The spreadsheet became the system of record.
Nobody decided this. It happened one tab, one workaround, one "use this version" at a time — until a file with 20 sheets and a macro no one will admit to writing was quietly running a real part of the business. Here's what the research says about that file, and what it has cost the people who trusted it.
It isn't that spreadsheets sometimes break. It's that they usually do.
Thirty years of independent audits — academic and commercial — keep landing on the same uncomfortable numbers. The file driving the decision should be assumed wrong until proven right.
Mid-market firms typically run weaker spreadsheet governance than the large enterprises these studies measured. The real exposure is likely worse, not better.
The experts ship the worst ones. They expect to be error-free.
Not interns. Kodak, JPMorgan, Fidelity, a national health service, a sovereign wealth fund — and, when the stakes drop, a national security service and the people who name human genes. The pattern isn't incompetence. It's a manual workflow with no controls, run by people too senior to be double-checked.
The "London Whale."
A new risk model, built in Excel with manual copy-paste between sheets, divided by the sum of two rates instead of their average — halving reported volatility and masking the real risk. The trading loss passed $6.2 billion; the bank was fined $920 million.
The cases that fell off the bottom of the sheet.
COVID results were aggregated in a legacy .xls file capped at 65,536 rows. When the data exceeded it, 15,841 positive cases were silently dropped from daily totals — and from the contact-tracing pipeline — for days.
One missing minus sign.
An accountant transcribing the year-end dividend estimate omitted the minus sign on a $1.3 billion net capital loss — booking it as a gain. The estimate was off by $2.6 billion, and a $4.32/share distribution had to be retracted.
A cut-and-paste that misaligned the rows.
Bids pasted into a template landed one row off — high bids on low-demand contracts. The power company won contracts it didn't want at prices it never intended, wiping out 10% of annual profit.
One wrong date in a benchmark.
The world's largest sovereign wealth fund lost NKr 980 million (~$92M) on its mandated benchmark calculation — because an incorrect date was entered into the spreadsheet that computed it.
Hidden rows became binding contracts.
An associate converting a Lehman asset list to PDF past 11:30pm re-sized the rows — un-hiding 179 contracts marked hidden because Barclays did not want them. Barclays found itself legally committed to acquire all 179.
Source: Computerworld; ABA Journal ↗
Too many zeros.
Kodak had to restate two quarters after a spreadsheet typo — too many zeros on a single employee's severance accrual — overstated expense by $11 million. It was caught only at the next quarter's reconciliation.
They renamed human biology to fit Excel.
Excel kept auto-converting gene symbols like SEPT1 and MARCH1 into dates. After ~1 in 5 genetics papers were found to carry corrupted gene names, the HUGO Gene Nomenclature Committee renamed 27 human genes — because changing the genome's vocabulary was easier than fixing the spreadsheet.
Source: Ziemann et al., Genome Biology (2016); The Verge (2020) ↗
Keyed 20,000 instead of 10,000.
A single mistyped figure in a ticket-availability sheet oversold four synchronized-swimming sessions. Roughly 3,000 ticket holders had to be re-accommodated — about 200 of them upgraded to the men's 100m final.
Source: ESPN/AP; CNN (Jan 2012) ↗
A formatting error bugged the wrong phones.
A spreadsheet replaced the last three digits of target phone numbers with "000," and the UK security service requested subscriber data on 134 unrelated numbers.
Source: Interception of Communications Commissioner report ↗
Different disasters. The same five ingredients.
Across two decades of horror stories, the mechanism barely changes. It's never one careless person — it's a workflow with no place for the controls that software gets for free.
Manual data entry — a number is typed, and typed numbers are wrong ~1–5% of the time.
Copy-paste between files — the moment rows misalign and no one notices.
Manual reconciliation — two sources, checked by eye, under deadline.
No version control — "FINAL-v7(use this one)" is a confession, not a filename.
No audit trail or formula testing — code gets reviewed; spreadsheets almost never do.
A spreadsheet isn't a bad tool. It's a brilliant one, asked to do a job — system of record for a business — that it was never built to hold safely. The fix isn't more discipline. It's software that makes the safe path the default one.
For thirty years, living with it was the rational choice. That math broke.
Replacing the spreadsheet with software fitted to your business used to mean a seven-figure budget and an eighteen-month build — so you lived with the risk, because fixing it cost more than the pain. AI-native development changed the economics. The capability that retires the spreadsheet is now buildable in weeks, against an outcome you name first.
Statistics: R. Panko, "What We Know About Spreadsheet Errors," J. End User Computing (1998, rev. 2008). · F1F9 / CNBC, "Spreadsheet blunders costing business billions" (2013). · Ventana Research, "Spreadsheets for Today's Enterprise" benchmark. · Lawrence & Lee, project-finance audit (via L. Pryor, 2003). Cases: Kodak Form 10-Q/A (SEC) & MarketWatch (2005); JPMorgan Management Task Force Report (2013); BBC / The Conversation (PHE, 2020); CIO Magazine & Fidelity shareholder letter (Magellan, 1994); The Globe and Mail (TransAlta, 2003); Computerworld & ABA Journal (Barclays/Lehman, 2008); Financial Times via EuSpRIG (Norway, 2011); Ziemann et al., Genome Biology (2016) & The Verge (2020). Several headline losses (e.g. the London Whale) reflect broader control failures in which the manual workflow was a contributing cause.