Sufficient and necessary conditions in the adoption of generative artificial intelligence by entrepreneurs: findings in Spain

The expansion of generative artificial intelligence (GenAI) has had a profound impact on economic, organisational, and business processes (Polak and Anshari, 2024). AI-assisted financial tools foster green finance engagement and, indirectly, sustainable consumption behaviour (Matta et al., 2025). Its influence is of particular interest for entrepreneurs, as it expands their operational capacity and facilitates the early scalability of new business projects (Abaddi, 2024). GenAI facilitates the generation of business ideas, the development of business plans, the exploration of opportunities, and the reduction of the time associated with the initial stages of business creation (Abaddi, 2024). In addition, it can reduce certain barriers to market entry, since it allows some projects to be developed with lower initial capital, smaller teams, and shorter product development cycles (Fernandes et al., 2024).

GenAI can improve business decision-making by processing large volumes of information, identifying patterns, and supporting strategic planning (Spaniol and Rowland, 2023). Its capacity to process and structure information efficiently enables firms to detect emerging market patterns, evaluate uncertain future conditions, reduce costs, and manage complex operations more accurately (Biloslavo et al., 2024). It can also contribute to the delegation of routine administrative and monitoring duties to automated systems, allowing managers and staff to devote more time to value-creating tasks (Ponnock, 2025).

The use of GenAI also poses important risks in entrepreneurial activities. Dependence on external technology providers can limit firms' strategic autonomy and create vulnerability to changes in prices, access conditions, regulation, or model availability (Dwivedi et al., 2023). Its adoption may lead to a certain competitive homogenisation, reducing the possibilities for strategic differentiation (Spaniol and Rowland, 2023). Added to this are risks associated with dependence on complex technological systems, model errors or hallucinations, the generation of overly generic strategies, and the ethical and legal problems derived from the use of biased, incorrect, or unverifiable information (Dwivedi et al., 2023). Organisational change management constitutes another relevant challenge, since both the lack of workforce training and concerns associated with its potential displacement effects may hinder the adoption of these tools (Humlum and Vestergaard, 2024).

These reasons motivate the present research, which analyses the use of GenAI among entrepreneurs in Spain. Unlike much of the literature on technology acceptance, which focuses on behavioural intention, this study analyses the actual use of GenAI. This choice is justified because intention and behaviour do not always fully coincide (Berkowsky et al., 2017). In addition, GenAI is embedded in commonly used technologies, such as messaging services, social networks, and web browsers. This immediate availability considerably lowers the access threshold and allows the use of these tools to occur in a relatively unplanned way, through habit, in response to contextual opportunities, or as an immediate solution to a specific need. Under these conditions, actual use may be a more appropriate indicator than behavioural intention for understanding the real adoption of GenAI (Viglia et al., 2024).

The proposed explanatory model, summarised in Figure 1, is structured around two axes. The first is based on the Technology Readiness Index (TRI) (Parasuraman, 2000; Parasuraman and Colby, 2015). The TRI measures individuals' general predisposition to adopt and use new technologies. This predisposition combines facilitating and inhibiting factors. The former includes technological optimism and personal innovativeness, which favour technology acceptance. The latter include technological scepticism and insecurity, which may generate resistance or delay adoption.

The second axis is linked to the Theory of Planned Behaviour (TPB) (Ajzen, 1991) and seeks to establish the drivers and barriers associated specifically with the use of GenAI. From this perspective, the performance of a behaviour depends on the attitude towards that behaviour, perceived behavioural control, and social influence. Following Pérez-Portabella et al. (2025), the perception of positive effects of GenAI on the economy (ECU) and perceived risk (RISK) are considered to contribute to shaping attitudes towards this technology. For its part, knowledge about GenAI (KNOW) is interpreted as an element associated with behavioural control. The perceived need for regulation (NREG) is approached as a normative dimension, insofar as it reflects social, institutional, and collective concerns about the appropriate use of GenAI (Pérez-Portabella et al., 2025). Whereas ECU and KNOW may be considered potential drivers of GenAI use, RISK and NREG may act as inhibitors.

Based on this theoretical framework, two research objectives (RO) are proposed:

This study contributes to the literature on GenAI adoption by focussing on entrepreneurs' actual use rather than behavioural intention, integrating general technology-readiness dispositions with GenAI-specific behavioural factors, and considering sociodemographic and occupational characteristics such as sex, age, educational attainment, and type of work. In addition, it combines net-effect and necessity-based approaches, showing that the factors associated with a higher probability of GenAI use do not necessarily coincide with the conditions that act as bottlenecks for reaching higher levels of use. Thus, the proposed framework and analytical approach make it possible to assess whether TRI- and TPB-related factors play distinct roles in entrepreneurial GenAI adoption. Instead of treating all variables merely as additive predictors of use, the study differentiates between factors associated with variations in use intensity and conditions that may limit the achievement of specific levels of use. Accordingly, TRI and TPB are integrated within a sufficiency–necessity framework to examine whether general technological predispositions and GenAI-specific beliefs contribute differently across levels of adoption.

The next section describes the conceptual groundwork and the research hypotheses. The survey and the data analysis methods are then presented, followed by the results. Finally, the paper concludes with a discussion of the findings, the main implications of the study, its limitations, and possible avenues for future research.

Research on artificial intelligence acceptance has developed mainly from established technology-adoption frameworks, particularly the Technology Acceptance Model (TAM) (Davis, 1989), the Unified Theory of Acceptance and Use of Technology (UTAUT) (Venkatesh et al., 2003), and their extensions. Extended versions of TAM are the most frequently applied frameworks, and variables such as perceived usefulness, performance expectancy, attitude, trust, and effort expectancy are generally positively associated with the intention to use AI-based technologies (Kelly et al., 2023). However, this literature relies predominantly on cross-sectional surveys and focuses much more frequently on behavioural intention than on actual use. Moreover, much of the available evidence concerns consumers, students, professionals, or employees rather than independent entrepreneurs (Kelly et al., 2023).

