August 31, 2021

An AI-Driven Approach to Business Model Architectures

Use case
Decision Augmentation


One key conceptualization highlighted by scholars to understand the mechanisms through which entrepreneurs can successfully commercialize new technologies is the business model concept (Chesbrough and Rosenbloom 2002);(Chesbrough 2007). Among entrepreneurs, the business model has been established as a blueprint to identify and assess opportunities for market exploitation (George and Bock 2011), especially in digital contexts (Nelson and Metaxatos 2016). Business models are systems of interrelated elements (Afuah and Tucci 2003; Massa et al. 2017; Afuah 2014; Baden-Fuller and Morgan 2010). Understanding their nature requires to look beyond the isolated elements towards their configuration (Klang et al. 2014; Lindner et al. 2010; Wirtz 2011). In this perspective firms achieve competitive advantage and superior financial performance when there exists “fitness” between these elements (Morris et al. 2005; Afuah and Tucci 2003; Teece 2010; Foss and Saebi 2017, 2018).

However, little is known on the role of configurations of single business model components and their complex interactions in influencing the performance of a firm. Thus, neither a theoretical nor a practical rational for how to make choices in the design of a business models. Although business model literature has arrived at a certain degree of consensus about the generic components, a more-fine grained and domain specific ‘design choice’-level as the primary unit of analysis is required when relating business model concept to firm performance or for building business model tools. Such an analysis is crucial for understanding what configuration of design choices distinguishes successful from non-successful business models and providing decisional guidance to entrepreneurs with the design paradigm of hybrid intelligence. 

The purpose of this research is therefore to provide a fine-grained, domain specific examination of design choices and investigates how configurations of design choices are related to firm performance. By conducting an inductive multiple case study design, the uses scalable machine learning techniques such as clustering and classification methods for data analysis to overcome methodological limitations of existing studies and providing an in-depth analysis of 188 IoT business models along 108 dimensional characteristics. I initiated my research by a taxonomy development (Nickerson et al. 2013) to identify the relevant components of IoT related business models. This initial step comprised the analysis of 188 IoT ventures and extant theories from management research. In a second step I applied a clustering analysis for identifying certain types of configurations. I then used another clustering over those types to identify successful business model archetypes. Finally, I used an ensemble of classification trees to identify pattern that distinguish successful from un-successful business models and the relevance of certain components in doing so.

For the context of this thesis, this offers several contributions that are required to address the conceptual DP of CBMV systems in Section 5.3 as well as to propose DPs and an implementation of the HI-DSS in Section 6.5.First, I provide a fine-grained and domain specific taxonomy of design choices for business models and their components. This allows to use the taxonomy as a cognitive schema to communicate the entrepreneurial opportunity between the entrepreneur and the crowd. Moreover, it provides a formal conceptual representation that can be used as data ontology for providing ML supported guidance.

Second, I identified four archetypes of successful business models design pattern that are context-specific and are identified based on an array of organizational features. This allows me to examine attributes of firm and their effect on performance, thus, providing decisional guidance based on real attributes of a firm.

Finally, I investigated the relevance of entrepreneurial design choices in determining success. Therefore, I can use this as an input to use the identified success pattern for providing decisional guidance through data-driven approaches. 

Business Model Design Configurations

For this thesis, I view business models as a configuration of design choices that entrepreneurs make to create and capture value. 

Prior research on configuration theory conceptualizes entrepreneurial firms from a systemic perspective understanding them as multidimensional constructs of interrelated design choices (Fiss 2011; Fiss et al. 2013). An organizational configuration describes the commonly observable co-occurrence of conceptually distinct attributes that collectively cause a certain outcome (Ketchen and Shook 1996; Meyer et al. 1993). In contrast to the theoretically computationally intractable number of possible design choice configurations these attributes reveal a tendency to occur in real-world organizations as coherent design patterns that are causal dependent (Meyer et al. 1993). Although, configurational view stresses equifinality, i.e., the possibility of a system to reach the same state through different configurations (Katz and Kahn 1978). I argue that such design pattern can be used to guide entrepreneurial decision making. This is in line with recent research that has focused on the relationships between a business model’s elements to explain firm performance and competitive advantage (Klang et al. 2014).


For my research, I apply a multiple case study approach to inductively build a theory on business model design from empirical cases (Yin 2017; Eisenhardt 1989). Building theory from empirical case evidence allows to generate theoretical constructs and midrange theories (Eisenhardt 1989). In my context, the overly phenomenon-driven research question is grounded in the emerging relevance of business models for digital innovation and the lack of plausible existing theory that provide explanatory justification on the design of business models. Little is known about how to configure characteristics of business models to succeed. Based on its replication logic, the theory building process leverages each individual case as distinctive analytic unit that allow to extend nascent theories in a field (Eisenhardt 1989). This process of creating theory is then achieved by recursive loops between the empirical evidence, related literature in the field, and the emerging theory itself (Eisenhardt and Graebner 2007). In contrast to laboratory experiments, examining cases in a real-world context enables to investigate and reason about the phenomenon in its realistic complexity. Moreover, case studies allow to gather rich, empirical descriptions of a phenomenon in a real-world context as well as the integration of various data sources (Yin 2017). This is especially valuable when the phenomenon is emerging, and theoretical rationales are still nascent.  Using multiple cases also ensure the creation of more generalizable and deeper grounding in empirical evidence.

For conducting my multiple case study approach, I followed the principles of the process of building theories from case study research Eisenhardt (1989). Given the limited theory, I relied on a multi-method approach to inductively build theory about organizational configurations of digital business models. To apply scalable principles for data analysis and cross-case pattern identification, I applied several machine learning techniques. Data triangulation, the use of multiple investigators, and the combination of qualitative and quantitative methods enables me to gain deep insights in each of my 188 cases along 108 theoretically specified constructs as well as their relation across cases. Therefore, I conducted the following steps.

