Which Jobs
Does AI Threaten?
Microsoft Study 2026

What the Microsoft Research Study Actually Says ▲

In July 2025, a team led by Kiran Tomlinson at Microsoft Research published the study "Working with AI: Measuring the Applicability of Generative AI to Occupations" on arXiv. The current version from December 2025 is one of the most thorough investigations into how AI is actually used in day-to-day work.

The methodology: 200,000 anonymized conversations with Microsoft Bing Copilot, evaluated by three criteria — how often AI is used for an occupation, how successfully, and what share of the work activity it covers. The result is a score for 785 occupations, together representing roughly 150 million US workers.

An important clarification: many German-language outlets cite the study as "Microsoft + OpenAI." That's factually wrong. Microsoft Bing Copilot does run on OpenAI models, but OpenAI isn't a co-author of the study. All six authors work at Microsoft Research. If you cite the study, name it correctly: the Microsoft Research study.

The study doesn't say "these occupations will disappear." It says: "AI is especially applicable to these occupations." That's a fundamental difference — and it starts with the next look at the list.

Which 12 Occupations Are on the List — and Why Exactly These? ▲

The Microsoft study itself lists 785 occupations. Fortune published the top 40 in July 2025, and DACH media distilled from that the 12 occupations relevant for Germany, Austria, and Switzerland. Occupations like "bridge tender" or "dredge operator" from the US list drop out because they barely exist in the German labor market.

Here are the 12 DACH-relevant occupations by AI-applicability score (source: marktundmittelstand.de, skill-sprinters.de, the-decoder.de):

1. Interpreters and translators (score 0.49)
2. Historians (score 0.48)
3. Travel agents
4. Stock traders
5. Journalists
6. Statisticians
7. Language teachers
8. Data scientists
9. Financial advisors
10. Database developers
11. Management consultants
12. Market researchers

The pattern is clear: all of these occupations produce digital output — text, data, analysis, translations, recommendations. And all have a high share of language or text processing. Prof. Peter Buxmann of TU Darmstadt puts it succinctly: "Every job whose output is digitizable is at risk."

At the other end of the scale are hands-on, physical occupations. The Microsoft study lists as least AI-applicable: roofers, plumbers, crane operators, foundry workers, loggers. The DACH translation: tradespeople, caregivers, construction workers, firefighters. These are occupations where the output isn't digitizable — a repaired roof, a painted wall, a person cared for.

But careful: that doesn't mean robotics won't move in. More on that later.

Why the Study Doesn't Mean Jobs Are Disappearing ▲

The Microsoft study measures applicability, not replacement. That's the crux most headlines miss.

The International Labour Organization (ILO) published a working paper in 2023 that makes exactly this distinction. In high-income countries (Germany, Austria, and Switzerland among them), 5.5% of employment is potentially exposed to full automation — and 13.4% exposed to augmentation. Augmentation means: AI strengthens people, it doesn't replace them.

Concretely: a bookkeeper who uses AI to categorize receipts doesn't get fired. Instead, she has more client conversations because she saves time on routines. A financial advisor who uses AI to write product proposals doesn't lose his job — he advises more clients because he needs less time for admin.

The Kiel Institute for the World Economy confirmed this empirically in January 2026. In a policy brief using data from Denmark, Portugal, and Sweden spanning 2010 to 2023, the authors write: "Overall progress in AI over the last decade has had little to no impact on overall employment levels."

Overall employment: stable. But: activity profiles are shifting. Routines shrink, demand for higher-skilled workers rises. That's not job loss — that's transformation.

McKinsey Global Institute reached a similar conclusion in 2017 in "Jobs Lost, Jobs Gained": 3 to 14% of global employment will need to switch occupational categories by 2030. Not become unemployed — switch occupational categories. The case worker becomes a customer success manager. The bookkeeper becomes a financial analyst.

That's not reassuring — transformation is strenuous. But it's not an apocalypse either.

What This Really Means for SMB Owners ▲

Let's take three concrete examples from the SMB world:

Bookkeeping

About 80% of a bookkeeper's tasks are digitizable — scanning receipts, categorizing, posting entries, DATEV data entry. AI can take that over. What remains: client conversations, follow-up questions on unclear receipts, advice on tax-optimal structures. These are the tasks that really keep the client loyal.

