Generative artificial intelligence and decision-making: evidence from a participant observation with latent entrepreneurs Open Access

The decision-making process is widely discussed in many fields – including management, psychology, sociology and political science – aiming to understand how humans make decisions (Gilovich and Griffin, 2010; Hastie, 2001; Shepherd et al., 2015). In brief, the decision-making process concerns how human beings select the best options (Beach, 1993). Mainly, several studies investigated how decisions are made under uncertainty and chaotic conditions (Ganter and Hecker, 2013; Makowski and Kajikawa, 2021), becoming a challenging topic in entrepreneurship research (Gartner, 1985; Alvarez and Barney, 2007; Sarasvathy, 2001; Shane and Venkataraman, 2000; Shepherd et al., 2015).

Part of the conversation tradition focuses on identifying the main models used to trace decision-making processes (Beach and Lipshitz, 2017; Saaty, 2008; Saaty and Kearns, 2014). On the other hand, some studies focus on the main stages that make up this process (Beach and Lipshitz, 2017; Doya, 2008; Lunenburg, 2011; Simon, 1972). Remarkably, the academic debate mainly concentrates on how the limited rationality of human beings can influence their outcomes (Simon, 1957, 2000; Edwards, 1954), as their memory and reasoning capabilities are not proportionate to the complexity of the problems to be addressed (Simon, 1978). Therefore, intrinsic factors such as motivation, emotion, intuition, courage and previous experience significantly influence decisions (Lerner et al., 2015; Sircova et al., 2014; Baron and Tang, 2009). An individual's previous knowledge and experience lead them to trust what they instinctively think about a specific choice that must be made or a task to be accomplished (Simon, 1997).

In this context, the role of digital technologies – such as artificial intelligence (AI), machine learning and big data–emerged as a remedy to reduce “heuristic distortions” in the decision-making process derived from the limited rationality of human beings (Werner et al., 2018; Hassani and Silva, 2023; Hao et al., 2024). Notably, the analytical and predictive capabilities of AI-based solutions – particularly Generative Artificial Intelligence (GenAI or GAI) - could contribute to improving the effectiveness and efficiency of business decisions (Kraus et al., 2020; Giuggioli and Pellegrini, 2023) from a human–machine interaction perspective (Brynjolfsson and McAfee, 2017; Bughin et al., 2017; Jeremiah, 2024), thus transforming individuals into “Human+” entities (Bartoloni et al., 2022; Gravili et al., 2023) and addressing some issues related to potential biases typical of human reasoning (Cristofaro et al., 2024; Cristofaro, 2017), ethical concerns (Russo-Spena et al., 2022), the reliability of content generated by such technologies (Sedaghat, 2024; Azaria et al., 2024; Mitchell, 2019; O'Neil, 2017), as well as substitution issues (Huang and Rust, 2018). Despite the key role of digital technologies in the decision-making process, the literature requires further investigation on how humans adopt them in making decisions, specifically through empirical exploration, severely hindering the accumulation of knowledge in this critical area of investigation. This is particularly relevant in complex scenarios as an entrepreneurial one, where decision-making is crucial to create new market dynamics (Shane and Venkataraman, 2000; Shepherd et al., 2015). Therefore, this paper aims to bridge this gap by answering the following research question (RQ): How do humans interact with Generative Artificial Intelligence to make decisions related to an entrepreneurial task?

The authors adopted an exploratory approach to tackle the research question, conducting participant observation. As previously mentioned, one of the primary settings where decision-making is crucial is entrepreneurship (Shane and Venkataraman, 2000; Sarasvathy, 2001). In this context, the weight of the above-mentioned factors varies based on individuals' current stage of the entrepreneurial path. Indeed, literature identified several types of entrepreneurs from the latent to the serial ones (Dew et al., 2009; Audretsch et al., 2022). Particularly, latent entrepreneurs are defined as individuals without previous entrepreneurial experience who simply express their preference to be self-employed instead of being a salaried employee (Vamvaka et al., 2020; Grilo and Irigoyen, 2006; Kolvereid, 1996; Verheul et al., 2012).

The body of knowledge about the latent entrepreneurs is still limited, especially when it comes to GenAI and decision-making (Audretsch et al., 2022). Thus, the latent entrepreneurs' decision-making process is qualified as a fertile exploration ground to observe human-GenAI interaction in making decisions to solve highly specialised tasks. To let hidden dynamics emerge in the decision-making process GenAI realm, it is crucial to mitigate individuals' reliance on previous experience and background in the field. Consequently, being a latent entrepreneur is the main selection sample criterion in this research. In doing so, the authors selected a group of individuals attending a master's course on entrepreneurship. Particularly, the master's attendees were chosen because of their latent entrepreneurial nature tested by the master committee. Additionally, the authors decided to set an entrepreneurial task for participants. Indeed, tasks encourage individuals to produce the kind of processes that they produce naturally (Salaberry and Comajoan, 2013; Mukherjee, 2019).

The analysis led to the development of the following main themes: (1) Human-machine interaction: Taking or Giving? (2) Humans as leaders of the conversation and (3) Perspective on GenAI: Operational tool vs Learning source. Authors based on these themes developed a conceptual framework to contribute to the current academic debate on innovation, decision-making and entrepreneurship (Korteling et al., 2021; Cristofaro and Giardino, 2025; Cristofaro et al., 2024; Cucari et al., 2023) by offering several implications for researchers, policymakers and practitioners. Particularly, this paper supports the decision-making and entrepreneurship literature by improving the understanding of how individuals without previous experience interact with GenAI to make decisions in a highly specialised decisional context as an entrepreneurship-related task. Additionally, authors shed light on latent entrepreneurs, considering them as the first individual category approaching entrepreneurship by implementing entrepreneurial activities.

