MBABench: Evaluating LLM Agents on End-to-End Spreadsheet Tasks in Finance

Researchers introduced MBABench, a new evaluation framework for testing LLM agents on end-to-end financial spreadsheet tasks—a capability increasingly demanded by enterprises but not yet adequately measured by existing benchmarks. The study found that even top-performing models like Claude fall short of professional finance standards, struggling with complex multi-step workflows and degrading sharply in quality as task difficulty increases.

The introduction of MBABench addresses a significant gap between enterprise expectations and current AI capabilities. Financial institutions rely heavily on spreadsheet-based workflows for modeling, forecasting, and scenario analysis, yet existing benchmarks measure only narrow tasks like formula corrections or simple Q&A. This new evaluation framework reflects real-world requirements by assessing outputs across three dimensions—Accuracy, Formula, and Format—mirroring how finance professionals actually evaluate deliverables.

The research emerges as AI labs race to develop enterprise-grade agents capable of autonomous work. Frontier models have achieved notable progress on general reasoning tasks, but specialized domains like finance demand both technical correctness and professional presentation standards. The gap identified here is particularly important because financial spreadsheets are collaborative artifacts that undergo multiple rounds of stakeholder review, making readability and modifiability critical beyond raw accuracy.

The benchmark's findings have immediate implications for enterprise AI adoption timelines. Organizations considering delegating financial modeling to AI agents must now recalibrate expectations—current systems cannot reliably match human professional standards on complex workflows. This extends deployment timelines and suggests that AI will augment rather than replace financial analysts in the near term.

Looking forward, MBABench establishes a productivity frontier that motivates continued development. The sharp performance degradation beyond simple calculations indicates specific architectural or training limitations that AI labs can now target. Finance teams should monitor improvements on this benchmark as an indicator of genuine enterprise-readiness.

• →Current LLM agents produce professional-quality spreadsheets on simple tasks but fail on multi-step financial workflows requiring chained calculations
• →MBABench introduces the first comprehensive evaluation framework measuring end-to-end spreadsheet creation rather than isolated formula tasks
• →Claude models lead competitors but still fall short of human finance professional standards across accuracy, formula structure, and output formatting
• →Financial institutions should expect AI agents to serve as augmentation tools rather than autonomous replacements in spreadsheet-based workflows
• →Performance degradation patterns suggest addressable architectural limitations that will drive next-generation agent improvements