time constraints: M = 0.65, SD = 0.48); t(154) = 0.14, p = 0.44. The t-test results for the continuous measure of strategic foresight were similar (no time constraints: M = 0.13, SD = 0.32, time constraints M = 0.12, SD = 0.34); t(154) = 0.12, p = 0.45.

LLMs

We also tested our hypotheses regarding the influence of LLMs on mental representations and strategic foresight using t-tests. The results are displayed in Table 4.

Table 4. Independent t-Tests Between the No LLM and LLM Conditions in Both the No Time Constraints (Left) and Time Constraints (Right) Conditions

Confirming Hypothesis 2a, we observed that LLM use increases breadth under both the no time constraints and the time constraints conditions. In the no time constraints condition, LLM use increased the average breadth of mental representations from M = 4.10 (SD = 1.48) to M = 9.42 (SD = 4.90); t(170) = −9.70, p < 0.01. In the time constraints condition, breadth likewise increased with LLM use from M = 3.23 (SD = 1.33) to M = 4.15 (SD = 2.56); t(174) = −3.14, p < 0.01.

In line with Hypothesis 2b, we found LLM use to significantly reduce depth in both the no time constraints and the time constraints conditions. In the no time constraints condition, depth declined from M = 2.39 (SD = 0.96) to M = 1.59 (SD = 0.59); t(170) = 6.62, p < 0.01. In the time constraints condition, we observed a significant decline from M = 2.43 (SD = 1.08) to M = 1.48 (SD = 0.73); t(174) = 6.45, p < 0.01.

Consistent with Hypothesis 2c, LLM use significantly reduced consensus under both the no time constraints and the time constraints conditions. In the no time constraints condition, consensus declined from M = −1.01 (SD = 0.38) to M = −2.00 (SD = 0.89); t(170) = 9.60, p < 0.01. In the time constraints condition, consensus declined from M = −0.87 (SD = 0.31) to M = −1.23 (SD = 0.62); t(174) = 4.98, p < 0.01.

Finally, Hypothesis 2d expected LLM use to improve strategic foresight. Similar to time constraints, we found no significant evidence for this hypothesis. In the no time constraints condition, LLM use slightly increased the binary measure of strategic foresight from M = 0.62 (SD = 0.49) to M = 0.66 (SD = 0.48). However, the t-test revealed that this increase in strategic foresight was not significant; t(170) = −0.52, p = 0.30. Regarding the continuous measure, LLM use slightly decreased strategic foresight from M = 0.16 (SD = 0.38) to M = 0.13 (SD = 0.32); however, the t-test was again not significant (t(170) = 0.65, p = 0.26). In the time constraints condition, the positive influence of LLM use was more pronounced but remained insignificant. Specifically, LLM use increased the binary measure of strategic foresight from M = 0.54 (SD = 0.50) to M = 0.65 (SD = 0.48). The t-test was not significant; t(174) = −1.39, p = 0.08. Moreover, LLM use slightly increased the mean of the continuous measure of strategic foresight from M = 0.08 (SD = 0.34) to M = 0.12 (SD = 0.34). The t-test was again not significant; t(174) = −0.81, p = 0.21.

Additional Analyses

We conducted several additional analyses to investigate the cognitive mechanisms triggered by time constraints and LLM use.

First, we investigated whether using LLMs increased participants’ perceived time pressure. Even in the no time constraints condition, participants perceived a significantly higher level of time pressure due to LLM use (no LLM: M = 2.05 (SD = 0.90); LLM: M = 2.61 (SD = 1.29); t(170) = −3.29, p < 0.01). Similarly, LLM use significantly increased perceived time pressure in the time constraints condition (no LLM: M = 4.26 (SD = 1.66); LLM: M = 5.56 (SD = 1.21); t(174) = −5.65, p < 0.01).