Thus, from a theoretical point of view, the study contributes to the innovation management debate in tracking the human-GenAI dynamics, highlighting the different individual approaches in using innovative tools. This also supports the research stream on the Human + paradigm, confirming that GenAI could support humans without replacing them. As a consequence, it is reflected in new insights for practitioners and policymakers, reinforcing the idea that GenAI-based tools could be a lever to support humans in their daily activities, especially in complex settings, considering individuals' intrinsic nature in doing so.

A decision is the output achieved through deliberation leading to a consequent implementation of an action (Buchanan and Connell, 2006; De Winnaar and Scholtz, 2020). Individuals make most decisions under uncertainty due to the complex conditions that characterise a business environment and society. Moreover, such complexity means that the volume of required insights to decide is massive, making it hard for humans to acquire all the necessary knowledge to avoid the probability of failure. Besides, not all the effects of such decisions are predictable (Zimmermann, 2011). Decision-making literature is wide, providing different evidence on the nature of such cognitive processes (Beach and Lipshitz, 2017; De Winnaar and Scholtz, 2020; Smith and Semin, 2004; Mitchell et al., 2011). Particularly, according to the academic debate (Doya, 2008; De Winnaar and Scholtz, 2020), a decision-making process is at least composed of four steps: recognition of the present state, evaluation of the immediate options, selection of an action based on one's needs and re-evaluating the option based on the outcome. Given its logical structure, theories explored this topic by considering decision-making as a rational mental pathway. According to the global rationality perspective, decision-makers own the necessary insights to make the right choice. They compute the expected value of utility associated with each alternative and then choose the one maximising the expected utility (Simon, 1997). Rationality implies that people of their choices (Edwards, 1954; De Winnaar and Scholtz, 2020). The theory of rational choice identifies six stages from which further theories and models can be developed (De Winnaar and Scholtz, 2020): problem identification, alternative generation, alternatives evaluation, alternative selection, decision implementation and decision effectiveness evaluation. However, over the years, researchers also considered that decision-makers, as human beings, do not always make the most rational choices without maximising their utility. This is the core assumption of Bounded Rationality (BR) theory (Simon, 1972, 1978, 1991, 2000), for which individuals' decision-making capabilities are “influenced equally by their internal environment as by their external environment, thus preventing individuals from making completely unbiased decisions” (De Winnaar and Scholtz, 2020, p. 1287). According to it, it is also important to understand factors that drive the decision-making process, together with rationality, such as the decision-makers’ needs and desires, their perceived risk regarding the available alternatives, and, finally, the perceived amount of time spent deciding (Doya, 2008). Researchers (e.g. Eisenhardt and Zbaracki, 1992; Pettigrew, 2014; Klein, 2008) suggest that rationality decreases when decision-makers face threatening situations in which uncertainty is high. In literature, it is stated that high levels of uncertainty and risks disincentivise individuals from deciding (Lipshitz and Strauss, 1997). This type of attitude is explainable because of the existence of individuals' decision frames; they represent their perspective on the specific situation, influencing the way they approach the decision and the possible alternatives, leading them to manifest a tendency to prefer normative or prescriptive models (Tversky and Kahneman, 1981; De Winnaar and Scholtz, 2020). This is also true because personal factors such as emotion and cognition, past experience are very influential in decision-making (Lerner et al., 2015; Elster, 1998, 1999).

The Fourth Industrial Revolution highlights how the increasing diffusion and integration of digital technologies, such as AI, into organisational processes contributed to redefining business strategies, competitive positions, organisational structures and decision-making processes (Appio et al., 2023; Khvatova et al., 2023; Shrestha et al., 2019; Haefner et al., 2023; Agrawal et al., 2018; Gorry and Scott-Morton, 1971; Pietronudo et al., 2022). Particularly in contexts characterised by chaos and disorder (Ganter and Hecker, 2013; Makowski and Kajikawa, 2021), AI emerges as an effective tool for managing complexity and uncertainty in making decisions (Agrawal et al., 2018; Kakatkar et al., 2020; Makowski and Kajikawa, 2021) by automating cognitive processes, elaborating vast amounts of data and enhancing interactions (Davenport and Ronanki, 2018).

Additionally, since individuals rely on “cognitive heuristics” (Tversky and Kahneman, 1981; Busenitz and Barney, 1997), resulting in systematic biases in judgments and decisions (Werner et al., 2018; Cristofaro, 2017), AI-based systems, particularly those leveraging GenAI (Sætra, 2023), present a potential solution to these biases (Cristofaro and Giardino, 2025). Thanks to their ability to autonomously extract information from large datasets, they serve as valuable tools for mitigating heuristic biases, overcoming the limitations of human rationality in information retrieval and selection and guiding decision-making toward fairer and more comprehensive choices (Hassani and Silva, 2023; Hao et al., 2024; Rani et al., 2024).

