Abstract

Software project planning involves defining the scope, requirements, stakeholders’ roles, task dependencies, duration of each task and potential risks. For the most part, these activities are carried out manually and highly depend on the experience of project managers. Fuzzy requirements, incorrect estimations and dependency issues can make a project fall behind schedule and decrease the quality of delivery. Generative AI, in particular, Large Language Models (LLMs), can solve this problem by analyzing a project’s textual description, extracting requirements, proposing the architecture, decomposing the work into tasks and producing the planning documentation. However, LLMs can produce false statements, wrong estimates and contradictions and thus the need for manual oversight is still necessary. The paper describes a framework that involves the generation of planning data by LLMs and the validation of the result. A validation layer evaluates the completeness of requirement coverage, rationality of dependencies and feasibility of schedule before the output is shown to the project manager. Ten selected papers are discussed, the comparison of the existing approaches, research gaps and several metrics are presented.

Keywords

How to Cite This Article

Rights