Problem definition: This paper examines the influence of an artificial intelligence (AI) application – facial recognition technology at airports – on flight on-time performance. While facial recognition at airports has the potential to save time during check-in and boarding procedures, flight departures could be delayed due to recognition errors and system inaccuracies of this immature technology. Therefore, the impacts of facial recognition on flight on-time performance remain uncertain and require rigorous empirical investigation. Methodology/results: In this study, we exploit the first terminal-wide implementation of facial recognition in the U.S. and examine its impact on flight on-time performance. Our analysis of flight-level data reveals approximately a 16% reduction in departure delays and a 6% reduction in arrival delays but no increase in early departures or early arrivals. Interestingly, the improvement in on-time performance is smaller for flights to destinations in Asia and Africa, which tend to have a higher proportion of non-Caucasian passengers, but more pronounced for flights with larger seat capacity. Managerial implications: These findings demonstrate that implementing AI tools such as facial recognition can enhance operational efficiency and reduce flight delays on average. However, the magnitude of benefits varies across flight destinations and aircraft sizes. These findings offer valuable insights for firms that are considering the deployment of AI technologies for operational efficiencies.
Problem definition: Digital educational technologies have the potential to address educational inequality by providing affordable and accessible learning resources. However, it remains unclear whether access to a wide range of learning resources through digital technologies corrects or exacerbates existing disparities, in both the short and the long run, among children from different socioeconomic backgrounds. Methodology/results: Using data from a unique large-scale field study conducted by a digital reading app for K-12 children, we apply a staggered difference-in-differences design to identify the causal effect of subsidizing a broad scope of digital reading resources. Our results show that providing free access to a broad scope of digital reading materials leads to an immediate increase of 428% in daily reading time, with the largest short-term effects observed among children from less developed cities. However, this initial boost in engagement declines sharply over time, particularly for children in less developed areas. We find suggestive evidence that the long-term difference in reading patterns reflects differing levels of parental involvement in the education of children of different socioeconomic status. The long-term contrast between children from poor and rich cities is stronger during the weekend and holidays when parents are more likely to be present. Compared with children in rich cities, children from poor cities perform worse in selecting the “right difficulty” content to read and are less able to sustain long reading sessions, especially with materials that demand more cognitive resources and parental support. Managerial implications: Our findings provide operational insights for EdTech firms and policy recommendations for policymakers, highlighting the critical role of parental involvement in fostering children’s learning persistence and their long-term participation in subsidized educational programs
Problem Definition. Rapid medical response is critical for out-of-hospital cardiac arrest (OHCA) cases. Using drones to deliver Automated External Defibrillators (AEDs) can significantly enhance the chances of survival by reducing delivery time. This paper aims to optimize the strategic deployment of drones and disposable AEDs within a budget-constrained Emergency Medical Services (EMS) system, using incomplete OHCA data. Unlike previous research, we focus on maximizing the number of timely AED deliveries within a critical window, rather than improving average or tail delivery time metrics. Methodology/Results. We frame this problem as a Modular Capacitated Maximal Covering Location Problem (MC-MCLP), incorporating a time constraint for AED delivery within a narrow therapeutic window. Our model can help alleviate resource imbalances across diverse service regions. We address demand variability using a distributionally robust optimization approach, which enhances decision resilience amid real-world uncertainties. Extensive testing reveals the impact of key parameters on the model, highlighting trade-offs between operational efficiency and both reliability and fairness. A case study of OHCA incidents in Virginia Beach demonstrates our model’s effectiveness in significantly increasing the number of patients reached within the critical time period. Managerial Implications. Our framework ensures prompt assistance to OHCA cases within the vital intervention window while promoting equitable resource allocation across regions. This approach addresses the primary challenges in EMS planning by improving response times within the crucial timeframe and establishing backup emergency resources. Our proposed methodology will enhance OHCA survival rates and optimize EMS resource distribution.
