Risk-Tiered Classification of Large Language Models in Public Sector Decision-Making

Abstract

Large Language Models (LLMs) are increasingly embedded in public sector workflows, ranging from administrative drafting and citizen information services to professional advisory systems, eligibility support, and rights-affecting decision processes. This study proposes a risk-tiered classification framework for assessing LLM deployment in public sector decision-making. The framework evaluates six governance dimensions: decision criticality, system autonomy, data sensitivity, harm potential, explainability demand, and oversight adequacy. Using a simulated public sector use-case mapping, the analysis classified 103 LLM applications across five decision domains. Administrative support represented the largest share with 34 mapped use cases, followed by citizen information with 27, professional advisory with 21, eligibility support with 13, and rights-affecting support with 8. The results show that administrative support was predominantly classified as minimal risk, with 28 out of 34 cases falling into this tier, while citizen information systems were mostly classified as limited risk, with 18 out of 27 cases. Professional advisory systems showed a mixed profile, with 10 limited-risk and 9 high-risk cases, indicating sensitivity to institutional reliance. Eligibility support was dominated by high-risk classification, with 9 out of 13 cases, while rights-affecting support produced the most severe profile, with 5 out of 8 cases classified as unacceptable risk. The study contributes an operational governance model that connects LLM risk classification with concrete responses, including documentation, disclosure, logging, human review, audit, appeal mechanisms, redesign, and prohibition.