The problem of designing and developing information system applications is like sitting on the fence because the knowledge of two very different domains must be aligned, i.e. the business domain the application refers to and the information system engineering domain responsible for the application’s design and development. In this article ‘activity-based cost GHG accounting’ is selected due to its recent appearance for solving the design and development problem by establishing a ‘real-time add-on’ digital GHG accounting information system (GHG-AIS) application with the ‘object-oriented’ domain engineering methodology. The ‘real-time’ capability is the key feature of the digital AIS application that distinguishes it from a ‘manual’ – e.g. spreadsheet – application, and the ‘add-on’ feature assures compatibility to the production information systems incorporated in existing ERP systems.
Machine learning models are increasingly used to explore survival and treatment outcomes for stroke patients. This study addresses that gap by applying causal inference methods to estimate the survival impact of different stroke treatment strategies by using real-world clinical data. The investigation was conducted by analysing records from 944 stroke patients treated at the Clinical Centre of Montenegro. The data variables included demographic characteristics, clinical status, stroke type, and treatment information. The proposed method addressed the observed problem of high deviation and differences in sensitivity across medical treatment types within patient groups with the same stroke type diagnosis. Propensity Score Matching was used to construct comparable patient groups based on demographic and medical data and to reduce bias inherent in observational data. The causal forests machine learning method was applied to explore causal relationships and estimate individual patterns that affect survival and recovery across patient subgroups. The applied method revealed age, gender, and health status effects, highlighting differences in patient responses to treatment, as compared to the comparative evaluation without group matching and causal effect research.
Business process execution can deviate from expected behavior, resulting in anomalous events or cases. Detecting such anomalies is important for ensuring compliance, reducing risks, and improving process reliability. Existing research has mainly addressed this problem in offline settings using unsupervised learning approaches. However, such approaches detect anomalies retrospectively on complete logs and do not leverage the anomaly labels provided by domain experts. In this paper, we propose STAMP, an approach for Semi-supervised sTreaming AnoMaly detection using next activity Prediction. STAMP integrates a next-activity prediction model with an anomaly classification model trained on a limited number of anomaly labels from domain experts. Both models are continuously updated in a streaming setting to capture recent process executions. We evaluate STAMP on benchmark event logs generated with three different noise levels. Our findings demonstrate that STAMP can improve recall in early anomaly detection compared with a fixed-threshold baseline while requiring only a modest number of labeled anomalies. These findings show how semi-supervised learning can leverage scarce expert feedback for anomaly detection in streaming process monitoring.
Honey robbing in honey bees is a fast, destructive event between colonies that can rapidly deplete food reserves, cause colony loss, and facilitate the spread of parasites and pathogens within an apiary. Early detection remains challenging because visual signs at early stages are often indistinguishable from normal foraging activity. This paper proposes a data-driven early warning approach for honey robbing based on online IoT monitoring of smart hives, with a focus on data quality, reproducibility, and readiness for integration into beehive monitoring systems. The study is based on a homogeneous network of ten AmoHive smart hives observed during a single nectar season, including one confirmed robbing case and nine reference hives. The paper contributes: (i) an open dataset combining raw IoT logs with standardized, cleaned, and quality-controlled time series together with a documented preprocessing workflow and provenance information; (ii) a formalization of a hidden early warning window from 10 to 1 days before the overt event, during which destructive weight loss accumulates without obvious external symptoms; and (iii) a lightweight and interpretable cumulative detector based on standard deviation, inter-hive normalization, and cumulative negative deviation, conceptually linked to approaches from statistical process control and negative cumulative summation. In the documented case, the system generated stable warnings several days before the overt robbing phase without false alarms in the studied network, demonstrating practical feasibility for precision beekeeping.
Public sector organizations increasingly procure AI-enabled ICT systems to support decision-making and service delivery. Although ethical AI frameworks emphasize transparency, accountability, and human oversight, these principles are rarely translated into explicit requirements in procurement processes. Consequently, human-AI interaction (HAI) is often left to vendor design choices. This paper conceptualizes HAI as a procurement-critical design dimension and proposes a taxonomy of interaction requirements tailored to public sector ICT procurement. The taxonomy enables contracting authorities to specify and govern interaction properties through procurement instruments, supporting both ethical compliance and sustainable value realization.
