Navigating the Jagged Technological Frontier: Field Experimental Evidence of AI's Effects on Knowledge Worker Productivity and Quality

This article was downloaded by: [199.94.42.12] On: 17 March 2026, At: 08:00 Publisher: Institute for Operations Research and the Management Sciences (INFORMS) INFORMS is located in Maryland, USA Organization Science Publication details, including instructions for authors and subscription information: http://pubsonline.informs.org Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality Fabrizio Dell’Acqua, Edward McFowland III, Ethan Mollick, Hila Lifshitz, Katherine C. Kellogg, Saran Rajendran, Lisa Krayer, François Candelon, Karim R. Lakhani To cite this article: Fabrizio Dell’Acqua, Edward McFowland III, Ethan Mollick, Hila Lifshitz, Katherine C. Kellogg, Saran Rajendran, Lisa Krayer, François Candelon, Karim R. Lakhani (2026) Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality. Organization Science Published online in Articles in Advance 11 Mar 2026 . https://doi.org/10.1287/orsc.2025.21838 This work is licensed under a Creative Commons Attribution 4.0 International License. You are free to copy, distribute, transmit and adapt this work, but you must attribute this work as “Organization Science. Copyright © 2026 The Author(s). https://doi.org/10.1287/orsc.2025.21838, used under a Creative Commons Attribution License: https://creativecommons.org/licenses/by/4.0/.” Copyright © 2026 The Author(s) Please scroll down for article—it is on subsequent pages With 12,500 members from nearly 90 countries, INFORMS is the largest international association of operations research (O.R.) and analytics professionals and students. INFORMS provides unique networking and learning opportunities for individual professionals, and organizations of all types and sizes, to better understand and use O.R. and analytics tools and methods to transform strategic visions and achieve better outcomes. For more information on INFORMS, its publications, membership, or meetings visit http://www.informs.org ORGANIZATION SCIENCE Articles in Advance, pp. 1–21 https://pubsonline.informs.org/journal/orsc ISSN 1047-7039 (print), ISSN 1526-5455 (online) Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality Downloaded from informs.org by [199.94.42.12] on 17 March 2026, at 08:00 . For personal use only, all rights reserved. Fabrizio Dell’Acqua,a,* Edward McFowland III,a Ethan Mollick,b Hila Lifshitz,a,c Katherine C. Kellogg,d Saran Rajendran,e Lisa Krayer,e François Candelon,e Karim R. Lakhania a Digital Data Design Institute, Harvard Business School, Boston, Massachusetts 02134; b The Wharton School, University of Pennsylvania, Philadelphia, Pennsylvania 19104; c Artificial Intelligence Innovation Network, Warwick Business School, London CV4 7AL, United Kingdom; d Sloan School of Management, Massachusetts Institute of Technology, Cambridge, Massachusetts 02142; e Henderson Institute, Boston Consulting Group, Boston, Massachusetts 02110 *Corresponding author Contact: fdellacqua@hbs.edu, https://orcid.org/0000-0002-1998-0542 (FD); emcfowland@hbs.edu, https://orcid.org/0000-0001-5249-7117 (EdM); emollick@wharton.upenn.edu, https://orcid.org/0000-0001-6231-496X (EtM); hdiginnovation@gmail.com, https://orcid.org/0000-0002-3461-003X (HL); kkellogg@mit.edu, https://orcid.org/0000-0003-4372-3498 (KCK); saravanan0822@gmail.com (SR); krayer.lisa@bcg.com (LK); francois.candelon@seven2.eu (FC); klakhani@hbs.edu, https://orcid.org/0000-0002-5535-8304 (KRL) Received: December 18, 2025 Abstract. We introduce and study the concept of a “jagged technology frontier” to Revised: January 23, 2026 describe the uneven impact of artificial intelligence (AI) capabilities, where AI assistance Accepted: January 27, 2026 improves performance for some tasks but worsens it for others, even within the same Published Online in Articles in Advance: knowledge workflow and with a seemingly similar level of difficulty. In collaboration with March 11, 2026 the global management consulting firm Boston Consulting Group, we have developed realistic management consulting tasks and examined the human performance implications https://doi.org/10.1287/orsc.2025.21838 of using AI to perform complex and knowledge-intensive work. The preregistered experi­ ment involved 758 knowledge workers. After establishing a performance baseline on simi­ Copyright: © 2026 The Author(s) lar tasks, subjects were randomly assigned to one of three conditions: no AI access, GPT-4 AI access, or GPT-4 AI access with a prompt engineering overview. For each one of a set of 18 realistic knowledge tasks within the frontier of AI capabilities ranging from creative to analytical tasks, subjects using AI outperformed those not using AI, completing 12.2% more tasks and completing them 25.1% more quickly on average while also delivering solutions of significantly improved quality. However, for a complex managerial task selected to be outside the frontier, subjects using AI were 19% less likely to produce correct solutions compared with those without AI, pointing to potential limitations of AI support­ ing knowledge workers. We discuss the positive and negative implications of AI-aided human performance in knowledge-intensive tasks. Open Access Statement: This work is licensed under a Creative Commons Attribution 4.0 International License. You are free to copy, distribute, transmit and adapt this work, but you must attribute this work as “Organization Science. Copyright © 2026 The Author(s). https://doi.org/10.1287/orsc.2025. 21838, used under a Creative Commons Attribution License: https://creativecommons.org/ licenses/by/4.0/.” Funding: Financial support of the Harvard Business School Digital Data Design Institute and Division of Research and Faculty Development is acknowledged. Supplemental Material: The online appendix is available at https://doi.org/10.1287/orsc.2025.21838. Keywords: technology and innovation management • organizational economics • economics and organization • organization and management theory • research design and methods • field experiments • implementation of new technology • organizational processes 1. Introduction (Geerling et al. 2023, Noy and Zhang 2023, Boussioux The capabilities of artificial intelligence (AI) systems to et al. 2024, Brynjolfsson et al. 2025). This represents a sig­ aid humans in a variety of tasks have improved rapidly, nificant shift in human-machine augmentation, with especially since the release of OpenAI’s ChatGPT, hybrid aptitudes increasingly relevant across a broad which made large language models (LLMs) widely range of knowledge work—including some highly crea­ available for public use. Recent research has shown that tive and highly educated professions (Eloundou et al. these systems are unexpectedly capable of augmenting 2024, Zhou and Lee 2024). human abilities in the completion of creative, analytical, As generative AI’s capabilities increasingly over­ and writing tasks as well as achieving top scores in aca­ lap with human skill sets, the integration of AI into demic aptitude tests at graduate and professional levels human work presents new fundamental challenges, 1 Dell’Acqua et al.: Navigating the Jagged Technological Frontier 2 Organization Science, Articles in Advance, pp. 1–21, © 2026 The Author(s) opportunities, and questions, especially in knowledge- Figure 1. The Jagged AI Frontier intensive scientific and professional fields. In this paper, we investigate the impact of generative AI on management consulting—a crossdisciplinary field where top graduates from leading universities tackle diverse projects across industries. We examine this issue using a randomized, controlled laboratory-in- Downloaded from informs.org by [199.94.42.12] on 17 March 2026, at 08:00 . For personal use only, all rights reserved. the-field experiment, in which highly skilled manage­ ment consultants performed tasks resembling their standard work assignments. Our results demonstrate that AI capabilities cover an uneven set of knowledge work, and we introduce the con­ cept of a “jagged technological frontier” to characterize the performance of AI when used by knowledge workers. Tasks that appear to human knowledge workers to be of similar difficulty may be performed either better or worse by humans using AI. Within this jagged frontier, AI can complement human work. However, outside the frontier, AI output is inaccurate, is less useful, and can degrade human performance. AI assistance improves human per­ formance only for tasks within current AI capabilities— within the jagged technological frontier—and worsens human performance outside of it. We find that workers Notes. This figure provides a conceptual illustration of the jagged tech­ who skillfully navigate this frontier in their use of AI sys­ nological frontier of AI. The dashed line represents tasks of roughly tems gain substantial quality and productivity benefits. equal perceived difficulty to human knowledge workers. Tasks that Figure 1 provides a conceptual representation of this appear similar in difficulty to humans may fall on opposite sides of the frontier such that AI assistance improves performance for some tasks jagged technological frontier, illustrating how tasks that while degrading it for others. The figure is stylized and not intended to appear similar in complexity for humans may none­ depict specific tasks or empirical effect sizes. ChatGPT produced this theless fall on opposite sides of the boundary where AI image starting from the authors’ prompts. assistance helps versus harms performance. Notably, jaggedness reflects systematic differences in how well tasks selected and evaluated by industry experts to rep­ AI capabilities align with different tasks compared with licate the real-world task assignments of knowledge expectations based on human capabilities. Because the workers. Most knowledge work includes a set of inter­ capabilities of AI