No. Recent benchmarks show generative AI cannot reliably read, calculate or repair financial spreadsheets on its own; performance drops sharply with multiple worksheets and compound calculations. So use AI only to structure data, and enforce the figures with deterministic calculation functions, formula checks and human approval.
The occasion is a series of academic benchmarks that test the ability of language models to read and calculate financial data, spreadsheets and financial statements. In FinSheet-Bench, ten model configurations were tested on financial spreadsheets modelled on private-equity structures. The best-performing model reached 82.4% accuracy, but on complex aggregations the accuracy of the three best models got no further than 33.3%. The researchers conclude that such standalone models are not suitable for unsupervised use in professional financial workflows, and recommend separating document understanding from deterministic calculation. These results come from the FinSheet-Bench study; they should not be attributed to FinQA, which is a separate benchmark for numerical reasoning over financial data.
That signal does not stand alone. Two further studies describe the same pattern, and a study titled Generative AI in Financial Reporting is described in the supplied research record as examining risks from inaccurate generative-AI content in financial reporting, including wrong amounts or data; the cited publisher page was not accessible during verification.
What exactly do the new benchmarks show about AI and financial spreadsheets?
They show that generative models make errors structurally as soon as financial tables go beyond a single, clear calculation. The error lies not only in addition, but also in locating the correct cells and repairing formulas.
In SpreadsheetBench 2, eight frontier models were tested on 321 end-to-end spreadsheet tasks, with an average of 11.8 worksheets and 593.5 cell changes per task. The best model reached 34.89% task accuracy; debugging existing formulas got no further than about 12%. The dominant errors were insufficient inspection of the workbook and selecting the wrong target cells. The results align with earlier benchmarks for numerical reasoning over financial data; the new spreadsheet benchmarks moreover show that performance declines with more complex workbooks and multi-step calculations.
Why do generative models fail specifically with complex, multi-layered financial tables?
Our reading of the benchmarks is that performance declines when tables contain dependencies, multiple periods or compound metrics. The studies report vulnerabilities in temporal alignment, structural table processing and the absence of formula hints; SpreadsheetBench 2 additionally reports errors in locating and changing target cells and formulas.
The study on financial reasoning over long financial statements describes a clear knowledge bottleneck when formula hints are removed and a structural bottleneck with tables containing multiple metrics. For one tested model, performance on table tasks dropped from 70.70% with hints to 38.22% without hints. Our assessment: multi-year statements, consolidated models and chained calculations are among the situations in which the benchmarks reveal the greatest vulnerability.
The study Generative AI in Financial Reporting in the Journal of Accounting Research adds the consequence: as recorded in the supplied research, inaccurate content, such as wrong amounts or data, creates risks for investors, litigation and sanctions, and the record states that generative AI already occurs in financial reporting; the publisher page itself was not accessible during verification. Our editorial conclusion is therefore that verification is not only a theoretical matter, but a current control need.
What does this mean for directors, lawyers and CISOs who deploy AI on figures?
Our analysis: because even the best-performing model in FinSheet-Bench stayed below one third accuracy on complex aggregations, unsupervised use of a standalone chatbot for critical financial models is insufficiently controlled; therefore every critical outcome should pass through a deterministic calculation layer before it ends up in a report. Our analysis: because SpreadsheetBench 2 describes models selecting wrong target cells and debugging formulas poorly, lawyers and directors may face a demonstrability problem when an erroneous entry later surfaces in annual accounts or a due-diligence file. We therefore advise recording the source of every displayed figure — worksheet, cell and formula — from the outset, and not reconstructing it after the fact. Because the accounting study links incorrect amounts to risks such as litigation and sanctions, it is in our analysis necessary that a qualified financial professional formally approves exceptions and critical outcomes before figures are used externally. The precise allocation of legal responsibility depends on the applicable rules and the concrete context. Our analysis: because performance drops sharply once formula hints fall away, it is risky to let model output flow unprocessed into downstream systems. We therefore advise treating a discrepancy between model and calculation engine as a process block rather than as a non-binding warning.
That calls for a separated architecture, not a larger language model. See also the GDPR accuracy principle for generative AI and, for chained processing, how errors between linked AI agents cascade through a chain.
How do I set up a verification protocol that enforces reliable figures?
Let the model structure data, but never enforce the figures. The arithmetical truth comes from deterministic code or a spreadsheet engine, and ultimate responsibility remains with a human.
- Separate extraction from calculation. Let AI first extract schemas, cell values and source locations; do not let it add, average or compute ratios itself.
- Calculate deterministically. Perform sums, averages, medians, ratios and checks with fixed code or a calculation engine that always gives the same outcome for the same input.
- Check identities and dependencies. Test accounting identities and formula dependencies automatically, so that a wrongly chosen target cell stands out immediately.
- Record evidence. Retain input, model version, intermediate results, formulas used and every discrepancy, so that an outcome can be reconstructed later.
- Enforce human approval. Have a qualified financial professional approve exceptions and critical outcomes; make that a hard step, not a formality. How to make that oversight testable is described in our explainer on making human oversight of AI-supported decision-making demonstrable.
A caveat about the evidence belongs here: FinSheet-Bench uses synthetic data and a text representation of spreadsheets, SpreadsheetBench 2 tests within one specific agent setup, and the study on long financial statements is still a preprint. Our analysis: although the studies use different test designs, they point in the same direction, while the exact percentages are bound to each test setup. Anyone wanting to go deeper into controllable AI output will find more in our topic hub on AI verification and controllable output.
Sources and references
Sources: The article draws on the academic benchmarks FinSheet-Bench, SpreadsheetBench 2 and a study on financial reasoning over long-horizon statements, plus a study titled Generative AI in Financial Reporting.