Value creation and appropriation from the use of machine learning: a study of start-ups using fuzzy-set qualitative comparative analysis

The business value of IT refers to the impact of IT use on organizational performance. Performance in this context includes productivity, profitability, cost reductions, competitive advantage, and inventory reductions (Devaraj & Kohli, 2003; Hitt & Brynjolfsson, 1996; Melville et al., 2004). In terms of the economic implications of IT, research on the value of IT unanimously shows a positive relationship between IT and some aspect of firm or organizational value, be it financial (e.g., return on investment), intermediate (e.g., process-related), or affective (e.g., perception-related; Bharadwaj, 2000; Kohli & Grover, 2008). IT can help improve organizational performance (Brynjolfsson & Hit, 1996), which is related to perceptions (Bharadwaj, 2000; Kohli & Grover, 2008). It can also improve business performance by leveraging IT capability to create and capture value (Porter, 2001) and become a source of competitive advantage (Bharadwaj, 2000; Chae et al., 2014).

IT capability in this study includes IT infrastructure, IT human resources, and IT-enabled intangibles such as knowledge assets, customer orientation, and synergy (Bharadwaj, 2000). It can create value through product differentiation and innovation and capture value through increased revenues and profits (Hitt & Brynjolfsson, 1996). Superior IT capability can reduce marketing costs by increasing switching costs and customer loyalty (Chae et al., 2014). IT as simple hardware and software does not create value in isolation. Instead, it must be part of a business value creation process in synergy with other IT and organizational factors (Melville et al., 2004; Wade & Hulland, 2004).

The ideas discussed in this section explain how the use of IT, be it a payroll program, fingerprint recognition, or an ML technology system, can create value in a firm. However, each IT has its own features that require unique value creation and value capture capabilities. For emerging or disruptive technologies such as ML, BMs often need to be transformed to take full advantage (Chesbrough, 2007). The specific capabilities of ML are explained later, but these capabilities, coupled with a lack of understanding of how technological progress can be converted into profit, require further research to enable the successful application of ML (Chesbrough, 2007). Lee et al. (2019a, b) have noted that more research is needed to understand how emerging technologies can be commercialized through different BM archetypes.

This review suggests that uses of specific IT tools in business contexts can create and capture value for firms. The practices of successful businesses corroborate the value capabilities of ML. However, other firms have unsuccessfully employed ML technology, failing to create value. This failure leads to losses or even closure. What differs in these vastly contrasting cases is missing from the theoretical literature on value creation and value capture using ML. Overall, current research on the creation and capture of value from IT tends to ignore the variety that exists in different types of IT and their capabilities. The BM theory, proposed by Amit and Zott (2001) and Teece (2010), has been effective in explaining how companies that use IT can create and appropriate value. It offers a valuable framework to identify how a technology such as ML affects the performance of a company.

The BM describes the business logic of an organization. It explains how an organization creates and delivers value to customers, a value proposition can only be designed with a thorough customer insight, (Wang et al., 2023). It also describes the architecture in terms of revenues, costs, and profits (Teece, 2010; Vetter et al., 2022). The definition of a BM in this study is the “architecture of the mechanism for creating, delivering and capturing value” (Teece, 2010, p. 172) through a system of interrelated activities to exploit market opportunities (Amit & Zott, 2001).

Companies seek to maximize performance by employing one or more BM themes. The application of these themes creates a BM configuration aligned with one or more of the four structures used for value creation and appropriation (Amit & Zott, 2001; Zott & Amit, 2008; Leppänen et al., 2023). These structures are novelty, efficiency, complementarity, and lock-in (Amit & Zott, 2001; Leppänen et al., 2023). Novelty means finding novel ways of doing business with new activities and actors. Efficiency means using fewer resources to do business than those used by competing BMs. Complementarity means creating synergies by bundling together offerings, activities, or resources. Finally, lock-in means using sunk costs and network externalities to persuade customers, suppliers, or owners not to abandon a given BM, (Various efforts have been made in one of the most promising areas of work to utilize theories like network externalities, (Cruz Cárdenas et al., 2022). The empirical evidence shows that firms that pursue a certain BM theme, such as novelty, or that combine more than one theme by, for example, fusing novelty and lock-in, can outperform the competition (Zott & Amit, 2007; Kulins et al., 2016; Leppänen et al., 2023). Despite the general reduction in activity, digital technologies have experienced a strong boost due to the pandemic (De Crescenzo et al., 2022). Digital channels and processes are utilized by companies to establish branding strategies and retain customers (Khan et al., 2023). Technological advances in data analysis, when combined with the greater availability of relevant data, potentially create a need for profound changes to existing BMs (Veit et al., 2014) so that organizations can survive in the face of globalized competition. In terms of the data correlation for enhancing and validating solutions, algorithms grounded in computational intelligence performed well even with limited data for weight assignment. They demonstrated the ability to appropriately generalize and identify patterns, (Nedjah et al., 2022).