are rapidly evolving and poorly dependent tasks, some of which may be a good fit for understood, it can be hard ex ante for knowledge work­ current LLMs, whereas others are not. We examine mul­ ers to grasp exactly what the boundary of this frontier tiple types of interdependent knowledge tasks and might be at a given moment in time. Indeed, within the build on recent studies to suggest ways of understand­ same knowledge workflow, some tasks are beyond the ing the impact of generative AI on real-world task frontier, whereas others remain within it, making effec­ assignments, under which circumstances knowledge- tive AI use challenging—especially without diligent based organizations may benefit but also face negative examination of the workflow to identify deviations outcomes, and we consider how this impact might from expectations and the potential for redesign. change as the technology advances. The potential benefits of AI for knowledge work have Three factors make it difficult for knowledge workers been a subject of considerable interest, although much of to know what the value and downsides of using genera­ this research focuses on AI prior to the release of GPT-3.5 tive AI may be ahead of time. First, LLMs have surprising (e.g., Brynjolfsson et al. 2018, Çalli et al. 2021, Davies et al. capabilities that they were not explicitly or intentionally 2021, Monisha et al. 2021). The general availability of developed to have—capabilities that are growing rapidly GPT-3.5 in the form of ChatGPT since November 2022 over time as model size, computational capability, and has changed both the nature and urgency of the discus­ algorithmic performance increase. Although trained as sion (Berg et al. 2023, Eloundou et al. 2024, Zhou and Lee general models, LLMs demonstrate specialist knowledge 2024). Studies of previous generations of AI (e.g., Agra­ and abilities as part of their training process and during wal et al. 2018, Furman and Seamans 2019, Brynjolfsson normal use (Reed et al. 2022, Boiko et al. 2023, Moor et al. et al. 2021) and of recently released LLMs (e.g., Noy and 2023, Singhal et al. 2023). Although considerable debate Zhang 2023, Choi and Schwarcz 2024, Brynjolfsson et al. remains on the concept of emergent capabilities from a 2025) suggest that these systems can considerably impact technological perspective (Schaeffer et al. 2023), the effec­ worker performance. Our study focuses on complex tive capabilities of LLMs and the ways that they are used Dell’Acqua et al.: Navigating the Jagged Technological Frontier Organization Science, Articles in Advance, pp. 1–21, © 2026 The Author(s) 3 are novel and unexpected, widely applicable, and in­ limitations and become overly reliant on AI’s output, creasing rapidly over short time spans. Recent work has potentially overlooking its limitations (Dell’Acqua 2022). argued that AI performs at a high level in knowledge The future of understanding how AI impacts work work contexts ranging from medicine to law (Ali et al. involves understanding how human interaction with AI 2023, Lee et al. 2023) and that the human-AI combination changes depending on where tasks are placed on this outperforms humans alone on many measures of innova­ frontier. This work investigates how humans navigate tion (Boussioux et al. 2024, Meincke et al. 2024). Addition­ this jagged frontier and the subsequent performance Downloaded from informs.org by [199.94.42.12] on 17 March 2026, at 08:00 . For personal use only, all rights reserved. ally, scores on various standardized academic tests, albeit implications. an imperfect measure of LLM capabilities, have been For this study, we collaborated with a global manage­ increasing substantially with each generation of AI mod­ ment consulting firm (Boston Consulting Group (BCG)), els (OpenAI 2023). advising them on the design, development, and execu­ The general ability of LLMs to solve domain- tion of a preregistered randomized experiment to assess specific problems leads to the second factor differenti­ the impact of AI on knowledge workers.2 We specifically ating LLMs from previous forms of AI: their ability to worked with them to develop tasks that closely resemble directly improve the performance of knowledge workers the workflow of their management consultants. Subse­ who use these systems and more generally, their unprec­ quently, the author team received the data that the com­ edented usefulness in knowledge-intensive domains. pany collected for this experiment and conducted the Early studies of the new generation of AI suggest direct analysis presented in this paper. Online Appendix F performance increases from using AI, especially for com­ shows a detailed analysis using Bureau of Labor Statistics plex work, like writing tasks (Noy and Zhang 2023) and data that contextualizes our experimental occupation software programming (Peng et al. 2023), as well as for within the broader landscape of knowledge work, dem­ ideation and creative work (Boussioux et al. 2024, onstrating how management consulting responsibilities Meincke et al. 2024). Consequently, the effects of AI are align with the multifaceted demands of modern profes­ expected to be higher in knowledge-intensive domains sional environments. featuring highly educated workers (Felten et al. 2023, We tested two distinct sets of tasks: one situated Eloundou et al. 2024). beyond the frontier of AI capabilities and one situated The final relevant characteristic of generative AI is its within the bounds of that frontier. The experiment relative opacity. Although previous generations of AI aimed to understand how AI integration with human- are often black boxed, which can be frustrating for users performed tasks might reshape the traditional work­ (Lebovitz et al. 2022), LLMs have been shown to pro­ flows of highly skilled professionals in a knowledge duce incorrect but plausible results (hallucinations or domain. Both sets of tasks were developed to be realistic confabulations) in a way that users find accessible and and were designed by senior professionals at BCG who “believable,” leading to inaccurate outputs.1 The advan­ had experience in the same roles as the participants dur­ tages of LLMs, although substantial, are frequently ing their careers and were currently supervising work­ unclear to users. Despite performing well at some tasks, ers in those roles. A senior managing director and LLMs fail in other circumstances in ways that are diffi­ partner at the company commented that these tasks cult to predict in advance. Adding to this opacity, devel­ were “very much in line with part of the daily activities opers provide no guidance regarding the best way to of the consultants” involved. Notably, some forms of use these AI systems, which appear to be best learned these tasks are also used by the company to screen job via ongoing user trial and error and the exchange of applicants, typically from elite academic backgrounds experiences and heuristics on various online forums. (including PhDs), for their highly coveted positions. These three factors together suggest that both the Online Appendix G uses a survey of 11 managing value and downsides of using generative AI may be dif­ directors and partners at BCG, among the most senior ficult for knowledge workers to know ahead of time. executives at the firm, to confirm that the underlying Some unexpected tasks (like idea generation) are easy competencies assessed by our experimental tasks are for AIs, whereas tasks that seem as though they should critical for recruiting, job performance, and career pro­ be easy for machines to complete are challenging for gression at BCG. some LLMs (e.g., basic math). Our research suggests Our experimental findings demonstrate mixed results that this feature of LLMs as used by humans creates a regarding the quality and performance effects of humans “jagged frontier,” where tasks that appear to be of simi­ integrating AI into their task flow. For inside-the-frontier lar difficulty are performed better or worse by humans tasks, we show that this generation of LLMs is highly using AI. Because of the “jagged” nature of the frontier, capable of significantly increasing quality and productiv­ a knowledge worker may face tasks in an assignment ity, leading to substantial productivity benefits across that are on both sides of the frontier without a priori multiple dimensions when used by humans. However, knowledge regarding each task’s position with respect quality drops substantially when knowledge workers to that frontier. Knowledge workers may fail to see AI’s deploy AI to solve tasks that are beyond AI’s capability Dell’Acqua et al.: Navigating the Jagged Technological Frontier 4 Organization Science, Articles in Advance, pp. 1–21, © 2026 The Author(s) frontier. Moreover, the location of the frontier of AI capa­ define as nonroutine, cognitively demanding activities bility is not immediately obvious to knowledge workers, centered on the creation, application, or integration of even though the tasks themselves are quite familiar. Con­ specialized intellectual expertise.3 Online Appendix F sultants tended to overrely on AI, leading to a decline in shows how management consulting exemplifies knowl­ combined human-machine performance when closer edge work by analyzing Occupational Information human supervision was necessary (Athey et al. 2020, Del­ Network (O*NET) occupational data. Although BCG con­ l’Acqua 2022, Glaeser et al. 2024). This is especially rele­ sultants operate at the upper end of the complexity spec­ Downloaded from informs.org by [199.94.42.12] on 17 March 2026, at 08:00 . For personal use only, all rights reserved. vant when we consider that our participants were highly trum, our comparative analysis reveals that the diverse specialized knowledge workers working within their tasks that they perform—from analytical problem-solving domain of expertise, emphasizing the need for