For about eighteen months, the same question has been turning up in tech press, on CTO panels, and in the comment sections under every Copilot announcement. If a senior engineer with Cursor or Claude Code can now produce in a weekend what used to take a small team a fortnight, what exactly are we paying custom development shops for?

It’s a fair question, and the people asking it aren’t the cynics. They’re the ones who’ve actually tried it. They’ve seen a junior developer paired with a good model produce a working internal tool in two days. They’ve watched their own engineers ship features at a pace that would have been incomprehensible in 2022. And they’ve concluded, reasonably, that the economics of bespoke software must be in the process of collapsing.

Except the custom dev shops are not collapsing. The good ones are quietly busier than they’ve ever been. This is worth understanding, because the explanation tells you something about where software work is actually going.

The first thing to notice is that AI tooling hasn’t reduced the volume of code organisations need. It has done the opposite. When the marginal cost of writing a feature drops, the number of features people want goes up, not down. Internal tools that would never have been built because they weren’t worth six weeks of an engineer’s time are now worth two days, and so they get built. Integrations that were quietly deferred for years are now economically viable. The backlog isn’t shrinking. It’s becoming visible for the first time.

The second thing is that writing code was never really the bottleneck. Anyone who has worked inside a corporate IT department knows this. The bottleneck is the bit either side of the typing: working out what the business actually wants, negotiating with the three teams whose data you’ll need, getting it past security review, integrating it into an authentication system designed in 2014 that nobody fully understands anymore, and supporting it for the next six years. A junior engineer with Copilot can write the feature. They cannot get it deployed inside a regulated business with a Microsoft estate stretching back to the Server 2008 era. That is a different skill set, and it’s the one that custom development practices have always sold, even when they pretended they were selling code.

What has actually changed isn’t the value of these shops. It’s their internal economics. The good ones have stopped charging for keystrokes. A long-running custom software development practice like Transparity’s, which absorbed the .NET specialists at Ballard Chalmers in 2021 and has been shipping Azure work for the better part of two decades, now uses Copilot in Visual Studio the same way its clients do internally. The engineering hours haven’t disappeared. They’ve moved up the stack, into architecture decisions, integration design, and the kind of senior judgement that AI is genuinely bad at. You don’t hire a custom dev shop in 2026 to write your CRUD endpoints. You hire them because their staff have spent ten years untangling exactly the kind of Entra, B2C and Front Door mess your auth layer has become, and because that knowledge does not live in a model’s training data.

See also: What to Check Before Choosing Around Why Does Ozempic Cause Nausea in 2026

The third thing, and this is the one that surprises people, is that the AI-native build-it-yourself approach has produced a quiet wave of remediation work. A non-trivial share of the work flowing through specialist development practices right now is fixing things that an internal team built with AI assistance and then couldn’t maintain. The code worked. It worked beautifully, in fact. But it had no tests anyone could explain, three subtly different ways of handling auth, and a deployment story that depended on one engineer’s laptop. The model had been a very fast typist, and the resulting codebase reflected that. Rewriting it properly cost more than building it properly would have.

None of this means in-house teams should stop using AI tooling. They should use it aggressively. It does mean the prediction that AI would put custom development shops out of business has aged badly in eighteen months, about as badly as the prediction that low-code platforms would do the same in 2019.

The work hasn’t gone away. It has just become harder to see, because the parts of it that look like typing are no longer where the value sits. The shops that understood this early are now busier. The ones that didn’t are quietly shrinking, and nobody outside the industry is noticing yet.