Problem definition: Growing environmental awareness is prompting consumers to consider reusable alternatives to disposable packaging, driving firms in the takeaway food and beverage sector to explore reusable packaging models. In addition to encouraging consumers to use their personal reusable packaging, some firms now offer firm-owned reusable packaging as an alternative reuse option. This paper examines how price incentives and convenience enhancements shape consumers’ packaging choices and the resulting environmental and profitability implications of reusable packaging models. Methodology/results: Using a game-theoretical model, we analyze a firm’s pricing and reuse-program decisions when consumers choose among disposable packaging, consumer-owned reusable packaging, and firm-owned reusable packaging. Our key findings are as follows. First, a more eco-conscious market does not necessarily strengthen the firm’s incentive to introduce a firm-owned reusable packaging program. When disposable packaging is inexpensive, the program’s main value lies in price discrimination rather than market expansion; as the market becomes more eco-conscious, this price-discrimination benefit weakens. Second, when disposable packaging is costly, the introduction of firm-owned reusable packaging can increase packaging waste, as the firm may strategically reduce the price discount for consumer-owned reuse. Finally, convenience improvements have asymmetric effects. Improving the convenience of consumer-owned reuse generally reduces packaging waste. However, making the return process more convenient for firm-owned reuse can increase packaging waste by shifting some consumers from consumer-owned to firm-owned reuse, which remains subject to nonreturn risk. Managerial implications: Firms in the takeaway food and beverage sector should jointly manage firm-owned and consumer-owned reusable packaging. Price incentives and convenience design should be evaluated based on how they shift consumers across packaging options and affect the tension between firm profitability and packaging waste reduction. Funding: This work was supported by the Major Research Project in Philosophy and Social Sciences of the Ministry of Education of China [Grant 2026JZDZ011]. Supplemental Material: The online appendix is available at https://doi.org/10.1287/msom.2025.0594 .
Problem definition: We study the inventory placement problem of splitting [Formula: see text] units of a single item across warehouses in advance of a downstream online matching problem that represents the dynamic fulfillment decisions of an e-commerce retailer. This is a challenging problem both theoretically, due to the computational complexity of the downstream matching problem, and practically, as the fulfillment team continuously updates its algorithm while the placement team lacks direct evaluation of placement decisions. Methodology/results: We compare the performance of three placement procedures based on optimizing surrogate functions that have been studied and applied: Offline, Myopic, and Fluid placement. On the theory side, we show that optimizing inventory placement for the Offline surrogate leads to an [Formula: see text]-approximation for the joint placement and fulfillment problem under any demand model that admits an [Formula: see text]-competitive fulfillment policy. We assume [Formula: see text] is an upper bound on how many warehouses can serve any demand location. The crux of our theoretical contribution is to use randomized rounding to derive a tight [Formula: see text]-approximation for the integer programming problem of optimizing the Offline surrogate. We further show how to extend this result to a multi-SKU setting, improving upon the best known approximation of [Formula: see text]. We use statistical learning to show that rounding after optimizing a sample-average Offline surrogate, which is necessary due to the exponentially-sized support, indeed has vanishing loss. On the experimental side, we evaluate how different combinations of placement and fulfillment procedures perform on a wide array of synthetic instances. When coupled with a good fulfillment procedure, optimizing the Offline surrogate performs best even compared to computationally-intensive simulation procedures, corroborating our theory. Managerial implications: Theoretical guarantees and extensive numerics both suggest that the placement team should optimize the (optimistic) Offline surrogate, assuming the fulfillment team has a good algorithm. Otherwise, the placement team could optimize the (pessimistic) Myopic surrogate instead.
Problem definition: In this paper, we examine how firms offering AI services can effectively acquire large volumes of training data from their consumers to improve the accuracy of the machine learning (ML) models that drive these services. Since consumers often incur privacy costs when sharing sensitive information, it is essential to design data-sharing mechanisms that balance data acquisition needs with consumers’ privacy. Methodology/results: Inspired by practice, we examine two fundamentally distinct data-sharing mechanisms: manual data-sharing, where consumers control the amount of data they share, and algorithmic data-sharing, where the firm’s algorithm redacts sensitive segments of data before using it to train the ML model. We obtain revenue-maximizing mechanisms for each approach and compare their impact on firm revenue, consumer surplus, and the volume of data collected. Our analysis highlights the conditions under which each mechanism yields superior outcomes in terms of revenue and consumer surplus. Managerial implications: Based on the comparative performance of the two mechanisms, we provide managerial guidelines that help firms choose the preferred data-sharing mechanism for different types of AI services and consumer characteristics.