knowl­ to creative ideation—mirror those found across other edge workers to have a nuanced understanding of AI’s high-skill professional domains that require a minimum capabilities as they navigate the evolving landscape of of an undergraduate degree for consideration to be eligible human-AI integration to optimize the balance between for entry into those professions. AI assistance and human expertise. Because this jagged frontier is expanding and chang­ 3. Methods ing, the overall results suggest that AI will have a pro­ We collected data from a randomized experiment to found impact on knowledge work, one that will become assess the causal impact of AI, specifically GPT-4 (then more pronounced as LLM capabilities expand. However, the most capable of the AI models available (June 2023)) the distribution of these impacts will be uneven. Our on highly qualified professionals otherwise working tra­ results not only highlight AI’s potential to transform ditionally without AI.4 We preregistered our study, knowledge work but also, underscore the importance detailing the design structure, the experimental condi­ of human involvement in machine-assisted knowledge tions, the dependent variables, and our main analytical workflows. The effectiveness of AI in knowledge work approaches.5 We aimed to determine how introducing will critically depend on human judgment—particularly this AI into the tasks of highly skilled knowledge work­ to discern which tasks within the workflow are suited to ers might augment, disrupt, or influence their perfor­ leveraging AI augmentation and where human expertise mance and workflow. Details about the experimental should be prioritized. structure are reported in Online Appendix A. The study included 758 BCG consultants (~7% of 2. Experimental Context their individual-contributor consulting workforce) per­ BCG is a premier global management consulting firm, forming knowledge work. This is an elite group of admitting only around 1% of applicants (Vlamis et al. knowledge workers with degrees from top universities 2024). Once hired, consultants enter a highly structured around the world, such as Harvard, Yale, and Oxford. career track in which promotions are largely determined Tables 1 and 2 display a breakdown of demographic by strict performance appraisals. Promotion decisions characteristics and experimental performance metrics often hinge on consultants’ abilities to excel in diverse for subjects across conditions. Table 3 reports the most tasks from creative idea generation to complex business common educational majors among our participants. modeling, making BCG an ideal setting to investigate Each participant completed the initial survey and was how AI might affect skill development and performance. randomly assigned to one of two distinct arms of the BCG also primarily promotes from within. experiment, each involving a unique type of task, with In close collaboration with senior BCG leaders, we designed tasks that capture core facets of consulting. no overlap between the groups.6 These task types closely capture the breadth of work that Approximately half of the participants (385 subjects) consultants commonly handle—ranging from innovative tackled a series of tasks that prompted them to conceptu­ solution proposals to methodical problem-solving. We alize and develop new products, focusing on aspects crafted each task to reflect realistic business scenarios and such as creativity, analytical skills, persuasiveness, and crucially, ensured that some tasks were well suited to AI writing. The other half (373 subjects) engaged in business support and others were not as described in Section 4. To problem-solving tasks using quantitative data, customer validate the real-world importance of these tasks beyond and company interviews, and a persuasive writing com­ our immediate author team, we surveyed 11 managing ponent. Both sets of tasks were developed to be realistic: directors and partners at BCG. Their feedback affirmed that is, representative of the tasks that these workers that the tasks that we designed were highly relevant for engaged in regularly. To achieve this, the tasks were day-to-day consulting work and aligned with the core designed by BCG professionals from the relevant sectors. competencies that BCG evaluates in both recruiting and Figure 2 provides a visual representation of our experi­ ongoing performance reviews (see Online Appendix G). mental design across the two arms of the experiment. BCG management consultants represent an ideal The two arms of the experiment followed a consistent population for studying knowledge work, which we structure. Initially, participants undertook a task without Dell’Acqua et al.: Navigating the Jagged Technological Frontier Organization Science, Articles in Advance, pp. 1–21, © 2026 The Author(s) 5 Table 1. Summary Statistics by Demographics and Experimental Conditions GPT + overview (A) GPT only (B) Control (C) p-values Variable N Mean SD N Mean SD N Mean SD (A � B, A � C, B � C) Panel A: Inside the frontier Female 126 0.27 0.45 129 0.34 0.48 130 0.34 0.48 0.22, 0.23, 0.96 English 126 0.57 0.50 129 0.57 0.50 130 0.58 0.49 0.93, 0.83, 0.76 Downloaded from informs.org by [199.94.42.12] on 17 March 2026, at 08:00 . For personal use only, all rights reserved. Tenure 126 0.56 0.50 129 0.53 0.50 130 0.52 0.50 0.56, 0.44, 0.85 Location 126 0.33 0.47 129 0.32 0.47 130 0.32 0.47 0.90, 0.86, 0.97 Tech Openness 126 0.59 0.49 129 0.65 0.48 130 0.61 0.49 0.30, 0.74, 0.47 Panel B: Outside the frontier Female 125 0.35 0.48 118 0.33 0.47 129 0.32 0.47 0.73, 0.57, 0.83 English 125 0.57 0.50 118 0.59 0.49 129 0.54 0.50 0.69, 0.69, 0.42 Tenure 125 0.56 0.50 118 0.50 0.50 129 0.55 0.50 0.35, 0.88, 0.43 Location 125 0.29 0.45 118 0.31 0.46 129 0.33 0.47 0.77, 0.52, 0.73 Tech Openness 125 0.57 0.50 118 0.64 0.48 129 0.63 0.49 0.28, 0.33, 0.90 Panel C: Inside � outside; p-values Female 0.16 0.86 0.72 English 0.96 0.67 0.50 Tenure 0.96 0.67 0.57 Location 0.52 0.83 0.86 Tech Openness 0.76 0.80 0.74 Notes. This table provides a breakdown of demographic characteristics for subjects across conditions. For each category, we report total sample size (N), mean values, and standard deviations (SDs) to detail the distribution and variability of the data. All p-values are from two-sided t-tests of equality of means. None are significant at the 5% level. the aid of AI, establishing a baseline for performance This platform, developed using OpenAI’s application and enabling within-subject analyses.7 Following this, programming interface (API), facilitated an interactive participants were randomly assigned to one of three con­ experience with OpenAI’s GPT-4 that mirrored the ditions to assess the influence of AI on their tasks, with dynamics of ChatGPT. Although the functionality these conditions kept consistent across both arms. The resembled that of ChatGPT, the platform enabled the first group (a control condition) proceeded without AI collection of all data logs, including participant prompts support. The second group (“GPT only”) had the assis­ and the AI’s corresponding responses, providing com­ tance of an AI tool based on GPT-4. The third group prehensive insight into the collaborative behaviors (“GPT + overview”) utilized the same AI tool but also between subjects and AI. All subjects used the same ver­ received a supplementary prompt engineering over­ sion of the tool, accessing GPT-4 as available at the end view, which increased their familiarity with the AI tool. of April 2023 and using default system prompts and These overview materials included instructional videos temperature. The same model was used for both sets of and documents that outlined and illustrated effective tasks. usage strategies. Other than thematic differences, the two sets of tasks Following the baseline assessment, subjects assigned differed in another key way. Although both were de­ to the various AI conditions had access to a custom- signed to be of comparable complexity and realistically built company platform for every experimental task. represent real-world consulting responsibilities, tasks in Table 2. Summary Statistics—Performance GPT + overview GPT only Control Variable N Mean SD N Mean SD N Mean SD Inside the frontier: Performance Quality 126 5.86 0.52 129 5.68 0.61 130 4.38 0.45 Completion 126 0.93 0.15 129 0.91 0.16 130 0.82 0.17 Timing 126 3,894 1,281 129 3,635 1,494 130 5,023 827 Outside the frontier: Performance Correctness 125 0.60 0.49 119 0.71 0.46 129 0.84 0.36 Timing 125 1,571 920 119 1,853 995 129 2,260 917 Subjective coherence quality 125 7.33 1.62 118 6.90 2.34 129 5.86 2.20 Notes. This table provides a breakdown of experimental performance metrics for subjects across conditions. For each category, we report total sample size (N), mean values, and standard deviations (SDs) to detail the distribution and variability of the data. Dell’Acqua et al.: Navigating the Jagged Technological Frontier 6 Organization Science, Articles in Advance, pp. 1–21, © 2026 The Author(s) Table 3. Most Frequently Reported Majors—Bachelor’s From their set of ideas, they identified the most viable Degree option and devised a comprehensive plan for its market debut. After this task, subjects moved to the main experi­ Major Count Subjects, % mental phase, and the context transitioned to the main Economics 104 14.10 experimental task. Finance 42 5.70 In this experimental task, participants were tasked Mechanical engineering 41 5.60 with conceptualizing a footwear idea for niche markets Business administration 41 5.60 Downloaded from informs.org by [199.94.42.12] on 17 March 2026, at 08:00 . For personal use only, all rights reserved. Chemical engineering 31 4.20 and delineating every step involved from prototype Business 23 3.10 description to market segmentation to entering the mar­ Engineering 17 2.30 ket. An executive from a leading global footwear com­ Industrial engineering 17 