Second, we investigated how time constraints and LLM use shaped the share of non-consumer items (number of non-consumer items/total number of items) in participants’ mental representations. Like prior research (Heshmati and Csaszar 2023), we considered all items in the categories industry structure, imitability and time to market, costs, operations, business model, and funding as non-consumer items. We assumed that time constraints would limit individuals’ ability to identify the more distant non-consumer cues, whereas LLM use would exert the opposite effect. We found a significant effect of time constraints on non-consumer items in the no LLM condition (no time constraints: M = 0.17 (SD = 0.12); time constraints: M = 0.13 (SD = 0.13); t(190) = 2.55, p < 0.01). In the LLM condition, time constraints also significantly reduced the share of non-consumer items (no time constraints: M = 0.42 (SD = 0.18); time constraints: M = 0.30 (SD = 0.25); t(154) = 3.42, p < 0.01). With regard to LLM use, we found that the share of non-consumer items increased significantly in both the no time constraints condition (no LLM: M = 0.17 (SD = 0.12); LLM: M = 0.42 (SD = 0.18); t(170) = −10.54, p < 0.01) and the time constraints condition (no LLM: M = 0.13 (SD = 0.13); LLM: M = 0.30 (SD = 0.25); t(174) = −6.05, p < 0.01).

Third, we suggested that time constraints diminished participants’ cognitive flexibility by limiting the awareness of alternatives and perceived cognitive control (Dennis and Vander Wal 2010). We found evidence for this suggestion in both the no LLM condition (no time constraints: M = 5.80 (SD = 0.81); time constraints: M = 5.21 (SD = 0.98), t(190) = 4.47, p < 0.01) and the LLM condition (no time constraints: M = 5.60 (SD = 0.87); time constraints: M = 4.53 (SD = 0.95); t(154) = 7.36, p < 0.00).

Fourth, we investigated whether LLM use reduced psychological ownership (Pierce et al. 2001), because LLM use may alienate participants from their pros and cons lists (Kosmyna et al. 2025). We found that LLM use significantly reduced psychological ownership in both the no time constraints condition (no LLM: M = 5.59 (SD = 1.17); LLM: M = 4.96 (SD = 1.42); t(170) = 3.17, p < 0.01) and the time constraints condition (no LLM: M = 5.16 (SD = 1.45); LLM: M = 4.45 (SD = 1.42); t(174) = 3.21, p < 0.01).

Fifth, we investigated whether time constraints and LLM use increased participants’ information overload, because time constraints reduce information-processing time and LLMs create additional information that participants have to process (Kosmyna et al. 2025). We found that both time constraints and LLM use increased information overload. With regard to time constraints, we observed this effect in both the no LLM condition (no time constraints: M = 1.59 (SD = 0.95); time constraints: M = 2.36 (SD = 1.43); t(190) = −4.29, p < 0.01) and the LLM condition (no time constraints: M = 2.36 (SD = 1.50); time constraints: M = 3.46 (SD = 1.82); t(154) = −4.11, p < 0.01). With regard to LLM use, we found a significant increase in information overload in the no time constraints condition (no LLM: M = 1.59 (SD = 0.95); LLM: M = 2.36 (SD = 1.50); t(170) = −4.08, p < 0.01) and the time constraints condition (no LLM: M = 2.36 (SD = 1.43); LLM: M = 3.46 (SD = 1.82); t(174) = −4.48, p < 0.01).

Sixth, building on findings from Kosmyna et al. (2025), we explored whether LLM use limited participants’ recall of items from their pros and cons lists. We investigated item recall for both human-generated and LLM-generated items across both startups. We found that participants remembered fewer human-generated items when using LLMs in both the no time constraints condition (no LLM: M = 1.90 (SD = 0.43); LLM: M = 1.68 (SD = 0.66); t(170) = 2.53, p < 0.01) and the time constraints condition (no LLM: M = 1.99 (SD = 0.10); LLM: M = 1.44 (SD = 0.75); t(174) = 7.48, p < 0.01). In general, recall was higher for human-generated than for AI-generated items in both the absence of time constraints (human item: M = 1.68 (SD = 0.66); AI item: M = 1.57 (SD = 0.74)) and the presence of time constraints (human item: M = 1.44 (SD = 0.75); AI item: M = 1.20 (SD = 0.80)).

