AI Belongs in Excel. Just Keep It Away From the Arithmetic.
There are few pieces of enterprise software as deceptively important as Excel. A spreadsheet might be a shopping list; it might equally be an insurer’s capital calculation, a bank’s reconciliation, or the model behind a multimillion-pound decision. Sometimes the spreadsheet is the system. And its error record predates AI: research by Ray Panko at the University of Hawaii has long found that a large majority of operational spreadsheets contain errors, and JPMorgan’s own task-force report on the 2012 London Whale losses traced part of the failure to a value-at-risk model run on manually copy-pasted Excel with a formula mistake. Into this environment, Microsoft is now embedding autonomous agents.
What Microsoft has shipped – and what it admits – is instructive. Agent Mode in Excel, launched through the Frontier programme in late 2025, plans, executes, validates, and iterates on multi-step spreadsheet tasks. Microsoft’s own evaluation put it at 57.2 per cent accuracy on the full 912-task SpreadsheetBench benchmark, against a human baseline of 71.3 per cent; publishing that gap is candour worth crediting. Its responsible-AI documentation is blunter still: the Word, Excel, and PowerPoint agents “can make mistakes, misinterpret facts, and/or produce inaccurate results”, and are explicitly ruled out for conclusions or decisions in financial, legal, medical, and similar sensitive domains without qualified human expertise. Read properly, none of this argues against AI in Excel. It specifies how AI must belong there.
The mistake would be letting the stochastic model become the calculation engine. A language model is excellent at interpreting intent – build me a twelve-month cash-flow forecast with a downside scenario – and at proposing formulas, mapping columns, spotting anomalies, and explaining what a workbook does. But once the accounting logic is established, deterministic software should execute it. The architecture that follows resembles a sandwich: probabilistic intelligence on top, handling interpretation, planning, and explanation; a governed semantic layer in the middle, carrying permissions, approved formulas, data lineage, and business rules; deterministic computation at the bottom. The AI then explains the result rather than inventing it.
The division of labour matters most as agents begin modifying workbooks autonomously. Microsoft’s account reconciliation agent in Dynamics 365 Finance points the right way: it monitors postings continuously, raises exceptions, recommends actions such as creating a journal entry, and records every suggestion and decision in an auditable timeline with an undo – the human accepts or overrides. Generalise that principle. Every consequential agent action should be reconstructable: which cells changed, which formula was introduced, which source data was referenced, which validation tests passed. The valuable product is not “Copilot can create a spreadsheet”; chatbots could produce formulas years ago. It is a spreadsheet agent whose work an auditor can challenge six months later.
Three moves follow for finance and technology leaders. First, draw the line in the architecture, not the policy manual: models propose, engines compute, and no free-text generation writes values into cells of record. Second, make reconstructability a procurement requirement – cell-level change logs, formula provenance, validation gates – and test it before deployment rather than after an audit finding. Third, retrain review, because the failure mode is shifting from wrong formulas to plausible ones accepted without scrutiny: treat agent-modified workbooks like code changes, with versioning and named reviewers. Microsoft spent decades convincing enterprises that Excel could be trusted with enormously important numbers. Generative AI can make those numbers easier to work with. The leader’s job is to ensure it does not also make them easier to fabricate.
Personal analysis. Views are mine, not my employer's.