2.30 pany verified that task design covered the entire process Civil engineering 15 2.00 that their company typically undergoes from ideation to Political science 14 1.90 Computer science 12 1.60 product launch.10 Participants responded to a total of 18 Politics 11 1.50 tasks (or as many as they could within the given time11). Chemistry 11 1.50 These tasks spanned various domains, categorizable Mathematics 10 1.40 into four types: creativity (e.g., “Propose at least 10 ideas Notes. This table presents the distribution of the most frequently for a new shoe targeting an underserved market or reported majors among our subjects’ bachelor’s degrees. The majors sport.”), analytical thinking (e.g., “Segment the foot­ are listed in descending order of prevalence, with the count and wear industry market based on users.”), writing profi­ percentage of subjects for each major provided. The percentages reflect the proportions of subjects within our study sample who reported ciency (e.g., “Draft a press release marketing copy for each major as their field of study during their undergraduate your product.”), and persuasiveness (e.g., “Pen an inspi­ education. rational memo to employees detailing why your prod­ uct would outshine competitors.”). This allowed us to the first experiment were designed to be within the collect comprehensive assessments of quality. All tasks potential technological frontier of GPT-4. In contrast, and details are reported in the Online Appendix. the second experiment was deliberately designed so The primary outcome variable in the experiment is the that human interaction and guidance were necessary to quality of the subjects’ responses. To quantify this qual­ perform successfully when using GPT-4. The second ity, we employed a set of human graders to evaluate each experiment also contained a persuasive component that question that participants did not leave unanswered.12 fell within the frontier.8 We conducted pretests for both Each response was independently evaluated by three sets of tasks, iterating several times to ensure that they human graders. We then calculated the mean grade were appropriately situated on either side of the frontier assigned by humans to each response. This gave us 18 before initiating the experiment. dependent variables (one per question). We subsequently Participant engagement and effort were encouraged averaged these scores across all questions to derive a through performance-based incentives designed to reflect composite “quality” score for use in our main analyses. BCG’s performance evaluation system.9 As outlined in As an additional assessment, we utilized GPT-4 to inde­ our preregistration, all participants who demonstrated pendently score subjects’ responses.13 Similarly to the honest effort received “office contribution” recognition— human grades, we produced a score for each of the 18 a designation with direct financial implications for annual questions and then, a composite “quality (GPT)” score. bonuses. The bottom 5% of performers forfeited this rec­ Figure 3 uses the composite human-grader score and ognition, whereas the top 20% received additional visually represents the performance distribution across acknowledgment to their supervisors noting their excep­ the three experimental groups, with the average score tional performance, potentially further impacting bonus plotted on the y axis. Comparing the dashed lines and calculations and advancement prospects. the overall distributions of the experimental conditions in Figure 3 clearly reveals the significant performance enhancements associated with the use of GPT-4, with 4. Results the GPT-only and GPT + overview groups both record­ 4.1. Quality and Productivity Booster—Inside ing superior performances to the control group. the Frontier Table 4 presents the results of the analyses using The inside-the-frontier experiment focused on creative response quality as the dependent variable and high­ product innovation and development. We defined lights the performance implications of using AI.14 Col­ inside-the-frontier tasks by pretesting them to confirm umns (1), (2), and (3) in Table 4 utilize human-generated that generative AI could produce useful, high-quality out­ grades as the dependent variable, whereas column (4) puts in the BCG client-consulting context from straight­ in Table 4 uses the composite grades generated by forward prompts. The initial assessment task asked GPT-4. Across all specifications, both treatments—GPT + participants to brainstorm innovative beverage concepts. overview and GPT only—demonstrate positive effects. Dell’Acqua et al.: Navigating the Jagged Technological Frontier Organization Science, Articles in Advance, pp. 1–21, © 2026 The Author(s) 7 Figure 2. Experimental Structure Downloaded from informs.org by [199.94.42.12] on 17 March 2026, at 08:00 . For personal use only, all rights reserved. Note. This figure illustrates the experimental setup utilized in our study. In column (1) in Table 4, GPT + overview leads to a 1.48 persuasiveness) and applied the same model detailed in increase in scores over the control mean of 4.37, repre­ column (1) in Table 4. Each of the regressions demonstrates senting a 33.9% increase, with GPT only producing a a significant effect of introducing AI on knowledge work­ 1.31 (or 29.9%) increase. Notably, columns (2), (3), and ers’ performance, with very sizable results for all question (4) in Table 4 incorporate performance metrics from the categories. We verified that the treatment effects are not assessment task, and the treatment coefficients that they driven by superficial spelling or grammar improvements. report remain very consistent.15 Column (4) in Table 4 Adding controls for the number and rate of such errors uses GPT scores as the dependent variable and shows leaves all main coefficients virtually unchanged (see coefficients of 1.35 above the control group for the GPT + Online Appendix H). Finally, we tested the reliability of overview treatment and 1.22 above the control group for these findings by replicating our analyses on a new set of the GPT-only treatment, corresponding to 19% and grades for the ideas generated in the first question. Our 17.1% improvements in performance, respectively.16 results are consistent as discussed in Section 5. The beneficial impacts of using AI remain consistent Another key observation from Table 4 is the differen­ across all our specifications.17 As a robustness check, we tial impact of the two AI treatments. Specifically, the isolated the quality of responses within each category (cre­ GPT + overview treatment consistently exhibits a ativity, analytical thinking, writing proficiency, and more pronounced positive effect compared with the Dell’Acqua et al.: Navigating the Jagged Technological Frontier 8 Organization Science, Articles in Advance, pp. 1–21, © 2026 The Author(s) Figure 3. Inside the Frontier—Performance Distribution subjects exposed to one of the two AI conditions into two distinct categories—top-half-skill performers (those ranking in the top 50% on the assessment task) and bottom-half-skill performers (those in the bottom 50%)—we observed performance enhancements in the experimental task for both groups when leveraging GPT-4. Note that the top-half-skill performers also Downloaded from informs.org by [199.94.42.12] on 17 March 2026, at 08:00 . For personal use only, all rights reserved. received a significant boost, although not as much as the bottom-half-skill performers. Our findings gain additional perspective through comparison with recent research in this area. Bryn­ jolfsson et al. (2025) explore customer support roles, Table 4. Inside the Frontier—Quality Notes. This figure displays the full distribution of performance in the experimental task inside the frontier for subjects in the three experimen­ (1) (2) (3) (4) tal groups (red for subjects in the GPT + overview condition, green for Variable Quality Quality Quality Quality (GPT) subjects in the GPT-only condition, and blue for subjects in the control GPT + overview 1.482*** 1.490*** 1.506*** 1.349*** condition). Each subject’s responses were rated by MBA graders on cre­ (0.061) (0.057) (0.061) (0.058) ativity, analytical thinking, writing proficiency, or persuasiveness depending on the specific question. Graders used a 1–10 scale, where GPT only 1.307*** 1.329*** 1.337*** 1.216*** higher scores indicate stronger performance. The x axis thus represents (0.067) (0.065) (0.065) (0.059) the average of these ratings across all applicable tasks for each subject, Assessment 0.174*** 0.171*** and the y axis shows the kernel density estimate of that distribution. (0.058) (0.060) Assessment (GPT) 0.167** GPT-only treatment. The bottom of Table 4 displays a (0.070) p-value that tests whether the effects of receiving GPT + Female �0.153*** �0.156*** �0.042 overview were equivalent to those of being assigned to (0.057) (0.057) (0.049) GPT only, showing this value to be below or around the English Native 0.066 0.070 0.097* conventional 5% threshold in all specifications. This (0.061) (0.063) (0.055) underscores the importance of the added overview for Low Tenure 0.045 0.045 0.018 enhancing the efficacy of AI assistance, with Table 4 also (0.052) (0.054) (0.048) highlighting various other factors, including gender, Location �0.029 �0.029 0.050 native English proficiency, tenure, location, and open­ (0.067) (0.069) (0.061) ness to new technology, none of which changed the sig­ Tech Openness 0.068 0.067 0.065 nificance of the impact of AI on performance.18 (0.056) (0.063) (0.058) Table 5 presents the results related to the percentage AI-exposure controls Yes Yes of tasks completed by subjects, this analysis’s dependent R2 0.610 0.642 0.646 0.639 variable. Across columns (1), (2), and (3) in Table 5, the GPT � GPT 0.014 0.025 0.021 0.049 two treatment conditions (GPT + overview and GPT) + overview only demonstrate a positive effect