DE
← All posts
Blog

AI Does Not Replace "the Occupation"

Um:bruch Editorial Team

Studies rank office work, accounting, call centers, translation, and software near the top. Why LLM exposure is still not a job loss forecast.

This article was created with AI assistance and reviewed by the Um:bruch editorial team.

Which occupations will generative AI replace first? Such rankings work well on social media. Scientifically, they are usually too crude. The more serious studies measure something else: Which tasks within an occupation could large language models technically support or take over? Between this exposure and an actual job cut lie product decisions, costs, error rates, legal frameworks, collective bargaining agreements, work organization – and human trust.

Our new data atlas therefore combines two perspectives: task exposure from the ILO–NASK Index 2025 and the number of employees subject to social security contributions in Germany.

To the interactive dashboard: 43 occupational groups in the LLM exposure check →
Circle area: Number of employees in Germany. Color: Simplified exposure level. Click: Sources, codes, and context.

What the major studies actually measure

The global ILO–NASK Index 2025 decomposes occupations into tasks and evaluates how strongly these tasks could be affected by generative AI. The researchers combined employee surveys, expert reviews, and model-based evaluations. Their global result: About one in four jobs exhibits some level of exposure, but only 3.3 percent are in the highest tier. Office and administrative occupations rank particularly high. At the same time, the ILO emphasizes that transformation is more likely than complete replacement.

The frequently cited 2023 OpenAI study “GPTs are GPTs” arrived at a similarly striking, yet often misunderstood figure for the US: For around 80 percent of workers, LLMs could influence at least one-tenth of their tasks; for about 19 percent, at least half. This, too, is an estimate of technical reach – not a forecast of 19 percent job losses.

A 2025 Microsoft Research preprint analyzed roughly 200,000 anonymized conversations with Bing Copilot. It found high practical applicability primarily in computer and mathematical occupations, office and administration, and sales. Information retrieval, writing, explaining, and advising were strongly represented. What was measured, therefore, was what people are already using AI for and whether the systems appear successful in doing so – not whether jobs were subsequently eliminated.

Which occupations rank high in the indices

In the ILO index, data entry, accounting, financial analysis, general office work, HR administration, and call center sales reach the highest exposure level. Close behind are translation, banking, marketing, journalism, insurance sales, system administration, and software development.

The common pattern is clear: LLMs primarily impact occupations involving a lot of language, structured information, recurring documents, and digital communication. Physically tied work ranks significantly lower. Nursing, childcare, cleaning, kitchen work, construction electrical work, automotive engineering, and masonry also consist of documentable and planable components – but their core takes place on people, locations, machines, and physical materials.

The phrase “most replaceable” is therefore misleading. Even for the highest-rated ILO occupation, data entry, the average exposure value remains below 1. An occupation is not a single prompt.

Germany: Scale changes political relevance

A small high-risk group can be severe for those affected. In terms of labor market policy, however, it makes a difference whether 7,600 employees subject to social security contributions in interpreting and translation are affected, or more than two million office and secretarial staff.

This is precisely why the dashboard displays employee numbers as circle areas. Large groups with strong or very strong exposure include office and secretarial staff, bank clerks, accounting, software development, marketing, and insurance. At the same time, very large groups such as nursing, childcare, retail sales, cleaning, or truck driving are significantly less exposed according to the narrow LLM metric.

Employee numbers originate from the statistics of the Federal Employment Agency, reference date September 30, 2025. Counted are employees subject to social security contributions at their place of work. Self-employed individuals, civil servants, military personnel, and exclusively marginally employed workers are not included. The circles therefore do not represent total employment.

Productivity is not the same as staff reduction

Field studies demonstrate why the direction of impact remains open. In a large customer service experiment involving 5,172 employees, an AI assistant increased productivity by an average of 15 percent. Less experienced workers benefited particularly strongly. The study by Brynjolfsson, Li, and Raymond thus demonstrates a real performance effect – but not an automatic replacement of entire teams.

The labor market track record to date is also more cautious than many headlines suggest. Humlum and Vestergaard found only minor effects on income and working hours in Danish register and survey data, despite rapid chatbot adoption. An Anthropic corporate analysis from March 2026 has so far found no systematic increase in unemployment in highly exposed occupations. However, it notes an initial, still uncertain signal of slower hiring for younger workers.

This is not an all-clear. Companies can invest freed-up time into better quality, shorter working hours, or higher output – or choose not to refill vacancies. The same technical productivity boost can lead to very different outcomes depending on institutional settings.

Three conclusions instead of a replacement list

  1. Analyze tasks, not job titles. The more standardized, language-based, and digital a task is, the more likely an LLM can assist or accelerate it.
  2. Combine exposure with employee numbers. For continuing education and codetermination, large groups with high exposure are more important than spectacular niche cases.
  3. Expose the distributional question. Whether productivity benefits employees, customers, or solely owners is not decided by any model architecture.

The most probable near future is not the unstaffed enterprise. It is an uneven restructuring of activities: some disappear, others become more control-intensive, new ones emerge. The central political question is therefore not just what LLMs can do. It is who decides on their deployment and who receives the gains.

Methodology and Transparency: The five dashboard levels consolidate the ILO categories “Not Exposed”, “Minimal Exposure”, and “Exposed Gradients 1–4”. “Non-automatable” in the dashboard exclusively designates “Not Exposed” in the ILO Index and is not an absolute statement. The mapping of ILO occupations (ISCO-08) to German occupational groups (KldB 2010, 2020 edition) was editorially reviewed but remains an approximation. Text type: EDT (Editorial/Analysis) · Author: Um:bruch Editorial Team · AI Support: GP · As of: July 15, 2026.

Sources

✉️ Write to us 📝 Contact form
ai & powerlabor marketllmautomationeditorial