The distinction between employees and entrepreneurs is substantive because they use AI under different conditions. Among employees, access to tools, training, technological infrastructure, technical support, and even the compulsory nature of use are partly determined by the organisation. For example, technological readiness was not significantly related to AI adoption among Nigerian librarians (Tunmibi and Okuonghae, 2023), while other studies have examined AI implementation among technology-firm employees and the role of leadership and trust in its use (Khan et al., 2025). These findings cannot be transferred automatically to self-employed workers, who make adoption decisions more independently and directly bear the associated costs, risks, and consequences.

Corporate entrepreneurs or intrapreneurs occupy an intermediate position. Although they may enjoy substantial autonomy, they remain embedded in organisations that provide resources, data, infrastructure, and protection against risk. In this context, AI may improve transparency and support decision-making under high uncertainty, while also introducing additional uncertainty into apparently more predictable tasks (Joussen et al., 2025).

Studies specifically focused on actual entrepreneurs remain relatively limited and are concentrated mainly in emerging economies. One of the closest precedents examines 278 young owners of high-growth firms in Indonesia through the Technology Readiness and Acceptance Model (TRAM) and generational cohort theory (Games et al., 2026). The study integrates technological readiness and AI acceptance through PLS-SEM, but restricts its sample to Millennial and Generation Z entrepreneurs and focuses primarily on their readiness and willingness to adopt AI rather than on explaining different levels of actual use. Other studies have treated AI adoption as a mechanism explaining entrepreneurial performance or particular psychological and organisational outcomes. AI adoption has been analysed as a mediator between compassionate entrepreneurship and cognitive workaholism among Indian entrepreneurs (Rayat et al., 2025). AI capabilities have also been linked to venture success through improvements in decision-making, innovation, risk reduction, and competitive advantage (Al-Mamary, 2025). Similarly, digital leadership has been found to mediate the relationship between AI applications and strategic decision-making in Omani SMEs (Jaboob et al., 2025). Although these studies confirm the entrepreneurial relevance of AI, they generally use adoption as an explanatory or mediating variable and do not examine the factors determining the intensity with which entrepreneurs use generative AI.

A smaller body of research has focused on the actual use of AI within the entrepreneurial process. Compatibility between generative AI capabilities and business tasks has been identified as an important predictor of both utilisation and perceived performance among early-stage Peruvian entrepreneurs (Marchena Sekli and Portuguez-Castro, 2025). However, that study was conducted in an educational setting, relied on a sample of only 78 participants, and applied a multiple-case design combining qualitative and quantitative techniques. Qualitative evidence also indicates that entrepreneurial decision-making logics shape how AI affordances are perceived and enacted. Entrepreneurs with a causal orientation tend to use AI as an outcome-oriented automation tool, whereas effectually oriented entrepreneurs are more likely to employ it as a collaborative counterpart in co-creative processes (Lupp, 2026).

From a theoretical perspective, the Technology Readiness Index captures relatively stable individual predispositions towards technological innovations. Optimism and innovativeness usually operate as adoption drivers, whereas insecurity, discomfort, or scepticism may inhibit adoption (Parasuraman, 2000; Parasuraman and Colby, 2015). TRAM integrates these predispositions with the central evaluations of TAM, particularly perceived usefulness and perceived ease of use, to explain how technological readiness shapes the acceptance of a specific innovation (Lin et al., 2007). This integration has subsequently been applied to AI-based technologies, showing that optimism and innovativeness influence perceived usefulness and ease of use, although the relevance of each factor varies across technologies and application contexts (Games et al., 2026; Torres-Coronas et al., 2026).

The broader AI-acceptance literature confirms this contextual dependence. The acceptance of AI-enabled products in mobility, household, and health applications has been explained through an extension of UTAUT2 that incorporates factors such as personal innovativeness, convenience, sustainability, and perceived security (Gansser and Reich, 2021). General technological predispositions are nevertheless insufficient by themselves to explain the use of a technology as specific as generative AI. AI acceptance also depends on evaluations closer to the behaviour itself, including the perceived advantages of the technology, perceived risk, social norms, and technological knowledge (Torres-Coronas et al., 2026). The absence of a significant relationship between technological readiness and AI adoption in some contexts further supports the need to complement TRI with beliefs directly related to the technology under examination (Tunmibi and Okuonghae, 2023).

Accordingly, as shown in Figure 1, the present study combines the TRI dimensions of optimism, innovativeness, scepticism, and insecurity with factors specifically related to the use of generative AI in business activities, which can be understood within the Theory of Planned Behaviour (TPB) (Ajzen, 1991). Attitudes towards generative AI use are shaped by the productive and economic benefits attributed to the technology and by its perceived risks. Knowledge captures the familiarity and competence required to use it, whereas normative beliefs reflect the perception that its use is accepted or expected in the professional and social environment. These variables bring the model closer to the specific decision to use generative AI rather than limiting it to a general predisposition towards technological innovation (Kelly et al., 2023; Torres-Coronas et al., 2026).

The analytical framework also incorporates entrepreneurs' personal and professional circumstances. Sex, age, and educational attainment may be associated with differences in technological familiarity, trust, perceived risk, and capacity to learn how to use emerging tools, although their effects are not necessarily uniform across contexts (Crnogaj et al., 2026; Kelly et al., 2023). Likewise, the extent to which the entrepreneur's activity is affected or may be transformed by AI can condition both its perceived usefulness and the actual need to use it, since opportunities for automation, content generation, information analysis, and decision support differ across activities and sectors (Marchena Sekli and Portuguez-Castro, 2025).

Consequently, the model presented in Figure 1 takes the integration of TRI and TPB as its starting point and extends it by incorporating the entrepreneur's sociodemographic and professional situation. These characteristics may condition technological familiarity, perceptions of usefulness and risk, and the practical opportunities to apply AI. Their relevance has been observed across different contexts of AI acceptance and use, both in the general population and among entrepreneurs (Crnogaj et al., 2026; Games et al., 2026).

Previous research has integrated technological readiness and acceptance with generational cohort theory among young Indonesian entrepreneurs, but it was limited to Millennials and Generation Z and focused on AI acceptance (Games et al., 2026). Other studies have incorporated age, sex, education, and sector to examine entrepreneurs' perceptions, but not the intensity of actual AI use (Crnogaj et al., 2026). Generative AI use has also been examined among the Spanish working population without isolating independent entrepreneurs as a specific analytical group (Torres-Coronas et al., 2026). The present study integrates these explanatory levels in a sample composed exclusively of independent entrepreneurs and uses ordered categories of actual generative AI use as its outcome.