University admissions increasingly rely on asynchronous video interviews, but manual review and transcription remain time-consuming and sensitive to transcription errors, particularly under accent variability. This paper presents a reproducible, privacy-oriented pipeline for processing authentic admissions interviews, including secure data preparation, automated personally identifiable information (PII) screening with human verification, dialect identification, and dialect-conditioned evaluation of automatic speech recognition (ASR) models. Dialect labels are inferred using a SpeechBrain ECAPA-TDNN accent identification model trained on CommonAccent. Transcription quality is assessed using word error rate (WER) for three transformer-based ASR systems: Whisper, HuBERT, and wav2vec2. Experiments are conducted on real applicant interviews and results are reported for dialect groups with more than 15 recordings. Across all evaluated dialects, Whisper yields the lowest WER (20–27
This article examines how the use of Cloud Business Intelligence (Cloud BI) systems relates to selected areas of enterprise operations. The study was conducted through an online CAWI survey among 400 medium‑sized and large companies in Poland that use Cloud BI. Using Spearman’s rank correlation, the analysis shows that neither the overall impact of Cloud BI nor the length of its use produces consistently strong or positive effects across all operational areas. However, several significant relationships were found within specific aspects of Cloud BI use, indicating its multidimensional and systemic nature. These results provide a clearer understanding of the role of Cloud BI in enterprise management.
Public grants are an important instrument for financing socio-economic development, innovation, and public policy objectives. However, identifying and comparing key information on available grants remains difficult due to the fragmented, heterogeneous, and largely unstructured manner in which grant information is published by entities administrating grants. Recent advances in LLMs enable new approaches to processing unstructured textual data and generating standardized representations that summarize and enhance key information and thus facilitate comparison. Nevertheless, their direct application in domains requiring high informational reliability is constrained by issues such as hallucination and limited reproducibility. This paper examines the practical applicability and limitations of LLMs for standardizing grant information originating from heterogeneous, human-created sources. We propose a hybrid algorithmic architecture that bifurcates the standardization process into two complementary pathways. Descriptive textual fields are generated using a tightly controlled, parameter-constrained LLM pipeline, while hard factual data are extracted using predetermined, rule-based methods. This design restricts the scope of generative models to mitigate hallucination risks while preserving flexibility in processing unstructured content. The proposed solution was implemented as a web-based system and successfully evaluated under real-world conditions. The results demonstrate that a constrained and selective application of LLMs can support effective standardization of heterogeneous information while mitigating the risks associated with hallucination and inconsistency. The proposed approach is not limited to the grant domain but is of a general nature and can be transferred to other business areas characterized by fragmented, unstructured, and inconsistently published information.
The transition toward a circular economy requires continuous tracking of material lifecycles through Digital Product Passports (DPP). Blockchain has been proposed for DPP implementations due to their immutable nature, but its application within the steelmaking sector remains largely unexplored. This paper investigates the technical feasibility of a permissioned blockchain framework to serve as a foundational infrastructure for steel DPPs. The proposed framework utilises Hyperledger Fabric with a hybrid storage strategy and cryptographic hash pointers to ensure data integrity and scalability. To validate the framework, we conducted a performance assessment simulating approximately 11.5 years of industrial activity across a 12-node distributed network. The simulation processed over 48 million transactions, displaying ledger stability and storage scalability under operational pressure. Our results display a consistent mean throughput of approx. 210 TPS and stable transaction latency even as the ledger grew linearly. The hybrid data strategy provided pruning that effectively managed storage overhead. These findings suggest that permissioned blockchains are technically feasible to serve as the foundational framework covering the requirements of DPP implementations within the steel supply chain.
The digital transformation of Higher Education Institutions (HEIs) has led to an influx of unstructured digital documents. Administrative repositories, containing orders, resolutions, and manuals, are difficult to navigate using keyword-based engines. This is problematic for the Polish language, where morphological complexity often renders simple BM25 algorithms ineffective. This paper presents the design and evaluation of an Agentic RAG system, deployed on-premise at the Poznań University of Economics and Business (PUEB). The proposed system employs an agentic workflow capable of multi-step reasoning. A comparative study of two locally hosted Small Language Models (SLMs) - the generalist open-weights OpenAI-OSS-20B and the native Polish SpeakLeash Bielik-11B-v3.0 - reveals trade-offs between reasoning capabilities and linguistic fluency. Based on these findings, a Heterogeneous Hybrid Architecture is proposed, where the generalist model handles logic, while the native model serves as a stylistic refinement layer.