on task completion. Control mean 4.375 4.375 4.375 7.098 On average, these coefficients indicate a 12.2% increase Observations 385 385 385 385 in completion rates. Notably, directly comparing the Notes. This table examines the effects of introducing GPT-4 on the two AI treatments reveals that the difference in their quality of the responses for the experimental task inside the frontier. Each column displays the results of a distinct linear regression model. impacts is not statistically significant. The control group Columns (1)–(3) have the average response quality as their dependent completed on average 82% of their tasks; those in the variable; each response was independently graded by three human GPT + overview group completed about 93%; and those evaluators. In contrast, column (4) uses the average response quality in the GPT-only group completed about 91%. Column in the experimental task as determined by GPT. Columns (2) and (3) incorporate the average response quality from the assessment task as (2) in Table 5 incorporates the performance metric from graded by human evaluators, whereas column (4) utilizes the GPT- the assessment, and column (3) in Table 5 extends the evaluated metric. Columns (3) and (4) add seven presurvey AI- analysis by including the same set of controls as in Table exposure controls: perceived current task automation (0%–100%); belief that one’s job could be automated; Likert-scale items on 4. The coefficients suggest that the integration of AI tools familiarity with text-generation tools, image-generation tools, prompt substantially enhances the rate of task completion while engineering, and large language model mechanics; and a five-point also increasing quality. measure of prior ChatGPT usage frequency (never to daily). Higher Figure 4 presents an important trend; the most signifi­ scores on all seven variables indicate greater AI familiarity or usage. The bottom of the table displays p-values from an F-test comparing the cant beneficiaries of using AI are the bottom-half-skill effects of receiving the GPT + overview treatment vs. the GPT-only subjects, consistent with findings from Noy and Zhang treatment. All regressions include robust standard errors. (2023) and Choi and Schwarcz (2024).19 By segmenting *p < 0:10; **p < 0:05; ***p < 0:01. Dell’Acqua et al.: Navigating the Jagged Technological Frontier Organization Science, Articles in Advance, pp. 1–21, © 2026 The Author(s) 9 Table 5. Inside the Frontier—Completion (1) (2) (3) Variable Percentage Completion Percentage Completion Percentage Completion GPT + overview 0.111*** 0.109*** 0.105*** (0.020) (0.020) (0.020) GPT only 0.090*** 0.088*** 0.082*** Downloaded from informs.org by [199.94.42.12] on 17 March 2026, at 08:00 . For personal use only, all rights reserved. (0.021) (0.021) (0.018) Assessment �0.007 0.006 (0.013) (0.011) Female �0.001 0.014 (0.018) (0.016) English Native 0.014 �0.003 (0.020) (0.019) Low Tenure 0.025 0.025* (0.017) (0.015) Location �0.021 �0.004 (0.020) (0.018) Tech Openness 0.017 0.009 (0.017) (0.017) Percentage Completion (Assess) 0.367*** (0.052) AI-exposure controls Yes R2 0.083 0.100 0.303 GPT � GPT + overview 0.282 0.265 0.216 Control mean 0.824 0.824 0.824 Observations 385 385 385 Notes. This table examines the effects of introducing GPT-4 on the subject’s task completion for the experimental task inside the frontier. Each column displays the results of a distinct linear regression model. The dependent variable across all columns is the percentage of subtasks (of 18) that subjects successfully completed. Columns (2) and (3) use the average response quality in the assessment task as evaluated by two human graders as a control. Column (3) additionally includes the percentage of completed questions in the assessment task and seven presurvey AI-exposure controls: perceived current task automation (0%–100%); belief that one’s job could be automated; Likert-scale items on familiarity with text-generation tools, image-generation tools, prompt engineering, and large language model mechanics; and a five- point measure of prior ChatGPT usage frequency (never to daily). Higher scores on all seven variables indicate greater AI familiarity or usage. The bottom of the table displays p-values from an F-test comparing the effects of receiving the GPT + overview treatment vs. the GPT-only treatment. All regressions include robust standard errors. *p < 0:10; ***p < 0:01. identifying productivity gains primarily among lower- Figure 4. Inside the Frontier—Bottom-Half Skills and Top- skilled workers, with negligible effects observed among Half Skills their higher-skilled counterparts. Choi and Schwarcz (2024) use law school examinations and find that stu­ dents at the bottom of the performance distribution benefited enormously from AI assistance, whereas higher-skilled students saw performance declines. Con­ versely, Otis et al. (2024b) report mixed outcomes from the deployment of generative AI among small-scale entrepreneurs in Kenya. Although high-performing indi­ viduals reap the benefits of AI-generated advice, low- performing individuals tend to fare worse when follow­ ing AI suggestions. In our study, centered on high-end knowledge work, we observe a universal benefit for workers involved in complex, high-level tasks, particu­ larly for tasks within the AI’s capability frontier. Notes. This figure shows the performance of AI-treated subjects, com­ Within this group of elite professionals, relatively paring those in the bottom half with those in the top half of the assess­ lower-skilled individuals gained the most, suggesting ment task performance. Assessment task results are in green, and that generative AI may serve as a significant equalizer experimental task results are in blue. Dell’Acqua et al.: Navigating the Jagged Technological Frontier 10 Organization Science, Articles in Advance, pp. 1–21, © 2026 The Author(s) in the realm of high-end knowledge work, narrowing Table 6. Inside the Frontier—Timing the performance gap between different skill levels; (1) (2) (3) however, those who were above average also benefited Variable Timing Timing Timing meaningfully through the use of the AI system. Together, these insights highlight the significance of GPT + overview �1,129.143*** �1,105.622*** �1,094.640*** task-specific and contextual considerations for effec­ (135.181) (136.614) (129.681) tively leveraging generative AI. GPT only �1,388.415*** �1,356.364*** �1,351.998*** Downloaded from informs.org by [199.94.42.12] on 17 March 2026, at 08:00 . For personal use only, all rights reserved. For tasks within the frontier of AI capacity, we pre­ (150.204) (152.252) (138.056) vented subjects from completing the experiment before Assessment 159.955* �0.376 the allotted time was over. To achieve this, the final (88.270) (81.973) question was especially long, asking participants to Female 53.466 �4.081 “synthesize the insights you have gained from the pre­ (141.422) (134.884) vious questions and create an outline for a Harvard English Native �70.594 144.690 Business Review-style article of approximately 2,500 (147.939) (143.001) words.” However, although participants were required Low Tenure �89.041 �18.741 to take the full time allotted to this task, we nevertheless (128.835) (117.147) tracked the amount of time that they took to reach this Location 45.125 12.766 last question, having completed the first 17 questions. (151.060) (144.315) Table 6 uses this Timing variable as the dependent vari­ Tech Openness �45.955 �96.996 able. The GPT + overview treatment reduces time spent (132.155) (132.153) on the first 17 questions by 1,129 seconds (18.8 minutes Timing (Assessment) 1.474*** or 22.5% faster than the control), and the GPT-only treat­ (0.183) ment reduces time spent on the first 17 questions by AI-exposure controls Yes 1,388 seconds (23.1 minutes or 27.6% faster than the R2 0.196 0.206 0.345 control). GPT � GPT 0.137 0.155 0.124 + overview Our results reveal significant effects, underscoring Control mean 5,023 5,023 5,023 the prowess of AI even on tasks traditionally executed Observations 385 385 385 by highly skilled knowledge workers. Not only did Notes. This table examines the effects of introducing GPT-4 on timing using AI increase the number of subtasks completed by the experimental task inside the frontier. The dependent variable more than 12%, it also enhanced the quality of the represents the total number of seconds that subjects took to reach the responses by an average of 32%. These effects support final question (question 18). Each column displays the results of a the view that for tasks that are clearly within its frontier distinct linear regression model. Columns (2) and (3) use the average response quality in the assessment task as evaluated by two human of capabilities, even those that historically demanded graders as a control. Column (3) additionally includes the timing intensive human interaction, AI support provides major necessary to reach the last question in the assessment task as well as performance benefits. seven presurvey AI-exposure controls: perceived current task automation (0%–100%); belief that one’s job could be automated; Likert-scale items on familiarity with text-generation tools, image- 4.2. Quality Disruptor—Outside the Frontier generation tools, prompt engineering, and large language model In refining the task within the frontier and recognizing mechanics; and a five-point measure of prior ChatGPT usage the substantial quality and productivity gains enabled frequency (never to daily). Higher scores on all seven variables by integrating AI, we sought to develop a task that parti­ indicate greater AI familiarity or usage. The bottom of the table displays p-values from an F-test comparing the effects of receiving the cipants could not have AI easily complete by simply GPT + overview treatment vs. the GPT-only treatment. All regressions copying and pasting our