However, based on the assumption that both GenAI and humans (Kamuri, 2022; Jeremiah, 2024; Cristofaro et al., 2024; Korteling et al., 2021; Mehrabi et al., 2021) are potentially biased in their decision-making processes, the industrial paradigm of Society 5.0 investigates the concept of “Human+” promoting the use of technology from a human–machine interaction perspective (Bartoloni et al., 2022; Gravili et al., 2023; Cucari et al., 2023). Particularly, Society 5.0, described as the “revolution of the human touch”, emerged in response to these concerns, aiming to refocus on human enhancement through a more sustainable and effective integration of digital solutions into production processes (Troisi et al., 2024; Bartoloni et al., 2022; Cucari et al., 2023). Thus, technologies are no longer selected exclusively based on efficiency and productivity but also on their ability to support and empower individuals (Troisi et al., 2024). Within decision-making processes, this concept envisions AI as an active collaborator with humans, generating crucial insights to define strategies, optimise operational efficiency and maintain market competitiveness (Brynjolfsson and McAfee, 2017; Bughin et al., 2017).

Although this era has redefined the concept of intelligence by promising greater efficiency and productivity, it is not without risks (Tsamados et al., 2021; Bubeck et al., 2023; Buttazzo, 2023). First, AI integration into decision-making manifests a lack of predictability (Shepherd and Majchrzak, 2022), redefining the boundaries between human intuition and AI. This transformation profoundly impacts human identity, requiring individuals to reassess their roles (Jeremiah et al., 2021; Jeremiah, 2024) and raising concerns about being replaced by machines (Huang and Rust, 2018). Additionally, integrating advanced technologies such as GenAI introduces significant ethical considerations, particularly regarding data privacy. This raises concerns about the reliability of AI-generated outputs, describing GenAI as a double-edged sword (Mitchell, 2019; O'Neil, 2017). Consequently, its outputs must be critically examined to ensure accuracy and reliability (Sedaghat, 2024; Azaria et al., 2024). Moreover, the competitive potential of AI could be undermined by disparities in access and digital literacy, granting disproportionate advantages to those with greater resources and technological expertise (Martinez Dy, 2022; Imjai et al., 2024).

In the academic debate, decision-making is an incredible interest among management researchers, especially in entrepreneurship and innovation (Shane and Venkataraman, 2000; Davenport and Ronanki, 2018; Pietronudo et al., 2022). Researchers studying entrepreneurial processes have focused on how, on the one hand, the internal organisation of firms can shape individuals' decisions regarding the pursuit of specific entrepreneurial opportunities (Wiklund and Shepherd, 2008) and, on the other hand, how the business environment – characterised by high uncertainty, ambiguity and risk–influences decision-making (Baron, 2008; Mullins and Forlani, 2005). Indeed, literature on decision-making widely investigates such issues within entrepreneurship, intending the entrepreneur as a decision-maker who must face challenges such as uncertainty and risks to survive in the market (Shepherd and Rudd, 2014). Indeed, evidence shows that the success of entrepreneurial businesses is related to the effective decision-making skills of the entrepreneur (Maine et al., 2015). Several studies in the field are interested in exploring how, when, where and by whom opportunities are created (Gartner, 1985; Alvarez and Barney, 2007; Sarasvathy, 2001) or “discovered, evaluated, and exploited under uncertainty” by different categories of entrepreneurs: mostly novice and serial entrepreneurs (Shane and Venkataraman, 2000; Shepherd et al., 2015). Thus, generally, the impact that entrepreneurs' decisions have on the market is one of the main reasons their decision-making is widely explored in literature, because the decision-making style they adopt affects the quality and outcomes of their entrepreneurial decisions. However, most of these studies approach entrepreneurship as some kind of dichotomous process where an entrepreneurs own an enterprise or not (Mickiewicz et al., 2017), focusing less on the stages that precede new business activities (Bennett and Chatterji, 2023). The risk is not reflecting the different decision-making processes carried out by different “types” of entrepreneurs (Minniti and Lévesque, 2010; Audretsch et al., 2022; Caiazza et al., 2020; Vamvaka et al., 2020; Grilo and Irigoyen, 2006; Verheul et al., 2012). In real life, it is not certain that all individuals interested in starting a business actually do so, and their decision-making is particularly relevant to the future consequences on market dynamism. According to the literature, they are considered “latent entrepreneurs” since they may either proceed to create a new enterprise or abandon their plans to do so (Audretsch et al., 2022; Caiazza et al., 2020). Thus, given the weight of both rational and irrational individual factors in decision-making, it is rational to assume that latent entrepreneurs, with no past experience nor interaction with the entrepreneurial challenges, would make decisions differently from nascent, emergent, novice, several, etc. (Looze and Desai, 2020; Vamvaka et al., 2020; Audretsch et al., 2022).

Since the decision's complexity depends on the level of thought behind it (Klein, 2008), global rationality is more feasible when the situation is simple, so the identification of the objective choice is evident (Simon, 1997). When this does not happen, evaluation errors can be made and decision-makers rely on heuristics to reach an end choice (Cristofaro et al., 2024; Cristofaro, 2017). However, heuristics are rules adopted to avoid making the wrong decision, but they can be biased and imperfect. Consequently, completely relying on them to make a choice can be deceptive (Kahneman et al., 1982; Bazerman and Moore, 2012; Hammond et al., 1999). Focusing on latent entrepreneurs' decision-making reveals a double face of heuristics usefulness because they cannot rely on them, and at the same time, they are not biased by these mechanisms.

