AI and Jobs: Which Tasks Change First, According to Researchers
Is your job next on AI's list?
AI is changing tasks, not whole jobs — and the first tasks to change are the ones that are digital, text-based, and repeatable.
That is the finding every major study of AI and work now agrees on. Researchers at Anthropic, Microsoft, Stanford, and OpenAI stopped asking which jobs will disappear and started measuring which tasks AI already handles. Their data points to the same pattern: the work that changes first lives inside documents, inboxes, and code editors — not on factory floors or in hospital wards.
This post uses their numbers, not headlines, to answer the question that matters to you: which parts of your job are exposed, and how can you tell?
The shift researchers made: measure tasks, not jobs
Almost no job is a single thing — so researchers now break each job into its tasks.
Anthropic's Economic Index, built from about a million real Claude conversations mapped to the U.S. Labor Department's O*NET task database, found that roughly 36% of occupations already use AI for at least a quarter of their tasks. Only about 4% of occupations use it for three-quarters or more.
OpenAI's 2023 study "GPTs are GPTs" reached a similar shape from the other direction: about 80% of the U.S. workforce could have at least 10% of their tasks affected by language models, while around 19% could see at least half of their tasks affected. "Affected" there means the task could be done at least 50% faster at the same quality — not that the worker is replaced.
The task lens matters because it changes the question. Instead of "will my job exist in five years," the useful question is "which of my tasks is AI already doing?"
The tasks AI is already doing at work
In real usage data, three kinds of tasks dominate: gathering information, writing, and answering questions.
Anthropic's March 2026 study introduced a measure called "observed exposure": not what AI could theoretically do, but what people actually use it to do at work. The most exposed occupations share one trait — their work is digital, language-heavy, and structured.
Microsoft's study of 200,000 real Bing Copilot conversations from 2024 tells the same story from a different angle. The most common thing people asked the AI to do was gather information from multiple sources — 19.8% of all Copilot goals, more than double that activity's 8.6% share in typical U.S. jobs. Writing and explaining came next. The occupations with the highest AI applicability scores were interpreters and translators (0.49), historians (0.48), and passenger attendants (0.47), followed by sales representatives, writers, and customer service staff.
The gap between "can do" and "does do"
AI could theoretically handle 94% of computer and math tasks. In practice, it covers about a third.
That gap is the most overlooked number in the whole debate. Anthropic's data shows the theoretical ceiling is far above real usage in every occupation group. Microsoft's researchers warn explicitly that their applicability score should not be read as a prediction of job loss — and their own data backs the caution: in about 39% of Copilot conversations, the AI's response did not match what the user asked for.
The labor market agrees, so far. Anthropic found no systematic rise in unemployment among workers in highly exposed occupations since late 2022, when generative AI spread widely. Capability is not the same as adoption, and adoption is not the same as displacement. The same pattern shows up with longer work: our earlier look at AI agents in 2026 found agents handle longer tasks every year, yet still complete only about 16% of real freelance projects to professional standard.
The first place the change actually shows up: entry-level hiring
The squeeze is hitting the youngest workers first — and it arrives as fewer hires, not more layoffs.
Stanford researchers Erik Brynjolfsson, Bharat Chandar, and Ruyu Chen analyzed ADP payroll records covering more than 25 million U.S. workers from January 2021 to July 2025. Workers aged 22 to 25 in the most AI-exposed occupations saw a 13% relative decline in employment since late 2022, after controlling for firm-level shocks. Employment of software developers in that age group fell nearly 20% from its late-2022 peak. Older workers in the same occupations kept growing.
Two details matter. First, the decline is concentrated where AI automates work rather than where it augments it — the tasks a junior employee used to do alone are the ones being handed to the model. Second, the adjustment runs through hiring: fewer entry-level openings, not more firings. A 2026 update of the same paper put the gap at 19% by mid-2026 on a simpler comparison. If you are applying now, the practical side of this shift is covered in our guide to AI tools for job seekers: let AI draft and rehearse, but keep the fact-checking for yourself.
