Will AI replace developers? What the research says about speed, quality and jobs.

The headlines swing between two stories. In one, AI writes the code and developers become optional. In the other, AI coding tools are an overhyped toy that slows serious engineers down. Both stories can point to a study.

So we went through the controlled trials, the big industry surveys and the official job data published up to October 2026. Every figure was checked against the original source. This is what they say, where they disagree, and what I think it means for anyone building a software team. The sources are listed at the end.

Speed: what the controlled trials found

The cleanest way to measure AI’s effect is a randomised trial. Give some developers the tool, keep it from others, and time the same work. Four trials stand out.

GitHub and Microsoft ran the first well-known one in 2022. Freelance programmers built an HTTP server in JavaScript, and those with Copilot finished in 71 minutes on average against 161 minutes for those without. That’s 55.8% less time. The task was small, self-contained and new, the kind of job autocomplete handles best. The authors also point out that they didn’t look at code quality.

Google repeated the idea inside its own walls in 2024, with 96 of its engineers and a realistic enterprise task. The group with AI took 96 minutes against 114. After adjusting for the developers’ backgrounds, the authors put the gain at about 21% and call their own confidence interval large.

The biggest study is a set of three field trials at Microsoft, Accenture and a Fortune 100 company. It covered 4,867 developers over two to eight months of normal work. Developers with Copilot completed 26% more pull requests a week, with the clearest gains among recent hires and junior developers. For long-serving senior developers at Microsoft, the effect wasn’t statistically significant.

Then came METR. In early 2025 the research group gave 16 experienced open-source developers 246 real tasks in large repositories they’d maintained for years. Chance decided which tasks could use AI.

With AI, tasks took 19% longer. Before starting, the developers expected to be 24% faster. Afterwards they still believed they’d been 20% faster.

The METR result gets quoted a lot, usually without the second half of the story. METR ran a larger follow-up between August 2025 and February 2026, with 57 developers and more than 800 tasks. This time the estimate pointed to an 18% saving for the developers who came back. The result wasn’t statistically significant, and METR itself called the data an “unreliable signal”. One of its reasons says a lot about the industry: many developers refused to take part, because they no longer wanted to work without AI.

Why the trials point in different directions

Read side by side, the trials agree more than the headlines suggest. AI helps most on well-defined work, in new or unfamiliar code, and for developers earlier in their careers. It helps least when an expert works in a large codebase they already know by heart. There the hard part is judgement and context, and typing speed was never the bottleneck.

People also misjudge their own speed with AI. METR’s developers felt faster while they were slower. GitHub’s participants underestimated their gain.

In a 2026 METR survey of 349 technical workers, the median answer was a speed-up of three times. METR warns that these answers are “not necessarily grounded in reality”. If you’re planning a team around how fast AI feels, measure it first.

Quality: where the extra speed goes

Writing code faster only helps if the code holds up. Here the evidence is less comfortable.

Google’s DORA programme surveys thousands of software professionals every year about how their teams deliver. In 2024, with nearly 3,000 respondents, it found that more AI adoption went with better documentation, slightly better code quality and faster reviews. For every 25% increase in adoption, it also went with a 1.5% drop in delivery throughput and a 7.2% drop in delivery stability. DORA’s explanation is that AI makes it easy to produce large changes, and large changes are what make releases slow and fragile.

The 2025 report, with nearly 5,000 respondents, shows how fast this moves. AI use is now linked to higher throughput. It’s still linked to more instability.

DORA sums it up in a line I keep coming back to: AI’s primary role in software development “is that of an amplifier”. Strong teams get stronger. Weak habits get worse, faster.

Studies of real codebases tell a similar story. Researchers at Carnegie Mellon compared 806 open-source projects that adopted the Cursor editor with 1,380 similar projects that didn’t. Commits jumped 55% in the first month and 15% in the second, then the effect faded to almost nothing. Over the same period, static analysis warnings rose 30% and code complexity 42%, and that build-up was linked to the later slowdown. The authors call it a “large, but transient increase” in development velocity.

Security is the clearest gap. In 2025 Veracode gave 80 coding tasks to more than 100 language models, without asking for security in the prompt. Of the code samples, 45% failed its security tests.

Its 2026 report puts the average pass rate at 56%. The code almost always compiles now. Whether it’s safe is still close to a coin toss for some kinds of flaw.