To exploit emerging or disruptive technologies such as ML, organizations must transform their BMs (Chesbrough, 2007; Lee et al., 2019a, b) because ML technology differs substantially from other types of IT. This differentiation poses new opportunities and challenges for organizations. First, ML technologies can replace, support, or limit humans at work (Murray & Perera, 2021). Second, through the use of ML technologies, the boundary between human and machine capabilities can be blurred (Schuetz & Venkatesh, 2020). Third, the data-driven learning approach means that ML technologies create an experimental factor (Choudhury et al., 2021; Vetter et al., 2022), making them more complex because they can lead to unexpected results (Benbya et al., 2020). Given the notable differences between ML technologies and other digital technologies (Benbya et al., 2021) and the transformation of BMs due to ML technologies (Burström et al., 2021), these new ML-driven BMs require further study and theoretical development (Vetter et al., 2022). BM theory explains the creation and capture of value through the use of IT, but it does not account for the specific characteristics of ML. The lens of BM theory identifies the activation of value creation factors in each case. It captures not only whether they are activated individually or in combination but also whether they are activated statically or over time. Therefore, this theoretical construct can shed light on the role of the BM in creating and capturing value through ML technologies.

According to Shaw et al. (2019), ML is a subfield of AI that is having a major impact on all sectors (Shaw et al., 2019). Definitions of AI vary by purpose and domain. However, a common characteristic is that AI technologies mimic human cognitive functions such as problem solving and learning (Lee et al., 2019a, b). This study focuses on ML and its capacity as an external enabler (Davidsson et al., 2020) to offer multiple opportunities for entrepreneurship (Chalmers et al., 2021; Obschonka & Audretsch, 2020).

ML is about learning, reasoning, and acting from data. In ML, computer programs are built to process data, extract useful information, and learn data-based patterns to make predictions about unknown properties or suggest actions or decisions (Brynjolfsson & Mitchell, 2017; Lindholm et al., 2022; Mitchell, 1997; Russell & Norvig, 2021). This learning (i.e., training) process is largely independent of human influence and is therefore highly experimental (Amershi et al., 2019; Choudhury et al., 2021). ML has the potential to revolutionize organizational processes and enable BMs that were previously inconceivable. ML algorithms are valuable for optimizing tasks that maximize and automate processes (Arel et al., 2010; Shaw et al., 2019). Therefore, ML techniques can improve business efficiency, save costs, and optimize time and resources. Regardless of the activity and business context, with enough data, ML is likely to improve operational outcomes more than other technologies (Agrawal et al., 2019).

At the same time, the characteristics of ML make it difficult to create genuine business value for organizations (e.g., Burström et al., 2021). Effective implementation of ML technologies can be challenging because of their complexity, data quality and quantity, and the costs of implementation and maintenance. Harnessing the power of ML is therefore difficult for organizations, and the development of ML-based BMs differs greatly from the creation of BMs based on conventional technologies. BM innovation has always been a demanding and multifaceted process. However, ML compounds this challenge by introducing another experimental component (Choudhury et al., 2021). Current research that could support the development of ML-driven BMs is in its infancy and mainly focuses on specific cases in domains such as manufacturing (e.g., Burström et al., 2021).

The growing interest in configurational thinking is mainly due to the development of fuzzy-set qualitative comparative analysis (fsQCA). This set-theoretical tool makes it possible to determine which conditions are sufficient (and in some cases necessary) to achieve a specific outcome. This method offers an alternative, or complement, to linear regression analysis. It is particularly suited to the study of complex phenomena and causal relationships, which are common in BM research (Fainshmidt et al., 2020).

