The 2%

About six months ago something changed about how I work.

I’ve been writing software for twenty years. I started in 2006 building automated credit decisioning systems for personal loans and mortgages. I worked on hypervisor code at a cloud company that’s now worth ten billion dollars. I built a performance testing company that HotStar used to simulate the two billion live streamers who tuned into the IPL in 2019, and later sold it. By rough count, I have spent something like thirty thousand hours in front of an editor.

For most of that time my output was a straight line. Hours in, code out. I got faster, but the relationship between effort and result was linear. If you’d asked me to draw it on a graph, I could have.

Six months ago the line broke.

Last week I needed an iOS app for one of our products. The old playbook would have been to pick a hybrid framework like React Native, accept a slightly worse experience on both platforms in exchange for not having to write two apps. The new playbook is to build the iOS app properly, in SwiftUI, with Claude. Then have Claude build a native Android version that follows Android’s own conventions. Two genuinely native apps. Three days.

Used to be a team. Maybe two teams. The better part of a year.

The interesting part isn’t the speed. It’s that the compromise went away. I didn’t pick a worse architecture to save time. I picked the right architecture for each platform, because the cost difference disappeared.

That’s a small example of something much larger.

About half of workers in advanced economies now use generative AI daily. Only one in eight say it’s changed how their organisation operates. And of those, only a fraction — call it 2% — have actually re-engineered how the work itself gets done.

The other 98% are using AI to type faster.

This is the part nobody quite says out loud. Most uses of AI right now are decorative. People draft an email and have AI tighten it. They write a brief and have AI summarise it. They produce the same outputs they always produced, but with fewer keystrokes. The underlying business runs the same way it ran two years ago.

The 2% are doing something different. They’ve stopped using AI as a tool and started using it to ask whether the job should still exist.

For a decade, every essay about which jobs AI wouldn’t replace had programming near the top of the list. Programming is technical. It takes years to learn. There’s a famous claim that mastery requires ten thousand hours, and that claim was almost certainly low for serious software work. Programming requires creativity, abstract reasoning, and judgement about messy real-world systems. If any white-collar job was going to hold the line, it was this one.

It didn’t hold the line.

I’m not saying programmers are out of work. Most of us are doing fine. But the shape of the work has changed completely. The job I trained twenty years for — typing code, debugging line by line, holding the model of an entire system in my head — that job is essentially gone. What’s replaced it looks more like what a creative director does: pointing at the product, applying taste, deciding what’s worth building, managing a flock of agents that do the actual construction.

If programming wasn’t safe, what is? Look at the work your team does and ask yourself: was programming meaningfully harder to automate than what your accountant does? Your operations manager? Your associate three years out of school? Your back-office team running invoices and onboarding?

The honest answer is no. Programming was the canary in the coal mine. The canary just died.

The instinct, when you hear “AI is coming for white-collar work,” is defensive. Protect your job. Protect your team. Treat AI as a threat. That’s only the right instinct if your work is the kind that gets replaced.

There are two kinds of work in any business. The first kind is linear: time is the product. One hour in, one unit out. Data entry, bookkeeping, document handling, scheduling, routine compliance. The work itself is execution — typing, processing, moving things from one place to another. AI doesn’t help with this work. AI replaces it.

The second kind of work is non-linear: time directs the product. One hour in, ten or a hundred units out. Strategy, system-level decisions, directing agents, client relationships. The work itself is directing taste — pointing it at the right problems, recognising what’s worth doing, knowing when an output is excellent versus merely correct.

AI doesn’t replace this kind of work. AI multiplies it.

There’s a term from cybernetics that fits here. In the 1950s a researcher named Ross Ashby coined “Intelligence Amplification” — the idea that the most powerful use of computing wouldn’t be to replace human thought but to amplify it. The term was largely forgotten, partly because the technology to do it didn’t exist yet. It exists now. The 2% aren’t using Artificial Intelligence. They’re using Intelligence Amplification.

I gave a version of this argument recently in Queenstown, where I live. Queenstown is a small town in New Zealand built almost entirely on selling experience to people who flew here for it. Tourism makes up about a third of our GDP, against four per cent nationally — an eight-times concentration in the part of the economy where the product is a human experience. Unemployment here is around 1.4%. Nationally it’s over five. We are operating at near-total labour capacity, which means the constraint isn’t demand. It’s people.

So when I make the case that AI is coming for white-collar work, the room hears it differently than it would in most rooms in New Zealand. For these businesses, AI isn’t a threat to the workforce. It’s a force multiplier for the workforce they can’t hire.

I think this is true of most businesses, not just ones in tourist towns. Clients don’t pay their wealth manager for paperwork. They pay for judgement. They don’t pay their architect for AutoCAD time. They pay for taste. They don’t pay their tour operator for logistics. They pay for the experience itself. The administrative work was never what they were buying. AI just makes that visible faster than it would have been otherwise.

When the cost of execution falls, the part of the work that’s left — the part the client was always actually paying for — gets more valuable, not less.

If you want to be in the 2%, the path is unglamorous. Don’t build an AI strategy. Don’t hire a head of AI. Don’t write a roadmap.

Find one task. One repetitive thing eating five or more hours a week somewhere in your business. Invoice coding, lead qualification, report generation, email triage — pick one. Hand it to an autonomous tool. Watch what happens to the people who used to do it. They won’t run out of work. They’ll move up the value chain.

Do it once. Learn what worked and what broke. Then do it again.

That’s the whole programme.

The 2% aren’t working harder. They aren’t running faster. They’ve just stopped doing the work that doesn’t require them.