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 cognitive bottlenecks imposed by time constraints. Theoretically, 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 decision-making (Moore et al. 2025). From a practical perspective, 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 foresight (see, e.g., Ahuja et al. 2005, Csaszar and Laureiro-Martínez 2018, Kapoor and Wilde 2023) in general and the use of LLMs in strategic decision-making in particular (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 different 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, implying 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 decision-makers instantly generate relevant cues (Csaszar, et al. 2024b), particularly when time constraints limit the cognitive flexibility to consider a broad set of cues (Tetlock and Gardner 2015, Laureiro‐Martínez 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 foresight. Previous research has stressed that representational characteristics are linked to strategic foresight (Gary and Wood 2011, Csaszar and Laureiro-Martínez 2018). We extend this prior work by emphasizing that whether mental representations improve strategic foresight could depend on how representations are generated. Specifically, our findings indicate that a deeper understanding of the cognitive processes involved in the generation of mental representations could be helpful for understanding when changes in representations translate into changes in strategic foresight. Thus, our findings also contribute to research on the interplay between internal (mental) representations and external (LLM) representations (Csaszar 2018, Csaszar et al. 2024a).

Fourth, we contribute to the growing literature on AI in strategic decision-making. Thus far, this literature has focused on how LLMs can generate and evaluate strategies (Csaszar et al. 2024b, Doshi et al. 2025); this focus has left a gap in our understanding of the relationship between LLM use and strategic foresight because we do not know how the interaction of decision-makers with LLMs relates to strategic foresight. Extant research on strategic foresight suggests that group interactions improve strategic foresight (Csaszar and Laureiro-Martínez 2018). Likewise, one may assume that the interaction with LLMs can improve strategic foresight. However, although we observed that LLMs change how participants represent strategic problems, we could not link these changes to better strategic foresight. Even when time constraints limited participants’ information-processing capacities (Simon 1997) and triggered perceptions of being pressured (Simon 1987), the ability of LLMs to instantly generate strategic cues (Csaszar et al. 2024b) could not provide relief. Instead, participants using an LLM under time constraints reported the highest perceived time pressure across all experimental groups. Given that these findings are specific to our experimental context, we suggest that our results should not be interpreted as a negative verdict on the effectiveness of LLMs in strategic decision-making. Rather, we believe they should be viewed as an invitation to further investigate the role of individual cognition and contextual factors such as time constraints in LLM-augmented strategic decision-making.

In summary, our findings contribute to a more nuanced understanding of how context and technology relate to mental representations and strategic foresight. Specifically, our findings suggest that the strategic decision-making literature should pay closer attention to the mechanisms by which mental representations are formed—not just their characteristics—and to how decision aids such as LLMs interact with cognitive aspects such as perceived time pressure.

Managerial Implications

Our findings offer guidance for managers adopting LLMs, particularly in time-sensitive strategic contexts. LLMs can be useful for quickly identifying otherwise overlooked cues, such as non-consumer cues. Yet managers should temper expectations because more is not always better. Using a greater breadth of cues does not necessarily result in better strategic foresight, and it can distract decision-makers from the average crowd judgment (e.g., of new products). Our results suggest that decision-makers should be mindful of situations in which LLM-generated input causes information overload and consider whether they truly understand why these cues make certain strategic outcomes more or less likely—otherwise, they risk relying on shallow “copy-and-paste representations.”

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 decision-making context may raise questions of generalizability. However, this design has proven to be useful in various contexts, such as group tasks (Csaszar and Laureiro-Martínez 2018) and strategy courses (Heshmati and Csaszar 2023), and has high ecological validity because it offers real-world ambiguity, complexity, and uncertainty. We made some adaptations to prior versions of the task (Csaszar and Laureiro-Martínez 2018, Heshmati and Csaszar 2023), which may limit comparability 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 different, 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.

