The Machine for Asking What If

In the spring of 1978 Dan Bricklin was a student at Harvard Business School, sitting in a lecture in which a professor worked through a financial model on a blackboard. The model was laid out in rows and columns, and every time the professor changed one number, such as a growth rate or a price, he had to erase and recalculate a long chain of others. Bricklin, who had worked as a programmer, began to imagine a different kind of blackboard: an electronic one, where you could change any number and watch every number that depended on it change too. He later described what he had in mind as an electronic blackboard with electronic chalk.

With his friend Bob Frankston, he built it. VisiCalc went on sale for the Apple II in 1979, and within a few years people who had no interest in computers were buying them in order to run it. The spreadsheet became the first program that made the personal computer indispensable to business, and it has remained, in one form or another, the place where most of the world’s commercial thinking is done. Budgets, forecasts, valuations, pricing plans, staffing models, sales targets, clinical trial logs and government statistics pass through grids of cells every day. Many of the numbers on which decisions are made were last touched in a spreadsheet.

This essay follows how the spreadsheet came to be what it is. Its story has the usual cast of a technology history: inventors, a market leader that fell, a platform war, lawsuits and a giant that won by waiting. But its deeper subject is a way of thinking. The spreadsheet taught organisations to treat a business as a model with adjustable assumptions, to ask “what if?” constantly, and to trust numbers that are often less certain than they look. To understand why the spreadsheet is designed as it is, and why it fails as it does, we need to start long before computers, with the ledger.

Before the computer: the ledger as a way of seeing

The word “spreadsheet” is older than software. In accounting, a spread sheet was a large sheet of paper ruled into columns and rows and spread across two facing pages of a ledger, on which an accountant laid out the figures of a business side by side: months across the top, accounts down the side, totals at the edges. The paper spreadsheet was the working surface of management accounting for most of the twentieth century. Its weakness was the one that frustrated Bricklin’s professor: change one figure and every dependent figure had to be recalculated by hand, with an adding machine and an eraser.

Behind the paper spreadsheet lay a much older invention. Merchants in the Italian trading cities developed double-entry bookkeeping over the thirteenth and fourteenth centuries, and in 1494 the Franciscan friar and mathematician Luca Pacioli described the Venetian method in his Summa de arithmetica, the first printed account of it. Every transaction was recorded twice, as a debit in one account and a credit in another, so that the books always balanced. If they did not, an error had been made somewhere. Double entry was a technology of checking. It allowed a merchant to see, from the books alone, whether the business as recorded was internally consistent.

It also changed what a business was, in the eyes of its owners. Profit could be calculated as a single figure; capital could be distinguished from income; the enterprise could be seen as an entity separate from the household of the merchant. Goethe, in his novel Wilhelm Meister’s Apprenticeship (1795–96), gave one character a famous speech in praise of double entry as one of the finest inventions of the human mind. A century later the German economist Werner Sombart argued that double-entry bookkeeping was essential to the rise of capitalism itself, because it made rational calculation of profit possible; Max Weber placed “rational capital accounting” among the defining features of modern capitalism. Historians have since disputed how much bookkeeping caused and how much it merely recorded. But the core point stands. Long before the computer, the grid of accounts was already a way of seeing an enterprise: as numbers, organised so they could be checked and calculated.

Early electronic grids

Computers took up accounting early, but for decades they processed transactions in batches rather than letting a manager explore a model. A few people saw the possibility sooner. In 1961 the accounting scholar Richard Mattessich published a paper on budgeting models and system simulation that described, in general terms, how the interlocking figures of an accounting spreadsheet might be computed by a program; his later books included a budgeting program written in FORTRAN.¹ In 1969 two engineers at Bell Canada, René Pardo and Rémy Landau, built LANPAR, a system that let budget analysts, rather than programmers, specify their models as arrays of cells with formulas. LANPAR could resolve the order in which cells had to be calculated so that every cell was computed after the cells it depended on, a problem now called natural-order recalculation. Its creators filed for a patent in 1970, which was eventually granted many years later.² Time-sharing services in the 1970s also offered financial modelling languages to large companies.

All of these ran on mainframes or minicomputers, owned by departments and operated by specialists. A manager who wanted to test a new assumption submitted a request and waited. The model belonged to the computer department. What changed in 1979 was not the idea of computing a spreadsheet, but who could do it, how quickly, and where.

VisiCalc: the magic blackboard

Bricklin and Frankston incorporated Software Arts in January 1979 and wrote VisiCalc for the Apple II, a small computer then sold mainly to hobbyists, schools and enthusiasts. A small publisher, Personal Software (later renamed VisiCorp), run by Dan Fylstra and Peter Jennings, marketed it from the middle of that year for about $100.³

The program’s design contained most of the ideas that still define the spreadsheet. The screen was a window onto a much larger grid of cells, labelled by column letters and row numbers, so that “B5” named a place. A cell could hold a number, a label or a formula referring to other cells. When any value changed, every formula that depended on it was recalculated at once. The user could scroll across the grid, copy a formula along a row so that its references adjusted, and see the whole model change in front of them.

That last property was the revolution. For the first time a manager could ask “what if?” and get an answer in seconds. What if prices rise by 5 percent? What if we hire two more salespeople in March? What if the loan rate is half a point higher? On paper each question meant an afternoon of recalculation; on VisiCalc it meant typing a number. The spreadsheet did not merely make the old calculation faster. It made a new kind of thinking cheap: exploring many possible futures instead of carefully computing one.

The effect on the computer industry was immediate. VisiCalc became the first program people bought a computer to use, the original “killer application.” Accounts differ on totals, but it sold on the order of 700,000 to a million copies over its life, and many buyers bought a $2,000 Apple II just to run it.³ Personal computers entered offices not because corporate computing departments chose them, but because managers bought them with their own budgets to run spreadsheets, often without asking anyone. The spreadsheet brought the personal computer into business through the side door.