Then, it is also important to understand factors that drive this process together with rationality, such as the decision-makers’ needs and desires, their perceived risk regarding the available alternatives, and, finally, the perceived amount of time spent deciding (Doya, 2008; Luszczynska and Schwarzer, 2005; Wyer and Shrull, 2014).

Giuggioli and Pellegrini (2023) also highlight that GenAI-based solutions are a valuable tool to solve heuristic issues. Despite the potential of GenAI, humans remain distinguished by their ability to reason by paradoxes, which allows entrepreneurs to “navigate in the realm of paradox” (Kamuri, 2022; Jeremiah, 2024), a defining characteristic of entrepreneurial activities (Rosado-Cubero et al., 2022). Moreover, certain entrepreneurial traits are inherently human as they stem from lived experiences over time (Sircova et al., 2014). For this reason, in the transition from Industry 4.0 to Society 5.0, the concept of “Human+” is central. This approach promotes the use of technology to enhance human capabilities (Bartoloni et al., 2022; Gravili et al., 2023). In the context of latent entrepreneurs Human + paradigm acquires a particular configuration, due to technologies' ability to not only mitigate heuristics but also to provide new knowledge and information on a specific topic to users within a complex setting. For this reason, it is consistent to consider that GenAI could represent a supportive tool when people act as latent entrepreneurs since it gives them the chance to try to face entrepreneurial challenges in a controlled setting without the constraint of making decisions in a real entrepreneurial context. Thus, individuals enrich their personal backgrounds about entrepreneurial themes.

To address the RQ, the authors adopted a qualitative case study methodology (Yin, 2018), selecting individuals enrolled in the Master's in Entrepreneurship and Innovation Management (MEIM) program at the University of Naples Parthenope as the unit of analysis, due to their latent entrepreneurial nature. An exploratory analysis (Meredith, 1998) was conducted to investigate how individuals interact with GenAI in their decision-making process. To guarantee the interaction, the authors used a conversational GenAI tool, ChatGPT. Particularly, the authors followed three steps. First, they conducted participant observation to collect primary data. In the second stage, they administered a post-observation survey to gather participants' perceptions of using GenAI and their feedback on its potential in decision-making. To ensure data triangulation, construct validity and reliability (Yin, 2009; Kaman and Othman, 2016), the authors identified secondary sources related to the use of GenAI as evidence, including: (1) online documentation, (2) reports and (3) press articles (see Table 1 in the Appendix). Particularly, the secondary sources were used for two main reasons. First, authors used these sources to align their knowledge on the research topics during the first step of data analysis. Indeed, in that phase, authors separately coded the data collected. Moreover, secondary data acted as a valid tool to ensure convergence between the authors. Indeed, it mitigated the potential conflicts (Hox and Boeije, 2005). The following sections provide further details on data collection and analysis.

As mentioned before, the research setting of the paper is the Master's in Entrepreneurship and Innovation Management (MEIM), an international program jointly promoted by the University of Naples “Parthenope” and the MIT Sloan School of Management. It is a twelve-month master's degree designed to mix theoretical knowledge with real-world approaches to entrepreneurship and innovation. The program provides courses both in Naples and at MIT Sloan in Boston, exposing participants to a high-density knowledge ecosystem and a globally connected entrepreneurial environment. The selection process is organised in two main stages to identify applicants with a strong entrepreneurial inclination.

Firstly, the master committee assesses candidates' curricula to map their academic background and experiences. Then, an oral evaluation is designed, planning two main parts: (1) an individual interview, to discuss candidates' interests, future goals and potential inclination to found a start-up; (2) a collective psycho-attitudinal interview, to assess individuals' problem-solving capability and teamwork orientation. This strict admission process ensures that the selected cohort possesses a marked entrepreneurial mindset and latent entrepreneurial characteristics. Indeed, master participants do not necessarily need to launch a new business at the end of the course, qualifying themselves as individuals who would rather be self-employed. For this reason, the authors selected the master's cohort as the observational sample due to their latent entrepreneurial approach in decision-making within a complex setting without guiding their decisions on past experiences. As a consequence, mitigating external factors that potentially could affect individuals' decision-making in facing entrepreneurial-related tasks guarantees putting the pure human-GenAI interaction at the centre of observation.

As mentioned before, participant observation was selected as a primary source of evidence (Musante and DeWalt, 2010). Remarkably, the authors spent two months (December 2024–February 2025) participating in master MEIM activities to record observations and collect data. This approach enhances the validity of a study as a data collection method, as it provides a deeper understanding of the context and the phenomenon under investigation (DeWalt and DeWalt, 2002). Additionally, it enables researchers to participate in spontaneous events, thereby further enriching data collection and interpretation (Kawulich, 2005; Musante and DeWalt, 2010). Indeed, since the decision-making process and human–machine interaction are both complex phenomena to explore, participant observation could improve such an investigation by identifying crucial aspects for an effective understanding of the dynamics characterising such phenomena.

