In 2003 I was 13 years old and I built my first website for money. A four-page brochure site for a dog breeder, hand-coded in HTML, and I charged about £500 for it. I remember being amazed that anyone would pay that for what felt to me like a couple of weekends of work.
I think about that site sometimes when I read the current wave of “AI will end paid software work” coverage. I’ve watched almost exactly this pattern happen twice already. Both times, someone in my industry said the next tool was going to shrink the work for developers. Both times, the next three years made them wrong, not because the work stayed the same, but because what counted as “the work” kept moving up the stack.
Round three is happening now. From the inside, it looks the same shape.
Round one and round two looked the same
Round one was WordPress. It launched on 27 May 2003, the same year I built the dog-breeder site, but it didn’t really matter to my work until the plugin system arrived in version 1.2 in May 2004 and the theme system in 1.5 in February 2005. By around 2007 the ecosystem was mature enough that small businesses started asking for “a website” and meaning something I would previously have called a content management system. Same client profile. Same broad budget category. But the deliverable was bigger: a CMS they could update themselves, a contact form that emailed them, a blog, basic search.
The cheque grew, but the scope grew faster. My WordPress sites might have come in at one to two thousand pounds against the old £500 floor, but the amount of functionality crammed into that fee multiplied by something like ten, because most of the heavy work was already handled by WordPress and its plugin ecosystem. I wasn’t reinventing the underlying wheels.
Round two was PHP frameworks. CodeIgniter shipped in February 2006 and Laravel followed in June 2011. Same shape as round one. The same client profile started asking, at a similar price band, for custom web applications: an inventory system, an e-learning platform, a custom order workflow. The cheque grew again, but the scope grew faster. The thing the client meant by “an app” had quietly redefined itself, because the framework took care of database scaffolding, authentication, routing, and most of the small dull pieces that used to dominate the build.
Two rounds, same shape. The price climbed. The scope climbed faster. What “a website” or “an app” meant kept expanding in the client’s head, because the per-feature cost on the developer’s side kept falling.
What I didn’t realise at the time
Looking back, what I thought was happening was “I can do more in the same time, so I’m getting more efficient.” What was actually happening was bigger than that. The cost-per-feature on the supply side was falling, but the floor of who could afford a custom build kept dropping along with it. Companies that couldn’t justify a £5,000 inventory system in 2006 could justify a £6,000 one in 2012 that did three times as much, because that fee now bought a custom-fit thing rather than a heroic hand-build from scratch.
That economic pattern has a name. The British economist William Stanley Jevons described it in 1865 in the context of coal: as steam engines got more efficient, total coal consumption went up, not down, because efficient engines opened new uses that hadn’t been economic before. The same shape applies to software almost cleanly. As the tools got better, the total amount of paid software work didn’t shrink. It grew, because cheaper unit cost expanded the market.
The current data backs the pattern continuing. Stack Overflow’s 2025 Developer Survey shows 84% of developers using or planning to use AI coding tools, and the labour market headline numbers from CNN’s April 2026 piece on developer employment point in the same direction: Indeed listings for software engineers up 11% year on year, the US Bureau of Labor Statistics projecting +15% growth in developer employment through 2034. The total work keeps going up.
Round three is eating SaaS, not developers
The freshest thing I’ve noticed in my own work in 2026 doesn’t fit the “AI will replace developers” headline at all. The layer it’s actually replacing is the one sitting above developers: software as a service. Specifically, the modular, single-purpose SaaS tools that small companies bolt on around their core product because building them in-house wasn’t economic.
A concrete example. I was using a service called Canny to track a public feature roadmap and bug requests for Podtastic. Last month they raised their pricing and disabled some of the features I was using. In a year that was 2024 my response would have been to shop around for a competitor SaaS, eat the migration cost, move on. In 2026 my response was to trigger an agent in the background to build a roadmap and bug-tracking system directly into Podtastic. A few database tables. A few extra screens on the app and the admin backend. Some status emails to users when a requested feature ships. Done in a working day rather than a fortnight, integrated with the user accounts I already had, no recurring fee.
That category of build wasn’t economic before. Building your own roadmap tool, your own help desk, your own analytics dashboard, your own A/B test framework, was the kind of thing a small team specifically avoided, because the unit cost of doing it well was higher than subscribing to a SaaS for thirty quid a month. The SaaS layer existed precisely because building the thing yourself was uneconomic at small scale.
What changed in the last twelve months is the unit cost of building those things yourself. It collapsed. Which means a category of software that used to be obviously buy is moving to obviously build for any team with the engineering capacity to steer an agent. That’s a market shift one layer up from the one the headlines are tracking. And it’s the same shape as the previous two rounds: tools got better, cost-per-feature fell, the floor of what could be built in-house dropped, and the unit of work got redefined upwards.
“We can do more with fewer people” is admitting something
The framing I find weakest in the current wave is the corporate line that goes “we can do more with fewer people, so we’re letting people go.” It’s a strange thing to say out loud. If you genuinely had 10x leverage on every employee through AI tools, the obvious move would be to ship 10x more value to customers, build 10x more internal software, accumulate 10x more institutional knowledge in code. The teams choosing to cash the gain as a smaller headcount are telling you their cap on the upside, which is that they didn’t have a 10x-output vision in the first place. Saving the cost line is what you choose when you don’t know what else to do with the productivity.
In fairness to the bear case, the productivity gains aren’t uniform. METR ran an experiment in mid-2025 on experienced open-source developers using AI tools on real codebases, and the headline finding was that developers expected a 24% speedup, were actually slower with the tools on, and still believed afterwards they had been sped up by 20%. That’s worth taking seriously. It’s also a specific shape of work: experienced maintainers, mature codebases, careful merge cultures, slow review cycles. From where I sit, building Podtastic and running multiple agents in parallel daily, the speed gains on greenfield product work are not in question. They’re sometimes uncomfortable to hold. Different domain, different mechanics, different conclusions.
In practice, AI helps owner-builders shipping new products faster than it helps maintainers polishing established ones. Some of the productivity-paradox concern is real. Most of the corporate “we can do more with fewer” line is balance-sheet management dressed as vision.
Three turns, same shape
Each round I’ve watched, someone in the industry has said “this means less work for paid developers.” WordPress was going to put freelancers out of business. Laravel was going to commoditise app builds. Now AI agents are going to end the profession entirely. Three turns. Same prediction. Three times the next three years have made the prediction wrong, because the total work kept going up and the unit of work kept getting redefined.
What would change my mind on round three is straightforward. If the SaaS-eating pattern stops at year three of AI tools rather than accelerating, and small teams keep paying for the same modular SaaS stack they were paying for in 2024, then the analogy breaks and AI was something different. From where I’m building, that doesn’t look like the direction. But I write that knowing each prior round had its own quiet critics who turned out to be wrong, and assuming the current critics are wrong without checking is exactly how you become the next round’s wrong person.