Contemporary economic systems, including global supply chains and corporate financial structures, are characterised by substantial uncertainty that complicates reliability modelling. Data on component reliability, interdependencies, and failure modes are often incomplete, vague, or derived from expert judgment, which challenges traditional probabilistic risk models. This paper introduces a novel methodology for data-driven construction of mathematical models of complex systems. We frame the problem as one of classifying a system’s overall state (e.g., “Stable,” “Vulnerable,” “Distressed”) based on the uncertain states of its constituent economic components. The core innovation lies in employing a decision-tree-based classifier capable of handling uncertain and categorical data, and algorithmically transforming it into a Multi-Valued Decision Diagram (MDD). The method remains compatible with fuzzy extensions, but the present implementation uses a crisp decision tree. The MDD serves as a formal, analyzable Multi-State System (MSS) model. The efficacy of the method is demonstrated through a case study on modelling the operational resilience of a multimodal logistics network, showcasing its utility for proactive economic risk management and decision support under uncertainty.
Innovative solutions are needed to meet the challenges of rapidly growing cities. In the resulting Smart Cities, the optimization of traffic flow impaired by parking search traffic is a central area of research. Like cities, regions with a strong focus on tourism also face such challenges. This study analyses parking behaviour and traffic perceptions on the German island of Sylt, a remarkably dense tourism hotspot. The analysis draws on a large-scale online survey encompassing residents and tourists alike. The final dataset enabled a differentiated analysis of seasonal perceptions, parking difficulties, and information needs. Results show substantial differences between residents and tourists regarding perceived congestion, parking-related stress, and typical search durations. Across groups, respondents expressed strong interest in digital parking information systems, especially real-time availability and location guidance. Significant relationships between local knowledge, visit frequency, age, and the willingness to adopt such systems are revealed. The findings provide empirically grounded implications for smart parking strategies and mobility planning in tourism regions, highlighting the potential of digital parking guidance and information systems to mitigate congestion and improve user experience.
Transdisciplinary innovation ecosystems are often organized through time-bound, multi-actor collaborations that operate without stable hierarchy structures commonly conceptualized as heterarchies. To date, efforts to recover relational knowledge in these transient heterarchies and to learn from the practices of historical projects remain limited. Existing tools for knowledge graphing and ecosystem mapping offer limited support for reconstructing relational data from these collaborations once actors disperse. Through our study, we demonstrate how a large language model (LLM) can assist in reconstructing the social network between actors by filling gaps left by actor mobility, organizational churn, and prior collaboration histories. We demonstrate this method through a single case study of Project Arrow, an all-Canadian electric vehicle concept project intended to mobilize national innovation capacity, and a transdisciplinary business heterarchical collaboration involving actors across industry, academia, and public-sector institutions. This paper contributes a methodological approach for retrospective relational knowledge reconstruction in transient innovation ecosystems and shows how LLM-supported prompt engineering can support the development of knowledge maps of heterarchical collaboration.
Standardized coding of laboratory data is essential for clinical interoperability, secondary data use, and multi-institutional research. Logical Observation Identifiers Names and Codes (LOINC®) provide a global standard for laboratory observations, but manual and automated mapping remain challenging due to heterogeneous local test descriptions, language differences, and evolving terminologies. We propose an agentic approach for automated LOINC coding that leverages a locally deployed large language model (LLM) interacting iteratively with the official LOINC Search API through a query-refinement loop. The method was evaluated on 151 frequently used laboratory tests extracted from German language routine clinical data. Performance was benchmarked against a retrieval augmented generation (RAG) baseline using multiple embedding models. The proposed approach consistently outperformed the RAG baseline, achieving a Top-1 accuracy of up to 85.4% and a Top-5 accuracy of up to 98.0% . The high Top-5 performance enables efficient expert-in-the-loop validation workflows. The agentic method demonstrated greater robustness to incomplete laboratory metadata and maintained competitive runtime while operating entirely locally. Agentic, API-driven LLM workflows offer a practical and transparent solution for automated LOINC coding in clinical settings. The approach supports interoperability, minimizes dependence on proprietary services, and is well suited for integration into routine hospital environments.
