AI Adoption in Small and Medium Enterprises: Complementarity Thresholds and the Dual-Pathway Problem
DOI: https://doi.org/10.5281/zenodo.21704615
Dr. Neha Bhatt, Ms. Pooja Rana, Ms. Shivali Yadav, Ms. Deepika Gehlan
Views: 55 | Downloads: 35
Keywords:
artificial intelligence; SMEs; technology adoption; complementary assets; absorptive capacity; digital transformation; productivity paradox; competitive advantage
Abstract:
AI has been touted as a technology that empowers small and medium enterprises (SMEs) to get into the game of doing what big business does. But the adoption evidence tells another story: AI adoption continues to be very firmly entrenched at the very top of the market, among the largest, youngest and most digital companies, and the adoption gap has not narrowed with the decreasing costs of tools. This paper proposes that the lack of progress towards closing the gap is not so much caused by awareness, cost and attitude the three factors that have been central in the SME adoption literature but by threshold effects in complementary assets that may influence whether this gap yields value or not. We create a Dual-Pathway Complementarity Threshold (DPCT) framework, separating embedded adoption (passive use of AI functions within vendor software) from deliberate adoption (reconfiguring processes and roles around the AI function). The two pathways have fundamentally different strategic implications: while the former (embedded) is accessible to almost all SMEs, and therefore does not provide a durable advantage after diffusion, the latter (deliberate) does provide an appropriable advantage, but would involve crossing thresholds of data adequacy, absorptive capacity, process formalisation, and governance that most SMEs would not cross. The framework generates six propositions, five structural asymmetries in the SME/large-firm adoption and a re-interpretation of the policy problem as complementary asset formation instead of technology diffusion. This paper does not gather firm-level data needed to produce credible estimates of these parameters; so no invented statistics are reported; instead, a clear simulation program is programmed which packages the threshold value function and rent dissipation dynamics with a template of estimation notation using placeholders.