What Makes Human-AI Collaboration Work in Organisations: An Empirical Framework of Four Coordinated Dimensions

Details

Supervisors

Chevalier, Philippe

Faculty

Louvain School of Management

Degree label

Master [120] : ingénieur de gestion, à finalité spécialisée

Abstract

enGenerative artificial intelligence has entered the workplace faster than almost any technology before it, yet aggregate productivity has not risen in proportion. This thesis treats that disconnect as a management problem rather than a technological one. Its empirical starting point is the meta-analysis by Vaccaro et al. (2024), which found that across 106 experimental studies, human-AI combinations performed, on average, worse than the better of humans or AI alone. The thesis asks how organisations can coordinate task allocation, calibration and interface and process design to sustain human-AI collaboration. The argument is structured by the automation-augmentation paradox of Raisch and Krakowski (2021), which holds that automation and augmentation are interdependent forces rather than a simple trade-off, so that sustainable productivity gains depend on managing the tension between them. From this lens the thesis derives three organisational levers: task allocation which maps work to the right actor along the jagged technological frontier (Dell'Acqua et al., 2023); calibration which concerns whether users hold accurate beliefs about AI reliability (Caplin et al., 2025); and interface and process design which configures systems to encourage appropriate reliance (Al-Refai et al., 2025). The three are integrated into a framework specified through six propositions. The framework is tested through an abductive design (Dubois & Gadde, 2002) drawing on 28 semi-structured interviews across approximately 22 firms and 14 industries, analysed with the Gioia method (Gioia et al., 2013). The original three-department sampling frame did not hold once data collection began and recruitment was redirected toward deployment-side voices with cross- departmental visibility. The framework survives confrontation with the evidence in a refined form: two propositions are reformulated, one is restated and a fourth dimension, the junior talent pipeline, emerges from the data. Two cross-cutting findings are named: human accompaniment as the binding condition under which the levers operate and pace-mismatch as the pressure exerted by rapid AI tool evolution. The thesis concludes that the tension is resolved not at the level of any single lever but across the four dimensions together, under the condition of human accompaniment.