Problem Definition: We study delay information disclosure policies for on-demand platforms serving two user classes (customers and providers) who seek matches using the platform. The platform's objective is to maximize the match rate by choosing the level of information—no information, binary information (indicating whether the wait is zero or non-zero), or occupancy information (indicating the expected delay based on the number of users currently in the system)—to disclose to each user class. Users of each class are strategic and decide whether to join or balk based on the delay information disclosed to them. Methodology/results: We consider two user types in each user class: patient users who are willing to wait for a match and impatient users who are not. We use continuous-time Markov chains to model the system as two-sided queues and employ equilibrium analysis to characterize users' joining behavior and the platform's match rate under each information disclosure policy. We show that the two-sided system decouples and disclosure decisions can be analyzed as two one-sided systems only if some level of information (binary or occupancy) is disclosed to both user classes. We find that disclosing binary information dominates disclosing no information or occupancy information to a user class when its patient users are sufficiently patient, while disclosing occupancy information dominates disclosing binary information to a user class when there are enough patient users. Numerical experiments show that disclosing occupancy information to both user classes, while often suboptimal, is typically not suboptimal by much. Finally, we find that compared to the platform's optimal disclosure choice, a user class may prefer more or less granular information for themselves or the other class. Furthermore, this utility analysis does not lend itself to decoupling. Managerial implications: Our findings hold crucial implications for platform managers: carefully evaluating the chosen information-sharing strategy is imperative, and guidelines from the one-sided literature are generally inadequate for making disclosure decisions
Problem definition: As the operational boundaries between two-sided platforms increasingly blur, we examine the strategic implications of pricing schemes on platform compatibility agreement. A dominant platform may offer a compatible interface for a competitive niche platform. In this setting, the niche platform strategically selects between a unified or a differential pricing scheme for the compatible channel relative to its exclusive channel. Methodology/results: We develop stylized game models to examine the coopetition interaction between the dominant and niche platforms. First, we find that compatibility does not always induce price reductions by either platform. Specifically, compatibility can soften competition under unified pricing but intensifies it under differential pricing, where the niche platform charges a lower compatible channel price while increasing its exclusive channel price. Second, compatibility can be sustained as an equilibrium under unified pricing only when the niche platform’s awareness is high and the compatibility-induced convenience gain is limited. In contrast, under the differential pricing scheme, compatibility becomes unstable as adverse profit redistribution erodes the dominant platform’s incentive to participate. Third, when comparing the effects of the two pricing schemes, we find that differential pricing provides the niche platform with greater leverage and shifts profits toward the compatible channel, whereas unified pricing mitigates competition and preserves a balanced channel structure. Finally, compatibility affects stakeholder welfare differently, redistributing benefits across consumers, providers, and platforms. Managerial implications: Our study provides guidance on the conditions under which platform firms profitably pursue compatibility. The strategic essence of compatibility decisions lies in the redistribution of surplus, which shapes the participation incentives of platforms. Managers are advised to calibrate compatibility agreements based on relative awareness advantages and the convenience gains associated with the compatible channel. Funding: This work was supported by National Natural Science Foundation of China [Grant 72402076]; Humanities and Social Sciences Research Project of the Ministry of Education of China [Grant 24YJC630123]; Shandong Provincial Natural Science Foundation of China for Young Scholars [Grant ZR2024QG123]; Taishan Scholar Project of Shandong Province of China [Grant tsqn202408212]; National Science and Technology Council [Grant NSTC 113-2410-H-005-038-MY3]; Hong Kong RGC [Grant C6020-21GF]; and Crown Worldwide Professorship of Business. Supplemental Material: The online appendix is available at https://doi.org/10.1287/msom.2024.0805 .