The potential use of GenAI to reduce uncertainty and complexity is widely discussed in the literature, highlighting its value for entrepreneurs. In this context, previous entrepreneurial experiences influence decision-making positively or negatively. Therefore, this research aims to investigate using GenAI, minimising its influence. Indeed, decision-making, together with problem-solving (Mia et al., 2025), is thus a cognitive process not only observed and boosted through specific tools when it comes to already existing entrepreneurs activating it. In fact, according to Mia et al. (2025), “people with good cognitive skills in decision-making and problem-solving have a better-developed entrepreneurial mindset”. Moreover, such cognitive processes are reinforced with increasing experience (Greenwald and Banaji, 1995). This also explains why literature (e.g. Polanyi, 2009) also confirms that the experts' decisional behaviour, when it comes to solving entrepreneurial issues, relies on intuition rather than analytical rationality, as they do not always decide by “calculating” before they act. This is explained by the human nature to trust their previous knowledge and experience, which guide them in addressing a specific task (Simon, 1997). Indeed, tacit knowledge stored in entrepreneurs' memories can increase the effectiveness of their problem-solving process. To this end, 20 MEIM participants were selected as representative latent entrepreneurs who do not rely on previous experiences when deciding about facing entrepreneurial tasks.

To ensure that the participants were involved in deciding, specific tasks, characterised by different levels of complexity, were developed by the researchers that recreated a realistic setting and enabled them to confront decision-making paths in entrepreneurship-related challenges. Additionally, tasks were formulated based on the main concepts faced during the master's to align the participants' background on the topic. Additionally, the authors assigned an entrepreneurial task to participants to engage individuals in processes that occur in real contexts. Indeed, tasks can elicit genuine forms of performance, allowing participants to reveal their own ways of reasoning, acting and making decisions (Salaberry and Comajoan, 2013; Mukherjee, 2019).

ChatGPT was selected as GenAI technology. Launched in November 2022, it has rapidly gained popularity for its ability to generate human-like responses and assist with content creation (Brown, 2024). Notably, they were asked to develop an innovative project that could be crowdfunded. Tasks were articulated in several steps to be accomplished within a specific time. In the first step, the participants used ChatGPT to generate the project idea, spending about ten minutes. Then, they interacted with ChatGPT to develop a crowdfunding campaign, writing a summary of the process. They also developed an essay explaining how ChatGPT helped them and what suggestions they followed or discarded. Additionally, they provide the transcripts of their conversation with ChatGPT, extracting it under researchers' control. To ensure that the observation process was conducted ethically, participants were informed about the purpose of the study, the duration of the stay and the researcher's identity. Permission to conduct the study was obtained from the master committee. The summaries and essays developed were qualitatively analysed following an inductive thematic analysis (Braun and Clarke, 2006).

As mentioned before, to reinforce the analysis, the authors administered a post-observation survey to evaluate perception towards ChatGPT. A total of 20 post-surveys were collected. The survey was developed by adapting Xue et al. (2024). Particularly, it includes five main sections: (1) participants' demographic information and background (see Table 2 in the Appendix), (2) General point of view, (3) Perception towards ChatGPT, (4) Intention of use and (5) Feedback. The questions were designed as Likert scale items ranging from 1 to 7, where 1 represents “Strongly Disagree” and 7 represents “Strongly Agree”, as shown in Table 3 in the Appendix. The participants' suggestions were then thematically coded.

Concerning reports and online documentation, the authors collected and analysed various reports and documents from leading consultancy firms on the adoption and use of GenAI in business operations. Additionally, statistical data and specialised online trade publications were examined to explore how GenAI could transform the modern competitive landscape (see Table 4 in the Appendix). In a first screening phase, ten sources of secondary data were selected. The authors then applied the following inclusion criteria: (1) focus on Generative AI and (2) focus on decision-making. Thus, only four sources were selected. In the next paragraph, the main steps of data analysis are described.

The empirical and exploratory nature of the study prompted the authors to adopt thematic analysis as a qualitative analysis technique (Braun and Clarke, 2006). It is particularly suitable to identify key patterns of meaning within the data, including affective, cognitive and symbolic dimensions (Joffe, 2011). Thematic analysis is characterised by high accessibility and flexibility, not bound to any specific theoretical framework. An inductive approach was adopted, using NVivo software (Joffe, 2011; Cohen et al., 2002; Braun and Clarke, 2006), due to its ability to let models or themes emerge directly from the data, rather than being shaped by pre-existing theoretical assumptions (Patton, 2014). Thematic analysis was driven by primary and secondary data. Additionally, the open-ended questions of the post-survey questionnaire were thematically analysed. This allowed authors to map participants' consideration in using GenAI, expanding the authors' understanding of the data gathered during the observation.

To enhance the validity and reliability of the study, the authors conducted separate coding before comparing their findings (Patton, 2014). To ensure rigour and clarity in the analysis, the study followed the six-step model proposed by Braun and Clarke (2006). First, the authors familiarised themselves with the data through a non-systematic review of the conversations with ChatGPT, eliminating non-essential words or phrases. In the second stage, they systematically explored the data, assigning codes through a descriptive and interpretive approach. Then, the identified codes (see Table 5 in the Appendix) that shared the same characteristics were reviewed and categorised into potential themes. Each was labelled to establish a clear focus, scope and purpose, with concise explanations provided. As previously stated, the participants' suggestions were subjected to a thematic analysis and integrated with the themes derived from the analysis of ChatGPT's conversations. Three main themes were selected: (1) Human-Machine Interaction: Taking or Giving? (2) Humans as leaders of the conversation and (3) Perspective on GenAI: Operational tool vs Learning source.

The following paragraph highlights the main findings.

Thematic analysis was used to understand the interaction between individuals and GenAI during the decision-making process. The following paragraphs describe in detail the above-mentioned themes.