The patent that was not filed

Bricklin and his colleagues did not patent the electronic spreadsheet, and the reason is instructive. Bricklin has explained that in 1979 software was rarely patentable in the United States, because programs were treated as mathematical algorithms. The publisher retained a patent attorney, who estimated about a 10 percent chance of success, even with techniques for disguising the software as a machine. “Given such advice, and the costs involved, we decided not to pursue a patent,” Bricklin wrote. Two years later the Supreme Court’s decision in Diamond v. Diehr made software patents far easier to obtain, but by then VisiCalc had been public for too long.⁴ Bricklin has said that if he invented the spreadsheet today, he would file for a patent, while also arguing that the late arrival of software patents caused problems for the industry.

The consequence was that the spreadsheet became, in effect, a shared idea. Anyone could build one, and many did.

Lotus 1-2-3: the spreadsheet as standard

In August 1981 IBM released its Personal Computer. It was more powerful than the Apple II, and above all it carried the name that corporate buyers trusted. VisiCalc was ported to it, but slowly and without taking advantage of the new machine.

Mitch Kapor, who had worked for VisiCorp, founded Lotus Development and, with the programmer Jonathan Sachs, built a spreadsheet designed for the IBM PC alone. Sachs wrote it in assembly language, the machine’s native instructions, which made it much faster than competitors written in portable languages. And it combined three tools in one, which gave it its name: a spreadsheet, a charting program and a simple database. Users could build a model, turn it into a graph and sort records without leaving the program. It also introduced macros, sequences of keystrokes that could be recorded and replayed, which let users automate repetitive work and, eventually, build small applications.

Lotus 1-2-3 went on sale on 26 January 1983. Its first-year sales were $53 million, against a business plan that had forecast about $1 million, and Lotus became one of the largest software companies in the world within a year.⁵ The IBM PC and 1-2-3 sold each other. For most of the 1980s, knowing 1-2-3 was a job skill listed in advertisements, and its slash-key menus, typed from memory by millions of office workers, became a kind of shared language. VisiCorp, embroiled in disputes with Software Arts, declined rapidly; Lotus bought Software Arts in 1985 and discontinued VisiCalc.

The spreadsheet takes over the office

Through the 1980s the spreadsheet changed the work of finance, planning and management. Steven Levy, in an article in Harper’s in 1984 titled “A Spreadsheet Way of Knowledge,” described the new breed of executives who modelled everything and the change in how decisions were argued. A line widely associated with his piece captures the shift: the spreadsheet is a tool, and it is also a world view, reality by the numbers.⁶

Several changes followed from that world view.

First, financial modelling became cheap enough to do constantly. Leveraged buyouts and the merger wave of the 1980s relied on spreadsheet models of cash flows and debt, which made it possible to test many deal structures quickly. Critics later argued that the ease of modelling had encouraged deals built on optimistic assumptions that nobody had examined closely enough. Whatever the balance, the speed of analysis changed the speed of finance.

Second, planning became negotiation over assumptions. Budget meetings increasingly turned into arguments about the inputs to a model, such as growth rates, conversion rates and headcount, because everyone understood that changing an input changed the conclusion. Managers learned to “work the spreadsheet” to reach a number they wanted, sometimes deliberately and sometimes not.

Third, the people who built the models gained influence. A junior analyst who understood the spreadsheet often understood the business’s assumptions better than the executives who signed off on the outputs. The spreadsheet moved some power from those who decided to those who modelled.

The platform war: Excel, Windows and OS/2

Microsoft had entered the market in 1982 with Multiplan, a spreadsheet designed to run on many computers. It lost decisively to 1-2-3 on the IBM PC. Microsoft drew a lesson from the defeat. Rather than build a better text-based spreadsheet, it built one for the next kind of computer.

Excel was released in 1985 for the Apple Macintosh, which had a graphical interface, windows, fonts and a mouse. Excel took full advantage of them: users could select ranges by dragging, format cells with different fonts, and see charts beside the numbers. On the Mac, it had little serious competition. In 1987 Microsoft released Excel for Windows, which at that point was a weak graphical layer on top of MS-DOS, used by few people.

The decisive choice was Lotus’s, and it is a good illustration of what this site calls moral luck: a sound bet judged harshly because of how it turned out. In the late 1980s IBM and Microsoft were jointly developing OS/2, which both companies presented as the future operating system of the PC. Lotus concentrated its development on a version of 1-2-3 for OS/2. Then, in 1990, Microsoft released Windows 3.0, which was a large commercial success, and the partnership between Microsoft and IBM came apart. OS/2 never achieved broad adoption (it is the subject of a road not taken on this site). Lotus’s Windows version of 1-2-3 arrived in 1991 and was widely judged disappointing.⁷ In 1990 Lotus still held about 54 percent of the spreadsheet market to Microsoft’s 12 percent. Over the following years, as Windows 3.1 spread and Microsoft bundled Excel with Word and PowerPoint in the Office suite, Excel overtook it.

It would be easy to call Lotus’s choice a blunder. It is more instructive to see it as a reasonable bet that went wrong. The two most powerful companies in personal computing had said OS/2 was the future. Lotus followed them. The mistake was not believing them; it was not keeping a strong option on the alternative. The company that won had made the opposite move: it had built for the graphical interface years before the market arrived there, and was ready when it did.

Owning the commands: Lotus v. Borland

While it lost the platform war, Lotus fought a legal one. By the late 1980s, 1-2-3’s menu commands had become something millions of people knew by heart: press the slash key, then “F” for File, “R” for Retrieve. Competitors built spreadsheets that offered the same commands so that users would not have to relearn them. Lotus sued. Against Paperback Software, which had closely imitated 1-2-3’s screens, Lotus won in 1990, in a decision that extended copyright to a program’s “look and feel.”

The case against Borland, whose Quattro Pro offered an optional mode with 1-2-3’s menu tree, went differently. Borland had not copied Lotus’s code, only the names of the commands and how they were organised. A district court ruled for Lotus, but in 1995 the US Court of Appeals for the First Circuit reversed, holding that the menu command hierarchy was a “method of operation,” which copyright law does not protect, comparing it to the controls of a car. In January 1996 the Supreme Court split four to four, which left the appeals court’s ruling in place without setting a national precedent.⁸

The case mattered far beyond spreadsheets, because it raised a question that recurs throughout software: who owns the knowledge users build up in learning a product? Lotus argued that the menus were its creative work. Borland argued, and the appeals court agreed, that once millions of people had learned to operate a program in a certain way, that way of operating could not be fenced off. The spreadsheet’s commands, like its basic idea, became common property.