Developers know this. In Stack Overflow’s 2025 survey of more than 49,000 developers, 84% use or plan to use AI tools. Yet 46% distrust the accuracy of what those tools produce, and only 33% trust it. Telemetry from Faros AI across more than 10,000 developers shows the knock-on effect. Teams that use AI heavily merged 98% more pull requests, and the time spent reviewing them rose 91%.

Where the human work moves

Put the speed findings and the quality findings together and the shape of the job becomes clear. Typing code is getting cheaper. Deciding what to build is not. Neither is breaking work into changes small enough to review, checking what the machine wrote, or owning what reaches production. Those parts are getting more valuable.

Gartner expects 90% of enterprise software engineers to use AI code assistants by 2028, up from under 14% in early 2024. It describes the developer’s role as moving “from implementation to orchestration”. In an October 2024 forecast it said 80% of the engineering workforce would need to upskill through 2027. Its long-term view is that organisations “will need even more skilled software engineers”.

There’s a warning in the research for anyone tempted to let AI do the learning. Anthropic ran a small trial in January 2026 with 52 junior engineers learning a new Python library. Those who used AI weren’t significantly faster, and they scored 50% on a quiz afterwards against 67% for those who worked by hand. The biggest gap was in debugging. Developers grow by doing the hard parts, and a team that hands all of them to a tool stops growing.

Jobs: what the employment data says

If AI were making developers redundant, the job data would show it by now. It shows something more interesting.

The US Bureau of Labor Statistics projects 10% growth for software developers, QA analysts and testers between 2025 and 2035. The figure for all occupations is 3.5%. That’s about 106,100 openings a year.

One of the reasons the bureau gives is the continued expansion of software for AI. In a 2025 article on its methods, the bureau reasons that AI productivity “may lower prices and increase demand for software products”. When something gets cheaper to make, people want more of it.

The same bureau expects jobs for computer programmers to shrink by 7%. Programming here means the narrower work of turning specifications into code, and the bureau says some of its higher-skilled tasks will shift to software developers. That’s the pattern in one sentence. Routine coding is being automated, while the job of designing and owning software keeps growing.

To be fair to the other side, the bureau has lowered its growth projection for developers three releases in a row, from 17.9% to 15% to 10%. It doesn’t put that down to AI.

Stanford’s Digital Economy Lab has found a real effect on the youngest workers. In its August 2026 update, people aged 22 to 25 in the jobs most exposed to AI are 19% behind less exposed peers. Most of the gap comes from companies hiring fewer of them. Experienced workers show no comparable gap, and where AI complements the work, employment is flat or rising. The researchers’ first finding is that there is “no evidence of widespread, economy-wide job displacement.”

Job postings tell the same story from the hiring side. Indeed’s Hiring Lab reports that US software development postings have risen almost 15% since late February 2025, while postings overall fell 7%. Senior roles accounted for 71% of that increase.

In the first quarter of 2026, senior roles made up 69.3% of software development postings and entry-level roles 4.5%. Postings are still about 27.5% below their pre-pandemic level, so this is a recovery and not a boom. It’s a recovery led by experienced developers.

The global picture points the same way. The World Economic Forum’s Future of Jobs Report 2025, based on more than 1,000 employers, lists software and application developers among the fastest-growing roles to 2030. Developers themselves aren’t panicking either. In Stack Overflow’s 2025 survey, 15% said AI was a threat to their job, and 64% said it wasn’t.

How companies are reshaping their teams

The companies with the most at stake aren’t cutting their engineering ambitions. They’re changing the shape of their teams.

Deloitte surveyed more than 500 business and technology leaders for its 2024 Global Outsourcing Survey. Of their organisations, 78% already run global in-house centres, and 73% of those plan to invest more in them. Access to talent is now the top reason to source abroad. The share of leaders naming cost savings as the main reason fell from 70% to 34% in four years. Everest Group counts more than 6,350 of these centres in offshore and nearshore locations, with new ones opening at about 7% a year.

The teams inside them are getting leaner and more skilled. Reuters reported in May 2026 that new capability centres in India plan 30% to 50% fewer hires than they would have before AI. They’re opening with smaller core teams. Workday’s February 2025 restructuring memo, filed with the SEC, put job cuts, a priority on AI investment and expanding its global footprint in the same document.

