Browsing by Division "IWI - Institute of Information Systems and Digital Business"

Process mining research has made tremendous progress in analyzing, visualizing, and predicting the perfor- mance of business processes through computational techniques. However, little attention has been brought to understanding why and how business processes behave as they do. Process mining results alone are not sufficient to arrive at meaningful interpretations about the dynamics and changes of a given business process. Rather, we need to account for contextual factors that underlie and explain the behavior of processes. In this paper, we make two central contributions. First, we develop a framework that depicts relevant factors to make sense of process mining results. The framework is intended to help researchers and practitioners explain why and how processes change across a variety of contexts. Second, we demonstrate the application of our framework within a real-world case: a customer onboarding process in a European financial institution.

Online communities are virtual communities where users exchange knowledge, organize tasks, and accomplish work. We focus on how individual users influence the ways others behave. We draw from mimetic theory and leverage computationally intensive theorizing to examine the influence of popular developers on other developers in GitHub, the largest and most popular open source software development community. Analyzing a subsample of 324 projects, we find that the behavior of rockstars –i.e. exceptionally popular developers– is imitated by other developers, and thus strongly influences overall work patterns in projects. We further find that this effect is stronger when a rockstar is more active in a project. Our findings offer important contributions for research on online communities, specifically by shedding light on the significant role that individual actors can have in such communities. Crucially, to our knowledge, our study offers the first empirical evidence that online communities actually change user behaviors through memetic processes.

In this study, we explore the potential of AI-generated podcasts as an educational tool in the evolving landscape of learning media. Podcasts have grown increasingly relevant in education due to their accessibility and ability to integrate learning into everyday life. With the advent of generative artificial Intelligence (AI), there is a unique opportunity for scalable and adaptable creation of learning media. However, with novel technology, there also come new challenges. Thus, we developed fine-tuned AI-generated podcasts using Google NotebookLM, our course materials, and a custom prompt. We conducted a one-month explorative evaluation in the field using a qualitative diary study. Our study reveals that students find the podcasts beneficial for flexible everyday learning but also point toward challenges like a lack of emotional engagement and technical non-English language issues. In sum, our study highlights the current benefits and challenges of AI-generated podcasts and presents an agenda for future research.

Technological advancements and evolving value orientations reshape future value creation and pose new requirements for service innovation. While a variety of disciplines are developing new approaches to drive service innovation, this is primarily done in isolation and generates only fragmented solutions. Sociological theory has proposed “boundary objects” as an effective umbrella for communication and cooperation among communities. Therefore, we introduce continuous value shaping (CVS) as a boundary object describing service innovation approaches along five principles. We reflect on this concept through the different disciplinary lenses of researchers in service marketing, information systems, service engineering, sociology of work, and innovation management. These perspectives highlight how the CVS principles already connect to discourses within the individual disciplines. However, the CVS concept will not only provide an umbrella to embrace existing activities in different academic disciplines. It also assists to identify research themes that will benefit from uniting the power of these disciplines, and it can serve as an integrating framework to conceptualize complex service innovation approaches. Thus, the CVS concept should guide both researchers and practitioners to develop and implement novel innovation and transformation efforts—in and across organizations.

Effective presentation slide creation is crucial for impactful communication, yet fully automating this task with AI is insufficient. Hybrid human-AI solutions often perform worse than pure AI or human creation due to overreliance on AI. To address this, we develop design principles for configuring human-AI hybrid systems in complex knowledge tasks using a design science research approach. Our prototype, NarrativeNet Weaver, leverages an underutilized corpus of existing presentation slides, applying generative AI advances in hybrid dense embedding and graph-based retrieval techniques. Evaluated through 15 think-aloud sessions and 73 user trials, users with NarrativeNet Weaver exhibit greater engagement and achieve equal or improved slide quality compared to those using a ChatGPT-based chatbot with a vector database. We contribute design knowledge for human-AI systems for complex multimodal content and offer a new approach to retrieving and visualizing existing slides, enhancing the utilization of valuable but underused resources.

“Digital futures” as a research field that examines diverse, long-term future(s) scenarios influenced by digital technologies has been proposed in information systems. Here, based on the emerging literature on digital futures, we define what this term means, delineate it from related concepts such as digital transformation, articulate why the information systems field should take note and consider the study of digital futures, and provide an overview of approaches.