Financial advisory

Germany currently has 178,791 registered §34d advisors (as of January 2026, DIHK). A recent study by the Arbeitsgemeinschaft für Finanzberatung (AfW) shows: 35% of independent financial advisors already use AI — double the previous year's figure (16%). What are they using it for? Automating product comparisons, writing proposals, preparing follow-up appointments. What stays with the human: the advisory conversation. The empathy, the trust, explaining why this retirement plan fits this life situation.

Administration

Government agencies, industrial administration, insurance claims processing. All activities where "digitizable output" applies. AI can review applications, reconcile data, draft standard letters. What it can't do: decide complex individual cases where human judgment is needed.

The pattern: routines disappear, advisory work stays. But only if you build the architecture right. If you deploy AI to eliminate routines and retrain people to spend more time advising — people become more valuable. If you deploy AI to cut costs and don't retrain people — you have a problem.

We ran the numbers with a trades business. An SMB with five employees that automates emergency calls and appointment requests saves about 15 hours a week. The question is: what do the employees do with those 15 hours? More jobs? More client conversations? More quality? That's what decides whether AI strengthens or weakens the business.

More on this in our AI architecture approach for SMBs.

Financial Advisors on the List — What Does the Complement Path Look Like? ▲

Financial advisors are explicitly on Microsoft's top-12 list. That's no coincidence — a large share of the work is digitizable: product comparisons, calculating risk profiles, writing proposals, sending follow-up emails.

But here's the complement perspective: 178,791 §34d advisors in Germany, an average of 2,300 clients per advisor (BVK structural data, 2025). A Deloitte study shows: 65% of advisors have no structured follow-up system. That means: 65% lose clients because they don't have the routines under control — not because they're bad advisors.

That's exactly where AI comes in. Not as a replacement for the advisor, but as a qualifier. AI takes over the routines — suggesting appointments, reminding clients of contract deadlines, clarifying standard questions upfront. The advisor focuses on what keeps the client: the advisory work.

Internally, we put it this way: "We qualify. You advise."

That's not a marketing slogan. That's the architecture. Our approach for financial advisory businesses looks like this:

1. Qualification AI pre-filters inquiries — which clients have genuine advisory needs, and which just have a standard question?
2. Follow-up automation ensures no client is lost because the advisor forgot to follow up in time.
3. Document automation generates proposals, comparison tables, follow-up appointment summaries.

The result: the advisor holds more qualified conversations because they need less time for routines. Advisory quality goes up, client retention goes up, revenue goes up.

More on this on our Financial Advisors pillar page.

Tradespeople Not on the List — But Where Robotics Is Actually Arriving ▲

Tradespeople, caregivers, roofers — they all sit at the bottom of Microsoft's AI-applicability scale. The reason: their output isn't digitizable. A repaired roof, a painted wall, a person cared for — AI can't deliver that.

But that doesn't mean robotics isn't coming in. Boston Dynamics, Figure AI, and others launched several pilot projects with humanoid robots in industrial manufacturing in 2025 and 2026. In the auto industry, in logistics, in large-scale manufacturing — a lot is happening there right now.

In service trades, it looks different. A tradesperson who drives to an emergency call, diagnoses on-site, picks the right tool, and improvises when something doesn't fit — that's a level of complexity robotics won't cover for the foreseeable future.

Where AI is already applicable in the trades today: emergency-call triage, appointment coordination, material reordering, invoicing. Everything that happens before and after the actual trade work.

We ran the numbers with a trades business (Müller Haus & Service, Calw). Five employees, 30 emergency calls a week, 80% of them standard questions. AI takes over phone triage, clarifies standard questions, books appointments directly. Result: 15 hours a week saved, more jobs, higher customer satisfaction.

The tradesperson isn't replaced. They're relieved. And that makes them more valuable.

More on this on our Trades pillar page.

Those Who Understand AI Fear It Most — and Why That's the Wrong Conclusion ▲

Prof. Peter Buxmann of TU Darmstadt published an AI monitor together with YouGov in 2026. The result is paradoxical: roughly one in two people with strong AI knowledge and digitizable output fears for their professional future. Among people with little AI knowledge, the worry is lower.