Problem definition: The renewable energy certificate (REC) market plays a critical role in ensuring renewable energy compliance by facilitating the matching of REC supply and demand. In this study, we focus on the problem faced by REC aggregators who act as brokers, making REC sales decisions on behalf of individual renewable energy generators. Methodology/results: We formulate the aggregator’s REC management problem as a discrete-time Markov decision process (MDP) and derive structural properties of the optimal policy for three prevalent service contracts in the marketplace. We develop a solution framework based on a deep reinforcement learning (DRL) algorithm that integrates state-of-the-art features, including dueling double deep Q-networks (DDDQNs) and graph neural networks (GNNs). Our solution algorithm leverages the structural properties of the problem to enhance the efficiency and performance of REC management. Managerial implications: Our simulation study, based on data from two representative U.S. REC markets, highlights the role and value of REC aggregation services for small-scale generators. Finally, we provide further insights into how REC markets and services perform under various business conditions. Funding: D. G. Choi acknowledges the support by H Energy Co. Ltd. M. K. Lim acknowledges the support by the Ministry of Education of the Republic of Korea and the National Research Foundation of Korea [Grant NRF-2022S1A5A2A01038230], as well as the Institute of Management Research, College of Business Administration, Seoul National University. Supplemental Material: The online appendix is available at https://doi.org/10.1287/msom.2023.0477 .
Problem definition: Vehicle routing problems (VRPs) with deadlines have received significant attention around the world. Motivated by a real-world food delivery problem, we assume that the travel time depends on the routing decisions, and we study a data-driven stochastic VRP with deadlines and endogenous uncertainty. Methodology/results: We use the nonparametric approaches, including k-nearest neighbor (kNN) and kernel density estimation (KDE), to estimate the decision-dependent probability distribution of travel time. To solve the resulting problem efficiently, we employ a logic-based Benders decomposition (LBBD) algorithm with several algorithmic enhancements. In particular, we propose a novel family of optimality cuts that includes the expected delay for all the subroutes. Moreover, we solve a total travel cost minimization problem to warm start the algorithm. We also use a local search procedure to improve the current routing decision and propose a machine learning–based lower bound heuristic to efficiently solve problems of realistic size. A practical case study for a food delivery routing problem using real-world data is conducted to show the efficiency of the proposed techniques and the advantage of the data-driven stochastic VRP in reducing the expected delay. Managerial implications: In our case study, we show that incorporating routing decisions into a nonparametric model outperforms a state-of-the-art data-driven parametric model by 23% on average in terms of the expected delay and the order-assignment decisions obtained from a robust model with travel-time predictors by 26% on average. Moreover, compared with the drivers’ actual routes and arc-based VRP models that ignore the endogenous uncertainty, our suggested routes can significantly improve the on-time performance of delivery services. We also quantify the value of the proposed routes with different service deadlines. Funding: S. Wang was partially supported by the Natural Sciences and Engineering Research Council of Canada [Grant RGPIN-2016-05208], IVADO, and a joint project between the Fonds de Recherche du Québec - Société et Culture (FRQSC) and the National Natural Science Foundation of China (NSFC) [Grant 295837]. She is also most recently supported by the National Natural Science Foundation of China [Grants 72501014, 72371022, and 72272014]. Supplemental Material: The online appendix is available at https://doi.org/10.1287/msom.2024.0899 .
Problem definition: Projects are often initiated by a single person—a principal—who then decides whether to form a team by sharing project value with an agent. We ask the following: When does a principal form a team, and which operating mode emerges in equilibrium—single execution by the principal, delegated execution to the agent, or joint execution? Methodology/results: We consider a coproductive principal-agent model with endogenous team formation. In the second best, joint execution is less frequent than optimal, whereas both single and delegated execution are chosen too much. With a linear contract, common with nonfinancial output (credits, coauthorship) or in entrepreneurial ventures, there is too much single execution; that is, the principal does not form a team often enough, and there is either too little joint execution when one of the workers has low productivity or too little delegation otherwise. Overall, the inefficiency in project execution (due to moral hazard) appears less severe than the inefficiency in the principal’s team formation decision (due to project hoarding). Managerial implications: Managerial under-delegation fundamentally stems from project hoarding; that is, principals do not partner enough. When principals choose to form a team, they then might delegate too much; paradoxically, this happens when agents have low productivity. Although the existing literature focuses on eliciting agent effort for a given operating mode, our analysis suggests the decision to form a team (or not) is a more critical issue. History: This paper has been accepted in the Manufacturing & Service Operations Management Frontiers in Operations Initiative. Supplemental Material: The online appendix is available at https://doi.org/10.1287/msom.2025.0016 .