First, a priori specification of constructs through taxonomy grounded in related literature (Eisenhardt and Graebner 2007). Although researchers agree on the main components of business models as conceptual models, there exists little consensus about the elements that constitute those components (Massa et al. 2017; Foss and Saebi 2017, 2018). Taxonomies represent an established mechanism to organize knowledge in a field by providing a set of unifying constructs that facilitate its systematic description (Nickerson et al. 2013). These characteristics make them particularly useful to analyse complex domains and hypothesize relationships among concepts. Taxonomies, hence, vitally contribute to theory building, especially in the context of configurational research. Taxonomies organize facts and data into meaningful sets, out of which theories can be developed (Dess et al. 1993).

Second, I gathered in-depth data by conducting a descriptive classification of each case as within case analysis (Yin 2017). For this purpose, three researcher gathered data on each of the 188 independently and classified each business model along 108 characteristics. Therefore, web data, CrunchBase descriptions, news articles etc. were used. The three independent classifications for each business model were then compared and discussed until consensus among all researchers was achieved.

Third, for identifying across case pattern, I applied an unsupervised machine learning approach (i.e. clustering). The clustering of characteristics allowed me to identify configurations of business models and find archetypes.

Finally, I applied an inductive machine learning approach to examine the relationship between configurations and their influence on firm performance. I therefore used a decision tree to discriminate successful and non-successful configurations of business model configurations. This inductive approach allowed me to identify the most important configurations that determine success and distinguish successful from non-successful business models.


I focused on the IoT as domain focus for my study due to its disruptive character among emerging technology domains (Manyika et al. 2015). IoT aims for the consolidation of the digital and the physical sphere by equipping physical things with sensors and communication technology that facilitate the collection and analysis of data on top of which digital services can be created (Yoo et al. 2010). This setting is particularly interesting when studying digital business models as it allows ventures to develop innovative value creation and capture mechanisms that serve as a key source for future competitive advantage (Porter and Heppelmann 2014).

My research primarily focuses on 188 ventures retrieved from the online start-up database CrunchBase ( There, each venture was categorized under the tag ‘Internet of Things’. I further used three criteria to identify potential case firms. First, each of the ventures had to meet the definition developed by the European Research Cluster on the Internet of Things referring to the IoT as a “[…] dynamic global network infrastructure with self-configuring capabilities based on standard and interoperable communication protocols where physical and virtual things have identity, physical attributes and virtual personalities and use intelligent interfaces and are seamlessly integrated into the information network […]”  (Vermesan and Friess 2014a). Second, the firm was active in terms of their business activities. Third, the venture discloses sufficient information to adequately assess their business model design. 

To improve the robustness and generalizability of my results, I aimed for variation in terms of industries, technologies, and market segments across the selected ventures (Eisenhardt and Graebner, 2007). Moreover, I just used ventures that were founded after 2014 as in this phase the business model has the most predictive power for entrepreneurial success.

Data Collection

For each venture I collected evidence of its business model design and financial performance. I relied on two main data sources: company websites and CrunchBase profiles. I triangulated these data sources with complementary data comprising news articles and official social media profiles (LinkedIn, Twitter, Facebook) to improve the robustness of my findings (Yin 2017).

I continued my data collection by interviewing top level managers from two additional ventures. During these interviews I provided them with a preliminary version of my taxonomy and let them explain their business model sub-layer by sub-layer. These interviews (each lasting around 2 hours) were particularly helpful for refining individual manifestations of the taxonomy. 

Taxonomy Development

For the initial step I relied on the taxonomy development method proposed by Nickerson et al (2012) that comprises inductive and deductive elements, and therefore incorporates both empirical and theoretical evidence. 

The first step of the taxonomy development method is to determine a meta-characteristic, that serves as the foundation of subsequent choice of characteristics. I defined the meta-characteristic as the main components of a business model reflected by the four dimensions proposed by Gassmann et al. (2014), Who? What?, How?, Why?. This step limits the odds of “naïve empiricism”, where many characteristics are defined, hoping that a pattern will emerge, and it reflects the expected use of the taxonomy (Nickerson et al. 2013).

The second step comprises the selection of objective and subjective ending conditions to terminate the iterative process facilitating usefulness of the taxonomy. I selected and adapted the following objective and subjective ending-conditions from Nickerson et al. (2013):

Objective conditions: 

  1. all the ventures of the sample are examined; 
  2. at least one object is classified under every characteristic of every dimension; 
  3. no new dimension or characteristics is added in the last iteration; 
  4. every dimension is unique and not repeated; 
  5. every characteristic is unique within its dimension and not repeated.

Subjective conditions: 

  1. conciseness: no unnecessary dimension or characteristic is included; 
  2. robustness: there are enough dimensions and characteristics to differentiate between the various ventures; 
  3. comprehensiveness: all IoT ventures can be classified in the taxonomy; 
  4. extendible: new dimensions and characteristics can be subsequently added; 
  5. explanatory: the information is valuable and contributes to characterizing IoT ventures; 
  6. information availability: the information is typically available or easily interpretable.

