Introduction

Recent years have seen an upsurge in start-ups that employ artificial intelligence (AI) and machine learning (ML), The automation of knowledge and service work has been greatly impacted by the advancements in Artificial Intelligence (AI) and its sub-fields, (Coombs et al., 2020). With the plethora of venture-backed startups launching in Silicon Valley, few people paid attention to SNSs that gained popularity elsewhere, even those built by major corporations, (Boyd & Ellison, 2007). Machine learning is an artificial intelligence technology that discovers and extracts hidden information from data warehouses (Lin et al., 2022). AI has potential in fields such as cognition and natural language processing. Applying big data thinking and artificial intelligence (AI) diagnosis technology in environmental governance can provide data and technical support for environmental issues, (Fu et al., 2023). ML start-ups have capitalized on data availability and advances in technology such as the Internet of things (IoT) and cloud storage, the learning curve demonstrates how a machine learning model's accuracy improves as the training dataset size increases, (Sáez-Ortuño, et al., 2023).These advances have allowed start-ups to optimize their business strategies and enhance their customer offerings (Jordan & Mitchell, 2015; Weber et al., 2022). There's a pressing need to incorporate advanced technologies like IoT, cloud computing, AI, and machine learning, (Gupta et al., 2023).

However, the use of ML also poses challenges for value creation and capture. Some start-ups that invest in ML technology succeed in creating unparalleled value. A notable example is Facebook (Joshi et al., 2021). Meanwhile, others fail. A long list of such businesses includes Quixey, Juicero, Amazon Fire Phone, and Google Glass (Klein et al., 2020). The following question therefore arises: What influences the success of value creation and value capture from the use of ML technology? The answer to this question could help start-ups seeking early-stage funding to develop strategies that highlight the value of their offerings (Singhal & Kapur, 2022). It could also improve value estimates by future investors (Leendertse et al., 2021). Understanding how start-ups transform business model innovation into superior firm performance in the digital economy is crucial, (Guo et al., 2022).

In an extensive study of value creation and capture through information technology (IT), Kohli and Grover (2008) described various factors that enhance value production. However, throughout this theoretical field, IT is presented as a uniform concept or an artifact that processes information in various ways. A detailed explanation of the implications of each particular type of IT in a firm’s value proposition is missing (Brynjolfsson & Hitt, 1996; Kohli & Grover, 2008). To capitalize on ML technology, firms must define how to capture the value they create using a suitable business model (Åström et al., 2022; Bouncken et al., 2021). According to Valter et al. (2018), research has not yet explained how companies successfully deploy technology solutions such as AI or ML through their business models (BMs), which are the bridge between technology and economic value (Chesbrough & Rosenbloom, 2002).

Despite insights from research, experts and professionals have not yet agreed on a definition of the BM or its core components (Rocha et al., 2018). The BM has two basic functions: value creation and value capture (Chesbrough, 2007). Value creation is typically analyzed in terms of economic profit (Shakina et al., 2013). BM theory has been able to explain how IT companies create and capture value by using different elements of their BM and activating one or more factors that drive value creation (Amit & Zott, 2001; Zott & Amit, 2007). These factors are novelty, efficiency, complementarity, and lock-in (Amit & Zott, 2001). However, such a theory only considers IT as a generic tool, obviating the unique characteristics of ML technology. Consequently, BM theory must be extended to account for value creation and appropriation through the use of ML technology.

The theory of data network effects explains how ML creates value for users (Gregory et al., 2021). The importance of the scale of data-driven learning and improvements made through ML technology is crucial (Costa-Climent, 2023). It focuses on three factors that modulate value creation: data management, user-centric design, and platform legitimization. Developed in the context of business platforms that leverage data network effects, these platforms process large volumes of user interaction data, detect patterns, and tailor offers based on interests. However, the theory does not clarify how value is captured when deploying ML (Clough & Wu, 2022; Gregory et al., 2022). Another limitation of the data network effect theory is that it does not take into account what happens with startups that use ML but have an extensive and competitive database due to their short existence or impossibility to fight the entry barriers of large companies.

