When Is a Spreadsheet Better Than an AI Agent?

When Is a Spreadsheet Better Than an AI Agent?

A spreadsheet wins when the logic is stable, the inputs are clean, and someone needs to see exactly how the answer was produced. An agent wins when the input is messy and the judgment is fuzzy. The mistake is assuming the spreadsheet is the safe option. Neither one is.

We get asked to automate things that a spreadsheet already does. Sometimes that is a good trade and sometimes it is a bad one, and the deciding factor is almost never how impressive the demo looks.

So here is the honest comparison, including the uncomfortable research on how often spreadsheets are wrong.

Why Is This Even a Debate?

Because spreadsheets are doing far more load-bearing work than anyone admits. Raymond Panko's EuSpRIG paper on spreadsheet errors cites a study by Gable, Yap and Eng that examined all 402 non-trivial spreadsheets in one organisation, where 46 percent were rated important or very important and 59 percent were used at least monthly.

These are not throwaway calculations. In a survey of 256 spreadsheet developers by Chan and Storey, also cited in that paper, 42 percent named their chief executive as the highest-level user of the spreadsheet's data. Real decisions ride on these files.

So when someone proposes replacing one with an agent, the question is not whether the agent is perfect. It is whether it is better than the thing already there, which is a much lower bar than people assume.

How Often Are Spreadsheets Actually Wrong?

More often than almost anyone believes. Panko's paper summarises seven field audits of real organisational spreadsheets covering 367 files in total, in which errors were found in 24 percent. That number is flattering, because most of the older audits used methods unlikely to catch a majority of errors.

The more rigorous audits are worse reading. Panko reports that audits from 1997 onward found errors in 91 percent of the 54 spreadsheets examined, and describes the most recent field audits as finding errors in at least 86 percent of spreadsheets audited. In controlled development experiments the figure was 51 percent, even though most of those spreadsheets were only 25 to 50 cells.

The cell error rates give a sense of scale. Panko cites rates of 1.1 percent, 2.2 percent and 2.5 percent of cells containing original errors in the most carefully audited spreadsheets, with a lower figure of 0.38 percent for formula cells in one tax auditing dataset. A one percent per cell error rate in a thousand cell model is not a small problem.

Can You Catch Spreadsheet Errors by Checking?

Partly, and it costs more than teams expect. Panko reports that individual code inspectors going cell by cell find only half or fewer of all errors, and that group code inspection, the only technique his paper says has been experimentally shown to be fairly effective, catches about 80 percent.

That last number deserves a moment. The best tested method for finding spreadsheet errors, done by a group, working slowly, still misses one error in five. And almost nobody does it, because it is expensive, exhausting and, in Panko's words, work that inspectors hate.

So the "safe, auditable spreadsheet" is mostly a story we tell. It is auditable in principle and unaudited in practice, which is not the same thing as correct.

How Reliable Are AI Agents by Comparison?

Capable more often than they are consistent, and the research community is starting to measure that gap properly. In "Towards a Science of AI Agent Reliability", first posted to arXiv in February 2026, Stephan Rabanser, Sayash Kapoor, Peter Kirgis, Kangheng Liu, Saiteja Utpala and Arvind Narayanan argue that compressing agent behaviour into a single success score hides its operational flaws.

Their paper proposes twelve metrics across four dimensions: consistency, robustness, predictability and safety. Evaluating 15 models across two benchmarks, they report that recent capability gains have produced only small improvements in reliability. In other words, newer models are better at doing the task and not much better at doing it the same way every time.

That is the number that should drive your decision. A process that must produce an identical result on every run is a bad fit for a system whose main weakness is variance between runs.

How Do the Two Fail Differently?

A spreadsheet fails silently and permanently. An agent fails loudly and differently each time. That distinction shapes everything about how you would catch the problem.

Question	Spreadsheet	AI agent
How does it fail?	One wrong formula, repeated forever	Varies between runs on the same input
How do you notice?	Usually when a number looks odd months later	Often immediately, because output shape changes
What fixes it?	Cell by cell inspection, which misses errors	Evaluation runs, logging and narrower scope

The practical reading is that spreadsheets are dangerous where consistency creates false confidence, and agents are dangerous where variance is unacceptable. Those are different risks, and you should pick based on which one your process can survive.

What Tasks Genuinely Belong in a Spreadsheet?

Arithmetic on clean, structured numbers, where the formula is the documentation and an auditor may need to trace it. Budgets, pricing models, simple forecasts and anything a finance team must be able to defend line by line.

Spreadsheets also win on speed of change. When the logic shifts weekly and the person changing it is not technical, a formula they can edit beats an automation they have to file a request for. That is a real organisational advantage and it is usually undervalued by engineers.

What matters is keeping the model small enough to inspect. The error research is clear that size is the enemy, so a spreadsheet that has grown to thousands of cells has quietly stopped being the safe option it started as.

What Tasks Genuinely Belong to an Agent?

Work where the input is unstructured and the output only has to be good, not identical. Reading a hundred inbound messages and grouping them by topic. Turning meeting notes into a structured summary. Drafting a first pass that a person will edit anyway.

These tasks share a property: variance between runs does not compound into a wrong number. Nobody audits the grouping of last quarter's support tickets against a second run. The output is an input to a human decision, not a figure on a board slide.

We went through how to pick these first candidates in our piece on what to automate first, and the test is close to the same one.

What About the Middle Ground?

Most real work lives there, and the answer is usually a script rather than either extreme. A script gives you the determinism of a spreadsheet with the scale of an automation, and it can be tested. We laid out that comparison in when to use an AI agent versus a script.

A common good pattern is an agent that handles the messy part and hands a clean, structured result to deterministic code that does the arithmetic. The model reads the invoice, the code adds up the numbers. Neither is asked to do the thing it is bad at.

The bad pattern is the opposite: a spreadsheet doing pattern matching through nested conditionals, or an agent doing arithmetic on figures it was asked to extract and total in one step. Both are systems asked to work against their own failure mode.

So How Should You Choose?

Ask what happens if this is wrong and nobody notices for three months. If the answer is serious, you want determinism and a real review process, whichever tool you pick. If the answer is mild, take the tool that gets the work done and spend your review budget elsewhere.

Then ask whether anyone would ever need to explain how the answer was reached. If yes, the logic needs to be readable by a person who was not there when it was built, which rules out a lot of clever prompt chains and a lot of clever spreadsheets too.

If you want help working out which of your team's processes are genuinely agent-shaped and which just need a tidier spreadsheet, we are happy to walk through it. You can reach our team at phoenix.studio.