instructions as a prompt. We include robust standard errors. designed the beyond-the-frontier task in collaboration *p < 0:10; ***p < 0:01. with BCG, using as a starting point the type of business cases that BCG uses for its highly competitive job inter­ participants would have to look at the quantitative data views. These cases are used to gauge the skill and using subtle but clear insights from the interviews. knowledge fit of these workers in the context of tasks of Although the spreadsheet data alone were designed to the kind that they would face in their work with the seem to be comprehensive, a careful review of the inter­ company. This meant creating a task at which knowl­ view notes revealed crucial details. When considered in edge workers would excel but AI would struggle, at totality, this information led to an incorrect conclusion least without extensive guidance and human interven­ regarding what would have been provided by AI when tion. We settled on a task based on an existing business prompted with the exercise instructions, the given data, case that used data from a spreadsheet and a file pre­ and the accompanying interviews. senting interviews with company insiders, which were In this second experiment, the primary objective was adjusted, anonymized, and adapted to the goals of this for subjects to offer actionable strategic recommenda­ experiment. To be able to solve the task correctly, tions to a hypothetical company. First, participants Dell’Acqua et al.: Navigating the Jagged Technological Frontier Organization Science, Articles in Advance, pp. 1–21, © 2026 The Author(s) 11 worked on the assessment task, which required them to Table 7. Outside the Frontier—Correctness analyze the company’s sales distribution channel per­ (1) (2) (3) formance. Using insights from interviews and financial Variable Correctness Correctness Correctness data, participants were asked to provide information and informed advice to the chief executive officer GPT + overview �0.245*** �0.245*** �0.248*** (0.054) (0.054) (0.054) (CEO). Their recommendations were to pinpoint which channel held the most potential for growth. GPT only �0.139*** �0.145*** �0.145*** (0.053) (0.052) (0.053) Downloaded from informs.org by [199.94.42.12] on 17 March 2026, at 08:00 . For personal use only, all rights reserved. Upon completing their assessment task, participants moved to the main experimental task. The focus transi­ Assessment 0.109** 0.117** (0.046) (0.047) tioned from the examination of the company’s distribu­ tion channels to brand analysis, with subjects having to Female �0.063 �0.073 (0.050) (0.053) analyze the company’s brand performance. Similarly to the assessment task, participants used insights from English Native �0.096** �0.097** (0.046) (0.048) interviews and financial data to provide recommenda­ tions to the CEO. Their recommendations were Low Tenure �0.118*** �0.118** (0.045) (0.046) designed to pinpoint which brand held the most poten­ tial for growth. Participants were also expected to sug­ Location �0.098* �0.104* (0.052) (0.053) gest actions to improve the chosen brand, regardless of the exact brand that they had chosen. Details of these Tech Openness 0.012 0.016 (0.050) (0.052) tasks are reported in Online Appendix B. For this task outside the frontier, our primary evalua­ AI-exposure controls Yes R2 0.051 0.099 0.109 tion metric is “correctness,” which is represented as a GPT � GPT + overview 0.082 0.095 0.088 binary variable, where a value of one indicates that sub­ Control mean 0.844 0.844 0.844 jects provided an accurate recommendation and zero Observations 373 373 373 signifies otherwise.20 Figure 5 visualizes the correctness Notes. This table examines the effects of introducing GPT-4 on the percentages for the different groups, highlighting a correctness of the responses for the experimental task outside the noticeable dip in performance among the AI treatment frontier. Each column displays the results of a distinct linear groups when juxtaposed with the control group. Sub­ regression model. The dependent variable across all columns is a binary indicator for whether the subject provided the correct strategic jects in the control group were correct about this exercise recommendation (1 � correct, 0 � incorrect). Columns (2) and (3) about 84.5% of the time, but the AI conditions scored at include a binary correctness metric from the assessment task as 60% and 70.6% (an average decrease of 19 percentage graded by human evaluators and a set of controls. Column (3) points when combining the AI treatment conditions and additionally includes seven presurvey AI-exposure controls: perceived current task automation (0%–100%); belief that one’s job comparing them with the control condition). could be automated; Likert-scale items on familiarity with text- Table 7 captures the impact of the AI treatments on the generation tools, image-generation tools, prompt engineering, and correctness of tasks in the outside-the-frontier experiment large language model mechanics; and a five-point measure of prior using linear regressions with correctness as a binary ChatGPT usage frequency (never to daily). Higher scores on all seven variables indicate greater AI familiarity or usage. The bottom of the dependent variable.21 Both AI treatments—GPT + over­ table displays p-values from an F-test comparing the effects of view and GPT only—show a significant negative impact, receiving the GPT + overview treatment vs. the GPT-only treatment. with the GPT + overview group recording a more pro­ All regressions include robust standard errors. *p < 0:10; **p < 0:05; ***p < 0:01. nounced decrease (24.5% versus 13.9%). Column (2) in Table 7 introduces the performance metric from the assessment, whereas column (3) in Table 7 further refines treatments, the difference in their impacts is statistically the analysis by incorporating the same set of controls as significant at the 10% level across specifications. in Tables 4 and 5. When directly comparing the two Table 8 examines the influence of the AI treatments on the time taken by participants to complete tasks in the Figure 5. Outside the Frontier—Performance outside-the-frontier experiment. The dependent variable here is Timing, which represents how long subjects spent on the task calculated in seconds.22 Column (2) in Table 8 further refines the analysis by incorporating the same set of controls as in Table 4. The findings for both AI treatments—GPT + overview and GPT only—document a reduction in the time spent: more than 11 minutes less for GPT + overview (a 30% decrease in time spent com­ pared with the control mean) and more than 6 minutes Notes. This figure displays average performance for the task outside the frontier. It reports the percentage of subjects in each experimental less for GPT only (a decrease of 18% in time spent com­ group providing a correct response in the experimental task. pared with the control mean). Comparing the two Dell’Acqua et al.: Navigating the Jagged Technological Frontier 12 Organization Science, Articles in Advance, pp. 1–21, © 2026 The Author(s) Table 8. Outside the Frontier—Timing graders were not informed of the correct strategic solu­ tion; their task was to assess the clarity, persuasiveness, (1) (2) (3) Variable Timing Timing Timing and coherence of reasoning. Our subjective coherence quality metric uses the average of these two grades. Col­ GPT + overview �689.191*** �671.526*** �677.139*** umn (1) in Table 9 shows that the treatment GPT + over­ (115.266) (94.987) (96.131) view leads to a 1.47-point-greater score (a 25.1% increase GPT only �407.329*** �279.837*** �287.775*** over the control mean), with GPT only increasing the Downloaded from informs.org by [199.94.42.12] on 17 March 2026, at 08:00 . For personal use only, all rights reserved. (121.833) (95.751) (97.945) score by 1.05 points (a 17.9% increase over the control Assessment Timing 0.681*** 0.681*** mean). Across all specifications, subjects using AI (0.046) (0.046) (whether GPT + overview or GPT only) consistently out­ Female 18.777 5.163 performed those not using AI in terms of subjective (87.671) (90.485) coherence quality, regardless of the correctness of their English Native �114.277 �118.673 answer. When we control for various factors in columns (85.932) (90.779) (2), (3), and (4) in Table 9, the positive impact of AI Low Tenure 82.151 86.135 remains robust. Column (3) in Table 9 repeats the specifi­ (81.736) (83.910) cation in column (2) in Table 9 for the subset of partici­ Location 57.024 48.935 pants with incorrect answers, and column (4) in Table 9 (95.524) (97.050) does the same for those with correct answers. In both Tech Openness 34.603 65.572 instances, the effects of using AI are positive. (85.090) (89.744) Figure 6 illustrates subjective coherence quality condi­ AI-exposure controls Yes tional on whether participants ultimately solved the prob­ R2 0.085 0.407 0.414 lem correctly. Across both the “correct” and “incorrect” GPT � GPT 0.022 0.000 0.000 + overview subgroups, AI-supported participants produced recom­ Control mean 2,260 2,260 2,260 mendations that were more coherent and persuasive, Observations 373 373 373 underscoring that AI improves presentation and argu­ Notes. This table examines the effects of introducing GPT-4 on the mentation even when the underlying analysis is flawed. time that subjects spent completing the experimental task outside the This finding underscores the multifaceted ways in which frontier. Each column displays the results of a distinct linear AI can influence performance in the workflow of highly regression model. The dependent variable across all columns is Timing, which is defined as the number of seconds that subjects spent skilled professionals. on the task before submission. Columns (2) and (3) include the timing Table 10 summarizes our results and presents the test of the assessment task and a set of