Automatic Speech Recognition (ASR) holds significant potential for reducing the workload of medical staff, primarily by automating documentation tasks. While numerous benchmarks exist for the English language, specific evaluations for the German-speaking medical context – particularly those considering dialects and accents – are still lacking. In this paper, we present a dataset of simulated doctor-patient conversations and evaluate a total of 29 different ASR models. Our test set encompasses open-weight models from the Whisper, Voxtral, and Wav2Vec2 families, as well as commercial state-of-the-art APIs (AssemblyAI, Deepgram). For evaluation, we employ three distinct metrics (WER, CER, BLEU) and provide an outlook on qualitative semantic analysis. The results reveal significant performance discrepancies among the models: while top-tier systems achieve very low Word Error Rates (WER), other models exhibit considerably higher error rates, especially when processing medical terminology or dialect-heavy speech.
Rapid technological development and diverse product offerings in the fashion e-commerce sector have led to higher customer expectations regarding an optimal product experience. This study analyses the potential of product experience management (PXM) solutions to improve the product experience for both Generation X and Generation Z consumers in Germany. A quantitative research design was adopted, involving a survey of 481 participants from both generations. The data were analysed using partial least squares structural equation modeling (PLS-SEM) to identify which elements of product experience management influence the product experience across generational groups. The results reveal that the expectations of Generation X and Generation Z differ less than previously assumed, although specific areas exist where product experience management can improve the experience for each group. Based on these insights, recommendations are provided for tailoring product experience management strategies to the needs of both generations.
This paper presents preliminary results from a survey study involving a financial organization that has started the adoption of the Generative Artificial Intelligence tool Microsoft CoPilot and the expectations their workforce has about the adoption of the tool. We concentrate on the task level and show results about the effort different types of tasks require and the expectations about the effort needed and the quality of the end-result when using CoPilot to do the same tasks. The results show that CoPilot users expect the tool to reduce the effort needed, while the quality is expected to remain at the same level or to be better. We identify many tasks for which CoPilot is perceived to be helpful. The results are new and relevant to knowledge-work focused organizations contemplating the adoption of Generative AI in finance and beyond.
Cost estimation in the early phases of product development remains challenging due to high uncertainty, limited data, and a strong reliance on expert knowledge. To address these challenges, this paper presents and evaluates a proof-of-concept study for accurate and explainable cost estimation using artificial intelligence (AI) to support decision-making in product development. The proposed approach applies machine learning to generate cost estimates and employs methods from the field of explainable AI to identify cost drivers and quantify their influence on the estimated costs. In addition, a large language model is used to suggest product design improvements to reduce costs. The evaluation shows that the proposed approach can outperform both expert-based cost estimates and rule-based cost estimates generated by software tools in terms of accuracy and speed, while also improving explainability. Overall, the paper lays the groundwork for integrating accurate and explainable AI-assisted cost estimation into decision-making in product development.
This study presents a novel approach to support management scenarios using a large language model (LLM) based decision support system (DSS) that integrates a Rational Decision-Making (RDM) framework with ChatGPT to evaluate alternative courses for action. The automated evaluation considers the problem, company context, confounding factors, and outcomes’ ramifications. 72 business professionals tested the system’s feasibility using their own management decision cases, assessing DSS output across six parameters: understanding, multifaceted deliberation, output reliability, explainability, originality, and applicability. Results indicate ChatGPT’s feasibility for the DSS task, with scores generally at or above 4 out of 5 across most parameters and domains. No significant differences were found across gender, business domains, or investment scenario types, suggesting consistent system performance. However, older and more experienced professionals assessed originality and applicability significantly more critically (Cohen’s d = 0.45–0.61), indicating that domain expertise raises expectations for creative and practically grounded recommendations.
Large language models (LLMs) are increasingly used to generate low-code workflows in platforms such as n8n, enabling rapid automation but also introducing structural flaws, unreachable paths, and mismatches between user intent and the produced workflow. We present VeriFlow, a multi-dimensional verification framework for LLM-generated workflows that combines three complementary analyses: (i) structural verification of workflow graphs, (ii) semantic verification that checks consistency between natural-language intent and workflow capabilities, ordering constraints, and parameter usage, and (iii) executable verification through a capability-aware sandbox that assesses reachability, parameter completeness, and execution readiness under a verification-oriented abstraction. VeriFlow produces interpretable diagnostic evidence, including missing capabilities, ordering inconsistencies, and parameter-level issues. Experiments on a benchmark of LLM-generated n8n workflows provide initial evidence that VeriFlow can reveal structural, semantic, and execution-related problems in an interpretable way, thereby supporting safer AI-assisted workflow engineering.