A key objective of the objective ending conditions is to generate dimensions of mutually exclusive and collectively exhaustive characteristics. While aiming for a high degree of mutual exclusiveness it became obvious that for certain dimensions characteristics may apply. A company can target several customer segments, simultaneously occupy multiple layers of the IoT ecosystem and use multiple approaches to monetize their products or services. This is not a unique property of the chosen unit of analysis and has been considered in ontology development research by the concept of slot cardinality that defines how many characteristics may maximally exist for any venture in a certain dimension. Multiple cardinality is contrary to the recommendation of Nickerson et al. (2013) but applied to accommodate the taxonomic nature of IoT business model components. Furthermore, to limit the cognitive load of the taxonomy user and make it more intuitive to use, I intentionally focused on a generalizable abstraction level. Hence, I kept the dimensions and characteristics to a reasonable number while preserving taxonomic completeness to achieve comprehension, application and ultimately usefulness of the taxonomy (Nickerson et al. 2013).

Following the definition of a meta-characteristic and the ending conditions, I began to derive dimensions and characteristics for the taxonomy. The process allows for two different approaches in each iteration. First, the empirical-to-conceptual approach creates characteristics and dimensions based on the identification of common characteristics from the sample. Second, the conceptual-to-empirical approach relies on extant theory to devise characteristics and dimensions, before validating these dimensions and characteristics on the sample. In total I conducted three iterations. 

Following the advice by Nickerson and colleagues (2013) I initiated this research step with a conceptual-to-empirical approach as I felt sufficiently knowledgeable about digital business models and the IoT but lacked data. I then examined a subset of 50 ventures to test the appropriateness of each dimension. Although each dimension aligned to the meta-characteristic, I noticed that the dimensions insufficiently described certain important aspects of the business model, indicating a lack of collective exhaustiveness. 

Rather than pursuing one of the two ideal typical approaches the second iteration can be described as a back-and-forth between the study of my sample ventures and theory until I had examined an enlarged subset of 100 firms. More specifically, in line with an empirical-to-conceptual approach I discussed notable business model related differences for the studied cases among the researchers, followed by an investigation of corresponding theoretical concepts. Once I felt that the iterated taxonomy was well suited for my subsample well, I conducted interviews with two ventures for evaluative purposes. The interviews revealed necessary adjustments for the characteristics.

In the third iteration I then applied the empirical-to-conceptual approach comprising a review of the full sample. I terminated the taxonomy development after I noticed that I were able to describe all ventures.

Cluster Analysis

I used cluster analysis as a valuable data analysis technique as it allows the classification of a huge number of cases (i.e. individual business models) along many characteristics (Ketchen and Shook 1996). For the applied case study research methodology, this approach supports inductively examining an open-ended, large-sample examination of cases on a fine-grained level. Clustering is a statistical technique to group a sample by minimizing the variance among cases grouped together while maximizing between-group variance (Ketchen and Shook 1996).

For clustering my data, I choose the characteristics along the conceptually defined subsets of business model dimensions (Ketchen and Shook 1996). As all the features are categorical, I use one hot encoding for categorical features, which results in a dummy variable for each characteristic, where each feature (i.e. characteristic) can take the values 0 or 1. This means that a business model can theoretically consist of 2108 unique configurations of characteristics. Those characteristics were then used as input features for clustering.

I applied the kModes clustering algorithm for categorical data with a Python implementation (Huang et al.; Huang 1997, 1998; Huang and Ng 2003). Other clustering algorithms such as kMeans use distance such as Euclidian distance measures between two objects, which is not possible for categorical data. Moreover, these approaches represent cluster centroids as means, which is not possible for nominal data. Moreover, this statistical approach captures the conceptual similarity between business models that can be defined as the degree of coincidence of elements (Rumble and Mangematin 2015).

kModes clustering can be seen an extension to the standard kMeans by applying a simple matching dissimilarity measure, using modes to represent cluster centres and updating modes with the most frequent values in each iteration (Huang 1998). This approach ensures that the iterative process converges to a local minimum.

The dissimilarity (Hamming distance) between two objects (i.e. business models) X and Y described by m characteristics is defined as d(X, Y)=  = δ (xj,yj) where δ (xj,yj) is    1, xj ≠  yj.

Thereby, xj and yj describe the values of attribute j in X and Y, where a higher number of mismatches of characteristics between X and Y express a higher dissimilarity. The dissimilarity between an object X and a cluster centre Zl  is then calculated as φ( ) = 1-   for      and 1 for   (Ng et al. 2007). 

Here   is the categorical characteristic of attribute j in   ,while    is the count of objects in cluster l and  is the number of objects whose attribute characteristic is r. 

The kModes clustering represents cluster centroids as the vectors of modes (i.e. the most frequent value) of categorical attributes. This means that a data set of m categorical features has a mode vector Z of m categorical values (z1, z2, ..., zm). This mode vector of a cluster then minimizes the sum of between object distances within a distinctive cluster and the cluster centroid (Huang 1998). 

Defining the most appropriate number of clusters is one of the biggest challenges in this approach (Ketchen and Shook, 1996). I combined both a statistical (i.e. Silhouette score) as well conceptual a priori constraints to set the number of clusters (Hair et al. 1992). For this purpose, I used a grid search over the possible search space of hyperparameters (i.e. number of clusters). Therefore, I conceptually constraint the number of possible clusters between 2 and 30. I then calculated the average Silhouette score for each cluster number and choose the local maxima of this value. The Silhouette score is an approach for cluster interpretation and validation of consistency by indicating how well each object lies within its cluster (Rousseeuw 1987). This value indicates the similarity of an object to its individual cluster (cohesion) in relation to all other clusters (separation). The silhouette score ranges from −1 to +1, where values close to 1 indicates that the object is properly matched to the right cluster. When most objects have a high score, the clustering configuration is optimal. Therefore, I calculated the average score across all objects per number of clusters.