Problem definition: Service quality is assessed with objective and subjective measures. Objective measures include queue length/wait times, service times, and service failures, whereas subjective measures include service evaluations and social media commentary. Subjective measures are often easier to collect, requiring no sophisticated tracking technology, and can capture multiple dimensions of the service experience in a single response. However, subjective data are prone to bias. We focus on bridging the gap between objective and subjective measures of service quality by reducing bias in service evaluations. The structure of service evaluations is consistent across organizations: star ratings are collected before comments. We propose a simple intervention-reversing this order-to mitigate bias in star ratings because the process of writing provides time and space for the evaluator to reflect on their experience. Methodology/results: We conducted an experiment where participants received a sequence-based service, with the same overall level of service, from servers who varied in their demographic characteristics. Following the service, participants evaluated the performance of the servers with the order of star ratings and comments randomized. We find no evidence of demographic bias toward the servers but find that the sequence of good and bad experiences in the service leads to biased star ratings. Most importantly, we find that collecting comments prior to star ratings mitigates the sequential bias in star ratings because of participants reflecting on the service experience. Managerial implications: We show that star ratings can be biased if collected in the traditional manner and that this bias can be reduced if comments are collected first. Implementing this change to service evaluations could help to ensure that the job performance of human servers is more accurately and fairly assessed. For artificial intelligence service systems (i.e., AI servers), this change can provide a simple way to debias training data.
Problem definition: A grocery retailer incurs expiration waste (EW) at its store when a perishable product crosses its expiration date without being sold. One frequent scenario accounting for EW occurs when units of a given product with multiple expiration dates are simultaneously available on store shelves. In such situations, a consumer is likely to purchase a later-to-expire unit, which in turn increases the likelihood of EW of a sooner-to-expire unit. To mitigate the occurrence of such multiple-dates-led expiration waste (MDEW), retailers undertake a variety of interventions, including a price markdown of sooner-to-expire units and in-store inventory rotation. Most retailers, however, are often unaware of the extent of MDEW in their stores and thus are constrained in mitigating its occurrence. Methodology/results: We provide the first large-scale evidence of the MDEW share of EW. We collaborate with a grocery retailer to compile a multicategory-multistore data set ([Formula: see text]15.3 million sales transactions) on grocery products with 3–14 days of shelf life. Across these products, at the product-store-week level, EW as a percentage of sales is 23% on average. To quantify MDEW’s share, we propose a novel and easy to implement methodology for computing its lower and upper bounds. In our retailer’s context, the MDEW’s lower and upper bounds equal 25% and 52% on average of the generated EW, respectively. Managerial implications: Our study highlights MDEW’s material share in generating EW; thus, it provides a solid premise for future in-depth academic investigation on MDEW management. Furthermore, for practitioners, it provides an immediately actionable methodology to measure MDEW in their store operations. Empowered with such a measurement ability, retailers can better plan their EW waste management interventions. Funding: The authors gratefully acknowledge financial support. A. Kabra received support from the University of Maryland, College Park and the Nanyang Technological University Business School [Start Up Grant 025540-00001]. V. Karamshetty received support from the National University of Singapore School of Computing [Seed Fund A-8003889-00-00]. Supplemental Material: The online appendices are available at https://doi.org/10.1287/msom.2023.0607 .