In 1995 IBM bought Lotus for about $3.5 billion. It wanted Lotus Notes, the collaboration software, more than 1-2-3. The spreadsheet that had defined the PC was by then a declining product.

A better idea on the wrong machine: Improv and the pivot table

One of the most interesting spreadsheets ever designed came from Lotus at the moment of its decline. Lotus Improv, released in 1991 for Steve Jobs’s NeXT computer, was designed by a team led by Pito Salas, who had studied how people actually built financial models. He found that ordinary spreadsheets mixed three different things in the same cells: data, formulas and the labels that organised them. That made models fragile. Insert a row in the wrong place and formulas silently pointed to the wrong cells.

Improv separated them. Data sat in a grid organised by named categories, such as region, product and quarter. Formulas were written separately in terms of those names (“Profit = Revenue − Costs”) rather than cell addresses. And users could rearrange the categories by dragging them, so that a table of products by quarter became, instantly, a table of quarters by region.⁹ It was, many users felt, how a spreadsheet should have worked from the start.

Improv sold poorly. NeXT computers were rare and expensive, and the Windows version that followed in 1993 arrived in a market Excel already dominated. But its central idea survived in another form. Excel 5, released in 1993, introduced the PivotTable, which let users summarise and rearrange data by dragging category fields, in a feature that owed much to Improv’s design. The pivot table became one of the most valuable features in business software. The rest of Improv’s philosophy, the separation of formulas from data and the use of names instead of cell addresses, did not catch on for most users, and the fragility it addressed is with us still.

Excel becomes the world’s grid

Excel 5 also introduced Visual Basic for Applications, a real programming language inside the spreadsheet, which let organisations build complete applications on top of Excel. Through the 1990s and 2000s, Excel became more than a calculator. It became a database for people without databases, a reporting tool for people without reporting systems, a project tracker, a scheduling system, a CRM and, in many trading rooms, a pricing engine.

This was possible because the spreadsheet occupies a space between using software and writing it. Writing a formula like =SUM(B2:B13)*1.2 is programming: it specifies a computation that will be re-executed whenever its inputs change. But it does not feel like programming. The values are visible, the computation is laid out in space rather than hidden in a file of code, and the result appears immediately. The anthropologist and computer scientist Bonnie Nardi, in her 1993 study A Small Matter of Programming, argued that spreadsheets were the most successful end-user programming systems ever built, precisely because they let people with domain knowledge build their own tools without learning to think like programmers.¹⁰ An Excel product manager later put it more simply: Excel is, if you think about it, the most widely used programming language.¹¹

That success created a phenomenon that information technology departments came to call “shadow IT” or “end-user computing.” Critical business processes ran on spreadsheets that nobody in IT knew existed, built by one person, understood by one person and often maintained by copying and pasting. The European Spreadsheet Risks Interest Group, founded in 1999, began collecting cases of what went wrong.

The grid moves into the browser

In June 2006 Google opened Google Labs Spreadsheets to a limited number of testers. It was built on XL2Web, a web spreadsheet developed by a small company, 2Web Technologies, which Google had acquired; one of its founders, Jonathan Rochelle, became the product’s manager.¹² It was much less capable than Excel. It did one thing Excel could not do well: several people could edit the same spreadsheet at the same time, from any computer, without emailing files.

Google’s spreadsheet, later named Google Sheets and combined with Docs and Slides, did not take Excel’s place in finance departments, where its depth, speed and features mattered. It did change what a spreadsheet was for many people: a shared, living document rather than a file passed around. Schools, small businesses, start-ups, volunteer groups and journalists adopted it. Microsoft responded by putting Excel online and adding real-time co-authoring. The same shift happened to the word processor, as described in the essay on word processors: the document became a place rather than an object.

The 2010s produced a further family of descendants. Airtable, founded in 2012, combined a spreadsheet’s grid with a database’s structure. Notion, Coda and others built documents in which tables, formulas and text mixed. “No-code” tools let people build applications from spreadsheet-like data. Each borrowed the spreadsheet’s central promise, that people who understand a problem should be able to build their own tool for it, and each tried to fix some of its weaknesses.

When the grid is wrong

The spreadsheet’s greatest strength is also the source of its characteristic failures. Anyone can build one, and almost no one tests it. Errors in formulas are invisible unless someone clicks on the cell. References can point to the wrong row. Data can be pasted over a formula. A range can stop one row short. Unlike software written by professional programmers, spreadsheets are rarely reviewed, version-controlled or tested.

How common are errors? The information systems researcher Raymond Panko spent years collecting studies of spreadsheet errors. In field audits using careful methods, errors were found in the large majority of spreadsheets examined, at least 86 percent in the more recent studies, and in experiments the rate of errors per formula cell averaged a few percent, similar to error rates in professional programming before testing.¹³ The difference is that professional software goes through testing; most spreadsheets do not.

Some of the consequences have been very public.

In 1994 an accountant at Fidelity Investments copying the Magellan fund’s figures into a separate spreadsheet left out a minus sign, turning a $1.3 billion capital loss into a gain. The fund told investors to expect a large year-end distribution that did not exist, an error of about $2.6 billion, and had to withdraw it.¹⁴

In 2010 the economists Carmen Reinhart and Kenneth Rogoff published an influential paper suggesting that economic growth fell sharply when public debt exceeded 90 percent of a country’s output. In 2013 Thomas Herndon, a graduate student at the University of Massachusetts Amherst, tried to replicate it and, with his professors Michael Ash and Robert Pollin, found that a spreadsheet formula had left five countries out of one average. Correcting that and some disputed methodological choices raised the average growth figure for high-debt countries from −0.1 percent to 2.2 percent.¹⁵ Reinhart and Rogoff acknowledged the spreadsheet error and disputed the other criticisms. The paper had been cited in debates over government austerity, and the episode became a lesson in how a single cell can enter public policy.

