Why a Spreadsheet Is the Wrong Foundation for AI Finance (And What Replaces It)

A widely cited review of 35 years of spreadsheet research, published in 2024, put the figure at 94 percent: that share of business spreadsheets examined across decades of studies contained at least one error. Separate work has found that roughly half of the models mid-sized and large firms run for actual decisions carry material defects. These are not exotic edge cases. This is the substrate most finance teams already use to budget, forecast, price deals, and brief the board.

Now picture an AI assistant sitting on top of that. It reads the sheet, writes back to it, and answers questions about it in plain English. The pitch is that the spreadsheet finally talks. The reality is that the AI inherits every broken reference, every hard-coded override someone forgot to remove, every cell that says 2024 but pulls 2023. It does not clean the foundation. It builds faster on a cracked one.

The spreadsheet was never a source of truth

A spreadsheet feels authoritative because it is the thing the controller looks at. But authority and reliability are different properties. The grid mixes inputs, calculations, and presentation in the same cells, with no enforced separation between a figure that came from the general ledger and a figure someone typed over it at 11pm before a deadline. There is no native lineage. Ask “where did this number come from,” and the honest answer is often “scroll up and trust the formula bar.”

That ambiguity is survivable when a human owns the sheet. People build local knowledge: which tabs are stale, which cell not to touch, which total is “the real one.” The model has none of that tribal context. It treats a poisoned cell with the same confidence as a clean one. So the first problem with AI-on-Excel is not that the AI is dumb. It is that the AI is faithfully reading a document that was never engineered to be a system of record.

What the benchmarks actually show

There is now hard evidence on how language models handle financial spreadsheets specifically. FinSheet-Bench, published in early 2026, built a benchmark from synthetic portfolio data modeled on real private-equity fund structures, then tested ten configurations from OpenAI, Google, and Anthropic on lookups and numeric reasoning over realistic layouts.

The best result was around 82 percent accuracy. That is roughly one wrong answer every six questions, from the strongest model on the board. And it got worse with complexity: on the largest file in the set, the average across all models fell to under 50 percent. The paper’s own conclusion is blunt: no standalone model reaches an error rate low enough for unsupervised use in professional finance.

Read that the way a CFO should. An 18 percent miss rate is not a rounding issue you cover with a disclaimer. If one in six figures in a forecast is wrong and you cannot tell which one, the whole forecast is unusable for any decision that carries consequences. The accuracy is high enough to look right and low enough to be dangerous, which is the worst possible combination.

And note where the errors come from. Some are retrieval failures: the model grabs the wrong cell on a messy layout with fund dividers and multi-line headers. Others are reasoning failures: it does the arithmetic itself and gets it wrong. Both fail silently. The output is a confident number with no flag, no source, no way to replay how it was derived.

Why writing back makes it worse

Reading a bad sheet is one risk. Letting AI write to it is another category entirely. The moment an assistant edits cells, it can propagate an error across linked tabs, overwrite a manual adjustment that existed for a real reason, or introduce a formula that looks plausible and is subtly off. You now have a spreadsheet that no single human fully authored, changing under automated edits, feeding a board deck.

There is a deeper structural point here. The LLM is a probabilistic system. A financial calculation is supposed to be deterministic: same inputs, same answer, every time. When you let the model both decide what to compute and perform the computation inside a free-form grid, you have wired a non-deterministic engine into the one part of the business that is legally and operationally required to be reproducible. That is not an automation upgrade. It is a control failure waiting for an audit.

What replaces the sheet

The fix is not a better prompt or a more careful model. It is moving the source of truth out of the grid entirely. Instead of an AI that edits a spreadsheet, you want an AI that queries a reconciled model underneath it.

Concretely, that means a layer that does three things the spreadsheet never could. First, it connects directly to where the numbers actually live (QuickBooks, Xero, NetSuite, Sage, SAP, Oracle, a warehouse, or uploaded statements) instead of inheriting a manual export. Second, it reconciles those sources into one model that ties out to the ledger, so there is a single figure for revenue, not seven slightly different ones across seven tabs. Third, it does the math in a deterministic engine, not in the model’s head. The LLM is allowed to reason, retrieve, and explain. It is not allowed to invent the arithmetic.

That architecture changes what an answer is. When the AI reports a number, that number was retrieved from a reconciled model and calculated by a deterministic engine, and it traces back to the source transaction. You can replay it. You can show an auditor the path. The accuracy problem the benchmarks expose does not get patched: it gets routed around, because the model is no longer the thing doing the counting.

This is also where scenarios finally land in the right place. CFOs do want multiple views: base case, downside, what-if on a hire plan. The mistake is letting those variations live as conflicting numbers in the underlying data. They belong as scenarios layered on top of one reconciled foundation, each one explicit about which assumptions changed. One set of facts, many modeled futures, not many sheets each claiming to be the facts.

The takeaway

Excel is an excellent surface for a human to think in. It is a terrible system of record for a machine to compute against, and the data on error rates has been clear for decades. Bolting AI onto the spreadsheet does not fix that; it scales it. The teams getting real, defensible value from AI in finance are not the ones with the cleverest assistant. They are the ones who stopped treating the grid as the source of truth and put a reconciled, deterministic modeling layer beneath it.

If you are evaluating AI for finance, the question to ask a vendor is simple: when your tool gives me a number, can you show me exactly where it came from and prove the math the same way I would prove a formula? Rexfin is built to answer yes. To see what that looks like against your own data, book a demo.

For the surrounding picture, start with the pillar on the reliable financial-modeling layer for AI. From there, MCP for finance is not enough explains why connecting an LLM to your ERP still gets the math wrong, and single source of truth vs. sufficient versions of truth covers where scenario variation actually belongs.