Coarsen or refine? coordination cost and the modular architecture of digital platforms under AI-driven complexity

Abstract Platform owners that build their own technology must decide whether rising, AI-driven complexity calls for coarser (more integrated) or finer (more decoupled) module boundaries. We develop a computational model in which a platform searches a rugged performance landscape under an endogenous coordination cost, and show that the answer depends on the source of that cost. When coordination cost is structural, scaling with unit size, the performance-maximising architecture coarsens as complexity rises; when it is transactional, scaling with the interdependence a change must renegotiate, the architecture refines, and a critical cost mix separates the two. We then make AI a decision variable. The platform chooses how much AI to deploy. Deployment raises interdependence and lowers decomposability, AI absorbs part of the coordination cost of the changes it handles, each use carries a cost, and capacity is limited. Against a no-AI benchmark and a treatment in which AI only changes parameters, deployment moves the optimal architecture to integration in both cost regimes, to about eleven of twelve decisions per module, while the parameter-only treatment leaves granularity unchanged. The usage cost governs this shift, and integration economises on scarce AI capacity. Along an adoption path, decomposability carries about 90% of the erosion of modular value. Modelling autonomous AI agents as look-ahead searchers yields governance rules, including a decomposability threshold for re-centralisation. The study recasts the modularity–complexity debate as two questions a platform can diagnose: which coordination cost dominates, and which one its AI absorbs.