The methodological approach also differentiates the study from previous research. Quantitative studies on AI and entrepreneurship rely predominantly on PLS-SEM to examine technological readiness and acceptance, AI adoption as a mediator of entrepreneurial outcomes, AI capabilities and venture success, or the relationships among AI, digital leadership, and strategic decision-making (Al-Mamary, 2025; Games et al., 2026; Jaboob et al., 2025; Rayat et al., 2025). Although PLS-SEM is well suited to testing direct effects, mediation, and moderation, these studies generally explain intention, acceptance, or performance rather than ordered levels of actual use.

The joint application of sufficiency- and necessity-based approaches remains exceptional. The combination of PLS-SEM, importance–performance analysis, Necessary Condition Analysis, and artificial neural networks has shown that variables that are not the strongest net predictors may nevertheless operate as necessary conditions (Ştefan et al., 2025). However, that study focuses on potential entrepreneurs and their intention to undertake AI-related entrepreneurial activities, rather than on actual generative AI use among entrepreneurs already managing their own businesses.

The TRI (Parasuraman, 2000) and its extensions, such as the Technology Readiness and Acceptance Model (TRAM) (Lin et al., 2007), argue that certain individual predispositions condition the way people evaluate and use innovations. These predispositions include technological optimism, scepticism, personal innovativeness, and technological insecurity (Parasuraman, 2000). These traits influence the perception of central attributes of an innovation, such as its relative advantage, complexity, or compatibility with user needs (Rogers, 2003), and may ultimately affect use, either directly or through mediating variables such as perceived usefulness and perceived ease of use (Lin et al., 2007).

Technological optimism (OPTI) is the belief that technologies improve people's control, efficacy, flexibility, and quality of life. This orientation favours a more open, active, and positive interaction with new digital tools (Parasuraman and Colby, 2015). Individuals with higher technological optimism are inclined to have a more positive attitude towards innovations, which facilitates their acceptance and use (Rogers, 2003). By contrast, technological scepticism (SCEP) expresses an attitude of caution, doubt, or distrust towards the adoption of new technologies. This disposition may lead individuals to question the real usefulness of the innovation, its consequences, or its reliability, thereby reducing the probability of use (Lin et al., 2007).

Personal innovativeness (INNOV) refers to the individual's willingness to experiment with emerging technologies and to adopt novel solutions before other users (Agarwal and Prasad, 1998). Individuals with higher INNOV usually view technologies as more beneficial, less demanding to operate, and more closely aligned with their requirements, thereby strengthening their willingness to adopt and use them (Lu et al., 2005). Higher levels of INNOV stimulates more favourable beliefs about its functioning, usefulness, and potential (Parasuraman, 2000).

By contrast, technological insecurity (INSEC) reflects distrust towards technology and concerns about its dependability, safety, operational performance, or potential to reduce the human dimension of decision-making, particularly when its application may have significant implications (Parasuraman, 2000). High insecurity can make individuals reluctant to engage with a new technology, even when its expected advantages are substantial. This reluctance stems from the ambiguity, perceived vulnerability, and risk that often accompany innovation. As a result, highly insecure people are more likely to delay adoption and to fall within the most resistant segments of users (Rogers, 2003).

TRI dimensions have been incorporated into established technology-adoption models through the Technology Readiness and Acceptance Model (TRAM) (Lin et al., 2007), which extends TAM (Davis, 1989). From this perspective, technological predispositions influence intention and use through beliefs such as perceived usefulness or perceived ease of use (Blut and Wang, 2020). The literature shows that TRI dimensions, particularly OPTI, contribute to the formation of favourable attitudes towards different technologies (Blut and Wang, 2020). This link has also been reported in the use of artificial intelligence-based technologies in the medical sector (Chalutz Ben-Gal, 2023) or library services (Tunmibi and Okuonghae, 2023). Thus, given that OPTI and INNOV are theoretically expected to facilitate GenAI use, whereas SCEP and INSEC are expected to act as barriers, we formulate the following hypotheses:

GenAI is applicable across a wide spectrum of uses, from routine daily activities to cognitively demanding tasks, making it a technology with substantial capacity to reshape economic processes and social practices (Mahmoud et al., 2025). In the literature on AI adoption, perceived usefulness, efficiency, and performance improvement expectations have proven to be particularly relevant factors (Camilleri, 2024; Ibrahim et al., 2025). This evidence has been observed in diverse contexts, such as the use of AI tools by employees in organisations (Yudhistyra and Srinuan, 2024), the acceptance of AI-based products (Gansser and Reich, 2021), or the adoption of artificial intelligence systems in education (Al Darayseh, 2023; Sobaih et al., 2024; Strzelecki, 2024).

In this study, the perceived usefulness of GenAI is approximated through the perception of its positive effects on the economy and productivity (ECU). GenAI tools can enhance efficiency and performance in knowledge-intensive activities (Benbya et al., 2024). Likewise, they can act as enabling technologies for industrial transformation through applications such as anticipating equipment failures, improving supply-chain efficiency, automating complex processes, monitoring quality standards, and supporting data-driven decision-making (Phatthanachaisuksiri et al., 2025). This potential also extends to sectors such as agriculture, where the adoption of smart technologies has been associated with improvements in production per hectare and with a better decision-making capacity among its agents (Alobid and Szűcs, 2025). Thus, ECU should be a driver in the adoption of GenAI:

The literature has also identified numerous risks of undesirable effects associated with the use of GenAI (RISK). Those related to privacy, data protection, and transparency are particularly relevant. In the absence of clear legal frameworks, GenAI may affect fundamental rights and undermine trust in institutions and companies (Fraisl, 2024).

There are also social risks linked to the increase in structural inequalities. These may be amplified by the presence of algorithmic biases, digital divides associated with cost or connectivity, and the growing concentration of technological capabilities, infrastructure, and decision-making authority within a small number of powerful firms (Fraisl, 2024; Madsen and Toston, 2025). In addition, GenAI may be used for malicious purposes, such as sophisticated forms of fraud, the production of synthetic media such as deepfakes, and the large-scale dissemination of manipulative content (INTERPOL Innovation Centre, 2023). Taken together, these risks may act as barriers to the use of these tools.

