Benchmarks show generative AI cannot calculate spreadsheets

You must enforce financial calculations through deterministic code, not generative AI output. Separate the model's role to data structuring only, and route every critical figure through a calculation engine with formula verification and human approval before external use.

The prompt is an analysis of 12 October 2026 of generative AI calculation errors in financial spreadsheets, which argues that language models cannot reliably perform calculations on complex financial data and should be confined to data preparation tasks. Recent benchmarks tested multiple frontier models on financial spreadsheets and multi-worksheet workbooks, finding that accuracy collapsed sharply once calculations became compound or dependencies emerged. In our assessment, this is not a marginal limitation but a structural control failure that affects any organisation using generative AI to process financial data for external reporting, due diligence or investor communication.

Why do models fail on financial calculations?

Generative models make errors that go beyond simple arithmetic. When tested on financial spreadsheets modelled on private-equity structures, the best-performing model reached 82.4% accuracy overall, but accuracy on complex aggregations fell to 33.3%. When eight frontier models were tested on end-to-end spreadsheet tasks averaging 11.8 worksheets and 593.5 cell changes per task, the best achieved only 34.89% task accuracy. Debugging existing formulas performed worse still, reaching only about 12% accuracy.

The failures cluster around three failure modes:

• Insufficient workbook inspection — models do not adequately scan the full structure before attempting changes.
• Wrong cell selection — models identify and target incorrect cells for data entry or formula repair.
• Formula repair failure — models cannot reliably debug or correct existing formulas, particularly in multi-step calculations.
• Temporal misalignment — models struggle with multi-year statements and period-to-period dependencies.
• Compound metric handling — performance degrades sharply when tables contain consolidated figures or chained calculations.
• Loss of accuracy without hints — when formula structure is not explicitly provided, accuracy drops by 30 percentage points or more.

Performance declines further as complexity increases. Multi-year consolidated models, statements with multiple metrics, and calculations that chain across worksheets are the situations in which benchmarks reveal the greatest vulnerability.

What are the practical consequences of relying on model output?

Inaccurate figures in financial reporting create direct risks: litigation exposure, regulatory sanctions, and loss of investor confidence. Once an erroneous entry surfaces in annual accounts or a due-diligence file, you face a demonstrability problem. If you cannot show the source of every displayed figure — the worksheet, cell and formula from which it came — you cannot defend the entry or trace where the error was introduced. Reconstructing that chain after the fact is both unreliable and insufficient as a control.

Because models select wrong cells and repair formulas poorly, unsupervised use of a standalone chatbot for critical financial models is insufficiently controlled. Every critical outcome must pass through a deterministic calculation layer before it enters a report.

What architecture separates data preparation from calculation?

The solution is not a larger model but a separated workflow. Let the generative model perform data structuring — extracting figures from documents, normalising formats, identifying line items. Do not let model output flow unprocessed into downstream systems or directly into financial statements.

Instead, enforce the figures through deterministic code. Route model output to a calculation engine — a spreadsheet application, a dedicated financial system, or a deterministic code layer — where formulas execute with certainty. Treat any discrepancy between model output and the calculation engine as a process block, not a non-binding warning.

You must be able to demonstrate the following controls:

1. Separate the data-preparation layer from the calculation layer — the model structures data; deterministic code enforces figures.
2. Record the source of every displayed figure — document the worksheet, cell reference and formula for each number used externally, from the outset.
3. Verify formula integrity before use — run automated checks to confirm that formulas match their intended calculation and that cell references are correct.
4. Require formal human approval for critical outcomes — a qualified financial professional must explicitly approve exceptions and all figures used in external reporting before they leave your control.
5. Block discrepancies as process failures — treat any mismatch between model output and the calculation engine as a halt point requiring investigation, not as a warning to be overridden.

Who bears responsibility when figures are wrong?

Ultimate responsibility for accuracy remains with a human professional, not the model. That responsibility cannot be delegated to a chatbot or transferred through process automation. The precise allocation of legal responsibility depends on applicable rules and your specific context, but the practical consequence is clear: you must be able to show that a qualified person reviewed and approved the figures before they were used.

Tooling can enforce the separation of layers and flag discrepancies, but it cannot replace professional judgement about what figures are reasonable, what assumptions are sound, or whether a calculation makes sense in context. Automation can make that judgement visible and auditable; it cannot make it unnecessary.