Innovationis a source of value creation through the introduction of new technologies and exploration of new markets (Okano et al., 2023). Disruptive technologies such as ML present opportunities to improve operational efficiency and accelerate innovation by managing large data sets. Crucially, they offer predictive capabilities to improve decision making (Lee et al., 2019a, b). However, the lack of understanding of how advances in the field of ML can be transformed into profit means that further research is needed to enable start-ups to apply ML technology successfully. For example, Leppänen et al. (2023) studied technology-based BMs, showing that a BM focused on novelty is insufficient to achieve high performance. Conversely, it is effective when combined with other elements of BM design.

Based on these gaps in the literature and the practical implications involved, we posed the following research question: How do business model themes and the moderators of data network effects combine to allow ML-based start-ups to create and capture value?

The primary objectives of this research encompass several dimensions. Firstly, it aims to contribute to the advancement of theoretical development by elucidating the mechanisms through which technological advancements, specifically the capabilities of ML, can be harnessed to generate tangible benefits. The enhancement in total productivity, stemming from technological advancements, will augment the overall output. This may lead to new entrepreneurial opportunities and the generation of jobs, (DeCanio, 2016) Secondly, this study seeks to underscore certain limitations identified within the existing body of literature that pertain to the valuation of IT utilization.

In addition, a significant goal of this research is to further deepen our understanding of the interconnectedness, as elucidated by Costa-Climent et al. (2023), between BM theory and the theory of network effects associated with data. This exploration endeavors to provide insights into how the application of ML can facilitate the process by which startups can both create and capture value in the contemporary business landscape. The study examines data on 122 European start-ups for the period 2019 to 2022. It links BM value drivers to the theory of data network effects. By presenting new findings on soft skills from the perspective of European employers, it will also contribute to the expansion of literature, (Ogunrinde, 2022). Under a neo-configurational approach (Misangyi et al., 2017), fuzzy-set qualitative comparative analysis (fsQCA) is used to link configurations of value drivers to funding outcomes (Fiss, 2011; Ragin, 2000, 2008a, b). The fsQCA approach necessitates constructing truth tables that showcase the distribution of sample cases across all potential combinations of causal factors (Apetrei et al., 2019). This study highlights the relationship between the financing of start-ups and the factors that affect value creation and capture, particularly efficiency and novelty. Although novelty is important for value creation, its ability to capture value is limited if it is not effectively combined with other elements.

Background

The background section of this study centers on three aspects. First, the value of IT is discussed, and recent developments are described. Second, research on the BM is examined. The unique features of ML in this area are highlighted. Third, the discussion turns to related research on the theory of data network effects and finally value creation and capture using ML.

The value of information technology

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.

Business model theory

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.

Unique features of machine learning

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).

Data network effects

The recently proposed theory of data network effects specifies the factors that individually and jointly condition the success of a learning loop (Gregory et al., 2021). According to the theory of data network effects, the utility that a user derives from a product or service depends on more than simply the size of the network. It is instead a function of the scale of data-driven learning and improvements made thanks to ML technology. Nonetheless, our findings indicate that a firm possessing greater user data becomes more competitive due to an additional externality that we have identified. This can be regarded as a supplementary effect of data network, (Schaefer & Sapi, 2020). The idea of data network effects is that there is a direct positive relationship between the AI capability of a platform (or provider) and users’ perceived value. The theory of data network effects posits that this relationship is moderated by data stewardship, user-centered design, and platform legitimization, as presented in Fig. 1.

Fig. 1
figure 1

Scatter plot of high funding against novelty: Positive outcome

Full size image

Data management conditions the relationship between AI capability and users’ perceived value. ML models rely heavily on the quality and quantity of data used for AI capabilities to enhance users’ perceived value by increasing the speed and accuracy of predictions. Businesses must refine and extract value from data through data stewardship, defined as the comprehensive management of a company’s data assets to help ensure the appropriate quantity and quality of data (Baesens et al., 2016). If data quality is low due to inaccurate measurements or insufficient data, the patterns that the technology identifies will be of poor quality, as will the predictions. An example is IBM’s inaccurate predictions in cancer diagnosis (Khoury & Ioannidis, 2014).