In 2012 JPMorgan Chase’s Chief Investment Office lost about $6.2 billion on a portfolio of credit derivatives, in the episode known as the “London Whale.” The bank’s own internal task force found that a new value-at-risk model used to measure the portfolio’s risk was run on spreadsheets maintained by hand, with data copied and pasted between them, and that one calculation divided by the sum of two numbers where their average was intended, which had the effect of understating volatility. The model’s reviewers had approved it while recommending that it be automated.¹⁶ The spreadsheet did not cause the trading losses, but it helped hide the size of the risk.

In October 2020 Public Health England announced that 15,841 positive Covid-19 test results recorded between 25 September and 2 October had been left out of the daily figures. Laboratory results had been loaded into a spreadsheet file format with a fixed maximum number of rows, and results beyond the limit were dropped. The patients were informed, but tracing their contacts was delayed during a rising wave of infection.¹⁷

And in genetics, the spreadsheet changed the language of science itself. Excel automatically converts text that looks like a date, so the gene symbols SEPT1 and MARCH1 became “1-Sep” and “1-Mar.” A 2016 study found such errors in about a fifth of published papers that included Excel gene lists. In 2020 the HUGO Gene Nomenclature Committee renamed 27 genes, so that SEPT1 became SEPTIN1 and MARCH1 became MARCHF1, among its reasons that the old symbols affected data handling. Microsoft later added an option to turn off automatic conversion.¹⁸ Geneticists changed the names of human genes because the world’s most widely used analysis tool would not change its behaviour.

The spreadsheet becomes a programming language

Microsoft responded to the spreadsheet’s role as a programming tool by making it a more complete one. In 2019 and 2020 Excel gained dynamic arrays, in which a single formula can produce a whole table of results, and functions such as XLOOKUP, FILTER and LET, which let users name intermediate values. In December 2020 Microsoft introduced LAMBDA, which lets users define their own functions in the formula language and call them recursively. Microsoft researchers, including the computer scientist Simon Peyton Jones, noted that with LAMBDA the formula language had become Turing-complete, able in principle to express any computation.¹¹

In August 2023 Microsoft announced Python in Excel, which runs code in the Microsoft cloud using libraries supplied by Anaconda, letting analysts use the programming language of data science inside a cell; it became generally available for business users in 2024.¹⁹ In the same period Microsoft added Copilot, its AI assistant, which can generate formulas, charts, pivot tables and analyses from plain-language requests. Google added Gemini to Sheets.

These changes complete a long arc. The spreadsheet began as a way for non-programmers to compute without code. It has become a full programming environment, and now an environment where an AI writes the formulas for users who may not read them. Each step has made the spreadsheet more powerful. None has addressed the problem that has dogged it from the start: the person who relies on the model often does not understand how it works.

What the spreadsheet is today

Today the spreadsheet is three things at once.

It is a calculator for models: the place where budgets, forecasts and valuations are built, where most of the world’s financial decisions are first expressed as numbers. Every finance department, consultancy and investment bank runs on it.

It is a universal database for people without databases: lists of customers, inventories, schedules, research data, sign-up sheets and trackers. Much of the world’s operational data is kept in grids that were never designed to hold it.

And it is a programming environment for people who do not call themselves programmers, increasingly assisted by AI. Formulas, macros, scripts, Python and now plain-language requests let domain experts build tools that would otherwise require a software team.

Its form has hardly changed since 1979: a grid of cells with letters across the top and numbers down the side, formulas that recalculate when inputs change, and a window onto a larger sheet. That stability is itself remarkable. Few interfaces in computing have lasted so long with so little change, because few have matched so closely a way of thinking that people already had.

How the spreadsheet changed the philosophy of management

The spreadsheet’s influence is not mainly technical. It changed how organisations think about decisions. Several shifts stand out.

From one answer to many scenarios

Before the spreadsheet, a forecast was a carefully calculated estimate. After it, a forecast became a set of scenarios: base case, upside, downside, with sensitivity tables showing how the result moved as each assumption changed. That is a genuine advance in thinking under uncertainty. It also changed what a plan is: less a prediction than a map of possibilities, which is more honest, and more easily manipulated.

From judgement to assumptions

Spreadsheets force judgement to be expressed as numbers. A belief that “the market will grow quickly” becomes “growth: 18 percent.” This makes assumptions explicit and debatable, which is valuable. It also gives them false precision, and it shifts debate from whether the strategy makes sense to whether the number in cell C4 is right.

From the specialist to everyone

The spreadsheet put modelling in the hands of anyone with a computer. Managers no longer needed a data processing department to test an idea. This was a real democratisation of analysis, and it created a vast, unreviewed body of models on which decisions depend.

From reading to recalculating

On paper, a model was read. In a spreadsheet it is manipulated. The result is a habit of mind in which analysis means changing inputs and watching outputs, rather than understanding the mechanism between them. Many spreadsheet users can tell you what happens to profit if prices rise; fewer can tell you why. The essay on databases describes the opposite discipline, in which structure is declared before any data is entered.

From the visible to the hidden

VisiCalc’s cells displayed their results and hid their formulas. That design choice, made for good reasons of screen space, has shaped the spreadsheet’s whole epistemology. The output looks like data. The logic that produced it is invisible unless you look for it. A number in a cell carries no trace of whether it was typed, computed correctly or computed wrongly.

What the history suggests

The spreadsheet’s history repeats a pattern familiar from the rest of this site. The first mover, VisiCalc, defined a category and lost it. The second, Lotus, won by betting on the platform that became the standard and lost by betting on the platform that did not. The eventual winner, Microsoft, won by being ready for the next interface before the market arrived there, and kept its position through compatibility: a file format and a set of skills that millions of people shared. The most innovative design, Improv, appeared on the wrong machine and survived only as a feature in its rival.

But the spreadsheet also shows something more specific. Its central idea, a model anyone can change and watch respond, has not been improved upon in forty-five years, only extended. And the problems it creates have not been solved either. The tools that would make spreadsheets safer, such as named variables instead of cell references, separated formulas and data, testing and version control, have existed since Improv and before. Most users do not use them, because the spreadsheet’s great appeal is that it lets you start without structure. Its strength and its risk are the same property.