Second, our sample relied on participants from Prolific. Using an online participant pool potentially raises concerns about domain expertise and generalizability. Yet prior research has shown that nonexperts can produce meaningful patterns in strategic foresight tasks (Doshi et al. 2025). Because the study was conducted online, we cannot say with certainty that no participants in the no LLM conditions consulted AI as an external source. However, several features of our design make external AI use unlikely; participants were not required to generate a minimum number of items, the task’s user interface intentionally made consulting external sources time-consuming, and we explicitly asked participants whether they had used outside information sources such as AI. Taken together, these factors provide confidence that our results were not affected by external AI use. However, future research in laboratory settings could definitively rule out the possibility of external AI use. In addition, recruiting participants from Prolific offered several advantages, such as familiarity with strategic evaluations, fluency in English, and residency in culturally similar countries. Still, we believe it may be useful for future research to study the effect of time constraints and LLM use on the strategic foresight of domain experts.

Third, and relatedly, relying on an online participant pool bears the risk that our participants differed in motivation compared with participants used in prior studies (Csaszar and Laureiro-Martínez 2018, Heshmati and Csaszar 2023). However, our main interest was in answering our research question based on the best possible study design. We decided to utilize Prolific for participant recruitment because this allowed us to set very high standards for participation (e.g., decision-making experience in business strategy). Consequently, our findings are less comparable to prior research, such as studies by Csaszar and Laureiro-Martínez (2018) and Heshmati and Csaszar (2023), whose participants had their performance publicly scrutinized. Nevertheless, to rule out concerns that differences in motivation might explain our results, we investigated two measures. First, effort (adapted 4-item scale from McAuley et al. 1989, M = 6.35, SD = 0.77, α = 0.83) did not significantly correlate with strategic foresight (r = −0.07, p = 0.16). Second, perceived diligence, a subscale of participant engagement from Meade and Craig (2012) (adapted 5-item scale, M = 6.45, SD = 0.62, α = 0.80), did not correlate with strategic foresight (r = −0.01, p = 0.86). These findings suggest that motivation did not systematically bias our results. Even if our participants were less motivated than those in Csaszar and Laureiro-Martínez (2018) and Heshmati and Csaszar (2023), such differences would not account for the patterns of strategic foresight we observed across experimental groups; random assignment to conditions strengthens our confidence that the results are meaningful in their own right.

Fourth, our study focuses on individuals without considering group-level processes or multiagent interaction. Yet prior work suggests that aggregation—whether across individuals or between humans and LLMs—could enhance strategic foresight (Csaszar et al. 2024b, Doshi et al. 2025). Future studies could explore how LLM-augmented mental representations perform when integrated in group deliberation and whether different representational characteristics become more valuable under such conditions.

Fifth, we excluded a relatively large share of participants from our analyses in the time constraints condition with LLMs (36.61%) compared with the other conditions (<8.00%). This comparatively high exclusion rate for participants that were most strongly affected by time constraints and the need to operate LLMs might have resulted in a bias (because the effect of LLMs might have been overestimated). To investigate whether this potential limitation influenced our results, we conducted a t-test with all participants in the time constraints condition, which showed no significant difference in strategic foresight (no LLM condition: M = 0.54, SD = 0.50; LLM condition: M = 0.61, SD = 0.49); t(224) = −0.96, p = 0.17). In light of these observations, we feel confident that our findings are robust in this respect.

Sixth, to classify participants’ pros and cons, we relied on LLM-based coding. This approach could raise the concern that the coding process introduced bias or lacked reliability compared with human coders. However, the high levels of agreement both between the LLMs we employed for coding and between human coders suggest that this is probably not a major methodological concern.

Seventh, we measured strategic foresight at only one point in time. This limitation prevents us from investigating how using LLMs affects mental representations and strategic foresight over time, and it does not allow us to explore, for example, whether LLMs benefit less skilled decision-makers and harm more skilled ones. Exploring these issues would contribute to research on the interplay between LLMs and individual differences in skills (Jia et al. 2024) and on strategic foresight in the context of learning over time (Peterson and Wu 2021).

