Syndicated from Paragon AI MAG on Substack: "Which Jobs Will AI Replace First?" — the four jobs AI hits first, the Klarna story, and a 5-minute job audit. Original with video: https://paragonaimag.substack.com/p/which-jobs-will-ai-replace-first

Which Jobs Will AI Replace First?

By Paragon AI MAG — 11 November 2026
Computer programmers have the highest AI exposure of any job: 75% of their tasks already show up in real AI use at work. That is the finding of the Anthropic Economic Index, published in March 2026, which measured 756 occupations against millions of real Claude conversations. The people who build artificial intelligence are the ones most exposed to it.
It is a strange place to start, but it is the honest one. The question "which jobs will AI replace first?" has a data-driven answer now, and it is not the one the headlines usually give.

The warning

In May 2025, Dario Amodei, the CEO of Anthropic, told Axios that AI could wipe out half of all entry-level white-collar jobs within one to five years, and push unemployment to 10–20%. He named technology, finance, law and consulting, and said AI companies have "a duty and an obligation to be honest about what is coming."
Note the word "could". It is a warning about a possibility, not a measurement. The measurement, from his own company's researchers, tells a more careful story.

The four jobs first

Three separate data sources point at the same four groups. The World Economic Forum's Future of Jobs Report 2025, based on a survey of more than 1,000 employers representing 14 million workers, lists the fastest-declining roles to 2030: postal service clerks down 34%, bank tellers down 31%, data entry clerks down 26%, cashiers and ticket clerks down 24%, administrative assistants down 19%, and accounting, bookkeeping and payroll clerks down 18%. Anthropic's index adds customer service representatives at 70% task exposure. Graphic designers appear on the WEF declining list for the first time, which the report attributes directly to generative AI.
What these jobs share is simple: their output is mostly routine text or numbers on a screen.
That is the rule that explains the whole labour market right now. AI replaces tasks, not whole jobs. A job is a bundle of tasks, and AI eats the routine, text-on-a-screen tasks first. Microsoft Research, which mapped Bing Copilot conversations to the US government's occupational database, puts it plainly: "A job is far more than the collection of tasks that make it up." Microsoft's study also carries an explicit warning: its data do not show jobs being eliminated, and should not be used to conclude that. Applicability is not replacement.

The twist: what AI can do vs what it does

Here is the number the headlines skip. Across all jobs, AI could theoretically handle 38.9% of tasks on average. Real observed use is only 9.3%. For every four tasks AI could do, only one is actually being done today. Out of 18,000 tasks in the occupational database, 92.5% show zero measurable AI usage at all.
The gap is not technical anymore. It is integration cost, corporate compliance, and above all a lack of human trust. Legal and healthcare tasks still need a human sign-off, however capable the model is. Workers and managers demand consistent, auditable outputs before they hand over consequential decisions.
The big forecasts disagree with each other for exactly this reason — they measure different things. Goldman Sachs (March 2023) said generative AI could "expose" the equivalent of 300 million full-time jobs worldwide, but its own split for the US was 7% of employment substituted, 63% complemented, 30% unaffected. By August 2025 Goldman had updated its view, writing that it "remains skeptical that AI will lead to large employment reductions over the next decade". The IMF puts global exposure at about 40%, rising to about 60% in advanced economies. The International Labour Organization estimates about one in four jobs worldwide has some exposure — and concludes transformation is more likely than outright replacement.

Klarna: the case that proves both sides

No company illustrates the gap better than Klarna. In February 2024 the Swedish payments company announced its AI assistant had handled 2.3 million customer-service chats in its first month — two-thirds of all conversations — doing the work of 700 full-time agents, with resolution times down from 11 minutes to under 2, and a projected $40 million profit improvement.
Fifteen months later, in May 2025, CEO Sebastian Siemiatkowski told Bloomberg the company had focused "too much on efficiency and cost" and quality had suffered. Klarna started hiring human agents again. By late 2025, on its first earnings call as a public company, Klarna said the assistant now did the work of 853 agents and saved about $60 million a year — while humans were back for the hard cases: disputes, fraud, hardship, nuanced judgment.
Two corrections matter. First, the "700 agents" was work Klarna would have needed to hire for, not 700 people fired. Second, the AI kept working throughout. Klarna's mistake was not using AI. It was trying to automate complex human judgment all at once.

The missing rung

The sharpest signal in the data is not about unemployment at all. US unemployment was 4.1% in August 2026, and Anthropic's own researchers found no systematic rise in unemployment among the most AI-exposed workers. But for workers aged 22 to 25 entering highly exposed roles, the job-finding rate has dropped 14% — a finding the paper calls "just barely statistically significant", and one that does not appear for workers over 25. Payroll-data research found a 16% relative employment decline since late 2022 for the same age group in roles like software development.
The pain lands on the young first, because the career ladder is losing its bottom rung. Junior programmers, junior analysts and junior lawyers used to learn by doing the routine work AI now does first. If the first rung disappears, where does the next generation start? No company, no government and no study has a proven answer yet.

Who is safe

Physical work in unpredictable places, work where a mistake is costly and a human must be accountable, and work built on trust between people — these remain the safest ground. Farmworkers are the single largest growing role in the WEF report, projected to add 34.9 million jobs by 2030. The fastest-growing roles overall are big data specialists (up 113%), AI and machine learning specialists, and fintech engineers. The new jobs are real, but they are different jobs, in different places, needing different skills than the ones being displaced. Even inside design, the split is visible: graphic designers are declining for the first time, while UI and UX designers are among the fastest-growing roles.
History says this pattern is normal, if not painless. Sixty percent of today's workers hold occupations that did not exist in 1940. When ATMs arrived, everyone said bank tellers were doomed; instead banks got cheaper to run, opened more branches, and tellers became salespeople. The job changed; the title stayed.

The 5-minute job audit

Run this on your own job. Three questions:

1. Is most of your output text or numbers on a screen?
2. Is the work highly routine — the same task, repeated?
3. If the AI makes a mistake doing it, is it cheap and fast to fix? Three yeses: your tasks are first in line for automation. Three nos: you have time. And three moves, starting this week. First, learn to direct the AI — become the reviewer and director of AI output, not the typist. Second, own the judgment part of your job: the decisions, the relationships, the accountability. Third, build one skill AI cannot copy from a screen: fixing things with your hands, reading a room, earning trust. One concrete first step, tonight: take the single most repetitive task in your week and make the AI draft it while you check it. ## Our verdict: Yellow / Watching The direction is proven: routine text work is being automated right now. Mass job loss is not proven: real AI use still covers only a small share of most jobs. Jobs change first; whole jobs vanish later and slower — except where a company bets everything on automation at once. The friction of integration, compliance and human trust is the moat protecting your career today. The new mandate is not competing with the machine. It is overseeing it. Watch the video version on the Paragon AI MAG YouTube channel, and hear the full deep dive — including the four forecasts that disagree, and the audit worked through for three different jobs — on the Paragon AI MAG podcast. Sources: World Economic Forum, Future of Jobs Report 2025; Anthropic, "Labor market impacts of AI" (Massenkoff & McCrory, 5 Mar 2026); Microsoft Research, "Working with AI" (arXiv:2507.07935); Goldman Sachs Global Investment Research (Mar 2023; update Aug 2025); IMF Staff Discussion Note SDN/2024/001; ILO Working Paper 140 (ILO & NASK, 2025); Klarna public disclosures (Feb 2024, May 2025, Q3 2025).