Understanding the philosophy

The rest of this essay examines the ideas beneath the spreadsheet, one at a time. Each is explained in plain terms, connected to the tradition of thought that has examined it, traced in the spreadsheet’s history, tested against its strongest objection, and related to how organisations decide today. The connections to philosophers, historians and sociologists are interpretations offered to make the ideas clearer. The spreadsheet’s inventors did not, as a rule, set out to answer them.

1. The ledger as a way of seeing

The idea in plain terms

A business is a tangle of people, goods, promises and hopes. A ledger turns it into columns of numbers that must balance. That transformation does not only record the business. It changes what the owners can see in it, and therefore what they pay attention to and decide.

The tradition

The economic historian Werner Sombart argued, in the early twentieth century, that double-entry bookkeeping was born of the same spirit as capitalism and essential to it, because it allowed the pursuit of profit to be calculated systematically and separated from the personal affairs of the owner. Max Weber, in Economy and Society, made “capital accounting” central to his account of modern economic rationality: an enterprise that calculates its capital in money, by bookkeeping, can compare the results of different courses of action. The literary historian Mary Poovey, in A History of the Modern Fact (1998), traced how double entry helped create a new kind of knowledge, the numerical “fact” that seems to describe the world neutrally. She argued that the balance of the books lent merchants’ records an appearance of accuracy and honesty, because the system’s formal consistency stood in for the truth of what was recorded. A ledger that balanced looked trustworthy, whether or not the entries were.

Where it shows up

The spreadsheet inherited the ledger’s grid and its authority. A model laid out in rows and columns, totalling neatly at the bottom, looks like a fact. The London Whale model and the Reinhart–Rogoff calculation both carried that authority, and in both cases the formal neatness concealed an error inside it. The spreadsheet also extended the ledger’s way of seeing to everything: not only money but sales leads, staff hours, customer satisfaction scores and clinical results became columns in a grid.

The strongest objection

Historians such as Basil Yamey argued against Sombart that many successful merchants did not use double entry at all, and that bookkeeping followed commercial development more than it caused it. The grid may be a useful tool rather than a way of seeing; people who use spreadsheets do not stop perceiving the world in other ways.

What it means today

Whether the grid causes a way of thinking or merely suits one, its effect on attention is real. What fits in a column gets tracked; what does not fit tends to be ignored. Leaders should notice what their models leave out, and should treat a model’s internal consistency as evidence that it adds up, not as evidence that it is true. Poovey’s point applies to every balanced spreadsheet: formal coherence is not accuracy.

2. The map is not the territory

The idea in plain terms

A model is a simplified picture of something complex. It is useful because it leaves things out. The danger is forgetting that it has left things out, and treating the picture as the thing.

The tradition

The phrase “the map is not the territory” comes from Alfred Korzybski, a Polish-American scholar who founded the discipline he called general semantics; he used it in the early 1930s to warn against confusing words and abstractions with what they describe. Jorge Luis Borges wrote a one-paragraph story, “On Exactitude in Science” (1946), about an empire whose cartographers made a map as large as the empire itself, point for point, which later generations found useless and left to rot in the desert. A perfect map is no map at all; usefulness requires omission. The statistician George Box gave the idea its best-known form in modelling: in 1976 he wrote that since all models are wrong, the scientist cannot obtain a “correct” one, and later added that some are nonetheless useful.

Where it shows up

A spreadsheet model of a company’s next five years is a map. It leaves out competitors’ responses, the morale of staff, the possibility of a pandemic, and most of what will actually decide the outcome. The leveraged buyouts of the 1980s, modelled in Lotus, sometimes failed because the map’s cash flows did not survive contact with the territory. The Public Health England episode showed another side of the problem: the map had a hard edge, a maximum number of rows, that the territory did not.

The strongest objection

Every decision uses some model, whether written down or held in someone’s head. A spreadsheet at least makes the model explicit and open to challenge, whereas intuition hides its assumptions. Criticising spreadsheets for being incomplete is criticising thought for being finite.

What it means today

The objection is right that the alternative to an explicit model is usually a worse implicit one. The lesson is not to abandon models but to hold them properly: to ask of every model what it leaves out, to stress-test it against events it did not anticipate, and to remember that the precision of the output says nothing about the completeness of the map.

3. Possible worlds: the discipline of “what if”

The idea in plain terms

To decide well you must imagine alternatives: what would happen if you did this, or that, or if the world turned out differently than you expect. The spreadsheet is a machine for imagining alternatives quickly.

The tradition

Gottfried Wilhelm Leibniz described the actual world as one of infinitely many possible worlds that God might have created, and argued that God chose the best of them. In the twentieth century the philosopher David Lewis developed a theory of counterfactuals, statements such as “if I had left earlier, I would have caught the train,” in terms of possible worlds: a counterfactual is true if, in the nearest possible worlds where its condition holds, its consequence also holds. In business, the strategist Pierre Wack developed scenario planning at Royal Dutch Shell in the early 1970s. Rather than predict one future, Shell’s planners described several coherent possible futures and asked how the company would fare in each. Wack argued that the purpose of scenarios was not to forecast but to change the mental models of decision-makers, so they would recognise a changing world sooner.

Where it shows up

VisiCalc’s defining feature, instant recalculation, made counterfactual reasoning nearly free. Sensitivity tables, scenario tabs and data tables are spreadsheets’ standard tools for exploring possible worlds. The best financial modellers use them as Wack intended: to understand which assumptions matter most and how the decision changes if they are wrong.

The strongest objection

Spreadsheet scenarios are not possible worlds in the rich sense. They vary a few numbers within a fixed structure. A real alternative future may change the structure itself: a new competitor, a new regulation, the collapse of a market. Varying growth from 5 to 15 percent explores a narrow corridor of the possible and may give false comfort that uncertainty has been considered.

What it means today

The objection identifies the most common misuse of the spreadsheet: mistaking sensitivity analysis for scenario thinking. Leaders should use both. Vary the inputs to see which assumptions matter, and then step outside the model to ask what events would change its structure. The most dangerous future is usually the one that does not fit in any of the scenario tabs.