A Taxonomy of Business Model Design Choices

The taxonomy builds on the four business model layers developed by Gassmann et al. (2014): What does the firm offer to target customers? Who are the target customers? How does the firm produce their offering? Why does the firm generate profit? The taxonomy is structured as follows:



The “what?” layer depicts the content of the value proposition of an IoT company, defined as “the benefits customers can expect from products and services” (Osterwalder et al. 2014: 6). Anderson et al. (2006: 4) identified three interpretations of value proposition: “all the benefits” for the customers, all the favorable points of differentiation from the competition, and a restricted number of key differentiation points. Accordingly, in the first sublayer of solution, we distinguish three dimensions: benefits (solution type), format (solution form), and differentiation (competitive strategy). The second sublayer is the ecosystem, which has four dimensions: the levels or layers in the IoT reference architecture at which a specific company creates value, its core function, the ownership of the ecosystem, and the combination possibilities with third-party solutions (interoperability).


The solution types are developed in reaction to the latent or explicit need of customers and users to remove hurdles or mitigate risks. Based on a survey conducted by Zebra Technologies for the Strategic Innovation Symposium at Harvard University, we identify four generic types of problems commonly solved by IoT solutions. The four generic problems can be defined as the “benefits and outcomes required, desired, expected or unexpected by the customers” (Osterwalder et al. 2014: 8).

The solutions may be offered under the form of goods, services, or a combination of both. Turber et al. (2014) describe these formats as the ‘carriers of competences.’ We adopt the set of distinguishing features to classify goods and services in the taxonomy.

Competitive strategies aim at creating a unique value proposition which differentiates a provider’s solutions from those of their competitors. Adapting Porter’s generic strategies (1980) – i.e. overall low costs, differentiation, and focus with empirical observations – we identify seven generic competitive strategies. Individual ventures may combine various competitive strategies, so they are not mutually exclusive.


The IoT combines multiple technologies in a complete stack that is essential to value creation (Püschel et al. 2016). In the taxonomy, we distinguish between seven IoT layers to situate each venture within the larger IoT ecosystem and specify the nature of the offering. An IoT company may be active across multiple layers. Therefore, the characteristics of this dimension are not mutually exclusive.

Based on empirical findings and theories developed by Porter and Heppelmann (2014), we define five core functions of IoT products. Each of the first four functions builds on the previous one; for example, the controlling function requires a monitoring function (Porter and Heppelmann 2014).

Interoperability, defined as the ability of solutions from different vendors to communicate and integrate with each other seamlessly, is a key driver of value in IoT. McKinsey estimates that interoperability could be responsible for 40% of the potential value created by the IoT (Manyika et al. 2015). The possibility for devices to communicate and operate together depends largely on the use of widely-accepted interface standards or translation schemes between operating systems and applications (Manyika et al. 2015). However, the competitive pressures and the constant evolution of technologies tend to impede the integration of devices in a homogeneous framework. In the taxonomy, we distinguish between three mutually-exclusive types of ecosystems: open, limited openness, and closed. 

Finally, the ecosystem ownership dimension differentiates companies developing their own ecosystem from companies leveraging existing ecosystems from third parties. Companies promoting their own ecosystems integrate the different layers of their offering under one proprietary umbrella. Alternatively, firms may restrict their solution to one or more layers, while collaborating with existing platforms or solutions from third parties. 


The business model layer “who?” defines the stakeholders for which value is created and the channels through which they are being reached (Gassmann et al. 2014: 55). This layer is split into two sub-layers: the market sub-layer describing the specific categories of targeted consumers, and the relations sub-layer which specifies the channels and intensity with which they are addressed.


The nature of the IoT broadens the traditional frontiers of market sectors. Manyika et al. (2015) argue that focusing on industry verticals provides a limited perspective on the value created by IoT because it fails to describe the interaction between systems which do not belong to the same market sectors. Therefore, they adopt the complementary lenses of the setting, defined as the “context of the physical environment in which systems can be deployed” (Manyika et al. 2015: 18). Specifically, they define nine IoT settings: human, home, retail, offices, factory, worksite, vehicle, city, and outside. Combining the industry verticals with settings provides for a very large number of potential combinations, especially since neither the market sectors nor the settings are mutually exclusive. We identify three dimensions of market: application ecosystems, customer, and market segment.

Based on the observations mentioned in the previous paragraph, we identify three generic application ecosystems: smart environments, smart industry, and smart health & well-being. They represent the three archetypical combinations of industry verticals and settings and are based on the domains proposed by Borgia (2014).

The customer dimension describes the type of customer towards whom the transaction is directed. Business-to-consumer (B2C) concerns companies that sell goods and services to the end customers. In business-to-business (B2B) markets, firms sell their offerings to other firms. In business-to-government (B2G), companies sell their products to governments and other public organizations in public markets.

The market segment dimension describes the size of the targeted customer segment. In the taxonomy, we adopt the five types of customer segments suggested by Osterwalder and Pigneur (2010).

Customer Relation

Customer relation includes two dimensions: interaction intensity and retention. Interaction intensity describes the magnitude of the interaction between a specific firm and their customers.

Recruiting new customers generally costs more to companies than selling to existing customers. The dimension of retention describes how the ventures retain their customers and increase the likelihood of repurchase. Based on customer retention and switching costs theories, we identify five general retention strategies.


In the How business model layer, we differentiate the value-creation mechanisms by separate resources, partners, and activities sub-layers.


We adopt the definition of a firm as “all assets, capabilities, organizational processes, firm attributes, information, knowledge etc. controlled by a firm that enable the firm to conceive of and implement strategies that improve its efficiency and effectiveness” (Barney 1991: 101).

The resource-based view posits that the internal resources of a firm are better contributors to competitive advantage than external factors (Barney 1991), and groups them in three generic categories: human resources, physical resources, and organizational resources. In most instances, companies rely on multiple resources. However, the aim of the distinction we propose is to identify which of these resources is the most critical in the value creation process.