This theme indicates the modality through which the individual interacts with the machine through prompts. Specifically, during the decision-making process, participants could interact with the machine by adopting two different behaviours classified as: (1) taking and (2) giving. The first occurs when individuals “take” information in the form of opinions and/or knowledge, while the second refers to a proactive attitude manifested when individuals “give” information in the form of their own opinions or prior knowledge. Specifically, taking opinions and/or knowledge can occur implicitly, because of the conversational nature of the specific technology analysed or explicitly.

Therefore, as it usually happens in dialogues among human beings when they must make a decision, also in human-machine interaction technology, the interlocutor provides its opinion while dialoguing without an explicit user's request to do it. Indeed, the following glance at a participant's conversation with ChatGPT illustrates such dynamic: “(1) Participant A: I must do a finance project on developing an innovative idea that could be funded by equity crowdfunding; My idea is to build up a start-up to streamline the lengthy Italian bureaucratic process of setting up a company; (2) ChatGPT response: That's a great idea! Bureaucracy in Italy is notoriously complex, and a startup that simplifies the process of setting up a company could be very attractive to investors and entrepreneurs alike.”

However, in most cases, the individual explicitly asked for the machine's opinion to drive their decisions in completing a task. As stated, the following statements were derived from participants' conversations: Participant B: “What is the best region in Italy to build up an Energy community of this type?”; Participant C: “Does it make sense for you to launch this idea through a crowdfunding campaign?”; Participant D: “Why people should support this project?”. Notably, participants treated ChatGPT as a more expert interlocutor, manifesting their perception of the existing information asymmetry leading them to consider machine opinion as more reliable than their one: “I have ten minutes to think about an innovative project that I believe could be successfully funded via CF. I am mostly interested in these fields: medicine, telemedicine, interior design, vintage fashion. You can change the topic since they are just suggestions”. This approach is a common behaviour among people interacting with sophisticated technologies since their knowledge about their intrinsic mechanisms and functionalities is often limited.

By observing the human–machine interactions, emerged from all the data a common behaviour among the participants. These last, indeed, never only “took” the AI opinion, they always also “gave” their own, manifesting their desire to be an active part in the decision-making process.

They did not take the machine's responses for granted but collaborated in the creation of debates (lasting as long as 3–4 prompts) to formulate more structured opinions, driving decisions. The following statements illustrated it: “I doubt what would be the community that will eventually support the project you suggested. Many people who shop vintage goods may be reluctant to spend money to verify a product and most of them may be in the sector for years so they will probably not need an agent AI to verify a product. Probably, they could do that on their own and I believe this is a huge piece of the community”.

The same dynamics occur when it comes to knowledge. Individuals requested ChatGPT to deepen some topics and insights provided by it when they recognised the need to enrich their knowledge baggage. Such added knowledge was not always necessary to complete the task but was perceived as useful to human to enhance their comprehension of a specific topic. For example, one of the participants not only requested which Italian region was most appropriate for the implementation of their business idea but also asked what made one region more appropriate than the others, even if it was not requested by the task. Thus, such information qualified as crucial in driving the human decision-making process even when this means contradicting users' ideas: “If you do not like any of them proposed, provide me a new example based on your knowledge and teach me something about it”. Thus, humans “took” and learned the knowledge given by ChatGPT.

Some participants, on the other hand, explicitly collaborated in this decision-making process by first sharing their knowledge by incorporating some previously learned theoretical notions into the prompt. In this way, individuals “gave their knowledge” to the machine, activating a reciprocal knowledge-sharing process.

The second theme developed is “Humans as leaders of the conversation”. Particularly, while interacting with ChatGPT, humans obtained responses which did not always meet their expectations, explicitly manifested within the prompt due to their higher proactiveness. This is consistent with the post-observation survey's results. Participants perceived AI's responses as not reliable without human interventions and supervision (see Table 6 in the Appendix). Additionally, it also emerged that some critical issues related to the overuse of AI without proper supervision regard the risk of reducing critical thinking ability and weakening creativity, leading to less original solutions and more superficial decisions.

For this reason, all the data attested that everyone manifested an intention to improve the conversation's output by taking control of it through the implementation of at least one of the following facilitative communicative mechanisms: (1) goal alignment, (2) role-playing and (3) training AI.

In the first case, the individual aims to monitor the machine's understanding of the prompts to ensure the alignment of the goals along the decision-making process. Goal alignment is a common activity of decision-making based on dialogue. As stated in the following sentence, indeed, the individual asks ChatGPT its goal: “Do you want to create a solar energy park for a community? Is this an energy community? Specify for me all the technical features for implementation also based on the provided material in this prompt”. In the same way, they communicate their intent to bring the machine to respond according to them: “I am the founder: an engineer passionate of energy efficiency and sustainable energy systems”.

The second mechanism occurs when participants ask ChatGPT to play the roles of a specific actor, ensuring its outputs are derived from a specific data background, as illustrated in the following prompt: “Act as an investor of a startup and explain to me if this is a good idea to create and if it will have success with the crowdfunding”. The third one regards AI training. Individuals acted as trainers by attaching scientific materials on which the machine had to base its responses. This is consistent with the post-observation survey's results. It emerges, indeed, that the reliability of ChatGPT's responses is believed to depend heavily on the accuracy of the prompts and how appropriately the human trains the machine.

The analysis of all the data led to the identification of two different purposes pursued by participants while interacting with the machine: Operational tool vs Learning source. In the first case, participants leveraged the machine's operational capability by using it as a simple executive tool. They provide ChatGPT with specific technical commands to execute some activities related to the task, supporting the decision-making process. The following participants' prompts show it: Participant A: “Give me a summary of all the possible costs I have to incur and provide me with a total”; Participant B: “Give me the Reward Structure for the Crowdfunding Campaign”.

