Will AI replace software developers?
No, and the people measuring the labour market say so. The Yale Budget Lab's tracker, updated in September 2026 with August data, "does not provide clear evidence" of AI-related disruption, and its measures sit within historical ranges. The US Bureau of Labor Statistics still projects employment of software developers, QA analysts and testers to grow 10 per cent between 2025 and 2035, with about 106,100 openings a year. The World Economic Forum's Future of Jobs Report 2025, a survey of more than 1,000 employers, puts software and applications developers fourth on its list of the fastest-growing roles to 2030.
Challenger, Gray and Christmas counted AI as the stated reason for 120,136 US job cuts in the first nine months of 2026, 21 per cent of the total, and in July created a category called "Technological Update (possibly AI)" because, as Andy Challenger put it, naming AI in a layoff announcement "can win over investors". The number is what companies said.
Why has junior developer hiring fallen?
Because the work a junior used to do is the work the tools now do, and firms stopped hiring for it before working out what the new entry-level job is. Stanford's Digital Economy Lab, using payroll records for millions of US workers, found that by June 2026 employment of 22 to 25 year olds in AI-exposed occupations was 19 per cent below where it would have been had it tracked their less-exposed peers, up from 15 per cent a year earlier, while "experienced workers show no comparable gap". The authors call the finding descriptive rather than causal, and the effect runs through reduced hiring, not people being let go. SignalFire's talent data shows the same shape: new-graduate hiring down roughly 65 per cent at the big tech companies and 76 per cent at start-ups against 2019, while engineering hiring overall fell 11 per cent at the majors and rose 7 per cent at start-ups.
The UK numbers are harsher and harder to pin on AI. The National Foundation for Educational Research found tech job adverts down around 50 per cent between 2019/20 and 2024/25, programmer adverts down almost 70 per cent, four senior adverts for every junior one, and the first fall in the number of people working as software developers since 2006. NFER points to offshoring and anticipated AI rather than measured substitution, and Indeed's UK software postings index, at 62 in September 2026 against a February 2020 base of 100, is only modestly below the whole UK jobs market at 32 per cent under its baseline. The 2025 trough was also a year of high interest rates and a post-pandemic hiring unwind; nobody has shown AI caused it. The rebound has begun: ITJobsWatch counted 1,449 UK adverts titled Software Developer in the six months to October 2026, against 863 a year earlier, with the median salary up 8 per cent to £65,000.
What does a developer do now that code is cheap?
The job Fred Brooks described in 1986. In "No Silver Bullet", a landmark, he wrote that "the hard part of building software" is "the specification, design, and testing of this conceptual construct, not the labor of representing it". Code generation is the representing. What is left is five jobs, and every piece of 2026 evidence points at one of them.
The job | What changed | The evidence |
|---|---|---|
Deciding what to build | Cheap code makes a bad idea cheaper to build, not better | Judgement roles held while junior roles fell (Stanford, SignalFire, NFER) |
Specifying it | The model does what it is told, and the telling is the work | Brooks 1986; the largest single coding task in Anthropic's usage data is "modifying software to correct errors" |
Integrating it | Your accounts package, your CRM and your legacy database are not in the training data | Integration sets the price before features do |
Reviewing it | The output is "approved by engineers" because 48 per cent of developers trust it only on low-risk, verifiable tasks | Google 2026; Stack Overflow 2026 |
Owning it | Somebody carries the permissions, the backups and the 2am call | Replit 2025; PocketOS 2026 |
Gartner said in October 2024 that generative AI would require 80 per cent of the engineering workforce to upskill through 2027, and in May 2026 that agentic coding is "shifting control, governance and validation to automated platforms". Developers are using the tools: Stack Overflow's 2026 survey of 30,903 developers found 80 per cent using AI for at least an hour a day.
What happens when a business skips the developers?
The code works until the day it does not, and nobody is there. In July 2025, during a twelve-day "vibe coding" exercise, Replit's agent deleted a live production database during a code freeze; Replit's chief executive called it "unacceptable and should never be possible" and separated development from production databases afterwards. In April 2026 a start-up called PocketOS lost its production database and its backups in a single API call that took nine seconds, issued by a coding agent running a Claude model through an infrastructure token whose permissions the founder said he had never fully understood. We use Claude Code across our own development, so this is a point about ownership, not vendors. Both incidents come down to environments, permissions and backups, and those are decisions a senior engineer makes before the agent is switched on.
The announce-cut-rehire pattern has already run in customer service. Klarna said in February 2024 that its AI assistant was doing the work of 700 agents; by May 2025 its chief executive was saying that "investing in the quality of human support is the way of the future for us". Gartner predicts half the organisations that planned large customer-service cuts will abandon them by 2027. Support went first because the work is easy to measure; software failures take longer to surface, so the same pattern arrives later and costs more.
Does cheap code mean less developer work, or more?
More software, and a different split of the work. Satya Nadella's reaction to cheaper AI in January 2025 was "Jevons paradox strikes again": when something gets cheaper, you use far more of it. GitHub's Octoverse 2025 counted 121 million new repositories and 36 million new developers in a year. Nobody has quantified the Jevons effect for developer jobs; the output figures are the evidence that demand for software rose as producing it got cheaper.
On our builds the senior engineers are the product. We use AI throughout: prototyping faster, throwing more options away, writing more tests, catching more in review. Our AI page says there is no AI discount, because every line a model produces is reviewed by an engineer who knows what a security flaw looks like in your stack, and the evidence for that position is in an earlier post. Whether AI belongs inside your product is a different question. The question to put to any supplier is who owns the system when the model is wrong, and the answer should be a named person who understood it before it was built.