The various technologies commonly used across the different layers of the IoT architecture can be complemented with other technologies which are not systematically related with the IoT. We call this blend of technology the technology combination and identify five technologies which are regularly incorporated in the IoT: Blockchain, Artificial intelligence, Robotics, and 3D printing.

Data has become a key asset for companies in the digital age (Bharadwaj et el. 2013). Therefore, it represents an important component of the business model. In the taxonomy, we extend the dimensions of data source and data usage identified by Püschel et al. (2016) to the complete IoT stack

The second dimension relating to data is the data usage; it describes how data is used. We adopt two characteristics (Püschel et al. 2016): transactional and analytical.


The creation and capture of value involves an increasingly large number of interactions between firms. Consequently, there is an incentive for firms up and down the value chain to align their interests and enter into explicit or implicit partnerships for value creation (Chesbrough and Schwartz 2007). Specifically, Dyer et al. (2018) have identified four key determinants for inter-company value creation: complementary resources and capabilities, relation-specific assets, knowledge sharing routines, and effective governance. In the taxonomy, the factor of resource and capability complementarity is reflected by identifying the key partners for the creation of final goods and services, the investors who contribute with the financial resources, and the institutional supports. However, since the taxonomy is developed exclusively with publicly available information, it does not contain any dimension about the assets and routines involved in the partnership, nor about the governance (i.e. the mechanisms to balance the power relationship) and nature (e.g. joint venture, strategic alliance) of the partnership.

To categorize the essential partners of a given firm, we distinguish between the different IoT layers at which partners supply hardware and software components to be embedded in IoT solutions. We also distinguish between channel partners and co-creation partners.

Given the limited access to resources and knowledge, the success of new ventures often relies on various institutional support. Based on our sample, we identified six types of support: incubators, accelerators, corporate programs, universities, boards of directors, and boards of advisors

In addition to these institutional supporters, new ventures nurture a close relationship with investors. Based on the classification from the startup database Crunchbase and on empirical evidences from the sample, we identified five main types of investors in IoT companies.


Key activities are the most important actions undertaken by an organization to create and distribute value (Osterwalder  and Pigneur 2010). The operational focus describes key activities performed by an IoT company. Porter differentiates between primary activities relating directly to the creation of the offering, and secondary activities which consist of all activities to support the primary activities (Porter 1980). The primary activities comprise inbound and outbound logistics, the operations transforming input into outputs, marketing and sales, and services. Secondary activities are listed as procurement, human resource management, research and development, and company infrastructure (a company’s support systems). As discussed above, the nature of value creation in the IoT scatters activities traditionally concentrated in one firm to an ecosystem of partner firms. By combining empirical observations with the distinctions proposed by Porter (1980), we retained four operational foci for IoT companies: operations, marketing & sales, services, and R&D.

Furthermore, all firms promote their solution with customer acquisition activities. They may use online and offline channels to raise product awareness, share technical and usage information, and increase willingness to buy. We propose five types of customer acquisition channels for the taxonomy.

In addition to their offering per-se, companies also compete on the customer service they provide. We distinguish between five types of customer services that depict commonly observable manifestations.

In addition to activities related to the creation of the core offering, companies need to deliver the created value through distribution, or marketing, channels. Osterwalder and Pigneur (2010) distinguish between types of distribution channels based on two characteristics: whether the company owns and manages its own channels or relies on partners, and whether there is a direct interaction between the company and the customers or not. From our sample, we identified four categories of distribution channels.


The “Why?” business model layer refers to the ability of the revenue streams and the cost structure to enable the commercial viability of a business model (Gassmann et al. 2014).


The revenue model describes the type of revenue that is generated by the selling of a solution. Osterwalder and Pigneur (2010) define two types of revenue models: transactional revenues, which involve a unique payment by the buyers, and recurring revenues, which generate multiple incoming payments over the lifecycle of the offering. We identified five IoT revenue models.

The dimension of pricing strategy describes how a price is presented to (potential) customers. We identified three types of mechanisms used by IoT companies to set prices.


Two direct components drive profitability: the contribution margin and the volume of units sold. Accordingly, the profit is either driven by large sales volume or by selling the product at a high margin. Additionally, for many services or “as-a-service”-solutions, the profit is dependent of the number of transactions or use. 

Design Pattern of Successful Business Models

For identifying and describing successful business model design pattern, I split the total data set in a subset of successful business models each characterized along its membership along the 9 dimensions. This approach resulted in a set of 31 ventures. I then used another kModes clustering to identify common pattern among those business models. I searched the possible space of design configurations between 2 and 20 clusters and identified 4 configurations as being most appropriate (Silhouette score = 0.053) (Rousseeuw  1987). Those four design patterns of successful business models constitute configurations of the nine business model dimensions identified in the previous section.

Design Pattern I

Design Pattern I creates value by offering Control solutions, which allow to influence the actions and behaviours of objects and persons. As an example, the start-up Avi-on provides clients a Control tool to manage and control lighting infrastructure. They describe their lightning control system as “[…] purpose-built from the ground up for commercial markets. […] [we] provide an easy to install, easy change, easy to maintain lighting experience”.

Design Pattern I solutions can take the form of both goods and services, a characteristic that is not unusual for IoT companies. For instance, the lightning control solutions of Avi-on comprise hardware and software elements. Another example is The Yield, a company offering smart agricultural services, which sells both the sensors and the software to measure and predict soil conditions. They describe their solutions as follows: “[…] [The Yield] isn’t measuring from a weather station 10, 20 or 100 kilometres down the road and leaving you to calculate the difference. It’s recording growing conditions under your feet and converting data into on-farm predictions that can help you plan activities with confidence […]”.