Prior knowledge of a technology is another relevant factor in adoption processes. Familiarity with a tool may strengthen user self-efficacy, that is, the perceived ability to use it appropriately (Compeau and Higgins, 1995). This self-efficacy influences perceived ease of use, which in turn favours both intention to use and perceived usefulness (Venkatesh and Bala, 2008). In the case of artificial intelligence, having general or specific knowledge (KNOW) is associated with a more favourable perception of its usefulness and with a lower perception of risk (Araujo et al., 2020). Along these lines, some studies show greater acceptance of AI among young people with university training in healthcare fields (Alshutayli et al., 2024). Similarly, research in health education indicates that a lack of training in AI hinders its adoption (Lokaj et al., 2023). Therefore, it is hypothesised:

The ethical, social, and legal risks of GenAI have increased public perception that these technologies need to be regulated (NREG). Concerns include issues such as privacy, disinformation, security, algorithmic discrimination, and lack of transparency (Alkamli and Alabduljabbar, 2024; Sebastian, 2024). These concerns may generate social distrust and slow down automation processes, especially when the systems used are opaque or difficult to audit (KPMG International, 2023).

As a consequence, social, political, and academic pressure has intensified to develop regulatory frameworks that ensure transparency, equity, data protection, and respect for human rights (Mahmoud et al., 2025). The literature also underlines the need to establish mechanisms that make it possible to audit models, identify biases, and reduce the structural risks derived from their use (Zhou et al., 2024). From this perspective, a greater perception of the need for regulation may reflect distrust, caution, or concern regarding GenAI, which would predictably be associated with lower intensity of use. Therefore, NREG is expected to act as a barrier to GenAI use, leading to the following hypothesis:

With regard to sex (SEX), men may be more likely than women to adopt and use GenAI in entrepreneurial activities because they tend to report greater confidence, familiarity, and willingness to experiment with emerging technologies. By contrast, women often display higher levels of AI-related anxiety, stronger risk perceptions, and greater concern about the ethical and employment implications of these tools (Otis et al., 2025; Russo et al., 2025). Evidence on actual adoption also points to a substantial gender gap. Aldasoro et al. (2024) found that approximately 50% of men, compared with 37% of women, had used GenAI, with differences in self-assessed knowledge explaining an important part of this disparity. Accordingly, gender differences in GenAI use may reflect unequal trajectories of technological familiarity, confidence, and risk perception rather than sex alone.

Generational differences may also shape GenAI adoption because individuals socialised in different technological environments develop distinct levels of digital familiarity and different approaches to technological change (Chan and Lee, 2023). Older cohorts (OLD), including Generation X and baby boomers, tend to evaluate new technologies more cautiously and to weigh their potential risks and benefits more deliberately (Kozak and Fel, 2024). By contrast, younger cohorts, particularly Generations Y and Z, have generally experienced earlier and more continuous exposure to digital technologies and may therefore perceive GenAI as more accessible, useful, and trustworthy (Games et al., 2026). Consequently, younger entrepreneurs may be more willing to experiment with and integrate GenAI into their professional activities, whereas older entrepreneurs may adopt it more selectively (Crnogaj et al., 2026).

Entrepreneurs' activity (ACT) may also be associated with GenAI use. Those engaged in knowledge-intensive or administrative occupations, such as scientists, lawyers, consultants, and other liberal professionals, may be more likely to use GenAI than those performing predominantly manual or trade-based work, such as artisans or installers. GenAI can improve productivity and work quality in knowledge-intensive activities, support pattern identification and the assessment of uncertain scenarios, and automate routine administrative duties, thereby freeing time for higher-value tasks (Benbya et al., 2024).

Educational attainment (ED) may likewise be associated with GenAI use. Higher levels of education may facilitate the acceptance and use of GenAI by strengthening digital literacy, information-processing skills, and users' confidence in evaluating AI-generated outputs. More highly educated individuals may also be better able to identify productive applications of these tools, thereby reducing perceived complexity and increasing perceived behavioural control (Gambacorta et al., 2025).

Thus, we postulate:

The empirical analysis draws on data from the 2025 survey (Centro de Investigaciones Sociológicas, 2025), targeting individuals aged 18 and above living in Spain. The survey was planned for 4,000 interviews and ultimately yielded 4,004 valid responses. Participants were selected through a mixed sampling strategy that combined random telephone dialling with quotas for sex and age, while stratification was based on autonomous community and municipality size. Fieldwork was carried out in the first half of February 2025.

The final sample consisted of individuals who qualified as entrepreneurs and reported being aware of what GenAI is, totalling 444 respondents. We define an entrepreneur as a self-employed worker, which is the commonly employed definition in the field, mainly because of the broader accessibility of statistical information and its greater consistency across sources and contexts (Bjuggren et al., 2012). Under this approach, an entrepreneur is considered to be a person who works for themselves, without an employment relationship of dependence, including both self-employed workers and owner-managers of firms, as well as members of cooperatives. Their sociodemographic profile is shown in Table 1.

The constructs were measured using items from the CIS survey (Centro de Investigaciones Sociológicas, 2025). Table 2 presents the items, response scales, variables and procedures used to construct the composite scores. The preprocessing and transformation of the items into composite indicators followed commonly accepted procedures in the literature on the construction of synthetic indices (Greco et al., 2019; Nardo et al., 2005).

The scores for OPTI, SCEP, INNOV, INSEC, ECU, RISK, and NREG were operationalised using the standardised score of the first component extracted through principal component analysis (PCA). This approach condenses the information contained in each group of items into a single composite indicator, thereby reducing the number of variables while retaining the variance shared by the items measuring each construct (Nardo et al., 2005). Table 3 shows that this procedure is robust. All factor loadings exceed 0.5, and Bartlett's test rejects the null hypothesis that the correlation matrix is an identity matrix, supporting the suitability of the data for PCA (Costello and Osborne, 2005). Likewise, the average variance extracted (AVE) exceeds 0.5 in all constructs, and Cronbach's alpha (α) exceeds the conventional threshold of 0.7 in all cases except INNOV. Nevertheless, the value obtained for INNOV (α = 0.675) is close to the conventional threshold and may be considered acceptable in the context of an exploratory study (Taber, 2018).

