Which Jobs Will AI Replace in Germany – and Which Won't?

I don't think Germany will wake up one morning without accountants, developers or translators. The more uncomfortable possibility is that the professions keep their names. Businesses simply hire fewer people to do them.

A robot taking someone's desk makes an obvious image. A different sentence worries me more: “We don't need that additional position after all.”

AI does not have to master an entire profession for that to happen. It only needs to remove enough work, reliably enough, to change a staffing decision. That is the distinction worth discussing: automating tasks, cutting positions and replacing a profession are three different things.

My assessment for the next few years: standardized screen-based tasks face pressure earlier. Work combining physical skill, personal relationships and decisions in unpredictable environments looks comparatively more resilient. This is not a timetable for the disappearance of individual professions.

First the tasks. Then, possibly, the positions.

A job contains multiple tasks. Processing an invoice, for example, involves more than extracting and transferring its data. There are questions, exceptions and decisions about whether something makes sense in the first place.

The ILO's May 2025 study on generative AI identifies clerical occupations as particularly exposed. Overall, it considers job transformation more likely than complete automation. Exposure means AI can affect tasks. It does not mean a particular worker will lose their job.

For a business, the practical question is whether the whole process becomes better or cheaper with AI, including review, integration and fixing mistakes. A good draft does not answer that by itself.

I would start with three questions. Is the information digital? Does the process repeat? Can the result be checked at a reasonable cost? Where all three apply, a change looks much more plausible to me than it does for an improvised repair on site.

Which jobs in Germany face pressure first?

The table below is my qualitative assessment of automation pressure on typical routine tasks over roughly 2026–2030. It is not an official ranking or a measured probability of unemployment. Two positions within the same profession can involve very different work.

The background is the ILO analysis and Microsoft's “Working with AI” study. The latter examines actual Copilot use and applicability to occupational activities, not layoffs in Germany. Information gathering, writing, computer work and office tasks feature prominently in its findings.

The important word is routine. Translation can also mean literary work or a delicate negotiation. Support can involve a conversation in which the real problem has yet to be identified. Software testing can mean discovering an entirely new category of failure. Replaying an existing sequence does not accomplish that.

“Plausible” also does not mean “ready to run unattended everywhere today.” A convincing answer can be wrong. A booking suggestion is not an executed travel booking. With sensitive data and consequential decisions, I would pay particular attention to permissions, verification and clear handoffs to people.

What German businesses actually expect

The ifo analysis of its May 2025 survey reports these expectations for the following five years. The population here is businesses already using AI or planning to use it.

These are shares of businesses, not shares of jobs disappearing. Expectations, not documented layoffs. The figures come from Table 4 of the report.

Why people starting their careers should pay attention

The Randstad-ifo HR survey published in January 2026 covers Q4 2025: 14% of surveyed businesses already used AI for entry-level tasks; 40% planned to within three years. Meanwhile, 65% expected stable entry-level headcount, 19% a decline and 16% an increase. Those staffing forecasts do not isolate AI's effect.

What concerns me is less the provocative question “Do we need juniors anymore?” and more the route to experience. When a company automates its simpler tasks, it should still be able to explain how new employees will learn the work.

An initial research assignment or a manageable programming task is not just cheap production. It can also be a chance to make mistakes under supervision, ask questions and develop judgment. I think removing that starting point is shortsighted when the same business expects to find experienced people later.

For someone starting out, I would not turn this into advice to abandon an apprenticeship or a degree. Instead: show more than a finished result. Explain your decisions, check your sources and demonstrate that you can identify and fix a mistake. That applies just as much to AI-generated results.

Developers: a working button is not a finished product

As a developer, I find this part particularly interesting. There is a considerable distance between “Create this form” and “Make sure our product works reliably.”

Imagine a generated checkout. The interface looks decent and the request works in a test. Who checks whether a customer can access someone else's orders? What happens when a payment notification arrives twice? How do we roll back a failed update? And are we building the right feature at all?

That is the distinction I would invest in: understanding requirements, examining systems and taking responsibility for an outcome, rather than delivering code as quickly as possible.

“Senior” is not a shield. A title does not make routine work less automatable. Equally, a junior who reasons carefully, questions tests and explains problems clearly is not automatically interchangeable. The actual task mix matters.