User-centered design can be a key success factor for companies that embrace consumerization, a process that involves the widespread adoption and diffusion of digital consumer technologies across society (Gregory et al., 2018). This type of design allows users to participate in value creation by contributing their feedback and personal data to the continuous improvement and refinement of a platform’s model and features. Consequently, enterprises must ensure that the changes brought about by digital transformation are understood by all the people involved and directly affected by them, particularly customers, (Méndez-Suárez & Danvila-del-Valle, 2023). Digitization has emerged as a catalyst for innovation and competitiveness, continually enriching its value through the collection of consumer data. This information is subsequently transformed into applicable knowledge that informs business decisions and addresses customer concerns (Miguel et al., 2022). This outcome is achieved from a combination of performance expectancy (Venkatesh et al., 2003) and effort expectancy (Venkatesh et al., 2003). Performance expectancy is the degree to which users assume that using the offering will provide gains in task performance. Effort expectancy is the degree to which users assume that using the offering will not be burdensome. Therefore, successful data network effect activation requires low effort expectancy and high return expectancy so that users are attracted to a given offer and thus become involved in the data network effect loop.

In the theory of data network effects, platform legitimacy is a balancing act performed by platform owners to satisfy the interests of various stakeholders by mitigating any potential risks of personal data use (Kroener & Wright, 2014; Wiśniewski et al., 2021) and the explainability of predictions (Coglianese & Lehr, 2019). Regarding the use of personal data, platform owners must demonstrate that data collection and use are morally acceptable. Meanwhile, greater explainability of predictions by algorithms increases the value perceived by users. Recent debates about Facebook’s use of ML algorithms reflect broader legitimacy challenges. Overall, the theory of data network effects describes a new category of network effects that offer a unique explanation of the creation of user value from the use of ML technology.

Creating and capturing value using ML

The increased use of ML gives providers the opportunity to create additional value by proactively managing uncertainty, thereby improving cost efficiency and boosting revenue (Cockburn et al., 2018; Laudien & Pesch, 2019). However, to capitalize on ML technology, adopters must understand how it can be commercialized using the right BM (Bouncken et al., 2021). According to Valter et al. (2018), existing research does not explain how companies can successfully deploy ML solutions through their BMs, which bridge the gap between technology and economic value (Chesbrough & Rosenbloom, 2002). BMs can be understood in terms of two of their most important functions: value creation and value capture (Chesbrough, 2007).

Value creation refers to situations where an offering (of a good or service) satisfies target customers’ needs (Chesbrough, 2007). Value capture, also referred to as value appropriation (e.g., Burkert et al., 2017; Mizik & Jacobson, 2003), refers to the mechanisms that ensure that created value is at least partially appropriated. Value creation and value capture are interrelated but distinct processes (Lan et al., 2019; Lepak et al., 2007; Schreieck et al., 2021; Sjödin et al., 2020) because the source that creates incremental value from a product or service does not always succeed in capturing any of this value (Lepak et al., 2007).

The theory of data network effects focuses on value creation for users. It does not explicitly consider value capture through the use of data network effects. Although the decentralized model is based on the exchange of value between parties, planning for value capture is not necessarily less relevant than planning for value creation (Clough & Wu, 2022). The ability of a BM to capture created value for stakeholders requires specific analysis. Value creation and value appropriation have been extensively researched in terms of their effects and determinants (e.g., Blocker et al., 2012; Obloj & Capron, 2011; Priem, 2007) or their interactions and trade-offs (e.g., Bowman & Ambrosini, 2000; Mizik & Jacobsen, 2003). However, these studies have predominantly focused on value creation rather than value appropriation (Reitzig & Puranam, 2009), even though value appropriation is arguably a firm’s main objective (Pitelis, 2009). The more effective a firm is at value appropriation, the better it can avoid value creep (Lepak et al., 2007; Parolini, 1999). Value appropriation is considered a fundamental organizational capability (Reitzig & Puranam, 2009).

A number of factors from the theory of data network effects condition value creation through ML. Examples include the quantity and quality of data, performance and effort expectancy, and the legitimacy of predictions (Gregory et al., 2021). However, no specific or moderating factors related to value capture, such as barriers to entry (Terlaak & Kim, 2021), bargaining power (Lepak et al., 2007), and pricing (Burkert et al., 2017), are able to explain value capture in traditional BMs.