Buxmann explains it this way: "Every job whose output is digitizable is at risk." Anyone who understands AI knows what it can do. And anyone who watches every day how well AI writes text, analyzes data, generates code — knows that their own work isn't unbreakable.

But here's the wrong conclusion: fear instead of action.

The right conclusion is: build architecture. Anyone who understands what AI can do shouldn't fear being replaced. They should build the architecture that embeds AI.

The Kiel Institute confirms this empirically: overall employment stable, but pressure to upskill rising. Routine administration shrinks, demand for higher-skilled workers rises. And "higher-skilled" in this context means: whoever understands how to build systems, rather than just execute routines.

A Goethe University Frankfurt study of 4,048 students shows the problem: 89% use AI tools, 45.5% worry about losing critical-thinking skills, 72% fall into "ghost role" usage (AI completes entire sub-tasks). Lead researcher Nora Hoffmann says: "Students still know how to write without AI and feel what excessive delegation costs them — that will change."

Exactly for that reason: anyone who understands AI should become an architect, not an anxiety patient.

Internally, we put it this way: "We build architectures, not models." We don't train our own AI models. We build the systems that embed existing models (Claude, GPT, Gemini) in a way that makes people more valuable.

More on this in our Anthropic Economic Index article — there we show what Claude power users are already automating today.

Our Complement Approach for SMBs — What We Actually Build ▲

When an SMB owner comes to us and says, "I'm afraid my employees will be replaced," we ask back: "Which routines will disappear in your business tomorrow — and what will your employees do with the time they gain?"

That's not a rhetorical question. That's the architecture question.

Our approach: a 5-layer AI architecture for SMBs. No training our own models (that's for Google, OpenAI, Anthropic). Instead:

1. Input layer: where does the request come in? Phone, email, form, WhatsApp?
2. Qualification layer: which requests are routine, which need real expertise?
3. Automation layer: which routines can AI handle immediately?
4. Escalation layer: which cases get passed on to a human — and with what context?
5. Feedback layer: what does the system learn from every request so it gets better next time?

Concretely:

• Financial advisors: 178,791 §34d advisors in Germany. We build systems that clarify standard inquiries, suggest appointments, prepare follow-up meetings. The advisor focuses on the advisory conversation.
• Tradespeople: ~1 million trades businesses in DACH, ~5.6 million employees. We build phone agents that triage emergency calls, answer standard questions, book appointments. The tradesperson focuses on the repair.

The result: people become more valuable, not redundant. Because they hold more qualified conversations, deliver more advisory work, bring more expertise — instead of working through routines.

Internally, we put it this way: "We build architectures, not models."

More on this in our 5-layer article.

FAQ — Which Jobs Are Threatened by AI? ▲

Which occupations are on the Microsoft list?

The top-12 DACH occupations by AI applicability: interpreters, historians, travel agents, stock traders, journalists, statisticians, language teachers, data scientists, financial advisors, database developers, management consultants, market researchers. All have digitizable output and a high share of language or text work.

Does that mean these occupations will disappear?

No. AI applicability doesn't mean job loss. The ILO sees 5.5% of employment at risk of full automation, but 13.4% augmented. The Kiel Institute finds no employment decline in countries comparable to DACH. Routines disappear, advisory work stays.

Are tradespeople safe?

Tradespeople aren't on the Microsoft list because their output isn't digitizable. But robotics is arriving in manufacturing. In service trades (emergency calls, on-site diagnosis), humans stay dominant — AI helps with phone triage, appointment scheduling, invoicing.

What about financial advisors?

Financial advisors are in the top 12. But that's exactly where the opportunity lies: anyone who automates routines (product comparisons, proposals, follow-up emails) gets more time for advisory work. 35% of independent financial advisors in Germany already use AI (AfW study, 2026).

How many jobs are really threatened?

McKinsey says: 3–14% of global employment will need to switch occupational categories by 2030 (not become unemployed). The ILO says: 5.5% at risk of full automation, 13.4% augmented in high-income countries. The Kiel Institute says: overall employment stable, but pressure to upskill is rising.

Why do AI-savvy workers fear the future the most?

TU Darmstadt's AI Monitor 2026: roughly one in two people with strong AI knowledge and digitizable output fears for their professional future. Anyone who understands AI knows what it can do. But the right conclusion isn't fear — it's building architecture. Whoever understands how to build systems will be needed.