Problem definition: We investigate how a network of healthcare facilities should manage nonpandemic and pandemic demand, asking whether each facility must operate in a specialized (i.e., treating a single patient type) or a generalized (i.e., treating both patient types) mode. Given the many drivers of the specialization-generalization decision, we focus on the value that specialization can create by reducing mix-variability in lengths of stay and limiting the scope and intensity of infection prevention measures. Methodology/results: We develop two optimization models that minimize the sum of patient waiting costs and facility infection mitigation costs under a static demand allocation policy. We provide an analytical characterization of the optimal allocation when facilities’ bed capacities are equal, and a highly effective heuristic when they are not. These results suggest that the optimal configuration contains at most one generalized facility. We also propose a simple expression that yields a conservative lower bound on the value of specialization, defined as the reduction in total cost when adopting the specialization-based configurations produced by our models instead of a configuration with all facilities being generalized. Managerial implications: For length-of-stay parameters from the first COVID-19 waves in England and the Netherlands, this lower bound equals [Formula: see text] and [Formula: see text], respectively. However, the actual performance gap could be substantially larger depending on infection mitigation costs and capacity imbalances, indicating a considerable potential value of facility specialization. For the special case of two equal-capacity facilities, we also show that a virtual pooling policy, as a representative dynamic policy, yields substantial savings over the best static policy only at very high traffic intensities; at lower intensities, static allocation can perform better. Finally, we show that adopting optimal static allocation policies can substantially reduce costs of a network facing an imminent pandemic by proposing a supply-demand management framework that combines capacity expansion with optimal demand allocation. Supplemental Material: The e-companion is available at https://doi.org/10.1287/msom.2024.1172 .
Problem definition: We study a paradox in which decision-makers deviate downward from proposals of an automated store replenishment system after a stockout. We argue that censorship bias explains this curious ordering behavior and it has negative performance implications. We compare the effect of censorship bias to that of anchoring bias to understand the relative importance of censorship bias in retail practice. Understanding the impact of certain biases on decision-makers’ inventory replenishment decisions in the presence of an algorithm is imperative to maximize the benefit from decision-makers' discretionary power. Methodology/results: We analyze data on perishable products from an upmarket supermarket chain by employing exclusion restrictions in our recursive bivariate probit model to account for the endogeneity of deviations. We complement our analysis with endogenous switching regression and seemingly unrelated regression models to further test the robustness of our results. We find that after a stockout, decision-makers’ likelihood of deviating downward is higher, a behavior aligned with censorship bias. We find that anchoring bias is more powerful than censorship bias in predicting downward deviations. Regarding the performance implications, we show that censorship bias is more detrimental than anchoring bias in terms of increasing the likelihood of a new stockout. Therefore, we suggest that this insight into censorship bias can be used to distinguish uninformed downward deviations from the informed ones. Managerial implications: To suggest actionable policies, we collect more data to test the idea of blocking downward deviations when censorship bias is suspected. With this additional data analysis, we show that by blocking the downward deviations susceptible to censorship bias, retail managers can reduce self-inflicted stockouts with reasonable inventory cost implications.
Problem definition: Motivated by passenger arrivals at the security checkpoint of the Raleigh-Durham International Airport, we develop methods to study arrivals to a system in which they are tied to scheduled events, such as flights. A key concept for modeling arrivals in such systems is the “show-up profile,” a probability distribution describing how far in advance passengers arrive for their flights. These profiles can be combined based on a known flight schedule to yield an aggregate passenger arrival forecast. Existing industry practice and academic work estimate show-up profiles using surveys or other data that are typically not available to U.S. airports. This motivates our study of an easy to implement and dynamic method for estimating show-up profiles. Methodology/results: We introduce an innovative solution for estimating show-up profiles using infrared-beam people-counting sensors and a structural estimation approach that does not require a mapping of passengers to flights. A direct maximum likelihood approach is intractable, but we propose a tractable approximation and prove that it yields consistent estimates of the underlying show-up profile parameters. Our approach produces forecasting results comparable to pure machine learning methods, yields significantly improved adaptive forecasts when combined with machine learning methods, and reveals empirical insights about passenger behavior variations across different times of day and flight destinations. Managerial implications: Our work presents a novel application of Internet of Things technology to service operations with incomplete data and demonstrates the value of integrating known operational structure with black box forecasting approaches. Show-up profiles are used at airports for decision making, for example, for crowd management, and our methodology has the potential to drive significant improvements in airport operations. The methods we develop can be readily applied at U.S. airports and other transportation hubs, and they can be adapted to other event-driven service environments such as theaters, healthcare facilities, and museums. History: This paper has been accepted in the Manufacturing & Service Operations Management Frontiers in Operations Initiative. Funding: This work was supported by the 2021 Triangle Impact Challenge. Supplemental Material: The online appendices are available at https://doi.org/10.1287/msom.2024.1575 .