KNOW was measured using a single item; therefore, its score was calculated directly by standardising responses on the original 1–10 scale.

A different methodological approach was adopted for USE, since, as Greco et al. (2019) note, the procedure for aggregating items into synthetic indicators has to take into account the theoretical meaning of the construct. In this case, the variable was defined as the highest reported frequency of use across the tools considered. The arithmetic mean or PCA may give a greater score to respondents who use several GenAI tools occasionally than to those who make intensive and recurrent use of a single tool. However, from a conceptual point of view, the latter profile reflects a more intense use of GenAI (Torres-Coronas et al., 2026).

Finally, as shown, the control variables related to sociodemographic and occupational characteristics were operationalised as dichotomous variables.

All procedures related to data screening, transformation, and variable construction were conducted in R 4.4.2. Data cleaning, item recoding, and standardisation of scores were performed mainly with functions from the tidyverse ecosystem, particularly dplyr and tidyr. Principal component analysis was implemented using the prcomp () function from the base stats package. The reliability of the multi-item constructs and the adequacy of the data for PCA were assessed with the psych package. Descriptive summaries and tables were subsequently prepared using dplyr, knitr, and related formatting tools.

To address RO1, an ordered logistic regression (OLR) was estimated, including the TRI- and TPB-based variables together with the sociodemographic and occupational control variables. Multicollinearity was assessed using variance inflation factors (VIFs). The proportional-odds assumption was evaluated using the Brant test, both globally and for each predictor. A partial proportional-odds model was also estimated as a robustness check, allowing non-proportional effects where required while retaining proportional effects for the remaining predictors. The predictive performance of the OLR and partial proportional-odds models was compared using the normalised Ranked Probability Score (RPS), with lower values indicating better ordinal probabilistic prediction. The statistical significance of the difference in RPS was assessed using a paired bootstrap based on observation-level score differences.

An a priori Monte Carlo power analysis based on 2,000 simulations was conducted for an OLR with 444 observations, 12 explanatory variables, and six outcome categories. Assuming a minimum detectable coefficient (B) with an absolute value of B = 0.25, equivalent to an odds ratio (OR) of 1.28 for a positive effect or 0.78 for a negative effect per standard-deviation increase in the predictor, and α = 0.05, the estimated statistical power was 0.808.

To address RO2, Necessary Condition Analysis (NCA) was applied in accordance with Dul et al. (2023) and was restricted to the TRI- and TPB-based factors. NCA was used to estimate the necessary-condition effect size of each explanatory variable (d) and to construct bottleneck tables expressed as normalised scores ranging from 0 to 100. Ceiling lines were estimated using the ceiling envelopment–free disposal hull (CE-FDH) method. Theoretical drivers were analysed using the upper-left ceiling, whereas theoretical barriers were analysed using the upper-right ceiling. Statistical significance was assessed through permutation tests with 10,000 replications.

Potentially influential observations located in the ceiling zone were identified using the NCA outlier procedure, which evaluates the change in the CE-FDH effect size after the removal of individual observations. These cases were not regarded as data errors and were therefore retained in the main analyses. As a sensitivity check, each condition was re-estimated after removing only the observations identified as potentially influential for that specific condition. The permutation tests were repeated with 10,000 replications in each re-estimation, and the bottleneck thresholds were recalculated using the corresponding variable-specific reduced samples. This procedure made it possible to assess whether the statistical significance and magnitude of the necessary-condition effects were sensitive to a small number of influential observations.

The statistical power of the NCA was assessed following Dul (2024b). Assuming uniformly distributed normalised values of USE and the explanatory factors, a sample size of n = 444, an effect size of d = 0.025, and a significance level of α = 0.05 yielded an estimated power of 0.99, indicating that the analysis was sufficiently powered to detect a necessary-condition effect of this magnitude.

All analyses were conducted in R 4.4.2. The OLR model was estimated using functions from the MASS package, the proportional-odds assumption was assessed with the brant package, and the partial proportional-odds model was estimated using the VGAM package. The NCA, permutation tests, outlier identification, sensitivity analyses, and bottleneck tables were implemented using the NCA package. Finally, knitr and complementary formatting tools were used to prepare the tables and descriptive outputs.

Table 4 shows that none of the TRI-based predictors reached statistical significance, whereas three variables grounded in TPB were statistically significant: KNOW, ECU, and NREG. KNOW had the strongest effect, with a coefficient of B = 0.949, z = 8.457, and p < 0.001. Its odds ratio was OR = 2.584. ECU also showed a positive and significant association with GenAI use (B = 0.489, z = 3.749, p < 0.001; OR = 1.630). By contrast, NREG was negatively associated with GenAI use (B = −0.232, z = −2.002, p = 0.045; OR = 0.793). RISK showed a negative coefficient in the hypothesised direction, although the association did not reach the conventional 5% significance level (B = −0.206, p = 0.084); therefore, H6 was not supported.

Regarding the sociodemographic and occupational variables, belonging to an older generation was associated with a lower likelihood of GenAI use (OLD: B = −0.428, z = −2.201, p = 0.028; OR = 0.652), whereas ACT was positively associated with use (B = 0.778, z = 2.933, p = 0.003; OR = 2.177).

Table 4 also indicates that the OLR model provides a reasonable fit. McFadden's pseudo-R² is 0.152, and the model is globally significant, as shown by the likelihood-ratio statistic, LR χ² = 207.42, p < 0.001. Moreover, the variance inflation factors for all explanatory variables are below 2, indicating no relevant multicollinearity concerns. Table 5 presents additional robustness checks. The omnibus Brant test rejected the proportional-odds assumption; however, the violation was attributable exclusively to KNOW, whereas all remaining predictors satisfied the assumption. A partial proportional-odds model was therefore estimated as a robustness check, allowing the effect of KNOW to vary across thresholds. Nevertheless, the Ranked Probability Score (RPS) analysis revealed no statistically significant difference in predictive performance between the ordered logit and partial proportional-odds models. Accordingly, for reasons of parsimony and interpretability, the proportional-odds specification was retained as the main model, including a single coefficient for KNOW.

Table 6 presents the NCA results separately, grouping the variables according to whether they are theoretical drivers or theoretical barriers. This distinction is necessary because, in NCA using free disposal hull, drivers and barriers are interpreted in the upper-left and upper-right quadrants, respectively, as shown in the NCA plots in Figures 2 and 3.

