Although integrating artificial intelligence (AI) into work processes is often expected to enhance productivity, such transformation may also undermine employee satisfaction, which is essential for the long-term success of AI adoption. As organizations implement various modalities of human–AI collaboration, it remains unclear how these arrangements affect employee satisfaction. This paper addresses this gap by delineating three distinct collaboration modalities: human-led, AI-led, and co-led. We examine how each modality influences user satisfaction with collaborative outcomes and collaborative processes, and how these effects vary across task environments. Our findings show that when task uncertainty is low, the human-led modality leads to higher outcome satisfaction; however, as uncertainty increases, differences in outcome satisfaction across modalities diminish. In contrast, process satisfaction remains consistently sensitive to collaboration modality regardless of task uncertainty, and this effect is further amplified by task complexity. Across all conditions, the AI-led modality produces the highest levels of process satisfaction. Moreover, we also reveal a tension between satisfaction and performance. Although the co-led modality produces lower satisfaction with both the outcome and the process, it achieves the highest decision quality. By shifting attention from collaboration performance alone to collaboration satisfaction, this study extends research on human–AI collaboration and offers practical guidance for designing AI-based work processes that support both effectiveness and sustainability.
Although organizations increasingly rely on artificial intelligence (AI) to support decision making, research consistently finds that decision makers exhibit varied responses to algorithmic advice. While prior research has examined when users are more likely to accept or reject AI advice, little is known about the cognitive factors associated with these responses or the broader psychological consequences of working with AI. Drawing on metacognition theory, this study explores how individuals monitor and revise their judgments when confronted with algorithmic advice in a behavioral experiment with 440 participants. The results indicate that metacognitive estimates of confidence are strongly associated with responses to algorithmic advice. However, individual differences in metacognitive sensitivity shape how consistently they rely on confidence when responding to AI advice. Specifically, when metacognitive sensitivity is low, confidence judgments are based on noisy internal states, which, in turn, lead to inconsistent patterns of advice-taking. We also find that exposure to contradictory AI advice is associated with lower retrospective confidence in individuals' subsequent decisions. This loss of confidence is more pronounced when individuals reject the AI advice, highlighting a hidden psychological cost of advice rejection. These findings contribute to the growing body of information systems (IS) research on algorithmic advice-taking by demonstrating that effective human-AI collaboration requires not only improvements in system design but also fostering users' metacognitive sensitivity.
Ethereum phishing detection is a decision-support problem: blockchain risk analysts must prioritize suspicious addresses under severe label scarcity and bounded review capacity. We present TRACE, a two-stage decision-support artifact that constructs dense Ethereum address graphs from Etherscan transaction records, pre-trains an edge-aware encoder with graph contrastive learning, and adapts it with validation-based threshold calibration and explanation analysis for reviewable phishing screening. Rather than proposing new learning primitives, the contribution is a domain-grounded configuration whose individual design choices are non-obvious for dense, heavy-tailed transaction graphs and whose benefit grows with graph scale. Across five graphs from 30,000 to 149,930 addresses, six label ratios, eight baselines, and five random seeds, TRACE achieves the best overall mean performance (F1 0.5094; AUC 0.9365). Component ablations, paired significance tests, precision–recall (AUPRC) reporting, and prevalence-sensitivity experiments down to sub-1% phishing rates confirm that each design element contributes and that the advantage is stable under realistic class imbalance. Performance improves with graph scale, reaching F1 0.6720 and AUC 0.9730 on D150k under full labels, and a smaller pre-training graph transfers to a larger target (F1 0.5670). The artifact logic generalizes to other scarce-label, interaction-rich risk-screening settings beyond Ethereum.