Learning by making decisions is typical in complex decision-making, where unpredictable situations occur, challenging humans in acquiring new knowledge to face them. This is even more true when the individuals did not have the opportunity to learn from previous experiences and background before making that decision. Thus, selecting individuals qualified as latent entrepreneurs as research participants led to another dynamic occurrence.

In interacting with the machine, individuals intended it as a learning source. Indeed, according to the data, the lack of previous experiences improves the value of the learning chance, pushing individuals to take this last to reduce the knowledge gap. This occurred whenever the individual required in-depth knowledge, as illustrated above. When the machine gave information, many participants asked for further explanations. Indeed, along the decision-making process, participants increasingly used AI to expand their knowledge baggage, beyond what is strictly required by the task instructions. Based on such expanded knowledge baggage, individuals made decisions. In some cases, they changed their mind on specific choices thanks to what they learned through the interaction, highlighting the relevance of GenAI on their cognitive processes.

Thus, findings suggest that the operational purpose occurred mainly in several cases: (1) when the individuals desired to save time, (2) when human oversight was not crucial to complete the activity and (3) when individuals were called to process a vast amount of information to complete the activity. Despite that, the more an individual's learning desire increases less they care about (1) saving time, (2) the relevance of their supervision and (3) the amount of information.

This study investigates human–machine interaction to explore how GenAI influences the decision-making of individuals qualified as latent entrepreneurs, contributing to the current academic debate on innovation, decision-making, Human+ and entrepreneurship (Korteling et al., 2021; Cristofaro and Giardino, 2025; Cristofaro et al., 2024; Cucari et al., 2023). When humans interact with GenAI, a virtuous cycle of knowledge exchange and better time management is activated by it (Figure 1).

Particularly, Figure 1 shows that: (1) human intention guides the whole decision-making process, (2) human decision-making is shaped by two main pillars' management (i.e. time and knowledge), (3) human-machine interaction enhances the humans' management of these pillars (e.g. GenAI as a learning source or GenAI as an operational tool). Consequently, this virtuous cycle leads to Human + decision-making. Technologies' role in usurping human capability and supremacy is widely discussed in literature, which leads to the urgency to promote human–machine collaboration to reduce negative technology impact, safeguarding human empowerment. Thus, this study shows that humans, in interacting with GenAI, activate several mechanisms to not lose their supremacy, improving their decision-making instead.

Indeed, according to the literature (Smith and Semin, 2004; Mitchell et al., 2011; De Winnaar and Scholtz, 2020), cognitive processes are not static but situated within individuals and situations characterised by different configurations of knowledge flow and time, confirming the key role of such pillars in shaping human decision-making. Individuals' momentary interactions with the context significantly impact their cognition. Therefore, this research investigated such interactions between an individual and an external source as AI, to detect the latter's influence on decision-making cognitive processes in addressing entrepreneurial-related tasks. Using ChatGPT presupposes that it acts by integrating and sharing information and opinions with humans, ultimately generating knowledge. Indeed, in real life, when it comes to dialogues among individuals, these last implicitly or explicitly seek the interlocutor's opinions and receive them. Findings reveal that such human behaviour is replicated during a GenAI–human debate. Indeed, post-observation survey results demonstrate that individuals recognised the machine's capacity to assist by offering suggestions and information that are essential for resolving entrepreneurial issues. According to the results, this is related to the enhancement of latent entrepreneurs in managing time and knowledge derived from human–GenAI interaction. In doing so, individuals can approach the machine both as an operational tool and a learning one. This expands the literature on latent entrepreneurs, paving the way for future research (Vamvaka et al., 2020; Grilo and Irigoyen, 2006; Kolvereid, 1996; Verheul et al., 2012; Audretsch et al., 2022; Caiazza et al., 2020).

Nonetheless, they also show that they do not think technology's suggestions are reliable without human intervention and supervision. Indeed, in the literature, the main concern relates to the proliferation of false information that is perceived as truth, described as the “Library of Babel” (Passmore and Tee, 2024), a scenario in which falsehood and truth overlap, undermining credibility (Liu et al., 2024). In addition, GenAI models are based on complex algorithms and unclear logic, creating a “black box” effect that can lead to reduced confidence in the reliability of outputs (Appio et al., 2020). As such, monitoring of GenAI is critical, requiring human oversight to determine whether suggestions are actionable or need to be restructured (Fridman et al., 2019). This aligns with the central role of human willingness and intentions, driven by rationality, highlighted by the framework, in guiding decision-making. This is consistent with the BR theory, which does not view decision-making as a purely instinctive and emotional process, as it acknowledges that individuals adopt rational approaches at certain stages of the decision-making process.

This human attitude towards technology has been widely investigated in the literature. For instance, it occurs in constructs such as the level of trust in AI (Glikson and Woolley, 2020), the perception of its output's reliability (Mitchell, 2019; O'Neil, 2017), fear of being substituted by AI (Huang and Rust, 2018), fear of losing entrepreneurial identity (Jeremiah et al., 2021, 2024) and, generically, perceived risks associated with technologies (Davis, 1989). Moreover, the intrinsic nature of GenAI integrated with conversational technology led the adopted tool to simulate human behaviours as autonomous thinking and proactiveness by going beyond human requests (Yan et al., 2024). For this reason, it does not always satisfy them, increasing the individual's perceived risk associated with AI use (Hasan et al., 2021). Additionally, findings confirmed literature highlighting that when an individual must decide never been made before or made within a completely different scenario, they first try to assess whether the situational conditions are like the past context and then seek more information to generate alternatives (Klein, 2008).