In the block of theoretical drivers, three variables exhibit significant necessary-condition status. The largest effect size corresponds to OPTI, with d = 0.149 and p = 0.028, indicating a medium necessary-condition effect. ECU also shows a significant, although smaller, effect, with d = 0.075 and p < 0.001, while KNOW presents a significant but small effect, with d = 0.044 and p < 0.001. By contrast, INNOV does not emerge as a necessary condition. Within the block of theoretical barriers, only SCEP displays significant necessary-condition status, although its effect size is small, with d = 0.045 and p = 0.014. INSEC, RISK, and NREG show no necessary-condition effects.

The outlier sensitivity analysis confirms that these conclusions regarding statistical significance remain unchanged, although the estimated magnitude of some effects is sensitive to observations located in the ceiling zone. OPTI is the most sensitive condition. After removing its three identified outliers, its effect size increases from d = 0.149 to d = 0.314, with p < 0.001. This increase is not only substantial in numerical terms but also implies a qualitative change from a medium to a large necessary-condition effect. The remaining significant conditions also show higher effect sizes after removing their condition-specific outliers, but their qualitative classifications do not change, and all remain small. KNOW increases from d = 0.044 to d = 0.067, ECU from d = 0.075 to d = 0.083, and SCEP from d = 0.045 to d = 0.088, with all three remaining within the small-effect category and statistically significant at p < 0.001. Although the relative increases are notable for KNOW and especially SCEP, their absolute effect sizes remain limited. Moreover, INNOV, INSEC, RISK, and NREG continue to show null effects. Overall, the identification of the significant necessary conditions is robust, but the magnitude of the OPTI effect should be interpreted with particular caution because of its sensitivity to a small number of influential observations.

Table 7 shows the bottleneck tables constructed from the scatter plots in Figures 2 and 3. The results indicate that, up to use level “2” (several times a year), no necessary conditions appear. To reach sporadic monthly use, the only necessary conditions come from the TRI constructs: a minimum level of technological optimism of 19.9 out of 100 is required, and a level of technological scepticism of 93.2 out of 100 must not be exceeded.

At higher levels of use, conditions linked to knowledge and perceived economic usefulness also emerge. Specifically, to reach a frequency of use of several times per week, certain minimum thresholds in KNOW and ECU must be exceeded. At the highest level of use (daily use of GenAI), three minimum conditions are required: a technological optimism level of 34.9 out of 100, an AI knowledge level of 11.1 out of 100, and a perceived economic usefulness level of 26.7 out of 100. In addition, in the case of barriers, a maximum threshold is observed for SCEP: to reach daily use, SCEP must not exceed 91.2 out of 100.

NCA provides a clearly different interpretation from the regression results. OPTI and SCEP emerge as earlier bottlenecks than KNOW and ECU, as they act as bottlenecks at lower levels of GenAI use. By contrast, KNOW and ECU become necessary only at higher levels of use. This is noteworthy because OPTI and SCEP were not significant in OLR, whereas both appear as necessary conditions in the NCA. Conversely, NREG, despite its significance in the regression, does not operate as a necessary condition. INNOV, INSEC, and RISK likewise show no bottleneck levels at any level of GenAI use.

The comparison in Table 7 with the bottleneck values obtained after removing the condition-specific outliers, shown in parentheses, confirms that OPTI is the most sensitive condition. Its bottleneck profile changes not only in magnitude but also in the use levels at which it becomes necessary. After outlier removal, OPTI already appears as a requirement for having used GenAI once or several times a year, and its thresholds increase substantially for monthly and weekly use. By contrast, the ECU bottlenecks change only modestly. The KNOW threshold remains unchanged for weekly use, although it increases for daily use, while the SCEP thresholds become somewhat more restrictive and emerge at an earlier level of use. Overall, the broad pattern remains stable. OPTI and SCEP operate as the earliest bottlenecks, whereas KNOW and ECU become relevant mainly at higher use levels. However, the precise OPTI thresholds should be interpreted with particular caution because of their sensitivity to influential observations.

This study analyses the explanatory factors of GenAI use among Spanish entrepreneurs. The variables considered are grouped into two complementary perspectives. The first refers to personal predisposition towards new technologies, measured through the Technology Readiness Index (TRI) (Parasuraman, 2000). The second includes variables related to the perception of GenAI, including perceived economic usefulness (ECU), perceived risk associated with undesirable consequences (RISK), perceived knowledge (KNOW), and perceived need for GenAI regulation (NREG). Following Pérez-Portabella et al. (2025), these variables are interpreted from the Theory of Planned Behaviour (TPB) (Ajzen, 1991). To these two groups of variables, we added factors that may also be linked to GenAI use, including sex (SEX), generational cohort (OLD), the entrepreneur's field of activity (ACT), and educational attainment (ED).

The first research objective (RO1) was to examine the sign of the relationship between the explanatory variables and GenAI use (USE). The approach adopted follows a logic of sufficiency: “higher X, higher use” in the case of drivers, and “lower X, higher use” in the case of barriers (Dul, 2024a). The results indicate that the TRI variables are not significant. By contrast, three GenAI-specific variables show statistically significant associations of varying strength. In particular, KNOW shows the strongest association with USE. The importance of this variable has already been observed in different contexts, such as the use of GenAI by healthcare workers (Alshutayli et al., 2024) and in educational contexts (Al Darayseh, 2023).

The positive relationship between perceived economic usefulness (ECU) and GenAI use can be interpreted in light of the importance attributed in numerous studies to perceived usefulness as a determinant of technology adoption (Al Darayseh, 2023; Gansser and Reich, 2021; Sobaih et al., 2024; Strzelecki, 2024; Yudhistyra and Srinuan, 2024). On the other hand, the negative association between NREG and USE can be understood from the relevance of normative, social, and institutional factors in the acceptance of artificial intelligence. Previous literature has pointed out that social influence is significantly associated with the adoption of AI in general (Kelly et al., 2023) and in the specific case of GenAI (Camilleri, 2024; Yudhistyra and Srinuan, 2024). Although RISK showed the expected negative association with GenAI use, the evidence was weak and did not reach the conventional 5% significance threshold. Moreover, RISK did not emerge as a necessary condition in the NCA. Its role in GenAI use should therefore be interpreted cautiously.

