GPTs Show Promise but Remain Unreliable for Spreadsheet Modeling

A recent study published on arXiv (2604.25689) examines GPT-based applications for creating analytical spreadsheet models. Researchers evaluated five GPT extensions and chose Excel AI from pulsrai.com for in-depth analysis. They conducted structured experiments using straightforward problem statements to measure Excel AI against the ERFR criteria (inputs in each cell; formulas; absence of hardcoded numbers; labels; precision). Findings indicate that while Excel AI can generate well-organized models, it often lacks consistency and reproducibility. The study highlights two main obstacles: "the problem of confidence" and "the problem of workflow," underscoring the necessity for proficient users to validate and modify GPT-generated spreadsheets. Despite the potential of GPTs to create draft models that could expedite development and reduce skill demands, the current tools are deemed unreliable for professional applications. The paper wraps up with suggestions for future research on prompting techniques.