Design Pattern I companies complement offered solutions with a competitive strategy that relies on turnkey characteristics and superior performance. Turnkey characteristics refer to solutions that are set up instantaneously and intuitively. Avi-on emphasis the easy-to-use, simplistic nature of its products and bases its competitive strategy on the brand message of “[…] Easy to Install. Easy to Use. Easy to Change […]”. Superior performance on the other hand is a cornerstone of Cognosos’ differentiation strategy. The asset tracker offers a “[…] smarter, more productive, way to keep up with all your equipment […]”. Cognosos’ solutions “[…] improve utilization, reduce labour costs, and decrease turnaround time […]”. To achieve increased efficiency, Cognosos follows Avi-on’s and The Yield’s example and provides customers with both the hardware and software to track assets and manage inventories.

Because of offering both goods and services, Design Pattern I companies operate on multiple layers of the IoT. As mentioned before, this is a common trait throughout all the observed design patterns. Design Pattern I core functions on these layers focus on monitoring, controlling, and optimizing processes and events. 

In the case of Cognosos, expanding solutions onto the Device layer, is essential to realize cost efficiency, as previously mentioned. On the software-based layers that Cognosos is active on, gathered data is stored, processed, and eventually converted into actionable insights. Processed data is controlled and optimized in Cognosos’ proprietary RadioCloud ecosystem.

While the core functions describe the purpose of an IoT product or service, the operational focus defines the types of activities performed by a company. For Design Pattern I companies the operational focus is set on Research and Development (R&D) and Services. R&D is a key activity to companies such as Zoox, which is currently in the testing phase of an autonomous vehicle solution. Services on the other hand, refer to activities which contribute to the value of the provided solution. As an integral part of their offering, The Yield provides installation and maintenance services for its smart agriculture solutions.

To be able to carry out core functions and eventually deliver a solution, companies are dependent on different types of resources. Physical resources play an important role for Design Pattern I companies. This type of resources is needed to create devices like Avi-on’s lighting control tools, The Yield’s sensors and Zoox’s car chassis. 

A resource all design patterns rely on is data. However, the sources of data differ throughout the four design patterns. In case of Design Pattern, I data is gathered based on the state, context, and usage of a company’s solution. This means that data gets originates from a company’s solution itself. 

Design Pattern I targets B2B clients from Smart Industry domains. Companies are addressing a diversified portfolio of clients. This focus on Smart Industry is unique to Design Pattern I. An example of this Design Pattern I configuration is Cloudleaf, which offers supply chain visibility solutions that are employed throughout multiple contexts and industries. In the words of CEO Mahesh Veerina: “[…] Today, I solve many of my in-door problems very well. In-door for me would be something like a large warehouse, distribution centre, a factory, or a warehouse. You are tracking your raw materials, supplies, and tools, so whatever you are tracking you want to have visibility on them. I offer that today […].” 

Concerning client interactions, Design Pattern I companies are distinguishable by a loose interaction intensity. This can be observed in the case of N.thing, a company which provides smart farming equipment online. N.thing’s customer interactions are limited to interactions through a conventional online store. In the aspect of interaction intensity, Design Pattern I is identical to Design Pattern III. However, in contrast to Design Pattern III, Design Pattern I solely utilizes financial switching costs as a customer retention mechanism. This mechanism is effective in cases like N.thing’s Planty Cube, a 7.5-ton container farm that enables year-long urban farming. The high acquisition costs and the option to stack modules, deters customers from switching provider.

However, even though physical resources are of high relevance for Design Pattern I companies, production and availability of hardware is dependent on Component partners. They play a major role for companies such as Flytrex, a drone delivery service that engages in partnerships with drone manufacturers such as DJI. The partnerships allow Flytrex to focus on its core service – autonomous drone delivery – while having access to equipment of industry leading drone manufacturers. In addition to Component partners, Network and Security partners are essential for Design Pattern I. They ensure that companies can connect their products and services while providing security for all users. Besides having partners on different levels of the value chain, companies rely on Investors, to secure sufficient funding for their business. Business Angels and Venture Capitalists are the investor categories that predominately invest in Design Pattern I companies. In the case of Flytrex, investments from angel investor Joey Low and venture capital firms Armada Investment AG and VI Partners AG resulted in a total funding amount of USD 3,000,000. 

Design Pattern I companies utilize Subscriptions as primary revenue model. Connecterra, a provider of dairy monitoring systems, charges monthly fees per cow equipped with a sensor. Flytrex prices the access to its drone Control Center with monthly rates. The monthly rates are feature dependent, which allows an improved targeting of different customer needs. Connecterra provides “Standard”, “Pro” and “Flex” subscriptions, which differ in monthly fees, start-up fees and additional features. Flytrex utilizes monthly fees that dependent on the type of drone you want to access via the Control Center.

Design Pattern II

In terms of offered solution, Design Pattern II exhibits multiple similarities to Design Pattern I and Design Pattern III. Remarkedly, all three design patterns provide control tools, which take the forms of both goods and services. Smappee exemplifies a control tool offering goods and services in case of Design Pattern II. The company does not just monitor household energy consumption, but also serves as a hub to manage energy, as it is described on Smappee’s homepage: “[…] you can use Smappee as a smart energy traffic controller, which decides when to use which energy source and to which appliance it should be assigned. As a result, you keep an overview of all the energy flows in the house, so you can save money more easily […].” So, like Avi-on in Design Pattern I, Smappee provides a tool to influence the actions of objects.