The idea is that implementing a novel element, such as the use of ML, can focus on improving efficiency, complementarity, and lock-in. It cannot be argued that the novelty of deploying ML as the sole driver of value is central to improving the performance of an innovative firm, as advocated in the literature on innovation and business strategy (Foss & Saebi, 2017; Zott et al., 2011). The three value drivers of efficiency, complementarity, and lock-in differ from novelty in that they focus more on value capture than on value creation (Almeida Costa & Zemsky, 2021).

In addition, start-ups may struggle to compete with large established firms with an innovative offering alone. They need to combine two or more value creation and appropriation factors and design a strategy that aligns value creation with value capture. BM theory explains value creation and value capture through the use of IT (Amit & Zott, 2001; Zott & Amit, 2007, 2008) but needs to account for the specific characteristics of ML. However, this theoretical lens identifies the activation of value creation factors in each case, whether they are activated individually or in combination and 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.

Casual empirical experience shows that data network effects can be crucial in value capture. However, start-ups need to be included in the explanation of the moderating factors of value creation through ML under the theory of data network effects. They are absent from the theory because, in their initial phase, they still need a vast data set to feed learning mechanisms and improve the accuracy and speed of their predictions (Gregory et al., 2021; Meinhart, 1966). The theory of data network effects refers to the moderating role of data stewardship in value creation. However, empirical experience (Valavi et al., 2020; Schaefer & Sapi, 2020) suggests that data-related factors or dynamics moderate or condition the creation and capture of value from the use of ML. Three examples are time dependency/data expiration (Valavi et al., 2020), variety (Schaefer & Sapi, 2020), and database size in the form of critical mass (Afuah & Tucci, 2003).

The combination of data and network effects creates high barriers to entry in online markets (Stucke & Grunes, 2016). For instance, Google blocks out potential competitors in part thanks to the massive amounts of data it obtains from searches. Nevertheless, some scholars have argued that the amount of data is not always a barrier to entry (Lambrecht & Tucker, 2015; Sokol & Comerford, 2016). Google’s advantage derives from not only the amount of data it collects but also the results of product-oriented experiments. According to Varian (2010), Google conducted 6,000 experiments on its search engine in 2008.

Similarly, start-ups, by definition, have no track record of transparent and appropriate personal data use or high-quality ML-based predictions. Without much data, ML-driven start-ups must create and capture value by activating BM themes in combination with moderating factors of ML-driven value creation using data network effects. They can thus enter a market and compete with incumbents. They can appropriate the value they create by securing external early-stage funding. Return expectancy, effort expectancy, and performance expectancy when using an ML-based offering (Venkatesh et al., 2003) are vital factors that fit with a BM’s value drivers of, for example, novelty and efficiency. An ML-based offering gives customers an innovative product that supposedly makes their life easier. Its efficiency saves customers cost and effort and increases performance expectancy by automating tasks, enhancing capabilities, personalizing recommendations, facilitating decision-making, and managing data, among other benefits (Brynjolfsson et al., 2021; Shaw et al., 2019; Arel et al., 2010).

In the case of start-ups, another key factor in value creation and capture is the ability to obtain early-stage financing (Singhal & Kapur, 2022). Funding is crucial for firms to grow, develop new products, invest in physical and human capital, and access new international markets (Giaretta & Chesini, 2021). It plays a critical role in survival and success (Croce et al., 2018; Werth & Boeert, 2013), but most start-ups perish within three to five years (Song & Vinig, 2012; Bandera & Thomas, 2017; Antretter et al., 2019). Hence, one focus of entrepreneurship research is to understand what conditions the success of start-ups. Successful start-ups can revolutionize their target market and provide large returns to investors (Singhal & Kapur, 2022), hence the importance of estimating the value of start-ups. The BM literature describes few strategic implications of how initial investment in start-ups influences their BM value creation and conditions the process that then affects outcomes by capturing value for stakeholders such as founders, investors, customers, and potential buyers (Chammmassian & Sabatier, 2020).

The issues discussed in this section lead to the following research question: How do business model themes and the moderators of data network effects combine to allow ML-based start-ups to create and capture value? To investigate this question, and based on the arguments presented in the preceding paragraph, we propose the following hypothesis:

H: A start-up can create and capture value at the onset of its activities by activating the moderating value factors related to novelty and expectations of efficiency and performance.

