Position: Adopting AI in Practice Does Not Guarantee the Productivity Boost

Synthesis: This ICML 2026 position paper argues that adopting AI in organizational practice does not automatically yield productivity gains — human and environmental factors critically moderate the relationship. Drawing on the partial equilibrium model of Gries and Naudé (2022), it identifies five key moderators that can attenuate or negate productivity benefits.

Five Moderating Factors

1. Human resource composition — the mix of skills, roles, and experience in the workforce
2. Baseline capability of individuals — pre-existing competence before AI introduction
3. Learning curve of practitioners — how quickly users adapt to AI tools
4. Incentives for fair use — motivations for appropriate and ethical AI usage
5. Flexibility of objectives — organizational ability to adjust goals with AI integration

What this means for practice

• Instructors. Budget for the learning curve before promising time savings: if practitioners cannot adapt quickly, AI tools may reduce rather than increase productivity — the same overestimation pattern documented in Cognitive offloading and the speedup illusion in human-AI interaction.
• Instructors. Set explicit expectations for fair and appropriate use so any efficiency gain does not come from surface-level shortcuts; the incentives moderator connects directly to Academic Integrity and to the risk that deeper learning is bypassed.
• Administrators. Treat human-resource composition and staff baseline capability as variables you manage rather than fixed conditions — the paper's central revision is that these five factors are endogenous organizational choices, not exogenous parameters.
• Administrators. Fund training, incentives, and the flexibility to revise objectives alongside AI deployment, since the framework predicts that adoption without attention to those human systems falls short of promised gains.

Limitations

• This is a position paper: it revises the partial equilibrium model of Gries and Naudé (2022) and presents no new empirical estimate of how much each moderator attenuates productivity gains.
• The five moderators are argued to be endogenous rather than measured, so the framework's predictions about training, incentives, and flexible objectives remain untested within the paper.
• Its larger claims rest on external evidence — cross-country firm-level findings that AI adoption concentrates in large, already-productive firms and Acemoglu's (2025) lower aggregate productivity estimates — rather than on data the authors collect.
• The authors acknowledge the limits may prove transitional rather than structural, citing electricity adoption, which required decades of factory reorganization before productivity benefits appeared; the framework cannot say how long the gap will persist.

Citation

Cho, W. I., Kim, S., & Kim, G. (2026). Position: Adopting AI in practice does not guarantee the productivity boost. Accepted at ICML 2026. cs.CY.