4. Trust in numbers

The idea in plain terms

Numbers seem objective. A decision justified by a spreadsheet seems less personal, less arbitrary and easier to defend than one justified by someone’s judgement. That is why organisations rely on them, sometimes more than the numbers deserve.

The tradition

The historian of science Theodore Porter, in Trust in Numbers (1995), argued that quantification is a “technology of distance.” Communities of experts who know and trust each other can rely on judgement. Large organisations and public institutions, which must justify their decisions to outsiders who do not share that trust, turn instead to rules and numbers, which promise impersonal objectivity. Numbers, in Porter’s account, are a response to weakness and suspicion as much as a sign of rigour. A related warning is often called the McNamara fallacy, after the US Secretary of Defense who managed the Vietnam War by quantitative indicators such as enemy body counts. The social scientist Daniel Yankelovich described its steps: measure what can easily be measured; disregard what cannot; presume that what cannot be measured is not important; and finally, say that what cannot be measured does not exist.

Where it shows up

The spreadsheet made quantification cheap, and therefore made it the default form of justification. A proposal without a model looks unserious; a proposal with one looks rigorous, even if its key inputs are guesses. The Reinhart–Rogoff figure entered policy debate partly because a precise number carried authority that a qualitative argument would not. JPMorgan’s risk limits were governed by a number produced by a model, and when the model understated the risk, the number reassured the people who should have been alarmed.

The strongest objection

The alternative to numbers is often the unexamined judgement of powerful people, which can be biased, self-serving and impossible to challenge. Numbers, whatever their flaws, can be checked, argued with and audited. Porter himself acknowledged that quantification can protect the weak against the arbitrary decisions of the strong.

What it means today

Numbers are best treated as one input to judgement, not a substitute for it. A leader should ask of any figure who produced it, from what assumptions, and how it could be wrong, and should keep a deliberate place in decisions for what the spreadsheet cannot hold: the quality of a team, the trust of customers, the risk that has not happened yet. The objection reminds us that the answer to bad numbers is better numbers and better judgement, not their absence. Readers interested in how a measure distorts the thing it measures will find the related idea at When a Measure Becomes a Target.

5. Seeing like a spreadsheet

The idea in plain terms

To manage something large, you must simplify it. Organisations, like states, need to make complicated realities readable, so they can be counted, compared and controlled. The simplification that makes something manageable can also destroy what made it work.

The tradition

The political scientist James C. Scott, in Seeing Like a State (1998), described how modern states made their societies “legible”: they standardised surnames, measures and land titles, and planted forests in neat rows, so that they could tax, administer and plan them. The simplifications were useful, but grand schemes based on them often failed, because they ignored mētis, the practical, local knowledge that makes real systems work. The scientific forests of eighteenth-century Germany, planted in uniform rows for maximum timber, flourished for a generation and then declined, because the messy undergrowth they had cleared away had been sustaining the soil. This idea is explored further at Legibility and Mētis.

Where it shows up

The spreadsheet is a tool of organisational legibility. It turns a sales team into a pipeline of numbered stages, a hospital into utilisation rates, a school into test scores. These views are indispensable for management at scale. They are also partial, and decisions made solely from them can strip out the informal knowledge that kept things running: the salesperson whose relationships do not appear in the pipeline, the nurse whose judgement does not appear in the utilisation rate.

The strongest objection

Legibility is how organisations become fair and accountable. A spreadsheet of salaries can reveal pay gaps that informal knowledge concealed. A spreadsheet of outcomes can show that a treatment works better than doctors believed. Scott’s critique risks romanticising the local knowledge of insiders, which can also be prejudice.

What it means today

The answer is to use legibility with humility. Before acting on what the spreadsheet shows, ask what it cannot show and who knows it. Before restructuring around a metric, find the people whose informal work the metric misses. The most expensive decisions are often those that optimise the legible part of a system at the expense of the illegible part that sustained it.

6. False precision and the illusion of control

The idea in plain terms

A spreadsheet can calculate a figure to many decimal places from inputs that are really guesses. The precision of the output disguises the uncertainty of the input, and the ability to adjust every number creates a feeling of control over a future that is not controllable.

The tradition

The economist Frank Knight distinguished in 1921 between risk, where the probabilities of outcomes can be known, and uncertainty, where they cannot. Many business decisions face uncertainty in Knight’s sense, but spreadsheets represent them as if they were risk, with a single number or a neat range. The psychologist Ellen Langer showed in 1975 that people feel more in control of chance outcomes when they are actively involved, for example when they choose their own lottery ticket, and that this “illusion of control” makes them more confident than the odds justify. The related idea at Risk and Uncertainty traces Knight’s distinction further.

Where it shows up

A five-year projection showing revenue of $47.3 million in year four looks like knowledge. It is a chain of guesses multiplied together. Analysts who build such models, adjusting every assumption by hand, often come to feel that they understand the future better than they do. The London Whale model gave a precise measure of risk that was precisely wrong.

The strongest objection

Precise models are not the problem; misreading them is. A good analyst knows that the decimals are an artefact of calculation and reports ranges. And modelling, by forcing assumptions to be stated, often reduces overconfidence by revealing how much depends on a few uncertain inputs.

What it means today

The objection describes good practice, which is not common practice. Leaders can enforce it: ask for ranges rather than single figures, round outputs to the precision the inputs justify, and ask which two or three assumptions drive the result. When a model seems to give control over the future, treat that feeling as a warning rather than a reassurance.

7. Programming for everyone

The idea in plain terms

The spreadsheet lets people who are not programmers build software. That is a great democratisation of computing, and it means much of the software organisations rely on was written by people who were never taught to test it.

The tradition

Bonnie Nardi’s A Small Matter of Programming (1993) argued that end-user programming succeeds when it lets people work with the concepts of their own domain, as spreadsheets do, rather than forcing them to think in the abstractions of computer science. She also showed that spreadsheet use is often collaborative: local experts help colleagues, and knowledge spreads informally. A broader tradition, going back to Ivan Illich’s Tools for Conviviality (1973), asked which tools enlarge people’s independence and which make them dependent on specialists. Illich called the first kind “convivial” tools. The spreadsheet may be the most convivial tool in the history of computing.