The methodology is based on a configurational analysis. The analysis captured the system of interactions between BM value drivers and value creation moderating factors from the theory of data network effects. A neo-configurational approach (Misangyi et al., 2017) built on fuzzy-set qualitative comparative analysis (fsQCA) was used (Fiss, 2011; Ragin, 2000, 2008a, b). Using qualitative comparative analysis (QCA), we then determine which combinations of factors lead to strategic agility, (de Diego Ruiz et al., 2023). The method is explained in the next section.

Method

The neo-configurational approach

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.

Research design and data collection

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.

The fsQCA method

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.

A fuzzy-set condition can be viewed as analogous to a continuous variable. It is carefully calibrated to indicate an empirical case’s degree of membership in a well-defined set. Calibration is only possible based on theoretical knowledge, which is essential for the specification of three qualitative cut-off points or thresholds (full membership, full non-membership, and maximum ambiguity). For example, cases in the lower ranges of a conventional continuous variable may be totally outside the set in question, with fuzzy membership scores truncated at 0. In contrast, cases in higher ranges of this same continuous variable may be totally inside the set, with fuzzy membership scores truncated at 1. To apply fsQCA in this research, fs/QCA software (version 4.0) was used.

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.

Results

This section presents the results of the fsQCA of the relationship between the activation of specific value creation and value capture factors and the funding performance of start-ups (Table 1).

Table 1 Descriptive analysis of the sample
Full size table

Calibration, in this research, indicates the extent to which businesses can be considered members of sets that vary according to their particular attributes. Therefore, each of these quantitative variables must be calibrated to assign degrees of membership or belonging to previously defined sets. For the calibration, different cut-off points of the variables are used, which are transformed into the values 1 (full member), 0.5 (ambiguous member), and 0 (non-member) in the new calibrated variable. In this research, the 95th percentile corresponds to the calibrated value 1 –full member–, the 75th percentile corresponds to the ambiguous calibrated value 0.5, and the median corresponds to the calibrated value 0, non-member (Table 2).

Table 2 Calibration
Full size table

Necessity analysis

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.

Positive outcome

With the exception of Useful, all individual conditions meet the requirement to be considered necessary (consistency greater than 0.80).Footnote 1 Thus, a high level of novelty, efficiency, or performance seems to be necessary for a high level of funding. The conditions that meet the requirement to be considered necessary are those that appear in at least two of the overall solution configurations in the sufficiency analysis. Specifically, novelty appears in three configurations, efficiency and performance appear in two configurations, and utility appears in one configuration. There is an important difference in the application of the subset principle when assessing sufficiency and necessity. To show necessity, researchers must verify that the outcome is a subset of the cause (Table 3).

Table 3 Results of analysis of necessity: Positive outcome
Full size table

In Fig. 1, a large proportion of companies lie below the diagonal line for novelty, which is a necessary condition for the outcome to occur. Even more companies would lie below the diagonal line for novelty if the consistency exceeded 0.90. Two-dimensional scatter plots representing the arithmetic relationship between the outcome and condition help visualize the concept of subsets, where the outcome is a subset of the condition.

Negative outcome

No individual condition meets the requirements of necessity (consistency greater than 0.80). However, the negative conditions have higher consistency values than the positive ones (Fig. 2) (Table 4).

Fig. 2
figure 2

Scatter plot of high funding against novelty: Negative outcome

Full size image
Table 4 Results of analysis of necessity: Negative outcome
Full size table

Sufficiency analysis

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).

Fig. 3
figure 3

Conceptual configurational model of antecedent conditions leading to funding success

Full size image

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.

Positive outcome

The resulting intermediate solution reveals four sufficient configurations for a high level of funding (Table 5).

Table 5 Reduced solution for the outcome of high level of funding
Full size table

Three common operations on fuzzy sets are negation ( ~), intersection (logical “AND” represented by the operator *), and union (logical “OR” represented by the operator +). In logical negation ( ~), the membership of set ~ M is equal to 1 minus the membership of set M. The logical “AND” (*) calculation is carried out by taking the minimum membership score of each case in the sets that are combined. The logical “OR” ( +) operation is carried out by taking the maximum membership score of each case in the sets that are combined.