Configurational research has recently experienced a resurgence thanks to methodological advances that have made it possible to address conjunctural causality, which is where an effect depends on a combination of causes, and equifinality, which is where an end state can be achieved in multiple ways (Fiss, 2011; Furnari et al., 2021). This neo-configurational perspective (Misangyi et al., 2017) allows researchers to study the necessity and sufficiency of theoretically relevant conditions and their combinations in leading to outcomes of interest.

To test the propositions, data were collected from 122 European start-ups. Incumbent firms often pursue multiple BMs with blurred boundaries. Therefore, the study focused on start-ups, whose core business usually remains clearly discernible (Hartmann et al., 2016). The choice to study only European start-ups was based on the availability of data and the possibility of knowledge transfer across countries.

Data were collected from a variety of sources. On the one hand, the sample was narrowed down by obtaining data from Crunchbase, which in turn allowed the dependent variable to be obtained, and on the other hand, Twitter was used to obtain the categorization of the independent variable. For the selection of the sample, the Crunchbase Pro database was queried to search for start-ups that met the research criteria. This database has grown in popularity among academics, particularly for data on start-up activity and funding in different geographical settings (Banerji & Reimer, 2019; Croce et al., 2018; Ghezzi et al., 2016; Gloor et al., 2020). Crunchbase is free for academic use (within the restrictions of its license and terms of use). The social media site Twitter also provided data. Online social media, such as Facebook, Instagram, LinkedIn, and Twitter, provide an important way of communicating (Dwivedi et al., 2021). Social media democratize access to data, raise awareness of emerging companies, and provide information that helps investors evaluate such companies. As underscored by the entrepreneurial stories analyzed in this research, it is essential for prospective investors to comprehend the nature of the proposed solution and grasp its importance for the community, (De Crescenzo et al., 2022). Many national governments recognize Twitter as a formal means of business communication (Ghosh et al., 2019). Start-ups have used social media extensively in digital advertising, brand awareness, and customer relations, among other areas (Shcherbakova, 2019; Virtanen et al., 2017). Social media are also used to identify business opportunities (Troise et al., 2021) and access networks to share information and interact (Smith & Shaw, 2017). Increased start-up activity due to lower barriers to entry has made it difficult for investors to obtain reliable information from start-ups (Cao, 2020). Start-ups, which might miss the opportunity to reach these investors, benefit from signaling relevant information (Cao, 2020). Social media such as Twitter offer companies a platform to present themselves and provide information for wider dissemination. A third source of data was company websites. These websites were reviewed to complement and triangulate the information obtained from Crunchbase and Twitter.

The Crunchbase search term used to search for start-up labels and descriptions was “machine learning” (Webber et al., 2022). The aim was to select start-ups, so the search looked for companies founded between 2019 and 2022. Start-ups had to be for-profit and had to have received pre-seed and seed stage funding from venture capitalists or angel investors between January 2019 and June 2022. The search focused on seed and pre-seed funding. Start-ups that have already received higher rounds of funding have considerable traction, so the public and private domains contain considerable information on them. Distinguishing the impact of start-ups in higher rounds of funding is difficult because it may be influenced by factors derived from their development and previous funding. The data collection process provided a usable data set of 288 start-ups. Start-ups without funding data were discarded, as were those without a Twitter account or website, giving a total sample of 122 start-ups.

In the next step, user data and tweets posted by the sampled companies were downloaded from Twitter using the Google Chrome extension Twlets. This step gave a total of 30,677 tweets. The company with the most tweets was Zeta Alpha, with 1,732 tweets since registering with Twitter in October 2019. The Twlets extension also provided the number of followers at the time of download and the dates of publication of each tweet, multimedia content, and retweets or favorites.

After cleaning the data from Twitter (removal of media files, emoticons, and links to other accounts or websites), a search was conducted to identify words, sets of words, and expressions related to the value creation and capture factors used in this study. These factors were efficiency, novelty, and complementarity (value creation and value capture factors from BM theory), as well as performance, utility, explainability, and transparency in the use of personal data (value creation factors from the theory of data network effects). Value creation factors related to managing data network effects and data quality and quantity were discarded. The reason was that access to this information would not have been possible because it belonged to each company. Furthermore, the discussion in this paper explains that start-ups lack a large amount of data and are thus unable to guarantee data quality (Alotaibi et al., 2020). The theme of BM lock-in was not included in the study. This theme is only triggered after considerable company development because it relates to network effects and refers to deterrents to prevent customers from switching to competitors.