Financial market segmentation has predominantly relied on supply-based methods that emphasize the corporate business market, while paying scarce attention to demandoriented schemes that highlight shared investor preferences and behaviors. This study narrows this gap by proposing an investor demand-oriented approach for financial market segmentation, which consists of three fundamental phases: financial network construction based on investor co-attention, network representation learning using network embedding techniques, and clustering over network embeddings. By implementing the approach in the U.S. stock market and cross-validating its usefulness in the Chinese stock market, we first show that it can reflect pairwise stock similarity, as network embedding similarity is significantly and positively associated with excess comovement between stocks. We then partition the stock market into different peer groups that demonstrate significant within-group comovement. The utility and efficacy of our approach are comprehensively evaluated by comparing it with traditional industry classifications, geographical location-based classification, text-based classification, and search cluster-based investment habitats. The results demonstrate that our approach not only achieves better performance in capturing greater withingroup homogeneity and between-group heterogeneity, but also outperforms standard classification schemes in predicting excess comovement and stock returns. This study adds to design science research on financial market segmentation by extending the important yet underexplored domain of demand-oriented methods. By systematically decoding latent investor preferences and behavioral commonalities, our approach can be integrated into automated decision support systems to improve the accuracy and efficiency of financial market analytics pragmatically.
While social media platforms enable the exchange of information and experiences across online communities, their extensive connectivity simultaneously accelerates the dissemination of misleading and potentially harmful content. Content moderation serves as a crucial mechanism for curbing such dissemination, with community rules playing a central role in guiding moderation decisions. Despite significant advances in automated content moderation, the explicit incorporation of community rules into model design remains rare. Among the different levels of social aggregation where community rules operate, macro norms establish standardized governance principles to ensure consistency across communities; however, they remain largely overlooked in automated content moderation research. To address these gaps, we propose TOpic-based Rule Affinity (TORA), a content moderation framework that draws on and adapts polycentric governance theory, algorithmic governance theory, and sociotechnical systems theory to incorporate rules from heterogeneous communities into moderation. Following the design science paradigm, TORA introduces two novel design artifacts: topic-based macro norms, which use Large Language Models (LLMs) to capture and standardize rules across diverse communities, and a topic-based rule affinity measure, which quantifies the relationship between user-generated content and those norms. We implement and evaluate TORA using data collected from multiple social media communities. Empirical results show that the framework consistently outperforms state-of-the-art baseline models, and each of its two key design artifacts contributes significantly to improved performance. These findings make significant theoretical and design contributions and have implications for improving societal outcomes by enabling more interpretable and transparent content moderation and supporting fairness and justice in online governance.
Personally Identifiable Information (PII) exposure on social media creates serious privacy risks, including identity theft and targeted scams, yet users and platforms lack effective tools to detect and assess such exposures in a timely manner. This study adopts a computational design science approach and proposes a decision support framework for social media privacy risk assessment, with automated PII extraction at its core. We develop MCA-PIIE, a deep learning–based PII extraction model that combines deep transfer learning with multi-context attention to address limited labeled data and capture semantic, syntactic, and PII-indicative cues in informal social media text. The framework is evaluated against state-of-the-art PII extraction methods using well-established public datasets and a curated Twitter ground-truth dataset. The results show that MCA-PIIE achieves an F1 of 72.34% on social media data, outperforming from-scratch and prompt-based baselines and matching the accuracy of a QLoRA fine-tuned 8B language model while using roughly 400 times fewer parameters. An exploratory user study examines how the proposed framework helps users identify PII exposure on social media and how their privacy-protection decisions align with the system-assessed risk levels. This study contributes a decision support perspective to social media privacy analytics and offers design insights for privacy-aware information systems.
Amid the growing threat of fake reviews, manipulated reviews of telemedicine applications (hereafter: apps) can mislead patients into choosing ineffective telemedicine services, compromising their well-being. Detecting suspicious reviews of telemedicine apps is crucial for safe services. However, it is constrained by the scarcity of domain-specific, annotated data. To overcome this challenge, we integrate the homophily principle and social presence theory to develop a dual-pronged approach to identifying app-level clusters exhibiting suspicious aggregate review patterns that warrant targeted investigation. The first prong is based on the homophily principle and employs a novel community-detection-based approach to generate communities of telemedicine apps. The second prong uses network and text analytics to compare clusters to identify those containing suspicious reviews. We validated our double-pronged approach on 166 telemedicine apps from the Google Play Store. In line with social presence theory and the deception detection literature, reviews in the suspicious cluster exhibited low social presence, with more formal, detached, mechanical, and superficial content, and a lack of detailed personal disclosure and in-depth cognitive reasoning. This study has important implications for the literature on suspicious review identification and for the practical design of systems that combat review manipulation in domains lacking high-quality annotated data.
