Regulating AI Procurement in the Public Sector: Lessons from Defence, Healthcare, and Criminal Justice

Governments are among the largest buyers of artificial intelligence systems, yet public sector AI procurement remains ad hoc, poorly standardized, and frequently captured by vendor marketing narratives that promise transformation while obscuring risk. This paper examines AI procurement practices across three high-stakes government domains - defence, healthcare, and criminal justice - identifying four systematic failures that collectively constitute what this paper terms the Public Sector AI Procurement Gap: evaluation criteria optimized for cost and delivery speed rather than for fairness, transparency, and long-term performance; vendor evaluation frameworks that lack the technical depth to assess model quality, bias, and robustness; contracts that fail to specify performance degradation thresholds, retraining obligations, or liability allocation for algorithmic harm; and accountability vacuums when deployed AI systems produce harmful outcomes with no clear institutional owner of responsibility. Drawing on procurement case studies from the US Department of Defense’s AI acquisition programs, the UK National Health Service’s AI deployment initiatives, and Indian criminal justice and social welfare AI pilot programs, the paper proposes the Public Sector AI Procurement Standard (PSAPS) - a five-component framework that includes mandatory algorithmic impact assessments, independent technical evaluation, AI-specific contract clauses, continuous performance monitoring, and structured accountability allocation. The paper argues that procurement reform is the most effective and politically feasible lever for improving public sector AI governance - more impactful than broad AI legislation (which takes years to enact), more specific than ethical guidelines (which lack enforcement), and more actionable than regulatory frameworks (which require institutional infrastructure most governments lack). The paper concludes with implications for government procurement officers, public policy researchers, AI vendors selling to government, and civic technology advocates.