I would apply the same mental model elsewhere. In legal and tax work, it separates preparation from case-specific judgment. In medicine, documentation from treatment and conversation. In teaching, producing materials from running a class and building relationships. In sales, preparing contacts from complex negotiations. In design, producing variations from deciding what a brand needs.

My article about AI in motion design looks at that creative distinction: an impressive draft and a result that fulfills a particular brief are not the same thing.

A degree alone does not make a job AI-proof

An IAB study from 2023 already identified especially high relative AI automation potential in degree-level occupations. It uses patent-based indicators and distinguishes AI from other software. It is not a current forecast of chatbot replacement. It does, however, challenge the simple idea that automation only affects lower-skilled work.

In an ifo publication from June 2026, nearly 20% of AI-using businesses considered replacing a degree-qualified worker with a non-degree worker using AI easy or very easy. For replacing experience, the figure was roughly 15%. The counterpoint matters: 55.4% and 62.7%, respectively, considered those substitutions difficult or impossible. These are assessments, not observed personnel changes.

I would not translate that into either “Degrees are pointless” or “Experience can never be replaced.” The more interesting question is whether someone can apply their knowledge to an unfamiliar case and judge the quality of an answer.

Which professions are comparatively more resilient?

Germany's Federal Employment Agency highlights care and social work, skilled trades, teaching and education as relatively difficult to replace completely. My assessment of similar task profiles:

A system can produce repair instructions. That does not fix the broken heating system in a cramped basement. A good lesson plan does not supervise a classroom.

More resilient does not mean unchanging or immune to economic trouble. Documentation and planning can be automated here too. Robotics could take over additional physical tasks. How quickly and economically that works in changing environments is a different question from a language model's capabilities.

That is why I would not base a career choice solely on a green rating in an AI table. Interest, working conditions, pay and personal suitability matter too.

The bigger effect could be a position that never gets advertised

Consider a deliberately invented example. A business currently needs three people for a bounded bookkeeping workflow. After changing the process, one person with software and AI handles the same volume at comparable quality.

This is a thought experiment, not a measured threefold productivity gain or a forecast for 2028. It illustrates the mechanism. The business could replace fewer departing workers. But it could also grow, clear a backlog or spend the time on more difficult cases. Time saved does not automatically become a job cut.

Pay can face pressure too. An ifo analysis published in August 2026, covering more than 3,000 AI-using businesses, found more frequent negative wage expectations for less experienced workers. It uses June 2026 responses about the next five years, not measured wage reductions. Depending on qualifications, between a third and half also expected stable wages.

For me, the number of open entry-level positions is therefore at least as interesting as a layoff headline. The possible change is quieter: fewer new hires, different assignments, different requirements. Attributing it to AI would still require checking other causes rather than turning every staffing reduction into an AI story.

What I would actually do

I would break down an ordinary working week. Which tasks repeat? For which ones is all the information available? Where do you need conversation, context, physical skill or a difficult decision? That gives you a more useful starting point than a job title alone.

Next, I would test a small workflow using approved or fictional data. Measure not just how quickly a draft appears, but how much checking and rework it needs. An hour saved only matters when it does not reappear somewhere else.

And I would deliberately practice what makes good work recognizable: subject knowledge, diagnosing mistakes, communication and understanding the entire process. “I know how to operate this tool” would not be enough of a long-term professional identity for me.

Employers should not put all the responsibility on workers either. My standard would be straightforward: a business automating work also organizes onboarding, verifiable quality standards and real opportunities to learn. Otherwise, short-term efficiency can become a long-term skills problem.

My take

I find this development technically pretty wild. At the same time, celebrating it purely as progress while ignoring career entry and income would be too convenient.

My rule of thumb: The more a position consists of repeatable digital processes, the more closely I would examine the pressure to change it. The more it combines real situations, physical skill and durable relationships, the less convincing I find the idea of quick, complete replacement.

The question is not just “Can AI do my profession?”

The better question is: “Which part of my work will someone still need when a first draft costs almost nothing?”

As of October 6, 2026. The linked studies use different survey periods and methods. Business expectations, technical potential and my own assessments do not form a single forecast of how many jobs Germany will lose.