The four configurations shown in Table 2 are sufficient to lead to a high level of funding in 85.6% of cases. They cover 79.5% of cases in the data set.Footnote 2 Consistency is the proportion of cases that are consistent with the outcome. In other words, of the cases that reflect a given configuration, the consistency is the percentage of cases that lead to the outcome. High coverage ensures that the cases where an outcome occurs cover a large proportion of total cases. It measures empirical relevance. The configurations with the highest raw coverage were USEFULfz*EFFICIENCYfz and NOVELTYfz*PERFORMANCEfz*EFFICIENCYfz. The efficiency condition is present in both. This finding suggests that it is a minimally sufficient condition for a high level of funding. These configurations have coverage scores of 62% and 68%, respectively. Novelty appears in three of the four configurations in the solution, also suggesting that it is important for financing.

When analyzing sufficiency, the membership scores of the outcome should be compared not only with the score of each individual causal condition (as in analysis of necessity) but also with the scores of all possible causal configurations. If all cases are above the diagonal, it indicates that the membership scores of the outcome are consistently higher than the membership scores of the causal combination. Hence, the causal configuration is a subset of the outcome, which is the set-theoretic way of expressing sufficiency.

To represent the consistency and coverage of a solution, scatter plots can be used to plot the outcome against the solution. A configuration where all (calibrated) scores are systematically less than or equal to the scores of the outcome (upper triangle) is said to be a subset of the outcome, and the consistency is high. Companies below the diagonal are those that are inconsistent with the outcome, and those above are consistent. However, within each group, there are degrees of relevance, depending on whether the combined membership score is lower or higher than 0.5 (upper right quadrant). Inconsistency in the red triangle (Yi ≥ 0.5, Xi > Yi) is more severe, and consistency in the green triangle (Xi ≥ 0.5, Xi ≤ Yi) is more relevant.

Inconsistent companies are those that lie below the diagonal. Severely inconsistent companies are those that lie in the red triangle. Consistent companies are those that lie above the diagonal. The most relevant companies are those that lie within the green triangle. In both configurations (Figs. 4 and 5), there is a large proportion of consistent companies lying above the diagonal (consistency > 0.85). Approximately a third of them lie in the green triangle (most relevant). These companies are from different countries. Turkey and Switzerland are present in the first configuration. The most severely inconsistent companies lie in the bottom right corner of the red triangle in the first configuration. They are mainly companies from the UK (Fig. 6).

Fig. 4
figure 4

Scatter plot of high funding against utility and efficiency

Full size image
Fig. 5
figure 5

Scatter plot of high funding against novelty and performance and efficiency

Full size image
Fig. 6
figure 6

Scatter plot of high funding against utility and efficiency or novelty and performance and efficiency

Full size image

The overall consistency of these two solutions together is 86.23%, and the coverage is 74.78%.

Negative outcome

One of the properties of fsQCA is asymmetry. Contrary to other quantitative analysis techniques, it is also informative to study which configurations of conditions lead to a low level of finance because the inverse of the solution for a given outcome does not always explain the negation of that outcome. The model for the negative outcome is shown in Table 6.

Table 6 Reduced solution for the outcome of low levels of funding
Full size table

The model has high coverage and acceptable consistency. The first two configurations have raw coverage scores of more than 40%, and the third configuration has a coverage of approximately 30%. All three have consistency scores of more than 80%. With an overall coverage of 58.9% and a consistency of 78.10%, the three configurations that are sufficient for a low level of funding are as follows:

USEFULfz* ~ EFFICIENCYfz + 

 ~ EFFICIENCYfz*NOVELTYfz + 

 ~ USEFULfz*EFFICIENCYfz* ~ NOVELTYfz.

Although the consistency of the solution (0.78) is below the threshold of 0.80, cases with severe inconsistency (red triangle) are in the minority (Fig. 7).

Fig. 7
figure 7

Scatter plot of low funding against utility and lack of efficiency or lack of efficiency and novelty or lack of utility and efficiency and lack of novelty

Full size image

Discussion

Research on value creation and capture through the use of ML by start-ups is in its infancy (Brynjolfsson & Hitt, 1996; Kohli & Grover, 2008). In this study, fsQCA was used to analyze data on start-ups created between 2019 and 2022. The study thus examined the relationship between funding received in early financing rounds and the activation of value creation and capture factors. This study corroborates the hypothesis raised in this research and contributes to the literature on value creation and appropriation in start-ups by adopting a neo-configurational approach to explain how a company’s BM interacts with various aspects to create and capture value using ML. Specifically, this study identifies two main elements of value creation and capture that start-ups primarily activate: data-driven innovation and operational efficiency. By examining the interaction between these elements and the company’s BM, this study provides a nuanced understanding of how start-ups can use ML to create and capture value sustainably. This study advances the literature on the business value of using ML, improving the current understanding of the impact of ML technologies on value creation and capture. The findings suggest that start-ups mainly activate two elements of value creation and value capture according to BM theory: efficiency and novelty. The results indicate that performance expectancy moderates value generation as per the theory of data network effects.

