Governing Large Reasoning Models in Enterprise Decision Systems: Transparency, Human Oversight, and Risk Classification

Abstract

The enterprise decision systems are increasingly being burdened with large reasoning models (LRMs), such as large language models used for complex inference, planning and judgement support. However, their applications in credit, compliance, procurement, the assessment of human risk, healthcare, and the making of operating decisions involve tricky governance issues related to transparency, man-in-the-middle and risk categorization. An enterprise governance model for LRMs is designed and tested with a single city based empirical design on Bengaluru (Karnataka), India. Six proposed constructs of the proposed model are shown: transparency, explainability, human oversight, risk classification, intention to use LRM decisions, and perceived quality of decisions, the latter being the model outcome. A carefully designed question paper was created to address professionals from the Bengaluru based companies with expertise in areas such as AI, Data Governance, Compliances, Product Management, Analytics, and Decision System. In this work development draft, a set of 300 responses was to be used for the empirical design, measurement model, and analysis technique of this manuscript-development draft. These included reliability analysis, exploratory factor analysis, correlation analysis, multiple regression analysis, moderation analysis, mediation testing, analysis of variance and governance risk-mapping. The findings reveal a relationship between transparency/explainability and trust in LRM decisions, but not a relationship between human oversight/risk classification and perceived decision quality. The influence of trust on the adoption intention is counterparts with perceived decision quality, and the relationship between the risk classification and perceived decision quality is strengthened by the influence of decision criticality. The study offers a context-sensitive governance framework for Indian enterprise decision systems as well as shows how the technical model assurance needs to be integrated with institutional accountability, documentation, escalation, and human-in-the-loop review. The paper argues that LRMs are not just software production systems or productivity systems but socio-technical decision infrastructures that need to be managed according to risk levels instead of solely like software systems.