None of these sources says AI is the reason companies move work offshore, and I won’t claim it either. What they show is two moves happening side by side. Companies are investing in AI, and they’re building global teams of experienced people to get the most out of it.

What I’d do if I were building a team in 2026

My view, after reading all of this, is simple. AI isn’t taking developers’ jobs. It’s raising the output of good developers, which makes software cheaper to build, which means companies will build more of it. Companies that cut their engineering teams on the strength of a productivity promise will be outpaced. The winners will be the ones that kept their developers and gave them better tools.

If I were hiring now, I’d hire for judgement first. The trials show AI gives the biggest lift on defined tasks, and the quality data shows what happens when nobody experienced is checking the result. I’d want developers who can split work into small changes, read a diff critically and say no to code that only looks right.

I’d give those developers the tools and the time to learn them properly. The DORA research is clear that AI amplifies whatever habits a team already has. So the habits come first: small batches, real code review, automated tests and security checks on every change.

And I’d measure delivery, not activity. More commits and more pull requests mean little if review time doubles and releases get less stable. Count what reaches users and how often it breaks.

It’s also how we build teams at Azendo. The developers we assign start at mid-level, and they’re trained to work with AI through our Talent Success and Academy programme. A service delivery manager tracks delivery with you every week, so faster never comes at the cost of what ships.

Companies that want to hire offshore developers this way can start with a dedicated software development team, managed by us from Chiang Mai and Bangkok. If AI features are on your roadmap too, a dedicated AI engineering team builds them. Our piece on AI-native engineering explains how we test for the skill.

Sources

Every figure in this article was checked against the original publication. Where a study was run or funded by a company that sells AI or developer tools, the text says so.

1. Peng, Kalliamvakou, Cihon and Demirer, The Impact of AI on Developer Productivity: Evidence from GitHub Copilot, February 2023.
2. Paradis et al. (Google), How much does AI impact development speed? An enterprise-based randomized controlled trial, October 2024.
3. Cui, Demirer, Jaffe, Musolff, Peng and Salz, The Effects of Generative AI on High-Skilled Work: Evidence from Three Field Experiments with Software Developers, 2025.
4. METR, Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity, July 2025.
5. METR, We are Changing our Developer Productivity Experiment Design, February 2026.
6. METR, Measuring the Self-Reported Impact of Early-2026 AI on Technical Worker Productivity, May 2026.
7. DORA, 2024 Accelerate State of DevOps Report, October 2024.
8. DORA, 2025 State of AI-assisted Software Development, September 2025.
9. He, Miller, Agarwal, Kästner and Vasilescu (Carnegie Mellon), Speed at the Cost of Quality: How Cursor AI Increases Short-Term Velocity and Long-Term Complexity in Open-Source Projects, MSR 2026.
10. Veracode, GenAI Code Security Report, July 2025, and the 2026 GenAI Code Security Report, July 2026.
11. Stack Overflow, 2025 Developer Survey: AI and Work, July 2025.
12. Faros AI, The AI Productivity Paradox Report 2025, July 2025.
13. Shen and Tamkin (Anthropic), How AI assistance impacts the formation of coding skills, January 2026.
14. Gartner, Top strategic trends in software engineering for 2025 and beyond, July 2025, and Generative AI will require 80% of engineering workforce to upskill through 2027, October 2024.
15. US Bureau of Labor Statistics, Occupational Outlook Handbook: Software Developers, Quality Assurance Analysts, and Testers and Computer Programmers, August 2026.
16. Machovec, Rieley and Rolen (BLS), Incorporating AI impacts in BLS employment projections: occupational case studies, February 2025.
17. Brynjolfsson, Chandar and Chen (Stanford Digital Economy Lab), Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence, revised August 2026.
18. Indeed Hiring Lab, AI and Job Postings: From Destruction to Creation? and The Labor Market Is Tilting Toward Seniority, July 2026.
19. World Economic Forum, The Future of Jobs Report 2025, January 2025.
20. Deloitte, Global Outsourcing Survey 2024.
21. Everest Group, Global Capability Centers Market Outlook for 2026, March 2026.
22. Reuters, Global centres in India slow hiring as AI reshapes work, May 2026.
23. Workday, Form 8-K and memo to employees, February 2025.