Gen-Plan: An Automated Framework for Generalized Planning in Finance Agent

Abstract

Automating planning for complex tasks remains a challenge for AI agents and limits the efficacy in specialized domains such as Finance. In this paper, we introduce Gen-Plan, an automated framework for systematically generating structured and detailed plans that support a coherent reasoning path to solve financial tasks. Gen-Plan employs a two-stage process, including a General Plan and a more comprehensive Extended Plan. The proposed principle is driven by a designed general research agent leveraging the Gemini 2.5 Pro model, augmented with Think tool to strategically decouple the planning process from generic domain knowledge into sub-steps and guiding zero-shot planning across diverse tasks. The prompt structuring strategy further refines the framework’s capabilities by integrating in-depth research into the planning and execution. Through our evaluation across general text classification (BBC News), financial sentiment analysis, and complex financial question-answering (FinQA) datasets, Gen-Plan consistently demonstrates significant improvements in the overall performance. Notably, Gen-Plan yields significant improvements in high-stakes financial tasks up to a 14% accuracy for the FinQA dataset. As a result, Gen-Plan demonstrated that structured, generalizable planning is not merely an enhancement but an essential prerequisite for agentic AI applications, particularly within precision-demanding sectors like finance.