Rapid advances in artificial intelligence (AI) are reshaping industry practices and creating new security vulnerabilities, raising questions about the effectiveness of firms' internal risk management capabilities under heightened technological pressure. Drawing on socio-technical systems theory, we investigate how top management team (TMT) IT readiness influences firm-level data breach risk under the external technological pressure of industry AI exposure, and how firm strategic positioning shapes that response. Using a comprehensive firm-year panel dataset, we first confirm, consistent with prior research, that TMT IT readiness is associated with lower data breach risk. Building on this baseline, our central findings show that this effect is not uniform but is conditioned by the firm's technological and strategic context. The breach-reducing effect of TMT IT readiness becomes stronger in industries with higher AI exposure, and this moderating effect is itself contingent on firm strategic positioning. Specifically, the effect is more pronounced among innovation-oriented firms and less consistent among efficiency-oriented firms. These findings advance the data breach literature by identifying the boundary conditions of the well-established link between TMT IT readiness and breach risk, showing that its effectiveness amplifies under higher industry AI exposure and depends on firm strategic positioning. This study also offers practical implications for managers by highlighting the need to tailor security governance to industry-level technological pressures.
Social recommendation leverages users’ social network information to enhance product recommendation. To mitigate the issue of low-quality social signals, prior studies have introduced graph denoising and graph disentanglement. However, due to the lack of strong supervision, these methods often yield suboptimal performance. In this study, we propose a new framework for explicit semantic disentanglement. Specifically, we first use a Large Language Model (LLM) to extract user intentions from reviews and construct intention-specific collaborative subgraphs, which serve as supervision signals to disentangle the social graph into multiple intention-specific social subgraphs. This enables the model to independently learn interaction and social signals under different intentions. Extensive experiments on three real-world datasets demonstrate that our model significantly outperforms state-of-the-art methods in both accuracy and explainability.
We propose the Expert-Augmented Semantic Embedding (EASE) weighting scheme, which integrates domain knowledge into machine learning-based sentiment analysis of corporate disclosures. EASE leverages the Loughran and McDonald financial sentiment lexicon to geometrically define a semantically meaningful direction in a word embedding space, enabling continuous and graded sentiment scoring of the full lexicon, including words absent from the original lexicon, without requiring labeled outcome data or model retraining. Using a large sample of 10-K annual reports, we show that EASE-augmented sentiment measures exhibit improved economic significance and explanatory power for market returns and future earnings relative to their non-augmented counterparts. Because EASE anchors sentiment weights to an expert-defined semantic axis rather than to historically observed outcomes, the resulting measures are potentially less susceptible to the influence of strategically deployed or otherwise shifting language in corporate disclosures. Our findings highlight the complementarity between human knowledge and machine learning in extracting economically meaningful sentiment from financial texts.
Effective communication between customer service representatives (CSRs) and customers is essential for maintaining customer relationships and improving operational performance. While firms increasingly use AI technologies to evaluate customer service communication effectiveness, existing approaches primarily rely on dialogue text and overlook the sequential use of communication strategies that reflect CSRs’ intentions. In addition, communication strategies are often quantified through manual coding, limiting scalability and explainability. Consequently, the effects of individual strategies and their combinations on communication effectiveness remain unclear. Following the Computational Design Science research paradigm, this study proposes an explainable Transformer-based model with a dual-input structure that integrates dialogue text and theoretically grounded communication strategy sequences. To enable automated strategy quantification, we design few-shot Chain-of-Thought prompts to guide large language models in annotating communication strategy labels from CSR utterances. Model explainability techniques are further employed to analyze the effects of individual strategies and their combinations at the dialogue-turn level. The framework is evaluated using a large-scale real-world customer service call dataset. Results show that the proposed model outperforms benchmark methods across multiple metrics and exhibits robust performance in ablation and stability analyses. The explainability results also provide actionable insights to support data-driven customer service training and performance optimization.
