dell-acqua-et-al-2026-navigating-the-jagged-technological-frontier 5c589c8c-fbb5-458f-b285-c944746cd717
The paper introduces the concept of a “jagged technological frontier” to capture the uneven impact of generative AI on knowledge‑intensive work: AI can boost performance on some tasks while impairing it on others, even when tasks appear similarly difficult to humans. Using a preregistered, randomized laboratory‑in‑the‑field experiment with 758 Boston Consulting Group (BCG) consultants, the authors compare three conditions – no AI, GPT‑4 access, and GPT‑4 plus a prompt‑engineering overview – across two sets of tasks. The “inside‑the‑frontier” tasks are designed to be within GPT‑4’s current capabilities (creative product ideation, analytical market segmentation, writing, persuasiveness). The “outside‑the‑frontier” task is a strategic case that requires nuanced data interpretation and judgment, deliberately placed beyond GPT‑4’s reliable performance. Results show that AI access increases task completion by 12.2 %, speeds work by ~25 % and raises quality scores by 27‑33 % for inside‑the‑frontier tasks, with the prompt‑engineering overview providing modest additional gains. Conversely, for the outside‑the‑frontier task AI use reduces correctness by 10‑30 % points and still speeds completion, indicating a performance trade‑off. The benefits are larger for lower‑skill consultants, suggesting AI can act as an equalizer. Grading was performed by MBA‑trained human graders using detailed rubrics focusing on creativity, analytical thinking, writing proficiency, and persuasiveness; GPT‑4 was also used to grade and showed >97 % agreement with human scores. Additional AI‑exposure covariates (perceived automation risk, familiarity with text‑, image‑generation tools, prompt‑engineering, and LLM workings, plus usage frequency) were included as controls. Binary demographic and role variables (gender, native English, tenure ≤1 year, Europe/Middle‑East location, tech openness) were also modeled. Robustness checks include alternative grading by shoe‑design experts (creativity‑focused) and replication of main effects using different estimation models (linear probability and logistic regressions). Qualitative follow‑up work identified three human‑GenAI collaboration modes (cyborgs, centaurs, self‑automators). The discussion highlights implications for firms: the need to map task‑level AI capability, provide training, redesign workflows, and manage over‑reliance, while acknowledging boundary conditions (incentive structures, adoption heterogeneity) and future research directions. The paper contributes to the emerging literature on AI‑augmented knowledge work by empirically documenting both the productivity gains and the risks associated with the jagged AI frontier.
Topics
Jagged Technological Frontier: Uneven AI Impact on Knowledge Work
Introduces the ‘jagged frontier’ framework, illustrating how generative AI can markedly improve performance on some knowledge‑intensive tasks while impairing it on others, even when tasks appear similarly difficult to human workers.
Experimental Design and Sample
Details the preregistered, randomized laboratory‑in‑the‑field experiment with 758 Boston Consulting Group consultants, covering the three conditions (no AI, GPT‑4 access, GPT‑4 + prompt‑engineering overview), the task sets, and the inclusion of AI‑exposure covariates and demographic binary controls.
Inside‑the‑Frontier Tasks: Productivity, Speed, and Quality Gains
Reports that AI access raises task completion by 12.2 %, cuts execution time by roughly 25 %, and boosts quality scores by 27‑33 % on tasks within GPT‑4’s capabilities (creative ideation, market segmentation, writing, persuasiveness). The prompt‑engineering overview adds modest incremental benefits.
Outside‑the‑Frontier Task: Performance Deterioration
Shows that for a strategically complex case deliberately placed beyond GPT‑4’s reliable performance, AI use reduces correctness by 10‑30 % points despite faster completion, highlighting the risk of over‑reliance on AI for tasks beyond its frontier.
Skill‑Level Heterogeneity: AI as an Equalizer
Analyzes heterogeneous effects, finding larger productivity and quality gains for lower‑skill consultants, suggesting that generative AI can narrow performance gaps among knowledge workers.
Grading Protocols and Robustness Checks
Describes the rubric‑based evaluation by MBA‑trained human graders, the parallel GPT‑4 grading with >97 % agreement, and robustness replications using shoe‑design experts and alternative statistical models (linear probability and logistic regressions) to rule out grader bias.
Organizational Implications and Recommendations
Discusses practical steps for firms: mapping task‑level AI capability, delivering prompt‑engineering training, redesigning workflows, and instituting safeguards against automation bias and over‑reliance.
Limitations, Boundary Conditions, and Future Research
Acknowledges constraints such as incentive‑structure dependence, heterogeneous adoption, and the need to extend the jagged frontier framework to other contexts and AI generations.