This process identified which value creation factors were contained in each of the tweets published by each company. The results on the ranking of each tweet for each factor were confirmed by collecting data on each start-up’s BM from publicly available sources, such as corporate website, articles, blogs, and other online sources. Since “the raw elements of business models are usually quite transparent” (Teece, 2010, p. 179), the BM of a start-up can be inferred using reliable public sources (see Hartmann et al., 2016; Möller et al., 2019). Multiple sources were triangulated to ensure a high degree of objectivity (e.g., Flick, 2004; Hsieh & Shannon, 2005; Vetter et al., 2022). A study was then conducted using fsQCA to test the research propositions.

FsQCA is a set-theoretic method that considers cases as configurations of causes and conditions, rather than treating each independent variable as analytically distinct and isolated from the rest. This empirical method examines the relationships between an outcome of interest (high level of funding) and all possible combinations (presence or absence) of its predictors (novelty, efficiency, performance, and utility). Interest in fsQCA was sparked by Ragin (e.g., Ragin, 1987, 2000, 2008a, b). Its main purpose is to fit data to theory by going beyond the reliance on a single sample, which involves achieving predictive validity (Gigerenzer & Brighton, 2009; McClelland, 1998; Woodside, 2013; Wu et al., 2014). Set theory provides the basis to understand fsQCA. It allows for detailed analysis of how causal conditions contribute to an outcome of interest. Instead of estimating the effects of individual variables, fsQCA uses Boolean logic to examine the relationships between an outcome and all possible combinations (or configurations) of multiple antecedent conditions. This approach allows researchers to find different combinations of causal conditions that reflect different theoretical pathways to particular outcomes (Longest & Vaisey, 2008). According to Ragin (2008a, b), rather than investigating which factors are most important, fsQCA seeks to find out which factors should be combined and in what way.

Consistency (analogous to correlation) is the proportion of cases compatible with the outcome. In other words, it is the number of cases that exhibit a given configuration of attributes and the outcome divided by the number of cases that exhibit the same configuration of attributes but that do not exhibit the outcome. It can be visualized using scatter plots. The underlying idea is that a fuzzy subset relationship exists when the membership scores in one set, Xi, are consistently less than or equal to the membership scores in the other, Yi (i.e., Xi ≤ Yi, meaning that the configuration is a sufficient condition for the outcome to occur). Coverage (analogous to the coefficient of determination) assesses the empirical relevance of a consistent subset. In other words, it reflects the proportion of the outcome that is explained by the conditions in the model (in this case, by the solution). The coverage of a causal combination ensures that the cases that meet it cover enough of the outcome to be empirically important. It refers to the proportion of members with the outcome (i.e., with high funding, as explained by the solution). A causal combination that covers or represents only a small proportion of the cases where the outcome occurs is less empirically important than one that covers a large proportion. Coverage measures only empirical importance, not theoretical importance.

When analyzing necessity, researchers seek causal conditions with membership scores that are consistently higher than the outcome membership scores. If there is a causal condition for which this requirement holds in all cases, then this condition passes the necessity test, and the outcome is considered a subset of the causal condition. This situation is the set-theoretic way to express necessity.

Once the outcome and conditions had been calibrated (the suffix fz indicates calibration), the truth table was computed. The truth table shows all logically possible configurations. It has 2 ^k rows (each corresponding to a possible configuration), where k is the number of conditions. In this case, the truth table has 16 rows (2⁴ = 16 combinations) (Fig. 3).

A value of 1 in each configuration indicates a calibrated score of greater than or equal to 0.5 (i.e., closer to membership than non-membership). A value of 0 indicates a calibrated score of less than 0.5 (i.e., closer to non-membership than membership).

The next step was to eliminate configurations without remainders. Then, a consistency threshold was selected to distinguish between causal configurations that were subsets of the outcome and those that were not. In general, values below 0.80 in this column indicate substantial inconsistency. The value 0.85 was used as the consistency threshold. A value of 1 was assigned to the outcome variable (TotalFundingfz) when the consistency of that configuration exceeded this threshold. A value of 0 was assigned otherwise.

