The Spreadsheet Moment
In 1979, a piece of software called VisiCalc turned the Apple II from a hobbyist toy into a business essential. Steve Wozniak later noted that small businesses, not the hobbyists he and Steve Jobs had expected, purchased 90% of Apple IIs.
We're living through another such moment. AI code generation tools are doing for software what VisiCalc did for financial modelling: putting creation power into the hands of people who understand problems, not just those trained to build solutions.
The VisiCalc Revolution
Dan Bricklin and Bob Frankston released VisiCalc on October 17, 1979, selling it for under $100. It would go on to sell over 700,000 copies. Harvard Business School later put up a plaque commemorating Bricklin in the room where he'd studied, declaring that VisiCalc "forever changed how people use computers in business."
Before VisiCalc, financial modelling required teams of clerks working for days or weeks with calculators and ledger paper. After VisiCalc, a single accountant could build sophisticated models in hours. As technologist Stewart Alsop wrote in 1984, VisiCalc "empowered thousands of business managers to escape the data processing department for mundane planning work."
Attractor state — the steady state that an industry is drawn to in response to a new innovation
The spreadsheet became what I'd call an attractor state for personal computing. Once VisiCalc demonstrated what was possible, the industry kept moving towards it until the vision became reality. Today, Excel has over 750 million daily users. The accountant's escape from the data processing department is complete.
The New Escape Route
By 2025, roughly 85% of developers were regularly using AI coding tools. More interesting than developers working faster is non-developers working at all.
The tools have proliferated — Cursor, Replit, Claude Code, and others — each lowering the barrier between having an idea and having software.
Over the holidays, I took a vague idea (a personalised news summary waiting on my printer each morning) and had it production-ready in under an hour. It was working software, for a customer of one. I'm already planning to add random family photos and connect it to my project management tools. The foundation is solid; iteration is easy.
The ultimate personalisation is custom-built software.
Just as VisiCalc let accountants escape the data processing department, AI coding tools let domain experts escape the engineering department. The person who understands a problem best can now build its solution. The consultant can create their own workflow tool. The operations manager can automate their own processes. The entrepreneur can prototype before hiring. (When that abundance becomes a problem rather than a gift, I've written about that separately as the end of simplicity.)
The new attractor state is coming into focus: personal, purpose-built software for everyone. Applications tailored to individual needs rather than subscriptions to generic SaaS tools, the digital equivalent of a custom spreadsheet that only you understand and only you need.
The parallel isn't perfect. Spreadsheets have bounded domains while code is limitless, and the failure modes are different. But the pattern of who gets to build is the same.
Easy to Create, Hard to Maintain
The productivity story is messier than it feels. A 2025 study by METR found that experienced open-source developers who believed AI had made them 20% faster had in fact taken 19% longer. The productivity gains are real but complicated.
These tools create what I'd call "good enough" software: functional, often impressive, but not always reliable. It's rather like Wikipedia: fine for quick answers, but you'd verify before court.
The deeper question is maintenance, and here the spreadsheet parallel becomes uncomfortable. Spreadsheets are famously easy to create and notoriously hard to maintain. Every organisation has mission-critical spreadsheets that terrify anyone who opens them. Built by someone who understood them at the time. Often that person has moved on.
I have dozens of files titled "business model" with version postfixes and dates. business_model_v2.xlsx. business_model_2024_final.xlsx. business_model_2024_final_FINAL.xlsx. When the assumptions changed, it was easier to start fresh than untangle the old version.
Engineering overreach — systems whose failure modes exceed their creators' ability to comprehend
AI-generated code will follow this exact pattern. If the news summary I built over the holidays breaks in six months, I'll probably rebuild it from scratch rather than debug it. And that might be fine for personal tools, for prototypes, for software with a customer of one.
The question is what happens when "customer of one" software quietly becomes "customer of many." When the operations manager's automation tool becomes load-bearing infrastructure. When the consultant's workflow app gets handed to a successor who can't prompt their way through the fixes.
Software isn't a document you write once; it's a garden you tend indefinitely. Unless it's easier just to rebuild.
Where This Leaves Us
We're in the early-adopter phase. The chasm, as Geoffrey Moore would say, hasn't been crossed, but the attractor state is visible.
Like spreadsheets before them, AI-generated applications will become invisible infrastructure. They'll power countless small businesses, side projects, and internal tools. Some will grow into real products. Many will be abandoned when their creators move on, leaving behind digital archaeology for future maintainers to puzzle over.
VisiCalc's creators couldn't have predicted Excel running the world's financial infrastructure. We can't predict what a generation of people building with AI will create.
AI will change who builds software. What matters is whether we're ready for a world where the answer is "everyone", and whether we're honest about where we actually are on the journey.