Paper Type

Complete

Abstract

As AI systems are increasingly embedded in organizational decision-making, their assured outputs risk conveying unwarranted interpretive authority—particularly where policies, contracts, and compliance documents intend ambiguity and delegated discretion. We reconceptualize policy automation as a human–AI co-agency problem and, using elaborated Action Design Research (eADR) methods, derive governance constraints for preserving traceability, bounding autonomy, and governing shared vocabulary evolution. We design a governance-oriented artifact that transforms unstructured organizational documents into structured, machine-readable rule representations while maintaining institutional control. The design separates semantic extraction from rule formation, constrains representation through a shared vocabulary catalog, and supports controlled, mixed-initiative expansion. Rather than resolving ambiguity, it preserves modality and handles uncertainty through under-specification and explicit uncertainty markers. Evaluation on real-world university policy documents demonstrates deterministic, traceable rule generation that maintains institutional discretion while enabling structured representation. This study contributes a governance-oriented design and transferable design principles for embedding bounded AI autonomy into systems operating over ambiguity-rich organizational documents.

Paper Number

1670

Comments

NEXTTRANS

Share

COinS
 
Aug 15th, 12:00 AM

Careful Policy Interpretation with RAG: A Design Science Research Study

As AI systems are increasingly embedded in organizational decision-making, their assured outputs risk conveying unwarranted interpretive authority—particularly where policies, contracts, and compliance documents intend ambiguity and delegated discretion. We reconceptualize policy automation as a human–AI co-agency problem and, using elaborated Action Design Research (eADR) methods, derive governance constraints for preserving traceability, bounding autonomy, and governing shared vocabulary evolution. We design a governance-oriented artifact that transforms unstructured organizational documents into structured, machine-readable rule representations while maintaining institutional control. The design separates semantic extraction from rule formation, constrains representation through a shared vocabulary catalog, and supports controlled, mixed-initiative expansion. Rather than resolving ambiguity, it preserves modality and handles uncertainty through under-specification and explicit uncertainty markers. Evaluation on real-world university policy documents demonstrates deterministic, traceable rule generation that maintains institutional discretion while enabling structured representation. This study contributes a governance-oriented design and transferable design principles for embedding bounded AI autonomy into systems operating over ambiguity-rich organizational documents.