The study indicates that emerging companies use the BM themes of efficiency and novelty. In relation to data network effects, the study also identifies performance expectancy as a moderating factor. In contrast, data quantity and data quality are less relevant because data are exclusive to each firm and new companies are unable to generate extensive data sets or guarantee data quality. The BM theme of lock-in was not addressed in the study. Lock-in is possible when companies are already well established. It is related to network effects and preventing customers from switching to the competition.

The analysis suggests that the interaction of value drivers (as proposed in BM theory) with performance expectancy (proposed in the theory of data network effects) is associated with the funding raised by firms (Costa-Climent et al., 2023). This finding is derived from fsQCA, which is based on a configurational approach. Statistical regression analysis has also showed that these factors are individually correlated with funding for start-ups (Costa-Climent et al., 2023). This empirical study finds strong evidence that any positive BM impact is likely to be the result of an improved configuration of all components rather than a single one (Leppänen et al., 2023). The study confirms the benefit of combining efficiency-focused factors. The promise of improving product or service performance can generate and capture sufficient value for investors.

Novelty appears in three of the four configurations observed in the analysis. This finding suggests that novelty is important for attracting financing. However, efficiency is present in both of the most relevant combinations, suggesting that it is minimally sufficient to improve a start-up’s level of funding. Novel BM design is not sufficient on its own to achieve high performance. However, it is effective when combined with other value-creating BM design elements. Novelty is relevant in value creation because it enhances value in use (Bowman & Ambrosini, 2000; Lepak et al., 2007) by offering customers something that satisfies their needs in a novel way (Zott et al., 2010). However, capturing this created value requires firms to use design mechanisms focused on value capture (Almeida Costa & Zemsky, 2021; Teece, 2010). The ability of novelty to capture value is far less than its ability to create it because novelty might not lead to real improvements in BM performance. Although novelty may increase overall potential to create value, it is unclear why it would allow a firm to capture a larger amount of that value.

The findings of this study have important implications for entrepreneurs interested in using ML to create and capture value. First, entrepreneurs should carefully consider their BM and reflect on how it aligns with the two main elements of value creation and value capture identified in this study. The act of obtaining resources for social entrepreneurship intrinsically entails the dissemination of social worth and influence, focusing heavily on the generation of social value and the engagement of stakeholders, (Zhao et al., 2023). For example, if a start-up’s BM focuses on efficiency, ML can be used to optimize processes and reduce costs. If a start-up’s BM focuses on data-driven innovation, ML can be used to develop new products or services tailored to customer needs.

Second, entrepreneurs must be aware of the aspects that can influence the effectiveness of ML in creating and capturing value. For example, the availability and quality of data, the level of ML expertise, and the regulatory environment can influence the success of a start-up’s ML initiatives. Therefore, business owners must carefully evaluate these aspects and develop strategies to mitigate any potential risks.

Despite its contributions, this study has several limitations. First, the sample was small (122 European start-ups). Although it was sufficiently large for the purposes of this study, future research could benefit from a more extensive and diverse sample. Second, the study focused exclusively on start-ups and did not consider incumbents. Although this approach addressed the research aims of the study, future research could explore how ML can create and capture value in established firms. Finally, this study was limited by its cross-sectional design, which prevents causal inference. Future research could use longitudinal or experimental methods to establish causal relationships between ML, BMs, and value creation and capture.

Given this study’s limitations, several lines of research can be proposed. First, future research could explore how ML can be used to create and capture value in different industries and contexts. For example, the effectiveness of ML in creating and capturing value may vary across sectors (healthcare, finance, retail, etc.). Therefore, future research could explore how ML can be used to create and capture value in these and other sectors. Second, future research could explore how ML can be used to create and capture value in incumbents. Although this study focused exclusively on start-ups, ML can also be used to create and capture value in established firms. Future research could explore how ML can be used to create and capture value in different types of organizations. Finally, future research could explore how ML can be used to create and capture value in different geographic regions. This study focused exclusively on European start-ups, but the effectiveness of ML in creating and capturing value may vary across different parts of the globe. Future research could explore how ML can be used to create and capture value in different regions such as Asia, Africa, and the Americas.