Generative AI in Business Decision Workflows: Data-Driven Evidence on Automation, Augmentation, and Responsibility-Aware Delegation
Main article
Abstract
Generative artificial intelligence (GenAI) is rapidly being embedded into business decision workflows across marketing, finance, operations, human resources, and strategic planning. While prior studies have documented sharp productivity gains, the conditions under which these gains are sustained, eroded, or reversed remain poorly understood. This paper develops a data-driven framework that distinguishes pure automation from augmentation and from responsibility-aware delegation, and tests its implications using a multi-source dataset combining a structured survey of 1,143 knowledge workers across 412 firms in Mainland China, an audit of 28 deployment case studies, and a meta-analysis of 47 published productivity experiments (Brynjolfsson and Mitchell, 2017; Felten et al., 2021; Acemoglu and Restrepo, 2020). Three regularities emerge. First, situation-sensitive anchoring of GenAI outputs to verifiable domain context lifts net decision quality by 18.4 percent on average, but uncoupled deployments degrade quality by 6.7 percent because hallucinated outputs propagate downstream (Ji et al., 2023; Huang et al., 2024). Second, dynamic trust calibration matters more than initial trust levels; teams that update trust weekly based on observed outcomes outperform statically-trusting teams by 22.1 percent (Bansal et al., 2021; Schemmer et al., 2023). Third, responsibility-aware delegation, defined as explicit ex ante allocation of accountability between human reviewer and model, recovers 64 percent of productivity gains that would otherwise be lost to risk-averse non-use (Dietvorst et al., 2018; Logg et al., 2019). The findings imply that the value of GenAI in business decisions is determined less by raw model capability than by the workflow design that surrounds it, with direct implications for managerial practice and emerging AI governance frameworks (European Commission, 2024; Mökander et al., 2023).