controls. Column (3) additionally statistics, p-values, and standardized effect sizes com­ includes seven presurvey AI-exposure controls: perceived current paring the treatment conditions with the control group task automation (0%–100%); belief that one’s job could be automated; Likert-scale items on familiarity with text-generation tools, image- across our key outcome measures. Depending on the generation tools, prompt engineering, and large language model outcome variables, we use different statistical tests (t- mechanics; and a five-point measure of prior ChatGPT usage test, chi-squared test, and MWU), which consistently frequency (never to daily). Higher scores on all seven variables show statistical significance across all our main results. indicate greater AI familiarity or usage. The bottom of the table displays p-values from an F-test comparing the effects of receiving Notably, all of these results remain statistically signifi­ the GPT + overview treatment vs. the GPT-only treatment. All cant after applying the Bonferroni correction to adjust regressions include robust standard errors. the significance threshold for multiple comparisons. We ***p < 0:01. are additionally reporting standardized effect sizes (Cohen d, Cohen h, and rank biserial) to convey the mag­ coefficients reveals that GPT + overview recorded a nitude of any significant differences. The effect sizes for more substantial—and statistically significant—decrease in time spent compared with the GPT-only group. our main findings range from moderate to large, with Table 9 further examines the persuasiveness and inter­ most falling in the medium to large range, underscoring nal coherence of recommendations in the outside-the- the substantive impact of our interventions. frontier experiment, independent of whether the under­ lying solution was correct. The dependent variable, Sub­ 5. Robustness Checks with Shoe jective coherence quality, captures how well arguments Design Experts were structured and supported as assessed on a 1–10 5.1. Individual Shoe Design Ideas scale. Two independent sets of graders evaluated each To further test the reliability of the findings presented in response: BCG consultants not involved in the experi­ the main paper, we replicate our analysis on a new set ment and business school students with prior grading of grading data. Specifically, we carry out this replica­ experience. Both groups followed a rubric developed by tion using the grades that artist evaluators assigned to the authors based on workplace evaluation practices (see shoe design ideas from the first question of the field Online Appendix E for the full rubric). Importantly, experiment. Dell’Acqua et al.: Navigating the Jagged Technological Frontier Organization Science, Articles in Advance, pp. 1–21, © 2026 The Author(s) 13 Table 9. Outside the Frontier—Subjective Coherence Quality (1) (2) (3) (4) Subjective Subjective Subjective Subjective Variable Coherence Quality Coherence Quality Coherence Quality Coherence Quality GPT + overview 1.475*** 1.330*** 1.448** 1.493*** (0.242) (0.229) (0.718) (0.248) GPT only 1.046*** 0.926*** 1.777** 0.724** Downloaded from informs.org by [199.94.42.12] on 17 March 2026, at 08:00 . For personal use only, all rights reserved. (0.289) (0.263) (0.805) (0.285) Assessment—Subjective 0.370*** 0.337*** 0.401*** Coherence Quality (0.053) (0.125) (0.056) Female �0.437* �0.884 �0.224 (0.238) (0.569) (0.255) English Native �0.102 �0.655 0.223 (0.218) (0.534) (0.238) Low Tenure �0.038 �0.610 0.229 (0.211) (0.556) (0.218) Location 0.297 �0.009 0.465* (0.231) (0.515) (0.245) Tech Openness 0.219 �0.150 0.344 (0.220) (0.508) (0.229) AI-exposure controls Yes Yes Yes R2 0.085 0.246 0.222 0.328 GPT � GPT + overview 0.098 0.103 0.531 0.007 Control mean 5.856 5.856 5.325 5.954 Observations 372 372 105 267 Notes. This table examines the effects of introducing GPT-4 on the subjective coherence quality of the recommendations provided in the experimental task outside the frontier. Each column displays the results of a distinct linear regression model. The dependent variable across all columns is Subjective Coherence Quality, which is measured on a 1–10 scale. Column (2) includes the Subjective Coherence Quality of the assessment task, a set of controls, and seven presurvey AI-exposure controls: perceived current task automation (0%–100%); belief that one’s job could be automated; Likert-scale items on familiarity with text-generation tools, image-generation tools, prompt engineering, and large language model mechanics; and a five-point measure of prior ChatGPT usage frequency (never to daily). Higher scores on all seven variables indicate greater AI familiarity or usage. Columns (3) and (4) run the same regression as column (2) using different samples. Column (3) takes into account only subjects who provided an incorrect response to the experimental task. Column (4) takes into account only those who provided a correct response. The bottom of the table displays p-values from an F-test comparing the effects of receiving the GPT + overview treatment vs. the GPT- only treatment. All regressions include robust standard errors. *p < 0:10; **p < 0:05; ***p < 0:01. 5.1.1. Data Collection. The data set of shoe design ideas five graders only evaluated half of the ideas. To account is derived from the first question of the “inside-the- for this constraint, each of these two graders was frontier” task. Participants were asked to “generate assigned a mutually exclusive, randomly selected subset ideas for a new shoe aimed at a specific market or sport comprising 50% of the data set. In total, every shoe that is underserved.” To conduct this analysis, we disag­ design idea received four different evaluations by four gregated each list into individual shoe design ideas, graders. anonymized participant identifications, and presented These new grading data differ from the initial data in the ideas to our graders in a randomized order.23 Gra­ four important ways. First, instead of distributing 18 ders were instructed to rate the creativity of each idea grades across 18 different questions as in the original on a scale from 1 to 10, disregarding grammatical and “inside-the-frontier” analysis, this approach leverages spelling issues. The graders were not informed about at least 10 independent grades on just 1 question, the study’s purpose or that some ideas were generated thereby limiting the analysis’s scope but preserving its with the help of AI. robustness. Second, the grades are based on creativity Five graders evaluated ideas. They were selected rather than general quality, allowing us to directly test from a pool of applicants through resume screening and the notion that AI tools enhance human creativity. interviews, and they were filtered based on their rele­ Third, the evaluators have extensive experience in foot­ vant experience. All evaluators were students majoring wear design rather than a general business background, in product design or industrial design at top programs suggesting that their creativity scores hold greater valid­ with demonstrated experience in footwear design. Three ity within this specific field. Fourth, this data set uses graders evaluated all shoe design ideas, and two of the four different evaluations instead of three. Dell’Acqua et al.: Navigating the Jagged Technological Frontier 14 Organization Science, Articles in Advance, pp. 1–21, © 2026 The Author(s) Figure 6. Outside the Frontier—Subjective Coherence the main analysis.25 In column (1) in Table 11, the GPT + Quality overview and GPT-only treatment groups had creativ­ ity scores that were 14% and 11.7% higher than the con­ trol group, respectively, as compared with 33.9% and 29.9% in the main analysis, respectively. 5.2. Full Replication with Shoe Design Experts Downloaded from informs.org by [199.94.42.12] on 17 March 2026, at 08:00 . For personal use only, all rights reserved. To further validate our findings and address feedback on the subjectivity of our evaluation metrics, we expanded our robustness checks with the shoe design experts. Although the earlier analysis focused on the creativity of ideas from the first question, this expanded analysis cov­ ers all 18 questions of the “inside-the-frontier” task. 5.2.1. Methods. We worked with three independent shoe design experts among those who evaluated ideas in Notes. This figure displays the average performance of subjects who were correct in the experimental task outside the frontier (on the left) Section 5.1 to evaluate responses for all 18 questions. To and those who were incorrect on that task (on the right). The green ensure high-quality and thorough assessments, each bars represent the subjective coherence quality of the control group, question for every participant was graded by two of the whereas the blue bars indicate the average subjective coherence qual­ ity of the treatment groups. The y axis denotes the average scores, three experts.26 The experts were compensated on an ranging from 1 to 10. hourly basis, and just like with our original graders, they were given unlimited time to complete their evaluations, 5.1.2. Results. The replication results are shown in removing any time pressure that might compromise the Table 11.24 As expected, both treatments—GPT + over­ quality of their assessments. As with the initial analysis, view and GPT only—demonstrate positive effects graders were blind to the experimental conditions. across all specifications. As with the original analysis, the GPT + overview treatment exhibits a more pro­ 5.2.2. Results. The results from this expanded expert nounced positive effect than GPT only, although this evaluation are presented in Table 12. The findings are difference is not statistically significant in any of the spe­ entirely consistent with the results reported in the main cifications. The magnitude of the effect is lower than in paper and in fact, demonstrate even