However, not only type and form of solutions in Design Pattern I, Design Pattern II and Design Pattern III am identical, all three design patterns utilize identical competitive strategies. Focus on superior performance and turnkey characteristics are the dominant strategies for the three configurations.

Design Pattern II follows Design Pattern I’s configuration regarding the core functions its solutions incorporate and the operational focus it sets, as both design patterns concentrate on R&D and Services. However, the key resources that are utilized to carry out core functions differ. Design Pattern II’s solutions build on human and organizational resources instead of physical ones. Human resources are of major value for companies like DSP Concepts, a company producing high-tech audio solutions for a variety of applications. For this reason, DSP Concepts’ “[…] engineering team includes some of the world’s top talent in numerous audio specialties, including microphone and speaker processing, automotive sound, and telematics systems, IoT applications, wireless technologies and more […]” (DSP Concepts, Services). Organizational resources, which comprise elements such as IT-infrastructure, organization structure and intellectual properties, are essential for companies such as Ushr. The Detroit-based company develops mapping systems to facilitate accurate driving of autonomous vehicles. Ushr’s main assets are the software and technologies that are utilized during the mapping process.

Clients of Design Pattern II companies are businesses from diversified settings. The environment Design Pattern II’s solutions are applied in however, differ from Design Pattern I’s application environment. In Design Pattern II, Smart Environment applications are now offered, such as the solution of Smappee, which ensures that “[…] you always have a clear overview of the energy flows in the house or company premises, wherever you are […]” (Smappee, My Technology). Another example is Connected Signals, a company which presents its solution as follows “[…] I eliminate the complexities of securely gathering real-time signal data and making it readily available in a standard format. I combine this data with map, GPS, and speed limit information, and then apply proprietary analytics and algorithms to predict upcoming traffic light behavior. That information is then delivered to vehicles via cellular networks […].” The traffic light predictions of Connected Signals are aimed at automobile manufacturers, municipalities, and navigation companies.

Design Pattern II closely interacts with customers. Complex, customized products like DSP Concept’s audio solutions, require close collaborations between the company providing the solution and its clients. DSP Concepts consults clients throughout planning and implementation phases of new audio technology. They promote their consulting and training services as follows: “I can explore your goals and constraints to help you decide the number of microphones, microphone topology, size and placement of speakers, and which processor to use. I will even work with you to help develop a proof of concept. 

In terms of partners, Design Pattern II exclusively collaborates with Component partners and turns away from Design Pattern I’s multi-partner approach. Corporate Venture Capital (CVC) entities are the main financial investors in Design Pattern II companies. DSP Concepts attracted investments from BMW’s CVC wing, who are interested in DSP Concepts’ audio solutions for automotive

For Design Pattern II companies, one-time sales are the preferred revenue model, instead of Design Pattern I’s subscription focus. The design pattern pursues a feature dependent pricing strategy. Smappee’s energy monitoring and controlling devices are an example for this configuration. A variety of monitors and smart plugs can be bought. Different monitors are available, depending on whether solar energy is measured additionally or not.

Design Pattern III

Design Pattern III’s solution and ecosystem characteristics are in my aspects identical to Design Pattern I and Design Pattern II. Besides providing their clients with control tools, Design Pattern III also emphasizes superior performance and turnkey characteristics as basis of its competitive strategies. 

While solution properties are very similar to other design patterns, Design Pattern III differs in core functions. These are restricted to monitoring and controlling, the aspect of optimizing, that was present with Design Pattern I and Design Pattern II, is absent. Silvair provides a perfect example for this. The lighting control system provides a platform to set up and control lighting systems and provides the firmware to set up tailored control systems. Organizational resources, such as the Bluetooth mesh-based lighting ecosystem, build the basis for Silvair’s offering. Together with human resources, organizational resources build the most important building block for Design Pattern III services. 

Along diverging core functions, Design Pattern III also exhibits a novel operational focus. Marketing and Operations are the activities with the highest relevance for Design Pattern III, a shift from R&D and Services that were the key operational focuses so far. The company Meural sets its operational focus on Marketing.  Meural creates a smart frame to seamlessly display art works and photographs. Meural relies on a multitude of social media channels, like Facebook, Instagram, Twitter, and Pinterest to promote its B2C-product. Design Pattern III companies rely on external data instead of data created by the solutions themselves. In case of Meural, external data in the form of pictures can be uploaded to the frame.

Meural remains representative of Design Pattern III companies, also in the field of client interaction intensity. Same as Atmoph, a company selling a smart frame acting as a digital window, Meural does not focus on intense customer interactions. On the contrary, both companies distribute their products through their proprietary online store what results in limited customer interaction. Their focus on private clients however, is an exception for Design Pattern III companies, as representatives of this design pattern primarily target B2B clients.

Component partners remain relevant for Design Pattern III affiliated companies. Meural, Atmoph and the baby monitor Cocoon Cam all produce devices that rely on physical components produced by partners. Besides component partners, network and security partners gain relevance for companies again, like Design Pattern I. CVCs represent the main investors in Design Pattern III companies. 

Feature based subscriptions are the revenue model of choice for Design Pattern III companies. 

Design Pattern IV

Design Pattern IV introduces an entirely novel type of solution. Instead of providing Control tools, Design Pattern IV offers Execution and Improvement tools. Execution tools enable the performance of an activity or the creation of a good or service. Seebo for example is an industrial IoT platform that empowers customers to “[…] create digital prototypes of […] industrial IoT use cases, then bring them to life with code-free tools for data connectivity, predictive analytics, automated root-cause analysis, and digital twin visualization […]” (Seebo, Platform Overview) Seebo’s platform acts as an enabler for users to carry out this multitude of functions. scriptr acts as another example for a platform as an execution tool. Like Seebo, scriptr provides developers the tools to build IoT applications.