Likewise, belonging to Generation X or an older cohort is negatively associated with GenAI use, in line with previous findings (Crnogaj et al., 2026). Conversely, engagement in knowledge-intensive or administrative activities (ACT) is positively linked with GenAI use, consistent with the evidence reported by Benbya et al. (2024).

To address RO2, a necessity logic was adopted, according to which variables that are not significant in regression may still be required to reach certain levels of use, while significant predictors may not be necessary conditions. NCA identifies only TRI-related bottlenecks for monthly GenAI use: a minimum level of OPTI and the absence of excessive SCEP. TPB-related variables become necessary only at higher levels of use; however, their necessity effect sizes remain below that of OPTI, which exhibits a medium necessary-condition effect in the main analysis and a large effect in the sensitivity assessment, according to the widely used benchmarks proposed by Dul (2016).

These results are consistent with Richter et al. (2020), who show, in a technology acceptance context, that PLS-SEM and NCA may produce non-equivalent rankings of explanatory variables. This difference reflects the fact that regression methods and NCA provide complementary information. The results also suggest that considering d > 0.10 as a strict threshold for identifying conditions with practical implications may be too restrictive. In our analysis, variables such as KNOW, SCEP, and ECU present values of d < 0.10, but act as bottlenecks for reaching regular levels of use. This result is especially relevant because regular use is what would be expected to characterise an entrepreneur who integrates GenAI into their everyday professional activity. Although this does not necessarily imply advanced strategic use, it may reflect an operational incorporation of the technology into recurring tasks such as clerical and monitoring activities, freeing up managers and employees to concentrate on more strategic and value-generating tasks (Ponnock, 2025).

The comparison between the OLR and NCA results confirms the value of combining sufficiency and necessity perspectives. Regression identifies the variables most strongly associated with the probability of reaching higher levels of GenAI use. From this perspective, KNOW and ECU show the greatest explanatory capacity. NCA, however, provides a different reading by identifying the minimum conditions that must be present for specific levels of use to be possible. Thus, the strongest predictors are not necessarily the most demanding bottlenecks.

From a theoretical standpoint, these contrasting results indicate that TRI- and TPB-related factors may fulfil distinct roles in entrepreneurial GenAI adoption. The TRI dimensions highlighted by NCA, especially technological optimism and the absence of excessive scepticism, seem to establish a basic enabling threshold for regular use. Although they are not the strongest net predictors, limited technological readiness may restrict the attainment of recurrent use. In contrast, GenAI-specific TPB-related factors, particularly knowledge and perceived economic usefulness, are more closely associated with variation in use intensity and emerge as necessary conditions only at higher levels of use. This pattern points to a layered structure in which general technology readiness provides an enabling foundation, while behaviour-specific knowledge and evaluations become increasingly important as use intensifies. Accordingly, integrating TRI and TPB extends beyond simply enlarging the set of explanatory factors, as it helps clarify their potentially distinct roles across different levels of actual GenAI use.

The NCA results indicate that a basic predisposition towards technology constitutes the main bottleneck for monthly or more frequent GenAI use. Specifically, OPTI and the absence of excessive SCEP act as necessary conditions for monthly or more frequent GenAI use. These bottlenecks are not particularly demanding: the main analysis indicates that a technological optimism score of at least 19.9 out of 100 is required, increasing to 32.5 in the sensitivity analysis after removing condition-specific outliers, together with the absence of very high SCEP. This indicates that entrepreneurs do not need an exceptionally favourable technological attitude to reach monthly or more frequent GenAI use; rather, the observed pattern is consistent with strong resistance or distrust acting as constraints.

However, the pattern observed across occasional, frequent, and daily use levels follows a different logic. At higher observed use levels, general technological predispositions are no longer the only relevant necessary conditions, as they are complemented by factors directly related to GenAI, particularly knowledge and perceived economic usefulness. Entrepreneurs need to perceive that GenAI can provide functional or economic value to their business and, above all, need sufficient knowledge to use it effectively. This is consistent with the regression results, where KNOW is the strongest predictor, followed by ECU.

From a business perspective, KNOW is especially important because intensive use is associated not only with a positive attitude towards technology, but also with an understanding of what GenAI can do, how it can be integrated into business routines, and where it can create value. Greater knowledge is linked to more systematic GenAI use across applications such as content generation, customer service, information analysis, process automation, and decision support. Likewise, higher perceived economic usefulness is associated with a greater probability of regular GenAI use. Finally, the negative association between NREG and GenAI use suggests that regulatory concern may reflect uncertainty and be associated with a lower likelihood of reaching higher levels of use. Therefore, promoting responsible GenAI adoption requires not only highlighting its benefits, but also providing clear guidelines and safe-use frameworks adapted to small firms and entrepreneurs.

These findings also support differentiated practical recommendations:

1. Public authorities should reduce regulatory uncertainty by providing accessible and regularly updated guidance on data protection, intellectual property, transparency, accountability, and the responsible use of GenAI in small businesses. They should also promote clear compliance frameworks tailored to entrepreneurs and small firms, which often lack specialised legal or technical resources. Financial incentives, pilot programmes, and support schemes could provide additional opportunities for responsible experimentation, particularly among firms with limited digital capabilities.

2. As far as entrepreneurship support agencies are concerned, these organisations should complement general awareness initiatives with practical diagnostics, sector-specific use cases, mentoring, and implementation support. Their role should be to help entrepreneurs identify where GenAI can generate economic value, select suitable tools, assess implementation costs, and redesign business processes accordingly. Support programmes should also include guidance on risk management, output verification, and the responsible handling of sensitive information.

3. Universities, business schools, vocational training centres, and lifelong-learning providers should place greater emphasis on applied GenAI literacy. Training should cover prompt design, output evaluation, ethical and legal risks, data protection, and the integration of GenAI into routine business activities. Courses should move beyond general awareness and include hands-on exercises based on realistic entrepreneurial tasks, such as market analysis, customer communication, content generation, process automation, and decision support.