Die Handlungsbroschüre beschreibt die Studienergebnisse des ZEVEDI Ad hoc- Vorhabens „Digitale Bildung in hessischen Schulen: Distribution von Forschungsergebnissen als Digitale Lernmaterialien“. Das Ziel der Studie bestand in der Analyse, wie die Distribution von neuartigem Wissen zu Digitalisierungsthemen als digitale Unterrichtsmaterialien in Schulen gefördert werden kann.

Unlearning outdated knowledge is crucial for organizations to adapt to environmental needs and foster innovation. While unlearning is an auspicious approach, little is known about how to design supporting tools. This paper explores how organizations can effectively implement unlearning interventions and presents an overview of design options for unlearning support systems. By combining deductive and inductive reasoning, we crafted an artifact in the form of a taxonomy. The taxonomy synthesizes insights from diverse research domains, takes a socio-technical stance, and thereby aims to bridge the theory-practice gap in implementing unlearning. To investigate the taxonomy’s applicability, we conducted evaluation sessions with experts and employed it to describe three illustrative unlearning cases. With our work, we seek to unleash the potential of unlearning by guiding the design of supporting digital tools and aligning existing insights beyond disciplinary boundaries.

Knowledge co-creation between enterprise-dominated online innovation communities and external individuals is a pivotal path for enterprises to improve innovation performance. However, little is known from existing research about how initiator characteristics, content characteristics, and community dimension affect co-creation performance in different co-creation settings. Therefore, based on 150 co-creation proposal samples from the Xiaomi online community, we use configurational and contingency theory to construct a theoretical framework and reveal configuration paths for shaping desired co-creation performance. The findings indicate that the multidimensional influences of knowledge co-creation can be combined into nine configurations to yield high or low co-creation performance in different co-creation settings. Across all solutions, ability and community responsiveness are the most common core conditions, and their complex interactions with other elements jointly affecting co-creation performance. Furthermore, these configurations are context-dependent, namely, the co-creation settings exert a moderating role between the multidimensional influences of knowledge co-creation and co-creation performance.

Human-AI delegation occurs when a human delegates work to an autonomous AI-based system. We report on a simulation study to examine how human-AI delegation dynamically changes in an organizational routine as it is enacted over and over again. We build on findings from previous research and examine the interaction of various human and AI-related factors. We compute the resulting dynamics in terms of complexity representing the degree of uncertainty as to whether delegation takes place. We find that online and offline learning capabilities interact with human willingness in various ways which leads to different, even non-linear changes in the dynamics of human-AI delegation over time. Our study yields implications for research on human-AI delegation, routine dynamics and business process management. We point to a number of practical implications and avenues for future research.

Information technologies are expected to improve organizational routines in a number of ways, yet their implementations often lead to unexpected, unintended, and even undesired effects. In this research, we investigate how IT-based changes affect the complexity of an organizational routine over time, that is, the number of ways through which the routine can be performed. We present the findings of a computationally-intensive research study of a customer onboarding routine at a financial institution in Central Europe. To this end, we investigate how IT-based changes in the associated low-code platform affect the dynamics of how the routine is performed over the course of 2 years. We explain the effects of IT-based change on the routine’s complexity along four core dimensions—the type of change, the strength of the effect, and the temporal unfolding of the effect and the permanence of effect—, where each dimension is characterized by different change patterns. We further distinguish between two types of IT-based effects: intended and unintended effects.

LLM-based agents improve upon standalone LLMs, which are optimized for immediate intent-satisfaction, by allowing the pursuit of more extended objectives, such as helping users over the long term. To do so, LLM-based agents need to reason before responding. For complex tasks like personalized coaching, this reasoning can be informed by adding relevant information at key moments, shifting it in the desired direction. However, the pursuit of objectives beyond interaction quality may compromise this very quality. Moreover, as the depth and informativeness of reasoning increase, so do the number of tokens required, leading to higher latency and cost. This study investigates how an LLM-based coaching agent can adjust its reasoning depth using a discrepancy mechanism that signals how much reasoning effort to allocate based on how well the objective is being met. Our discrepancy-based mechanism constrains reasoning to better align with alternative objectives, reducing cost roughly tenfold while minimally impacting interaction quality.

