Do we still need developers if AI writes the code?

Fewer jobs, no oversupply

According to figures from the digital association Bitkom published in September 2026, Germany is short of around 79,000 IT specialists. Three years ago the figure was 149,000. It has halved, precisely in the years in which AI entered software development. The Federal Employment Agency reports the same picture. In 2025 it had an average of around 13,000 vacancies for IT occupations on its books, 22 per cent fewer than the year before. The unemployment rate in these occupations rose from 3.7 to 4.5 per cent. In its shortage analysis, the agency no longer lists software development as a shortage occupation.

Anyone who concludes from this that developers are becoming superfluous has read only half the figures. Just 5 per cent of the companies surveyed by Bitkom see an oversupply, while 78 per cent still see a shortage. Of the companies with open IT positions, 97 per cent have difficulty filling them. Employment in IT occupations nevertheless rose by 2 per cent in 2025 to 1.15 million. The Federal Employment Agency gives two reasons for the drop in vacancies: the weak economy and structural change driven by AI and automation.

What AI takes off a software developer's hands

In 2025 Stack Overflow surveyed around 49,000 developers from 177 countries. 84 per cent use AI tools or plan to, up from 76 per cent the year before. Of 36 million new GitHub users, almost 80 per cent reached for the AI assistant in their first week, according to the platform.

The Stack Overflow survey also shows the flip side. 46 per cent of developers do not trust the accuracy of AI output, compared with 31 per cent in 2024. Two thirds are most annoyed by solutions that are almost right. Google's DORA report, based on just under 5,000 respondents, shows more throughput but less stability. In 2025 the research institute METR had sixteen experienced open-source developers work with and without AI. With AI they took 19 per cent longer, yet believed afterwards that they had been 20 per cent faster. METR warns against generalising from this. AI delivers code in seconds; it does not deliver responsibility for it.

Newcomers are hit first

Where demand is falling is shown by an analysis from Stanford University, based on wage data covering 4.6 million American employees. Among software developers aged 22 to 25, employment has fallen by 16 per cent since the end of 2022. Among those over 30 in occupations heavily exposed to AI, it grew by 6 to 12 per cent. In the August 2026 update, the gap for young employees in AI-related occupations stands at 19 per cent, while for experienced ones there is none. The effect comes from hiring that does not happen, rather than from redundancies. The researchers see no broad displacement by AI.

This fits with what Indeed reports for the United States. Since spring 2025, job adverts for software development there have risen by around 15 per cent, while the overall market fell by 7 per cent. 71 per cent of this growth is accounted for by senior positions, 37 per cent by roles with AI in the title. In Germany, a good one in five job adverts for software development had an AI connection in November 2025.

What companies themselves expect

Companies report little in the way of job cuts. 42 per cent of the companies surveyed by Bitkom use AI in their IT teams. Only 10 per cent have cut the odd position as a result, 85 per cent have not. Looking ahead, 32 per cent expect cuts, 65 per cent do not. 76 per cent expect the skills shortage to get worse.

One reason is obvious: more software is being built. According to Bitkom, the share of German companies using AI has risen from 17 to 41 per cent within a year. A third of them report higher costs than expected. Every one of these systems needs someone to connect it, secure it, run it and repair it when it does not do what it should. The US labour market agency expects employment for software developers to grow by 10 per cent by 2035, against 3 per cent across all occupations. One driver it names is the development of AI applications.

Code is cheap, judgement is not

I have been writing software myself since my school days, today with AI support. That work tells me what is technically realistic and what merely sounds good. And I have seen how digitalisation lands in companies and where it fails. It is rarely the code. Usually nobody has established beforehand who will work with the software, what becomes easier for those people and whether they will enjoy using the result. Good software does not begin with code, it begins with these questions, and a language model answers none of them.

That is exactly where the demand remains. Whether AI code fits the existing systems, reflects how the department actually works and will still be understood by someone in two years' time is judged by a human being. When I moved a solution to an AI-based version that produces mandatory documents automatically, review and approval by people remained a fixed part of the process. For AI agents that take over entire steps of the work, this applies all the more. AI has made code cheap. Judging whether it does the right thing costs as much as it did before.

What this means for companies

For managing directors and HR managers this means: anyone who cuts the development budget because AI writes the code is saving in the wrong place. The work shifts from writing to reviewing, integrating and operating, and that calls for people who know your own system. The DORA report describes AI as an amplifier of what is already there in an organisation: good teams get better, weak teams ship their problems faster.

The second consequence concerns the next generation. Anyone who barely hires newcomers because AI takes over their tasks will, in a few years, lack the experienced people who are in such demand today. Already, applicants in a third of companies lack knowledge of current technologies. The counter-argument that, thanks to AI, everyone will soon be able to program for themselves only goes so far. At Stack Overflow, 77 per cent say they do not use so-called vibe coding, that is programming by calling out instructions to the AI, in their professional work. Demand for developers is shifting towards those who understand requirements, assess results and answer for them.