What Spreadsheets Teach Us About AI: Democratized Power, Hidden Risk

The computerized spreadsheet is the best example of a tool that began as a productivity aid and quietly turned into a management philosophy. VisiCalc and Lotus 1-2-3 did more than speed up arithmetic. They changed who was allowed to model a business decision in the first place.

Before spreadsheets, financial modeling was slow, specialized, and tied to formal systems or painstaking manual recalculation. After spreadsheets, a manager, analyst, entrepreneur, or consultant could build a model, swap assumptions, test scenarios, and present conclusions with a speed and confidence that had not existed before.

That shift in who got to do the analysis ended up mattering more than the raw speed of the calculation. Analytical power moved outward, away from a small priesthood of specialists and toward anyone willing to learn the grid.

Lotus 1-2-3 shows the magnitude. VisiCalc had helped carry the Apple II; Lotus 1-2-3, with charting, graphing, and macros, quickly outsold it, brought in 53 million dollars in Lotus's first year, and went on to dominate business software through the mid-to-late 1980s.

It became one of the defining applications of the personal-computing era because it put computation, modeling, presentation, and user control in one place. It did not so much automate accounting as hand analytical production to everyone.

That is the thread running straight to AI. The headline revenue number is the least interesting thing about it. What matters is what happens to an organization when a capability that used to require specialists suddenly belongs to anyone.

When analytical power spreads, influence spreads with it. A junior analyst who was good with a spreadsheet could shape investment decisions, staffing plans, pricing, acquisition analysis, and budgets. The spreadsheet made the underlying business logic visible and easy to push on.

It made what-if reasoning fast in a way earlier systems never could. Organizations could model uncertainty, compare options, and produce persuasive quantitative stories on demand. Often, whoever built the model quietly shaped the decision.

AI extends the same dynamic. Whoever can coax a useful answer out of a model, frame the prompt, supply the right context, and package the result picks up influence the way the spreadsheet-literate analyst once did. Capability migrates to whoever handles the new tool well.

The same qualities that made spreadsheets powerful also made them dangerous. In a 1984 essay, Steven Levy noted that spreadsheets let users reproduce the relationships among the parts of a business and build models from them, and he warned that bad assumptions could hide not just in the data but in the formulas and relationships that drive the whole model.

That distinction matters more than it sounds. A spreadsheet can be wrong even when every number is typed correctly, because the structure itself misrepresents reality. The output looks rigorous while resting on fragile assumptions, and the polish tells you nothing about whether it is right.

Anyone who has inherited a sprawling financial model knows the feeling: the formatting is immaculate, the totals foot, and somewhere three sheets deep a single hardcoded growth rate is quietly doing all the work. The artifact carries an authority the underlying logic never earned.

This maps almost exactly onto AI, with the volume turned up. A spreadsheet can make a flawed model look precise; AI can make flawed reasoning sound fluent. Where the spreadsheet buried its assumptions in formulas, AI buries them in prose, code, summaries, and recommendations.

Executives trust a spreadsheet because it looks quantitative. Readers trust an AI answer because it sounds articulate and sure of itself. In both cases the form is signaling a rigor the content may not actually have.

None of this makes spreadsheets bad, or AI bad. The point is that both tools move the boundary between expert and non-expert production, and that fluency and accuracy are not the same thing. The more polished the output looks, the more deliberate the checking has to be.

When more people can produce sophisticated outputs, organizations need new review disciplines, and spreadsheets eventually forced the issue. In high-stakes settings they led to model review, audit practices, version control, access controls, and financial-model standards.

AI will need its own versions of those: source verification, model evaluation, human review at defined points, provenance tracking, output testing, and clearer accountability for AI-assisted work. The specifics differ; the principle does not. Democratized production demands review that scales with it.

The organizations that learned to govern spreadsheets did not ban them. They built the review layer that let non-specialists produce safely. The same play works for AI, and the playbook later in this series gets specific about how.

Spreadsheets also reshaped how managers think. Once anyone could model a scenario in an afternoon, management got more quantitative, more scenario-driven, more comfortable with abstraction. That brought real gains and a quiet bias with it: the tendency to assume that whatever could be modeled was what mattered most.

AI may pull management the same way. Once an organization can generate documents, analyses, code, and recommendations on demand, it starts to overvalue speed and polish and undervalue judgment, context, and verification. The artifact gets more convincing even when the thinking behind it has not improved at all.

The managerial job is to take the productivity and not hand your judgment over to the artifact. Use the tool to produce more, and faster, but keep treating fluent output as a draft to be checked rather than a conclusion to be trusted.