Go to ACL ARR 2025 May homepageTravelReimGPT: A Conversational Agent for Policy-Compliant and Reliable Travel Reimbursements
Abstract: We present TravelReimGPT, a user-centric, conversational AI system for automating travel reimbursement tasks under strict policy constraints. While recent LLM-based agents have shown promise in open-ended dialogues, they often struggle with numerically sensitive and rule-governed applications due to limitations in symbolic reasoning and consistency.
To address these challenges, we propose a Programming-Object-Centric Architecture (POCA) that transforms user inputs and documents into programming objects, which serve as the foundation for deterministic, logic-driven policy enforcement. A self-corrective object constructor, combining validation logic and iterative model-guided refinement, enhances the accuracy and completeness of extracted information. Through multi-turn interactions, TravelReimGPT gathers necessary inputs (e.g., receipts) and generates reimbursement reports that comply with complex policy rules. Experiments on real-world reimbursement cases show that, powered by GPT-4.1, our system consistently produces accurate reimbursement outcomes across all tested cases, exhibits robust conversational fluency, and achieves high user satisfaction. In contrast, prompting-based baselines occasionally yield inaccurate reimbursements, highlighting reliability and control limits. This work demonstrates a practical and extensible framework for building reliable AI agents for rule-intensive domains, with potential applicability to broader tasks such as auditing and budget compliance.
Paper Type: Long
Research Area: Dialogue and Interactive Systems
Research Area Keywords: task-oriented, evaluation and metrics, human-in-the-loop, applications, conversational modeling
Contribution Types: NLP engineering experiment
Languages Studied: English
Submission Number: 5051
Loading