Process mining holds substantial potential to discover and optimize processes utilizing event log data. However, current applications primarily rely on (semi-)structured data from process-aware information systems, limiting their capacity to incorporate multimodal data from diverse sources, particularly in domains like IT service management (ITSM). While existing stand-alone approaches can extract event log data from unstructured sources such as videos, documents, or bot logs, they fall short of leveraging the full range of real-world data available in ITSM. To address this gap, our research focuses on developing a reference architecture for constructing event logs from multimodal data. This architecture integrates diverse data types, construction functions, and process mining use cases. Following a design science research methodology, we aim to evaluate the architecture through a software artifact leveraging real-world ITSM data and incorporating state-of-the-art generative AI. In this study, we present the preliminary reference architecture and share early insights from expert evaluations.

The adoption of generative artificial intelligence (GenAI) and large language models (LLMs) in society and business is growing rapidly. While these systems often generate convincing and coherent responses, they risk producing incorrect or non-factual information, known as confabulations or hallucinations. Consequently, users must critically assess the reliability of these outputs when interacting with LLM-based agents. Although advancements such as retrieval-augmented generation (RAG) have improved the technical performance of these systems, there is a lack of empirical models that explain how humans detect confabulations. Building on the explainable AI (XAI) literature, we examine the role of reasoning-based explanations in helping users identify confabulations in LLM systems. An online experiment (n = 97) reveals that analogical and factual explanations improve detection accuracy but require more time and cognitive effort than the no explanation baseline.

The emergence of generative AI (GenAI) has transformed work by enabling humans to delegate tasks like writing and coding to GenAI agents such as ChatGPT. While existing studies highlight AI capability awareness and perceived competence as drivers of delegation, they overlook parallels between human-AI and human-human delegation. Our ongoing research proposes that human-AI delegation can be understood through a leadership lens, with leadership experience and traits as key predictors. Hence, we investigate whether individuals with leadership experience demonstrate higher delegation levels than those without such experience. In an initial online experiment (n=48), participants were grouped by leadership experience and AI transparency to decide whether to delegate or personally perform image classification tasks. Preliminary findings indicate that under a low-transparency condition, leadership experience results in higher delegation rates. However, leadership alone does not significantly predict delegation. Transparency in GenAI consistently leads to higher delegation, while greater domain knowledge corresponds to lower delegation rates. Our ongoing research seeks to deepen understanding of delegation behavior and its predictors in the age of GenAI.

Generative AI (GenAI) can enhance organizational processes and productivity. To realize these benefits, organizations must design GenAI agents that augment human work. A key challenge lies in making sense of the diverse forms of GenAI agents and aligning them with existing work processes. To address this, we propose a reusable co-creation process for identifying GenAI agents (GenAI-CoP) that enables organizations to involve domain experts in leveraging GenAI’s potential for their products and workflows. Grounded in action design research (ADR), our approach draws on collaboration engineering to develop GenAI-CoP. We iteratively refined and tested it through simulations, expert interviews, and pilot tests. Our research contributes to GenAI and collaboration engineering literature by introducing a reusable, bottom-up identification procedure for GenAI agents. GenAI-CoP packages facilitation expertise, allowing practitioners to execute it without prior training in tools or techniques. They gain actionable guidelines to identify augmentation potential and summarize it as GenAI agents.

The banking sector faces challenges in using deep learning due to data sensitivity and regulatory constraints, but generative AI may offer a solution. Thus, this study identifies effective algorithms for generating synthetic financial transaction data and evaluates five leading models - Conditional Tabular Generative Adversarial Networks (CTGAN), DoppelGANger (DGAN), Wasserstein GAN, Financial Diffusion (FinDiff), and Tabular Variational AutoEncoders (TVAE) - across five criteria: fidelity, synthesis quality, efficiency, privacy, and graph structure. While none of the algorithms is able to replicate the real data's graph structure, each excels in specific areas: DGAN is ideal for privacy-sensitive tasks, FinDiff and TVAE excel in data replication and augmentation, and CTGAN achieves a balance across all five criteria, making it suitable for general applications with moderate privacy concerns. As a result, our findings offer valuable insights for choosing the most suitable algorithm.

Coordination in human-AI ecologies is made possible by social norms for interaction. Managing human-AI ecologies, therefore, requires decisions about norm specification, while allowing for the emergence of norms in ecologies in such a way that enables coordination in unspecified, unstructured situations. Yet it is also important that emergent norms are not harmful and that they follow guardrails set by managers under consideration of values. We integrate predictive processing theory and social norm theory to explain how existing norms are enacted and reinforced through agents' predictive models but also how new norms emerge as agents update their predictive models in response to prediction errors in uncertain situations. Rooted in this perspective, we propose a set of design principles for managers to encode norms to evolve in human-AI ecologies, monitor outcomes, and intervene when necessary.