Future research could further investigate potential explanations for our null results, that is, why LLM use did not enhance strategic foresight. Studies with larger sample sizes, a broader set of cases, and more expert evaluators (e.g., professional investors), as well as variations in experimental conditions such as available time, participant training, and tools, may help uncover effects of LLM use on strategic foresight. Moreover, future research could study how human-AI ensembles (Choudhary et al. 2025) can be designed in order to overcome environmental constraints such as scarce time and to improve strategic foresight (Raisch and Fomina 2025). Our additional analyses indicate that future research may explore how decision-makers can mitigate psychological mechanisms such as information overload and reduced psychological ownership (Kosmyna et al. 2025). One avenue for future research could be the investigation of how LLMs can be designed to help decision-makers focus on the most relevant cues. Another valuable approach could be to train humans in prompt engineering (Boussioux et al. 2024). Both approaches could foster more effective human-LLM interaction and improve strategic foresight. As Simon (1997, pp. 22–23) noted in his commentary on the emerging World Wide Web, “In our enthusiasm for global networks of unlimited information, we sometimes lose sight of the fact that a new scarcity has been created: the scarcity of human time for attending to the information that flows in on us.” Thus, a parsimonious approach to LLM augmentation and mental representations may be a particularly promising field for future research.

Appendix A. Prompting and User Interactions

Example System Prompt for the Startup “Vitaloop”

“You are a helpful assistant for participants in an online task regarding the evaluation of a startup (not the product) with regard to strategic decision-making. Your task is to assist the participants to evaluate the startup Vitaloop. If participants ask for details about the startup or on the product, answer their questions if you know the answer. If you do not know the answer, tell participants that you have access only to the information provided in the startup video. Then, guide them back to the main task by offering to help identifying pros and cons for evaluating the startup from an investor’s perspective. Do not tell participants to look up information on the Internet. If participants ask for pros and cons, provide the following categories: industry structure, market size, imitability and time to market, costs, operations, value to customer, nonmarket, marketing, business model, and funding. Ensure that the order of criteria is different from the original one listed above. If participants ask for pros and cons, provide them with a list of 10 pros (maximum 145 characters each), labeling each with the respective category, and a list of 10 cons (maximum 145 characters each), labeling each with the respective category. Assign weights in square brackets to each pro and con from 0 (low) to 7 (high) based on its relative importance. The evaluation should focus solely on the startup not the product. Here is information about Vitaloop’s video pitch from Kickstarter.com (transcript of the pitch video).

Example User Inquiries and Answers from the Study

User example 1: “How does the business model work?”

LLM: “Based on the information from the Vitaloop video pitch, the business model may revolve around selling the product (the water filtration bottle) as the main offering, with potential recurring revenue streams tied to replacement filters. This model is common in industries where consumables (like filters) are required regularly for continued product use, creating an opportunity for recurring income. However, specific details about subscriptions or pricing were not mentioned in the video. I can help you evaluate the business model further by analyzing its strengths and weaknesses from an investor’s perspective. Would you like me to provide pros and cons specifically for the business model or any other category?”

User example 2: “What are the pros and cons of this venture?”