Based on the results of this study, there are several key findings that can be synthesized with existing work published on the topic of value creation and appropriation from the use of machine learning in start-ups.

Firstly, the study found that funding received in early financing rounds is positively associated with the activation of value creation and capture factors. This finding is consistent with previous research that has highlighted the importance of funding for start-ups to develop and implement machine learning technology effectively.

Secondly, the study identified several key elements of value creation and capture that are activated by start-ups using machine learning technology, including efficiency, novelty, and utility. These findings are consistent with previous research that has emphasized the importance of these factors for successful value creation and appropriation in start-ups.

Thirdly, the study highlights the importance of a firm's business model in explaining how machine learning technology can be used to create and capture value. Undoubtedly, digital transformation has become a crucial driver of innovation and has also profoundly altered profit models and business models, (Li et al., 2023). The digital transformation of companies presents a complex challenge that leaders must actively guide to succeed in a fast-paced and constantly changing environment, (Tagscherer, & Carbon, 2023). This finding is consistent with previous research that has emphasized the importance of a firm's business model in shaping its ability to create and capture value in a variety of contexts.

Overall, the findings of this study contribute to a growing body of literature on value creation and appropriation from the use of machine learning in start-ups. By highlighting the importance of funding, key elements of value creation and capture, and a firm's business model, this study provides valuable insights for entrepreneurs, investors, and policymakers seeking to support the development and implementation of machine learning technology in start-ups.

Theoretical contributions and implications

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.

Implications for practice

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.

Limitations and Future Research Direction

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.

Conclusions

Incorporating ML technologies into the business practices of start-ups has the potential to yield significant financial gains. Thus, it is imperative for start-ups to view the costs of ML as a strategic investment that can attract funding for further development of their offerings. By adopting an investment-oriented mindset, start-ups can emphasize the importance of securing future value through return on investment.

This study presents empirical evidence that highlights the correlation between combining efficiency and novelty with performance and utility expectations, leading to value creation and partial capture of that value through early-stage funding for ML-driven start-ups. This correlation introduces a novel understanding of how value can be created and captured through ML technologies and contributes to the conceptualization of this process.

The study's findings offer a comprehensive multi-theoretical perspective on how start-ups can strategically utilize ML technologies to create and capture value. This can open up further avenues for research and practical applications in this promising field. Therefore, start-ups that embrace ML as a strategic investment can position themselves to reap the rewards and stay competitive in an ever-evolving business landscape.

One of the key findings of the study is that start-ups that have both funding and a high degree of novelty in their machine learning technology are more likely to create and appropriate value than start-ups with only one of these antecedents.

This suggests that having access to funding is important for start-ups to be able to develop and implement innovative machine learning technology that can create value for users.

Another important finding of the study is that the business model of a start-up plays a critical role in determining the value that can be created and appropriated from the use of machine learning technology. Specifically, the study found that start-ups that focus on efficiency and novelty in their business model are more likely to create and appropriate value than start-ups that focus on lock-in. This suggests that start-ups should prioritize developing business models that emphasize efficiency and novelty in order to maximize the value that can be created and appropriated from the use of machine learning technology.

The study also found that data network effects can play an important role in creating perceived value for users of machine learning technology. Specifically, the study found that the activation of data network effects by start-ups can lead to the creation of perceived value for users. This suggests that start-ups should focus on developing machine learning technology that can activate data network effects in order to create value for users.

Overall, the findings of this study provide valuable insights for entrepreneurs, investors, and policymakers seeking to support the development and implementation of machine learning technology in start-ups. By highlighting the importance of funding, key elements of value creation and capture, and a firm's business model, this study can help stakeholders to better understand the factors that contribute to successful value creation and appropriation in this context.