How to check your own job in 10 minutes
Write down the five tasks that fill most of your week, then ask three questions about each one.
1. Is it digital and text-based?
If the task lives in a document, an inbox, a spreadsheet, or a code editor, AI can already reach it. If it needs hands, eyes on a physical site, or face-to-face presence, it cannot — yet.
2. Is it structured and repeatable?
Tasks with a clear input, a standard format, and a right answer change first: first drafts, summaries, data entry, routine replies. Tasks that start with an ambiguous problem and end with a judgment call change last.
3. Does it need judgment, accountability, or a relationship?
Someone has to own the decision, sign off on the result, or hold the client's trust. That part of your job is the hardest to automate, because the risk of a wrong answer lands on a person, not a model.
Score each task. The ones that answer yes, yes, no are the first to change — start using AI on those yourself, before someone else does it for you. The ones that answer no to the first two questions are where your time should move. The strongest defense is not avoiding AI; it is owning the judgment part of your work while AI handles the drafting.
What the next five years look like
Employers expect more jobs created than lost — but the new jobs are not the old ones.
The World Economic Forum's Future of Jobs Report 2025, based on a survey of more than 1,000 employers representing 14 million workers, projects 170 million new jobs created and 92 million displaced by 2030 — a net gain of 78 million. About 22% of today's jobs will be structurally transformed, and 39% of workers' core skills will change or become outdated by 2030.
The fastest-declining roles are the ones built on routine, repeatable steps:
- Postal service clerks: the steepest expected decline in the WEF survey.
- Bank tellers and related clerks: routine transactions are increasingly automated.
- Data entry clerks: the task AI already does cheapest.
- Cashiers and ticket clerks: self-service keeps spreading.
- Administrative assistants and executive secretaries: scheduling and document work moves to software.
The new jobs cluster elsewhere: AI specialists and big data analysts top the growth lists, alongside frontline roles such as farmworkers, delivery drivers, and construction workers. The net positive headline hides the real story: the jobs being created need different skills, in different places, from the jobs being displaced.
Key Takeaway
The research is consistent: AI changes tasks, not jobs, and the tasks that change first are digital, text-based, and repeatable. The clearest early signal is not mass unemployment — it is fewer entry-level openings in the most exposed occupations.
Check your own five tasks against the three questions, and move your time toward the judgment only you can own.
Frequently asked questions
Will AI replace my job entirely?
No study finds whole occupations disappearing at scale. Only about 4% of occupations see AI used across three-quarters or more of their tasks, and unemployment in highly exposed occupations has not risen systematically since late 2022. Jobs are evolving task by task, not vanishing overnight.
Which skills are hardest for AI to replace?
The studies point to the same short list: judgment under uncertainty, accountability for decisions, relationships and trust, and physical presence or dexterity. Occupations built on those — from skilled trades to senior professional roles — show the lowest AI usage in every dataset.
I'm early in my career. What should I do?
Aim for work where you own an outcome, not just produce a first draft. Learn to check AI output rather than generate it blindly, and build the judgment part of your role — deciding, verifying, and taking responsibility — because that is the part employers still pay a person to do.
Sources
- Anthropic — "Introducing the Anthropic Economic Index" (February 10, 2025) — anthropic.com
- Anthropic — Massenkoff & McCrory, "Labor market impacts of AI: A new measure and early evidence" (March 5, 2026) — Anthropic Economic Index
- Microsoft Research — Tomlinson et al., "Working with AI: Measuring the Applicability of Generative AI to Occupations" (arXiv:2507.07935) — arXiv
- Eloundou et al. — "GPTs are GPTs: An early look at the labor market impact potential of large language models" (2023) — OpenAI
- Brynjolfsson, Chandar & Chen — "Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence" (Stanford Digital Economy Lab, August 2025) — Stanford Digital Economy Lab
- World Economic Forum — "The Future of Jobs Report 2025" (January 2025)

Comments
Post a Comment