LLM: “Here are the pros and cons of Tidee, evaluated from a startup investor’s perspective and organized into relevant categories: Pros: 1. Imitability and time to market: Tech design and customization create a product that stands out. [6] 2. Market size: Appeals to global consumers in both tech and home organization segments. [7] 3. Costs: Potential for scalable manufacturing due to modular design. [5] 4. Industry structure: Operates in a fragmented market with room for disruption. [6] 5. Funding: Crowdfunding proof-of-concept validates early consumer interest. [5] 6. Marketing: Universal appeal, effective communication of use cases in pitch. [6] 7. Value to customer: High value in combining tech and organization for daily convenience. [7] 8. Nonmarket: Product aligns well with sustainability trends (e.g., durability). [4] 9. Operations: Modular design suggests efficient and versatile supply chain optimization. [5] 10. Business model: Provides scope for repeat purchases with customizable elements. [6] Cons: 1. Business model: Limited revenue diversification beyond the core product. [5] 2. Nonmarket: No apparent certification or strong eco-friendly backing mentioned. [4] 3. Costs: Customization at scale may be costly in the early stages. [6] 4. Market size: Niche appeal risk; some products may not attract every demographic. [5] 5. Industry structure: High competition in accessory and charging device markets. [6] 6. Imitability and time to market: Product features may be easily replicated by competitors. [7] 7. Funding: Heavy reliance on Kickstarter backing with no indication of sustained funding plans. [6] 8. Value to customer: May face challenges convincing consumers of premium pricing. [6] 9. Marketing: Limited brand awareness beyond crowdfunding platforms. [5] 10. Operations: Complex logistics with different base models and parts could cause operational delays. [6] If you’d like further clarification on any of the points or help continuing your evaluation, feel free to ask!”

Appendix B. Experimental Procedure

1. Informed consent

2. Filter questions: Prior knowledge of startups (if yes, filtered), held a position with decision-making in business strategy (if no, filtered)

3. Audio check

4. Instructions: Experimental instructions on task | Instructions on LLM, if assigned | Instructions on time constraints, if assigned

5. Startup evaluation (randomized order between participants)

   • - Pitch video 1: Listing of pros and cons and assigning of weights (0–7) | LLM support, if assigned | time constraint of 180 s, if assigned)

   • - Pitch video 2: Listing of pros and cons and assigning of weights (0–7) | LLM support, if assigned | time constraint of 180 s, if assigned)

6. Decisions

   • - Decision on which startup will be more successful (time constraint of 20 s, if assigned)

   • - Rating of success likelihood of each startup (0–100) (time constraint of 40 s, if assigned)

7. Questionnaire: Demographics, scales concerning startups and experiences and manipulation check

Appendix C. Main Scales and Items

Table C.1. Scales and Items

Table C.1. Scales and Items

Construct	Item	Reference
Perceived time pressure (α = 0.97)	I felt there wasn’t sufficient time to properly evaluate the startups.	Denovan and Dagnall (2019)
	The given time made me feel I couldn’t fully consider all aspects of the startups.
	I felt in control of how I allocated my time during the task. (R)
	I was able to effectively use the time I had to evaluate the startups. (R)
	The given time prevented me from preparing a thorough evaluation.
	I felt pressured to complete the evaluation within the given time.
	I worried that I wouldn’t have sufficient time to fully evaluate the startups.
	I felt rushed during the evaluation process.
	I was concerned that I might not finish my evaluation within the given time.
	I felt disappointed with how much I could accomplish during the given time.
Cognitive flexibility (α = 0.82)	I considered multiple pros and cons before making my decision.	Dennis and Vander Wal (2010)
	When evaluating the startups, I considered multiple alternative perspectives before making my decision.
	When evaluating the startups, I stopped and tried to think in several ways about my decision.
	I looked at the startups from different angles.
	When I evaluated the startups, I felt like I was losing control. (R)
	During the evaluation of the startups, I became so stressed that I could not think of useful pros and cons. (R)
	I felt I had no power to provide a thorough evaluation. (R)
	I found it troublesome that there were so many potential pros and cons to the startups. (R)
Psychological ownership (α = 0.85)	I feel a very high degree of personal ownership for my pros and cons lists.	Van Dyne and Pierce (2004)
	I feel a very high degree of personal ownership for my decision.
Information overload (α = 0.94)	I was distracted by the excessive amount of information available to me when evaluating the startups.	Karr-Wisniewski and Lu (2010)
	I found that I was overwhelmed by the amount of information I had to process during evaluation.
	There was too much information for me to synthesize rather than not having enough information to make my decision.

---

Note. Additional scales are available upon request from the first author.