Where it shows up

VisiCalc took modelling away from the data processing department and gave it to managers. Lotus macros and Excel’s VBA let them build their own applications. Google Sheets let them share those applications. Every generation enlarged what non-programmers could build. And every generation enlarged the risk: Panko’s studies show error rates comparable to those of professional programmers, without the testing that catches them.

The strongest objection

The real choice is not between amateur spreadsheets and professionally engineered systems; it is between amateur spreadsheets and nothing, because professional software teams would never have built most of these tools. The errors are the price of a vast amount of useful work that otherwise would not exist.

What it means today

The objection is correct, and it points to the right policy. Organisations should not ban end-user computing; they should identify the spreadsheets that have quietly become systems, those on which money, health or legal obligations depend, and give them the review, testing and ownership that systems need. AI assistants that write formulas will make the democratisation larger and the need for that discipline greater.

8. The glass box: why visible computation matters

The idea in plain terms

In a spreadsheet, every calculation can in principle be inspected. Click on any number and you can see how it was made, trace its inputs and follow the chain back to the data. That transparency is an unusual and valuable property, and it is now under pressure.

The tradition

The cognitive scientist Edwin Hutchins, in Cognition in the Wild (1995), studied how the navigation team of a US Navy ship computed its position, and argued that the cognition was distributed across people, instruments and charts. The tools were not merely aids; they were part of the thinking, and their visible representations let team members check each other’s work. Philosophers of science describe a similar ideal of “epistemic transparency”: a method is trustworthy when its steps can be followed and checked by others. The essay on large language models on this site, The Machine That Answers, describes a technology whose reasoning is largely opaque; the spreadsheet sits at the opposite pole.

Where it shows up

Herndon could find the Reinhart–Rogoff error because the authors shared their spreadsheet and its formulas could be read. JPMorgan’s task force could identify the faulty calculation because it was there in a cell. Spreadsheet errors are common, but they are also discoverable, which is more than can be said for many computational systems. The trend of recent decades, toward macros, scripts, Python and AI-generated formulas, has made spreadsheets more powerful and less transparent at the same time.

The strongest objection

Transparency in principle is not transparency in practice. Large spreadsheets contain thousands of formulas that no one will read, and the fact that an error could be found does not mean anyone looks. Most of the famous errors were found by chance, long after they mattered.

What it means today

The objection is a call to use the transparency the spreadsheet offers. For any model that matters, someone other than its author should trace its key calculations. As AI writes more formulas, organisations should keep the spreadsheet’s great virtue, that its reasoning is visible, by requiring that generated logic be readable and reviewed, rather than letting the grid become another black box.

Bringing the threads together

The spreadsheet inherited from the ledger a way of seeing an enterprise as numbers that balance, and gave it to everyone. It made counterfactual thinking cheap, which is its great contribution to decision-making. It also made models look like facts, assumptions look like knowledge, and the parts of a business that fit in a column look like the whole business. It turned millions of people into programmers without teaching them to test. And it offered something rare, computation that can be inspected, which its users have mostly not inspected. The spreadsheet’s history is therefore a history of a tool that improved judgement when it was held as a map, and replaced judgement when it was mistaken for the territory.

The machine for asking what if

Dan Bricklin wanted a blackboard that would recalculate itself. What he made changed how organisations think: every plan became a model, every assumption a cell, every meeting an argument about the inputs. It is hard to imagine modern management without the spreadsheet, and hard to find an important decision in which one did not play a part.

The spreadsheet’s history suggests some working principles for anyone who relies on it. Treat a model as a map: ask what it leaves out. Use scenarios to understand which assumptions matter, and then imagine the futures that no scenario tab contains. Distrust precision that the inputs cannot support. When a spreadsheet becomes a system, give it the review and ownership a system deserves. And keep the spreadsheet’s great virtue, visible reasoning, alive by actually looking at the formulas. The question the spreadsheet made easy, “what if?”, is the right question. It deserves to be asked of the model as well as of the world.

Selected bibliography

• Pacioli, Luca. Summa de arithmetica, geometria, proportioni et proportionalità. Venice, 1494.
• Goethe, Johann Wolfgang von. Wilhelm Meister’s Apprenticeship. 1795–96.
• Sombart, Werner. Der moderne Kapitalismus. 1902; revised 1916.
• Weber, Max. Economy and Society. 1922.
• Yamey, Basil S. “Scientific Bookkeeping and the Rise of Capitalism.” Economic History Review 1, no. 2/3 (1949): 99–113.
• Poovey, Mary. A History of the Modern Fact. University of Chicago Press, 1998.
• Mattessich, Richard. “Budgeting Models and System Simulation.” The Accounting Review 36, no. 3 (1961): 384–397.
• Bricklin, Dan. “Patenting VisiCalc” and “The Idea.” bricklin.com
• Levy, Steven. “A Spreadsheet Way of Knowledge.” Harper’s, 1984.
• Nardi, Bonnie A. A Small Matter of Programming: Perspectives on End User Computing. MIT Press, 1993.
• Lotus Development Corp. v. Borland International, Inc., 49 F.3d 807 (1st Cir. 1995), aff’d by an equally divided court, 516 U.S. 233 (1996).
• Panko, Raymond R. “Spreadsheet Errors: What We Know. What We Think We Can Do.” Proceedings of EuSpRIG (2000); arXiv:0802.3457.
• Herndon, Thomas, Michael Ash and Robert Pollin. “Does High Public Debt Consistently Stifle Economic Growth? A Critique of Reinhart and Rogoff.” Cambridge Journal of Economics 38, no. 2 (2014): 257–279.
• Reinhart, Carmen M., and Kenneth S. Rogoff. “Growth in a Time of Debt.” American Economic Review 100, no. 2 (2010): 573–578.
• JPMorgan Chase & Co. Report of JPMorgan Chase & Co. Management Task Force Regarding 2012 CIO Losses. 16 January 2013.
• Ziemann, Mark, Yotam Eren and Assam El-Osta. “Gene Name Errors Are Widespread in the Scientific Literature.” Genome Biology 17 (2016): 177.
• Bruford, Elspeth A., et al. “Guidelines for Human Gene Nomenclature.” Nature Genetics 52 (2020): 754–758.
• Gordon, Andy, and Simon Peyton Jones. “LAMBDA: The Ultimate Excel Worksheet Function.” Microsoft Research blog, 25 January 2021.
• Korzybski, Alfred. Science and Sanity. 1933.
• Borges, Jorge Luis. “On Exactitude in Science.” 1946.
• Box, George E. P. “Science and Statistics.” Journal of the American Statistical Association 71 (1976): 791–799.
• Lewis, David. Counterfactuals. Blackwell, 1973.
• Wack, Pierre. “Scenarios: Uncharted Waters Ahead.” Harvard Business Review, September–October 1985.
• Porter, Theodore M. Trust in Numbers: The Pursuit of Objectivity in Science and Public Life. Princeton University Press, 1995.
• Yankelovich, Daniel. Corporate Priorities: A Continuing Study of the New Demands on Business. 1972.
• Scott, James C. Seeing Like a State. Yale University Press, 1998.
• Knight, Frank H. Risk, Uncertainty and Profit. 1921.
• Langer, Ellen J. “The Illusion of Control.” Journal of Personality and Social Psychology 32, no. 2 (1975): 311–328.
• Illich, Ivan. Tools for Conviviality. Harper & Row, 1973.
• Hutchins, Edwin. Cognition in the Wild. MIT Press, 1995.