Problem definition: Whereas artificial intelligence (AI) technologies are increasingly becoming powerful and useful in operations, human workers often resist adopting algorithms, known as algorithm aversion. This aversion can undermine the algorithm performance in practice. Whereas numerous studies explore short-term mitigation strategies for such aversion, this paper investigates whether and why forced interventions can promote AI adoption and reduce algorithm aversion in practice. Methodology/results: Data from a leading online education company reveal that sales workers underutilize a new matching algorithm and often selectively use it on low-quality leads. The company conducted a field experiment in which sales workers were forced to use or not use the algorithm for three weeks. Experimental results show that forcing workers to use the algorithm during the experiment causally increases their algorithm usage over the month after the experiment by 15.8 percentage points. We develop a theoretical model to derive empirical strategies for exploring the mechanisms behind this improvement. Contrary to the traditional literature focusing on habit formation, our findings suggest learning is a key driver for algorithm adoption among workers over the month after the experiment. Specifically, forced algorithm use allows workers to experience the unbiased algorithm performance and positively adjust their beliefs about it. Consequently, after the experiment, workers use the algorithm not only more frequently but also more on high-quality leads. Managerial implications: The study empirically shows that forced intervention can effectively improve persistent algorithm use after the intervention, which is crucial for continuous development of the algorithm. More importantly, forced intervention breaks the vicious cycle of biased beliefs and selective usage by enabling workers to form unbiased evaluation of the algorithm efficacy and mitigate selective adoption on low-quality cases. This suggests that firms can implement extrinsic interventions or educational programs to help workers recognize the benefits of algorithms and develop unbiased beliefs about their capabilities, thus facilitating sustained algorithm usage. Supplemental Material: The online appendix is available at https://doi.org/10.1287/msom.2024.1137 .
Problem definition: Worker attrition is critical and costly, disrupting operations across industries and leading to significant productivity losses. In healthcare, nurse attrition poses even greater challenges, which are exacerbated by persistent shortages and increasing burnout. Despite its importance, nurse attrition remains underexplored in operations management (OM) literature, particularly concerning how different workload dimensions influence voluntary attrition. This study aims to address this gap by investigating how various workload dimensions, including nurse responsibility, overtime shift, emotional toll, and cumulative workload, affect voluntary attrition among ICU nurses. Methodology/results: Utilizing high-resolution data from a large U.S. hospital system, we analyze 26 months of operational, clinical, and HR records, capturing nurses’ dimensions of workload leading up to voluntary attrition decisions. Our findings reveal a nuanced relationship; whereas greater nurse responsibility during a shift reduces the likelihood of voluntary attrition, cumulative workload over time shows a U-shaped relationship with the likelihood of voluntary attrition. Additionally, both the emotional fatigue from handling patient death events and the burnout from attending overtime shifts heighten the likelihood of voluntary attrition. One more incident of emotional fatigue increases the odds of voluntary attrition by 54.3%. One more overtime shift increases the odds of voluntary attrition by 58.5%. However, supportive coworkers can help mitigate some negative effects, highlighting the importance of collaborative environments. Managerial implications: This research makes several contributions. First, we estimate the distinct effects of workload dimensions on voluntary attrition. Second, we demonstrate how supportive coworkers act as a buffer against burnout-induced attrition. Finally, we offer actionable strategies for managers to enhance workforce retention: monitoring workloads to prevent fatigue-driven attrition, implementing flexible scheduling to allow recovery, and fostering peer support systems. By addressing a critical issue in high-stakes environments like healthcare, this study enriches OM literature and provides practical insights for organizations seeking to retain talent in knowledge-intensive fields. History: This paper has been accepted in the Manufacturing & Service Operations Management Frontiers in Operations Initiative. Supplemental Material: The online appendix is available at https://doi.org/10.1287/msom.2025.0037 .