On the other hand, findings confirmed literature (e.g. Luszczynska and Schwarzer, 2005; Wyer and Shrull, 2014), explaining that participants were more engaged and committed towards decisions when they perceived that they were in control of the situation (Luszczynska and Schwarzer, 2005). Other facilitating mechanisms to lead the conversations are applied by the participants. Some of them gave AI a role like an entrepreneur or an investor, manifesting humans' intention to improve the quality and reliability of AI answers (Sedaghat, 2024; Azaria et al., 2024). Another human's conversational attitude is assuming the trainer role. Once again, human beings proactively approach AI, although some people use it primarily to perform and automate tasks (Davenport and Ronanki, 2018). Thus, GenAI must not be intended anymore as a mere tool for information organisers. The AI solutions investigated enable the human-machine interaction to be intended as a mutual learning opportunity. However, when humans use it to ask for knowledge and not mere operational content, GenAI starts to be intended as a valid tool for humans' learning processes. Findings, indeed, illustrate that the knowledge required is not always necessary to solve the specific phase of the task. Nonetheless, these people still want to learn more before deciding. It also happens that once the necessary knowledge is acquired, decision-makers change their preferences because of machine influences, confirming theories affirming that decision-making is affected by context's influences (De Winnaar and Scholtz, 2020; Mitchell et al., 2011; Randolph-Seng et al., 2014).

In this study, the authors have conducted an exploratory study to investigate how individuals without past experiences interact with GenAI when making decisions related to entrepreneurial tasks. Through participant observation authors aimed to explore how a sample of master's attendees – defined as latent entrepreneurs – use GenAI during their decision-making process to solve a specific task. The thematic analysis highlighted three main themes useful to develop a conceptual framework that laid the groundwork to map a Human + decision-making process and the main factors guiding it. The study contributes to innovation and entrepreneurship studies by providing new insights on human-GenAI interaction when it comes to latent entrepreneurs, an under-investigated category in the current academic debate. It shed light on how individuals with no previous experience approach to GenAI-based tool to improve their decision-making guided by their intention to safeguard their supremacy all along the process. Moreover, it emerged that human–GenAI interaction improves the individuals' management of the main pillars of decision-making in a complex scenario (i.e. time and knowledge). Indeed, GenAI characteristics lead individuals to approach it as an operational tool and a learning source.

Based on these premises, this study contributes to theory and practice, providing helpful implications. From a theoretical perspective, this study bridges the gap in understanding how individuals lacking previous entrepreneurial experience interact with GenAI to guide their decision-making processes within a highly specialised entrepreneurial domain. Although the transformative power of GenAI is widely attested, several studies focus only on its purely technical features without delving into how this technology is used by individuals.

Thus, from a theoretical point of view, the study contributes to the innovation management body of knowledge in mapping the human–GenAI interaction, shedding light on the different individual approaches in using technology. The conceptual framework reveals the potential role of GenAI in all six stages of decision-making identified in literature (De Winnaar and Scholtz, 2020). This also contributes to the research debate on the Human + paradigm, validating that GenAI does not necessarily replace individuals. Indeed, the conceptual framework delineates the main pillars guiding Human + decision-making. This work provides important and interesting food for thought for scholars of innovation and entrepreneurship, paving the way for future research. Particularly, future studies could explore in depth the different individuals' patterns, providing a vademecum in using GenAI to support complex decision-making processes when previous experiences are lacking. Scholars could replicate the study investigating the interaction between other kinds of entrepreneurs (e.g. serial entrepreneurs, novice entrepreneurs, etc.) to map similarities and differences among different categories. Additionally, a longitudinal study can provide more evidence on the evolution of human approaches in interacting with machines over time.

From a managerial point of view, this paper implies the need for managers to promote a human-centred AI culture in which competencies are integrated with a view to the co-creation of value, based on the GenAI role in supporting individuals' management of time and knowledge while making decisions. This is consistent with McKinsey's report on GenAI and its role in decision-making processes (Chui et al., 2022). Finally, this study offers food for thought for policymakers called upon to define guidelines for the ethical use of AI in decision-making processes, supporting the Human + paradigm.

Furthermore, for AI to become a strategic tool, policymakers must define training programmes aimed at developing AI interaction skills, so that individuals can identify potential biases in AI-generated outputs. Thus, digital technologies are not a threat but an opportunity to boost human cognitive processes if correctly implemented and regulated by policymakers.

This paper is not without limitations. First, results could be influenced using a specific type of GenAI. Future research could replicate the analysis by selecting other technologies.

Additionally, the results may be context-specific, which could reduce the generalisability of the findings. Future studies could compare individuals from different contexts or examine the effect of technology on long-term decision-making in a longitudinal study. Additionally, as mentioned by Cristofaro and Giardino (2025), future studies should investigate the synergy between humans and AI by adopting a granular, longitudinal approach exploring different industrial settings, also considering the potential biases and dark side of AI (Cristofaro et al., 2024; Cristofaro, 2017) occurring in such dynamics, limiting the power of such collaborative process.

The supplementary material for this article can be found online.