The Payroll Sheet

What was wrong

The checks that worked

• Totals against something outside the sheet. The time clock and Avery's log each took a few minutes and showed that something was wrong, and in which column.
• A test you can do in your head. One made-up week caught the overtime mistake that reading the numbers could miss.
• The payroll service's own check. It knew who the employees were, which the sheet didn't, and it added up the file itself.
• Adding a column yourself. A minute, and it caught the biggest mistake of the day.

Checking every row by hand would have taken 48 minutes, longer than the time you had, before you fixed anything. The quick checks didn't fix anything either. They told you where to look.

What's true outside the game

• News from this September. On September 1, 2026, British Columbia's government said spreadsheet errors had overstated its expected natural gas royalties (the money gas companies pay the province for the gas they take) by about 1.46 billion Canadian dollars over five years. The biggest error treated Canadian dollars as if they were US dollars. First Nations, Indigenous communities in Canada, had raised questions about the numbers in a letter in July. The minister called it a human error, and the Premier, the head of the province's government, asked the auditor general, the official who checks the government's books, to review it.
• A payroll upload went wrong this March. In March 2026, a field left blank on a spreadsheet uploaded to the City of Austin's payroll system overpaid 675 city employees about $1.4 million in all.
• A total that stopped short. A famous 2010 economics paper by Carmen Reinhart and Kenneth Rogoff, used in arguments about government debt, averaged a range of spreadsheet rows that left out five countries. In 2013 Thomas Herndon, a graduate student, and two professors at the University of Massachusetts found it while checking the paper's numbers. The authors admitted the spreadsheet error and stood by their main conclusion. It was one of three problems the critics raised.
• Pasted out of line. In 2003 the Canadian power company TransAlta lost about 24 million US dollars when bids it sent for US contracts that protect against the cost of crowded power lines didn't line up with the right prices after a spreadsheet paste. Its chief executive called it "a simple clerical error," a paperwork slip.
• A minus sign. In 1994 Fidelity's Magellan fund, a big pool of investors' money, told investors to expect a year-end payout of $4.32 for every share they owned. A tax accountant had left the minus sign off a $1.3 billion loss in a spreadsheet, turning it into a gain. The estimate was off by $2.6 billion, and the payout never came.
• Hidden data that wasn't gone. In 2023 the Police Service of Northern Ireland answered a public records request with a spreadsheet that had a hidden tab listing all 9,483 officers and staff. The UK's data regulator fined it £750,000. In 2008, a junior lawyer reformatting a spreadsheet to make a PDF made its hidden rows visible again. That added 179 contracts the bank Barclays never meant to buy to its deal for parts of Lehman Brothers, a bank that had just failed.
• Data cut off. In early October 2020, England found that 15,841 positive COVID cases from the previous eight days had been left out of the daily figures. Public Health England said the files had gone over a maximum size. Reports said they were loaded into an old Excel format that stops at 65,536 rows.
• Software that changes what you typed. A 2016 study of 18 leading journals about genes found that about one in five papers that came with gene lists in Excel had gene names turned into dates or numbers. The committee that names human genes renamed some of them in 2019 and explained why in 2020: SEPT1 became SEPTIN1 and MARCH1 became MARCHF1, so Excel would stop turning them into dates. Excel also drops leading zeros, like the ones in 0417; since 2023, newer versions of Excel let users turn some of these changes off.
• A formula wrong from the start. In 2012 JPMorgan's London office lost about $6.2 billion on trades. Its new risk model, a spreadsheet that estimated how much the trades could lose, was filled in by hand, and one step divided by a sum where it should have divided by an average. The bank's own review said that probably made the risk look smaller than it was. It wasn't the only cause: a US Senate report blamed bank officials who ignored the growing risk.
• Spreadsheet mistakes are common. In five careful checks (audits) from 1995 to 2000, collected by the researcher Raymond Panko, 91% of 55 spreadsheets had at least one serious error. That's a small sample, and Panko thinks the audits missed errors too. In 14 lab studies, people working alone got about 4% of cells wrong. In one of Panko's experiments, people checking a spreadsheet alone caught 63% of the planted errors, and teams of three caught 83%.
• AI in spreadsheets: check its work. Microsoft, Google, OpenAI and Anthropic all offer AI that works inside spreadsheets, and their own pages say not to take its work on trust. Microsoft says to review and verify anything Copilot creates before you rely on it. Google says its Gemini features in Sheets may suggest inaccurate information. OpenAI says AI can make mistakes, so verify important numbers. Anthropic says to always review Claude for Excel's changes before you finalize your work. In a public test of realistic business spreadsheet work, researchers tried eight leading AI models in March 2026, and the best one got about 35% of the tasks right. On a smaller set of 30 tasks they also tried four AI spreadsheet tools, Kimi Sheet, GLM in Excel, ChatGPT for Excel and Claude for Excel, and the tools did no better. Microsoft's and Google's own tools weren't in that test. Two other companies later posted scores of about half on the test's public scoreboard.

Now do it for real

1. Watch a spreadsheet change your data. In LibreOffice Calc, which is free and needs no account, type a short list: an ID that starts with zeros (0417), a gene-style code (MARCH1, which Excel turns into a date; see what yours does), a 17-digit number, and a column of hours where you type one of them with an apostrophe first ('38), which stores it as text. Before you press Enter on each, write down what you expect to see. Then add up the hours with =SUM() and see whether the one stored as text counted. Then add the same cells with plus signs (like =B2+B3+B4) and compare. If the apostrophe shows in the cell, the cell is already set to Text, so type 38 again without the apostrophe; it's still stored as text.
2. Plant three, swap, catch. Make a small sheet of made-up pay for ten people with a total at the bottom. Plant three mistakes from this lesson. Swap with a friend, set a timer for ten minutes, and use only the quick checks: totals against a list, one test case, a row count.
3. At your next job, ask one question: "When a spreadsheet goes out, who checks it, and what do they check it against?"

You will have seen a spreadsheet change your data with your own eyes, and caught a friend's mistakes on a clock. Time: 40 minutes. You need: LibreOffice, or any spreadsheet you're allowed to use, and a friend. No AI account.

Last checked November 11, 2026. Fernbrook Day Camp, Ledgerline Payroll, every counselor and everyone else in the story are made up, and so is every number in the sheet. Ledgerline's rule about holding a short payroll, the parent at 3:30 and how long each check takes are made up too, and taxes are left out of the math to keep it short. The sheet's AI helper and everything it says are made up too; it isn't any real company's product. Four of the mistakes are modeled on the real cases above; the fifth, the AI's overtime column, on what the AI companies themselves warn about. Torah Tech wrote this lesson with the help of Claude, an AI made by Anthropic. Any mistakes are ours.

Back to the Lab