Improvement tools on the other hand, increase the value of products and services. Passkit improves the process of building a wide array of mobile applications, by offering an easy to use online platform. They promote their service as follows: “[…] With expertise in mobile wallet, beacons, chatbots, blockchain, and CRM/POS systems, PassKit ensures that businesses of all sizes are able to access the latest innovations in technology through the PassKit platform, making it even easier for brands to have real engagement with customers that build real loyalty […]”

Design Pattern IV utilizes a competitive strategy that differs from the characteristics of other design patterns. Design Pattern IV competes based on turnkey and integration characteristics. While It is the purpose of integration characteristics to allow the user to easily integrate other solutions to his Design Pattern IV tool. This seamless incorporation is an essential feature of both Seebo’s and scriptr’s platforms. Thanks to integration features, Seebo’s customers can “[…]  cut sourcing time in half and reduce risk by working with pre-screened partners, ready and available for collaboration within the Seebo IoT Marketplace […]” This integration characteristic increases the value of execution tools, as it simplifies the connection to other devices.

Design Pattern IV engages in close interactions with its B2B customers. Solutions like Sensorberg require collaboration between clients and solution provider. Sensorberg equips buildings with connected appliances, which essentially enable the digital controlling of building processes and creates “smart spaces”. Close interaction between Sensorberg and real estate owners during the planning phase, results in clients being able to “[…]  ditch the extra keys or access cards and start using your smartphone to interact with digitally responsive environments: open doors, book meeting rooms and organize meetings with your colleagues […]” (Sensorberg, smart workspace).

Design Pattern IV concentrates on one type of partnership – component partners – and receives financial support from CVCs. Moreover, Design Pattern IV charges subscription fees for its solutions. Differences in fees are based on features of offered products.  

Design Choice Relevance for Success

Finally, for analysing the large number of cases and identify pattern within them I used another machine learning technique. Therefore, I use a classification tree. Tree-based machine learning approaches fit a relatively simple model on partitions of the feature space divided by a set of rectangles. Although, this approach is conceptually quite simple, it is very powerful in terms of both performance and interpretability of the model and its results (James et al. 2013). 

The aim of this approach is to use the cluster membership for each firm along the sub-dimensions of my business model taxonomy (i.e. solution, ecosystem, market, customer relation, activities, resources, partners, revenue, and cost) to define a model that discriminates between successful and non-successful business models. In other words, this means predicting if a business model is successful or not based on its configuration of dimensions and characteristics.  For this purpose, I used Random Forests with 1000 estimators (Breiman 2001). 

To draw conclusions from this approach, I used the measure of feature importance. The higher, the more important is the feature. The importance of a feature, also known as Gini importance, is computed as the (normalized) total reduction of the feature space caused by that feature (Breiman 2001). For my context, this measure indicates how important a configuration is to predict if a business model is successful.

Configurations within the whole data set of business models and their relationship to determine firm performance. For this purpose, I use binary dummy variable where 0 indicates that the start-up did not receive Series A funding, while 1 indicates the presence of Series A funding (i.e. success). Using Series A funding is a common success proxy for start-up firms and frequently applied in management research (e.g. Baum and Siverman 2004). As I focused on firms that are not older than four years and controlled for rivalry explanations for achieving Series A funding, I assume that the business model configuration can be used as explanatory factor for firm performance. The successful and non-successful business models did not indicate significant differences in firm age, technology trend, and location.

I used the configuration of business model design choices of each start-up along its components as input features for the decision trees. This means that each business model is characterized by its membership along the nine dimensions. Thus, each business model is a configuration of nine clusters (i.e. configuration of single design choices). For instance, the business model of the venture Accerion ( is characterized through having the solution configuration 3, ecosystem configuration 1, market configuration 5, customer relation configuration 4, resource configuration 2, partner configuration 2, activities configuration 2, revenues configuration 2, and cost configuration 4. The random forest consists of an ensemble of decision trees each of which is an inductive discriminative model that divides the feature space between successful and non-successful business models.
These results can be interpreted as the most important component configuration of design choices to separate successful from non-successful business models. The result show that the component configuration of customer relation design choices and the revenue model are the most important features to define successful business models. Those features are followed by the market addressed by the business model and the solution provided to the customer. Consequently, the four customer centric component configurations of design choices are most relevant in separating successful from non-successful business models in the IoT industry.

Discussion and Conclusion

Within this paper, I inductively explored what configuration of components constitute successful archetypes of business models, and what component configurations are most important in defining business model success. To the best of my knowledge this study is the first research that aims at examining a theoretical rational for designing business models on a large scale. Thereby, my study contributes to different streams of business model research in strategic management (Massa et al. 2017). 

For this thesis, this examination provides several contributions. First, it develops a context specific exploration of design choices (i.e. a taxonomy) that allows the application in real-world decisional guidance for entrepreneurial decision making. Therefore, it constitutes a cognitive schema that I use in Section 6.5 to create a shared understanding between the entrepreneur and the crowd as well as formal conceptual representation that allows to translate this cognitive schema in a data model to also create a common understanding between humans and machines for the HI-DSS.

Second, the identified design patterns allow me to examine decision model design choices as attributes of real firms, thus, providing ML supported analytical guidance for entrepreneurial decision making.

Finally, the investigated feature relevance of design choices in predicting entrepreneurial success highlight the relevance of stakeholder interaction such as the market and customer in business model design, as those are the most important factors in discriminating between successful and non-successful entrepreneurial ventures. Therefore, the conceptual arguments for including not only supply-side knowledge of experts but also demand-side knowledge of users in providing decisional guidance for entrepreneurs is empirically validated. 

See references

Original version