Research limitations

Bricklin’s account of the blackboard and of the patent decision is his own, published on his website. VisiCalc sales figures vary across sources between about 700,000 and a million copies. Lotus’s first-year revenue is widely reported; unit sales for the same period are inconsistent across sources and are not used. The line from Steven Levy’s 1984 article is quoted as it was attributed to that article by NPR’s Planet Money; the original text was not consulted, and the article’s month is given as reported in bibliographic listings, which differ. Market-share figures for Lotus and Microsoft in 1990 come from secondary sources. The account of Lotus Improv and its influence on the pivot table relies on secondary histories, which differ slightly on dates. The 2Web acquisition is dated 2005 in contemporary reports and 2006 in some later ones. Panko’s error rates come from his surveys of others’ studies, which vary in method. The JPMorgan task force report is described from contemporary press summaries; the quoted calculation error appears in those summaries. The Reinhart–Rogoff dispute involved methodological disagreements beyond the spreadsheet error, which the authors contested; this essay does not take a side on the economics. The Public Health England figures are from its own statement, and the role of the file format from contemporary reporting. The decisions in the decisions box are judged by the author, separating what could be known at the time (process) from what happened (call). The philosophical connections to Sombart, Weber, Poovey, Korzybski, Borges, Box, Leibniz, Lewis, Wack, Porter, Yankelovich, Scott, Knight, Langer, Nardi, Illich and Hutchins are the author’s interpretation; their works are paraphrased, apart from short, attributed phrases.

Footnotes

1. Richard Mattessich, “Budgeting Models and System Simulation,” The Accounting Review 36, no. 3 (July 1961); see the summary at the University of Maryland’s early history of spreadsheets. ↩

2. On LANPAR, see “A look back: LANPAR, the first spreadsheet” and historyofinformation.com. ↩

3. “Today in Apple history: VisiCalc,” Cult of Mac; Centre for Computing History; “The VisiCalc Dawn,” Strategic Finance, June 2019. ↩ ↩²

4. Dan Bricklin, “Patenting VisiCalc,” bricklin.com. ↩

5. Wikipedia, “Lotus Development”; “Lotus 1-2-3 hits the big 3-0,” InfoWorld; on assembly language, Encyclopedia.com, “Lotus Development Corporation”. ↩

6. Steven Levy, “A Spreadsheet Way of Knowledge,” Harper’s (1984); the line as attributed to Levy’s article in NPR, Planet Money, “Episode 606: Spreadsheets,” 25 February 2015, npr.org. ↩

7. “The dilemma of Lotus 1-2-3,” Korn Ferry Briefings; Dave Farquhar, on market share and the Windows version. ↩

8. Lotus Development Corp. v. Borland International, Inc., 49 F.3d 807 (1st Cir. 1995), aff’d, 516 U.S. 233 (1996); see Wikipedia. ↩

9. Wikipedia, “Lotus Improv”; “The history of the pivot table, the spreadsheet’s most powerful tool,” Quartz. ↩

10. Bonnie A. Nardi, A Small Matter of Programming: Perspectives on End User Computing (MIT Press, 1993). ↩

11. Andy Gordon and Simon Peyton Jones, “LAMBDA: The ultimate Excel worksheet function,” Microsoft Research blog, January 2021; Brian Jones’s remark as reported by TechRepublic. ↩ ↩²

12. Wikipedia, “Google Sheets”; “Google targets Excel with online spreadsheet,” TechCentral, June 2006. ↩

13. Raymond R. Panko, “Spreadsheet Errors: What We Know. What We Think We Can Do,” arXiv:0802.3457; see also Panko, “What We Don’t Know About Spreadsheet Errors Today,” arXiv:1602.02601. ↩

14. RISKS Digest 16.72 (January 1995), quoting Fidelity’s letter to shareholders; Wikipedia, “Fidelity Magellan Fund”. ↩

15. Thomas Herndon, Michael Ash and Robert Pollin, “Does High Public Debt Consistently Stifle Economic Growth?” PERI Working Paper 322 (April 2013); see The Conversation and Retraction Watch. ↩

16. JPMorgan Chase & Co., Report of JPMorgan Chase & Co. Management Task Force Regarding 2012 CIO Losses (16 January 2013), as summarised by James Kwak, “The Importance of Excel”, and Bloomberg. ↩

17. “Excel: Why using Microsoft’s tool caused Covid-19 results to be lost,” The Register, 5 October 2020. ↩

18. Mark Ziemann, Yotam Eren and Assam El-Osta, “Gene name errors are widespread in the scientific literature,” Genome Biology 17 (2016); Elspeth A. Bruford et al., “Guidelines for human gene nomenclature,” Nature Genetics 52 (2020); see GenomeWeb. ↩

19. “Python in Excel now generally available,” The Register, 18 September 2024; Microsoft Support, “Introduction to Python in Excel”. ↩