AI agents still can't reliably do the spreadsheet work that fills most corporate offices.
Researchers introduced SpreadsheetBench 2, a benchmark designed to test AI agents on realistic business spreadsheet workflows rather than toy problems. The dataset contains 321 tasks drawn from actual financial reports and corporate filings, each averaging 11.8 worksheets and requiring nearly 600 cell modifications. Eight frontier large language models were tested under a consistent multi-turn agent setup, with several commercial spreadsheet AI products included as additional baselines. The best model in the evaluation hit 34.89% overall accuracy — and debugging tasks were worse, bottoming out at 12%.
Most existing spreadsheet benchmarks grade agents on isolated operations: write this formula, edit that cell. SpreadsheetBench 2 instead demands full workflows with cross-sheet dependencies, which is what business users actually need. The failure analysis is telling: agents most often stumbled on two things — not inspecting the spreadsheet thoroughly enough before acting, and targeting the wrong cells when they did act.
For a category of AI tools that vendors routinely pitch as productivity multipliers for finance and operations teams, a 35% ceiling on realistic tasks is a candid reminder that "Excel copilot" marketing runs well ahead of what the underlying models can actually do.