4. Entrepreneurs should begin with clearly defined and relatively low-risk applications and assess their contribution to productivity, costs, time savings, service quality, and decision-making before extending their use to more complex activities. They should also establish basic governance practices, including the verification of generated content, the protection of confidential information, and clear criteria for deciding when human supervision remains necessary.

The most appropriate intervention may also depend on the entrepreneur's profile and current level of use:

1. Non-users and highly sceptical entrepreneurs may benefit primarily from practical demonstrations, trusted examples, peer experiences, and opportunities to experiment in controlled environments. For this group, the priority is to reduce excessive scepticism and uncertainty rather than to promote advanced technical skills.

2. Occasional users may require support in identifying relevant business applications, assessing their economic usefulness, and moving from exploratory use to more systematic adoption. Short practical programmes, mentoring, and sector-specific examples may be especially valuable for this group.

3. Frequent users are more likely to benefit from advanced training, workflow redesign, governance procedures, and mechanisms for monitoring the quality, reliability, and risks of GenAI-generated outputs. At this stage, the emphasis should shift from initial adoption to effective integration and responsible scaling.

4. Entrepreneurs in knowledge-intensive or administrative activities may be able to integrate GenAI more rapidly because their tasks are more directly compatible with information processing, content generation, analysis, and documentation. For these users, support should focus on advanced applications and process optimisation.

5. Entrepreneurs in less digitally mature firms or sectors may first require more basic technological and organisational support. In these contexts, adequate digital infrastructure, data management, employee skills, and confidence in technology may be required before more sophisticated GenAI applications can be adopted.

The use of GenAI among Spanish entrepreneurs is most strongly linked to knowledge and perceived economic usefulness. The OLR shows that KNOW is the strongest predictor of higher levels of use, followed by ECU. This suggests that more intensive GenAI use is associated less with a generally favourable attitude towards technology than with greater knowledge and stronger perceptions of its practical or economic value for the business.

The joint interpretation of the OLR and NCA results confirms the usefulness of combining the logics of sufficiency and necessity. Whereas KNOW and ECU are the main predictors in the regression, the NCA shows that the first bottlenecks are linked to technological predisposition. In particular, a minimum level of OPTI and the absence of excessive technological scepticism are identified as necessary conditions for attaining monthly or more frequent GenAI use.

The results suggest that the conditions associated with frequent or daily use differ from those observed at lower levels of use. The bottleneck analysis indicates that regular use is observed only when a basic technological predisposition is accompanied by sufficient knowledge and a clear perception of economic usefulness. Therefore, business support programmes should not be limited to raising awareness about GenAI, but should offer practical training, sector-specific examples, and guidance on how it can be applied to support productivity, decision-making, customer service, marketing, and administrative processes.

This study has several limitations that open up future lines of research. First, the empirical analysis is based on cross-sectional data. Therefore, although the results identify significant associations and necessary conditions, they do not allow strong causal conclusions to be drawn. Future research should therefore employ longitudinal designs to analyse changes in optimism, scepticism, knowledge, perceived economic usefulness, and regulatory concerns over time, and to determine whether the conditions required for initial adoption differ from those associated with continued or increasingly intensive use.

The dependent variable measures the self-reported frequency of use. This indicator is useful for analysing behavioural intensity, but it does not distinguish between types of use, quality of use, or strategic relevance. Using GenAI daily for routine personal tasks is not equivalent to integrating it into business planning, customer management, product design, accounting, or market analysis. Future research should therefore differentiate between personal, operational, managerial, and strategic uses.

The sample focuses on Spanish entrepreneurs, defined as self-employed workers. This definition is common and empirically operational, but it includes heterogeneous profiles, ranging from individual self-employed workers to business owners with employees. Future research could compare freelancers, microenterprises, SMEs, start-up founders, and owner-managers. Moreover, because entrepreneurial attitudes may vary substantially across countries and cultures (Pylaeva et al., 2026), the generalisability of the findings to other institutional and cultural contexts remains uncertain. Nevertheless, the distinction between probabilistic drivers and necessary bottlenecks is not confined to the Spanish context and may offer a useful analytical framework for examining entrepreneurial GenAI adoption elsewhere. However, the strength of the observed associations and, in particular, the NCA bottleneck thresholds are likely to depend on context and should not be directly generalised to other countries. Cross-country research could assess whether the identified bottlenecks vary according to regulatory frameworks, digital infrastructure, entrepreneurial ecosystems, and national attitudes towards technology. Further studies should also compare entrepreneurs across sectors, firm sizes, levels of digital maturity, and types of business activity to determine whether distinct adoption pathways and necessary-condition profiles emerge.

Some constructs are based on the items available in the CIS survey rather than on scales specifically designed to analyse the business adoption of GenAI. It would therefore be advisable to develop more specific measures of GenAI self-efficacy, perceived usefulness for business tasks, legal uncertainty, trust in AI providers, and perceived competitive advantage. Although the analysis includes several contextual variables, such as sex, age, business activity, and education, other potentially relevant factors could not be considered because they were not available in the survey. These include years of entrepreneurial experience, firm size, digital maturity, prior exposure to AI tools, access to technological resources, and participation in training or business-support programmes. Their omission may leave part of the heterogeneity in GenAI use unexplained and may also affect the estimated associations of the included variables. Future research should therefore incorporate a broader set of individual, firm, and context-level characteristics to assess whether the identified predictors and necessary conditions remain stable across different entrepreneurial profiles and organisational settings.

Potential endogeneity and reverse causality cannot be entirely excluded. The relationship between knowledge, perceived economic usefulness, and GenAI use may run in both directions: greater knowledge and perceived usefulness may be associated with more intensive use, while more frequent use may, in turn, coincide with higher levels of knowledge and stronger perceptions of usefulness. The available cross-sectional survey does not provide suitable instrumental variables to identify the direction of these relationships. Future longitudinal or panel-data studies could help establish their temporal ordering and assess whether the observed associations remain stable over time.

The data used in this study are publicly available from the Centro de Investigaciones Sociológicas (CIS), Study No. 3495, Inteligencia artificial. The study webpage provides access to the microdata file, questionnaire, technical documentation, tabulations, and results: Link to the website. The R code used for data preprocessing and statistical analyses is provided in the Supplementary Materials. A README file is also included, describing the supplementary scripts, their required inputs, execution order, and generated outputs.

The supplementary material for this article can be found online.