stronger positive Table 10. Test Statistics and Effect Sizes—Treatments vs. Control Test statistic Effect size Variable Treatment effect Name Value p-value Name Value Panel A: GPT only vs. control group Inside the frontier Quality 1.307 MWU 15,674 <0.000 Rank biserial 0.869 Completion 0.090 t-test 4.349 <0.000 Cohen d 0.528 Timing �1,388.415 t-test �9.244 <0.000 Cohen d �1.679 Outside the frontier Correctness �0.139 Chi squared 9.481 0.002 Cohen h �0.337 Timing �407.329 t-test �3.343 0.001 Cohen d �0.444 Subjective Coherence Quality 1.046 MWU 10,021 <0.000 Rank biserial 0.317 Panel B: GPT + overview vs. control group Inside the frontier Quality 1.483 MWU 16,079 <0.000 Rank biserial 0.963 Completion 0.111 t-test 5.583 <0.000 Cohen d 0.651 Timing �1,129.143 t-test �8.352 <0.000 Cohen d �1.365 Outside the frontier Correctness �0.245 Chi squared 17.874 <0.000 Cohen h �0.560 Timing �689.191 t-test �5.980 <0.000 Cohen d �0.752 Subjective Coherence Quality 1.475 MWU 11,400 <0.000 Rank biserial 0.414 Notes. This table shows test statistics and effect sizes for the treated conditions with respect to the control group. For rows with an ordinal outcome variable, a Mann–Whitney U test (MWU) is used for the test statistic, and rank biserial is used for the effect size. For rows with a binary outcome variable, a chi-squared test is used for the test statistic, and Cohen h is used for the effect size. All other rows use t-tests and Cohen d. Dell’Acqua et al.: Navigating the Jagged Technological Frontier Organization Science, Articles in Advance, pp. 1–21, © 2026 The Author(s) 15 Table 11. Inside the Frontier—Quality (Individual Ideas) work. Using a preregistered, randomized laboratory-in- the-field experiment, we test the potential dual role of AI (1) (2) (3) Variable Quality Quality Quality in augmenting the efforts of knowledge workers, recog­ nizing that it functions as both a booster—enhancing effi­ GPT + overview 0.566*** 0.584*** 0.581*** ciency and productivity on tasks within its capability (0.143) (0.138) (0.131) frontier—and a disruptor, which negatively impacts per­ GPT only 0.471*** 0.491*** 0.483*** formance on tasks beyond its frontier. Our findings under­ Downloaded from informs.org by [199.94.42.12] on 17 March 2026, at 08:00 . For personal use only, all rights reserved. (0.077) (0.073) (0.073) score the transformative potential of AI and offer insights Assessment 0.075*** 0.081*** into harnessing its capabilities for optimal outcomes. (0.024) (0.023) A crucial feature of our experiment was the nature of Female �0.002 �0.020 our experimental subjects. Specifically, we were able to (0.067) (0.069) recruit from a highly skilled and motivated population, English Native �0.007 �0.007 with participants who engaged in tasks that were devel­ (0.058) (0.060) oped in collaboration with our partner company. These Low Tenure �0.175*** �0.177*** tasks closely mirrored their professional activities as (0.036) (0.036) knowledge workers, allowing us to assess AI’s impact Location 0.049 0.042 on tasks that approximate real-world workflows while (0.065) (0.065) recognizing that they do not fully capture every com­ Tech Openness �0.073** �0.001 plexity of on-the-job work. (0.031) (0.037) We found that the utility of AI can fluctuate across a AI-exposure controls Yes professional’s workflow, with some tasks falling inside R2 0.015 0.018 0.019 the frontier, whereas others fall outside of it. Figure 7 GPT � GPT 0.290 0.302 0.216 + overview displays the treatment effect sizes in percentage change Control mean 4.043 4.043 4.043 for both inside and outside the frontier, showing the Observations 15,404 15,404 15,404 magnitude of our treatment effects conditions across all Notes. This table reports results for the inside-the-frontier quality outcome using only responses to question 1. The dependent variable is overall quality of the idea measured on a standardized scale and Table 12. Inside the Frontier—Quality (Design Experts) evaluated by human graders. Each column displays the results of a distinct linear regression model. Columns (2) and (3) incorporate the (1) (2) (3) average response quality from the assessment task and a set of Variable Quality Quality Quality controls. Column (3) includes seven presurvey AI-exposure controls. The bottom of the table displays p-values from an F-test comparing the GPT + overview 1.748*** 1.753*** 1.775*** effects of receiving the GPT + overview treatment vs. the GPT-only (0.067) (0.063) (0.066) treatment. All regressions include robust standard errors. GPT only 1.613*** 1.616*** 1.629*** **p < 0:05; ***p < 0:01. (0.072) (0.071) (0.071) Assessment 0.182** 0.175** effects of AI on performance. Across all specifications, (0.079) (0.081) both the GPT + overview and GPT-only treatments show Female �0.121* �0.130** a significant and substantial improvement in the quality (0.063) (0.064) of work produced by participants. The effect sizes are English Native 0.103 0.113 larger than those observed in the main analysis, reinfor­ (0.068) (0.071) cing our conclusion that AI assistance enhances the qual­ Low Tenure 0.039 0.038 ity of knowledge work on tasks inside the frontier. These (0.057) (0.058) positive effects remain statistically significant when the Location 0.054 0.050 analysis is disaggregated by question type (creativity, ana­ (0.073) (0.075) lytical thinking, writing proficiency, and persuasiveness). Tech Openness 0.078 0.074 These robustness checks support our main findings. (0.060) (0.067) Additionally, this convergence across diverse evaluator AI-exposure controls Yes R2 0.663 0.685 0.690 types and methodologies provides strong evidence GPT � GPT + overview 0.070 0.060 0.049 against systematic evaluator bias and the hypothesis Control mean 4.498 4.498 4.498 that graders are systematically fooled by AI-generated Observations 385 385 385 content. Notes. This table replicates the inside-the-frontier quality analysis using evaluations provided by a new set of design expert graders. The 6. Discussion dependent variable is overall quality measured on a standardized scale. The specifications parallel those in the main inside-the-frontier quality In this paper, we study the use of AI with a jagged capabil­ analysis. Robust standard errors are reported in parentheses. ity frontier by humans performing high-end knowledge *p < 0:10; **p < 0:05; ***p < 0:01. Dell’Acqua et al.: Navigating the Jagged Technological Frontier 16 Organization Science, Articles in Advance, pp. 1–21, © 2026 The Author(s) Figure 7. Summary—Effect Sizes Downloaded from informs.org by [199.94.42.12] on 17 March 2026, at 08:00 . For personal use only, all rights reserved. Notes. The figure displays the treatment effect size (in percentage change) for several metrics in two experimental conditions: GPT only (green dots) and GPT + overview (red dots). Subj. Coh., subjective coherence. of our key outcome measures. For tasks inside the fron­ incommensurability of knowledge work outputs. Never­ tier, our findings have substantial and positive perfor­ theless, subjective metrics may contain evaluation bias, mance implications. Across 18 realistic business tasks— representing a key limitation of our study and suggesting ranging from creative to analytical tasks—AI signifi­ important avenues for developing additional objective cantly improved performance and quality for every performance proxies in future research. model specification, increasing speed by more than Our findings should be interpreted within important 25%, performance by more than 30%, and task comple­ boundary conditions regarding incentive structures. Our tion by more than 12%. However, for a task outside the results apply most directly to knowledge work contexts frontier, subjects using AI were 19 percentage points where quality within allocated time frames is prioritized less likely to produce correct solutions. Although this over speed optimization, such as is the case at BCG and design highlights a single outside-the-frontier task—an similar professional service environments. In settings inherent limitation—it underscores that even a single where efficiency gains translate directly into rewards or complex challenge beyond AI’s current capabilities can where explicit accuracy-speed trade-offs exist—such as disproportionately reduce overall performance if users manufacturing, customer service with call-time metrics, rely too heavily on AI’s output. These results suggest a or piecework environments—the effects of AI deploy­ need to understand the frontier’s shape and position as ment may differ substantially. Future research should well as how it is perceived by knowledge workers to investigate how the jagged frontier manifests across dif­ better determine AI’s influence on work. ferent incentive structures in knowledge work. Our design employs different evaluation methods by A further caveat concerns who actually chooses to task type, directly mirroring how consulting work is adopt generative AI when its use is discretionary rather assessed in practice. For writing-intensive activities in­ than strongly encouraged, such as in our setting. Recent side the frontier, performance is evaluated through sub­ evidence shows sizable heterogeneity; early adopters jective rubrics, matching BCG’s standard procedures are disproportionately male and more technologically because no single correct answer exists. Conversely, confident, and the resulting performance gains can the strategy case outside the frontier features an objec­ widen pre-existing gender- and skill-gaps rather than tively correct recommendation that is measured dichoto­ close them (Otis et al. 2024a, Humlum and Vestergaard mously and supplemented with separate persuasiveness 2025, Blandin et al. 2026). If such adoption gaps persist, assessments. This hybrid approach reflects again the the technology could widen existing promotion Dell’Acqua et al.: Navigating the Jagged Technological Frontier Org