The Algorithmic Bureaucrat: Dismantling Clientelist Networks in Public Procurement Through Agent-Based Modeling

The transition from e-Government to Intelligent Governance promises to dismantle systemic corruption by replacing human discretion with algorithmic objectivity, yet empirical validation remains scarce given data opacity. Using Agent-Based Modeling (ABM) calibrated with confidential micro-data from a Mexican public agency, this quasi-experimental study simulates the structural impact of Artificial Intelligence (AI) on clientelist networks. The framework integrates rational choice with behavioral psychology (moral disengagement, Theory of Planned Behavior, inequity aversion), providing a conservative lower bound. We compare three regimes: a discretionary Status Quo, an ex-post AI Auditor, and an ex-ante AI Meritocracy. Under AI Meritocracy, the Herfindahl-Hirschman Index (HHI) collapsed from 1,752 (oligopolistic capture) to 218 (competitive market), dissolving the cartel without legal intervention; cost overruns fell from 19.24% to 10.69% and bribery from 95% to near-zero. Outcomes are invariant to the distribution of bureaucratic integrity: algorithmic allocation removes the human decision point that integrity governs. A sensitivity analysis with up to 20% “gaming” agents (suppliers inflating quality signals to manipulate the score) confirms deconcentration persists (HHI below 275 in all 150 runs) while efficiency erodes moderately (overruns reach 13.3%), making quality verification the binding design constraint. A bounded-memory learning check (Erev–Roth reinforcement with forgetting) leaves structural results unchanged but shows surveillance deterrence decaying as agents forget sanctions. These findings substantiate the “Automated Bureaucratic Model”: the most effective anti-corruption strategy is not surveillance but algorithmic removal of the discretionary market for favors. Findings are Mexico-calibrated; endogenous gaming coordination, cultural moderators, and psychological reactance remain priority extensions.