Despite an increase in research on social exclusion, a consistent definition and measure is lacking, especially for online and virtual environments. Additionally, previous measures cannot account for inconsistencies in the literature, such as social exclusion's relationship with trust and social factors. In this research, we present a review of the literature, clarify the conceptualization, and provide a definition. Using a student sample and another more general population, we develop a two-dimensional scale. First, we use focus groups to develop new survey items, then rely on subject-matter experts to assess content validity. After developing measurement items, an exploratory analysis for grouping the items and confirmatory analysis for validating the measure was conducted, followed by group & predictive analyses to further validate the measure. The new scale reliably and validly appraises individual online social exclusion. The scale resolves inconsistencies from the literature, showing it to be more discerning than previous operationalizations.
Reliable last mile delivery (LMD) of essential commodities continues to present significant operational challenges, particularly due to delivery disputes, black-market diversion, and the lack of reliable real-time verification at the point of service. To address this gap, this study develops and pilot-implements an AI–enabled Decision Support System (DSS) for secure delivery authentication in LPG cylinder distribution. The proposed DSS integrates a convolutional neural network–based image verification module and a geolocation verification module within a Django-based framework and is deployed as a Progressive Web Application (PWA). In operational use, the system supports distributors by automatically verifying whether cylinders are delivered to the intended customers at their registered delivery locations, thereby enabling evidence-based delivery assurance.,The system was pilot deployed at an LPG distribution agency in Chennai, India, using two delivery agents over a ten-day operational window. During live field operations, the DSS achieved an average end-to-end authentication time of approximately two seconds, with no authentication failures observed during the pilot. These findings demonstrate the operational feasibility of the proposed approach and its ability to provide low-latency, automated delivery verification without imposing additional burden on delivery personnel or customers.,This study contributes to the DSS and logistics literature by introducing a mobile-integrated authentication framework that embeds artificial intelligence within real-time last-mile operations to strengthen delivery accountability and digital trust. Managerial implications related to field deployment, scalability, and operational monitoring are discussed. While the pilot results are encouraging, broader multi-region validation remains an important direction for future research.
Credit card fraud detection remains a significant challenge due to extreme class imbalance, evolving fraud patterns, and asymmetric costs of misclassification. While existing machine learning methods often prioritize fraud detection performance, they frequently overlook model risk management, which is crucial for reliable decision making in high-stakes environments. To address this issue, we introduce DISCO-a framework that combines discriminative and robust representation learning with a conformal prediction-based mechanism for provable risk control. At its core, DISCO is established upon the principle of decoupling representation learning from risk-controlled decision-making. To this end, we first leverage deep metric learning (DML) to construct an embedding space that is inherently robust to class imbalance and resilient to concept drift. Subsequently, we employ a conformal risk control (CRC) mechanism to provide a formal statistical guarantee on the false negative rate (FNR) of the resulting model according to a user-specified FNR target. Computational experiments on a real-world dataset demonstrate that DISCO outperforms state-of-the-art methods in both classification performance and operational efficiency. To the best of our knowledge, this is the first work to integrate a DMLbased representation learning with a formal risk control mechanism for reliable credit card fraud detection. The proposed framework therefore offers a trustworthy and practical solution that addresses the dual imperatives of predictive performance and reliability.
Service failures are inevitable in chatbot-based customer service. In hybrid customer service systems, when a chatbot fails, a human agent typically takes over and resume service for the customer. However, little is known about how to effectively initiate this transition from chatbot failure to human-agent intervention. This study identifies four initiation methods varying in service proactivity: passive initiation, text prompt initiation, button prompt initiation, and automatic initiation. Extending social information processing theory with a cognitive decision perspective, we develop a theoretical framework to examine how consumers attribute the observed initiation methods to organizational motives, which subsequently influence their satisfaction. We conduct three online experiments to examine how these initiation methods influence recovery satisfaction, as well as the underlying psychological mechanisms and boundary conditions. The results reveal significant differences in recovery satisfaction across initiation methods, mediated by customers' inferred selfish intent. Furthermore, compared with low task urgency, high task urgency significantly amplifies differences in recovery satisfaction and inferred selfish intent among various human-agent initiation methods. These findings advance understanding of recovery strategies for chatbot service failures, offering actionable guidance for designing effective transitions from chatbots to human agents in hybrid service settings.
