{
  "slug": "dumai-zadachamy-ne-profesiyamy",
  "url": "https://neurodrift.org/en/blog/dumai-zadachamy-ne-profesiyamy/",
  "title": "AI Replaces Tasks, Not Professions: Decomposing Your Job into Tasks Shows Who Gets Automated First",
  "description": "AI doesn't replace professions — it takes bites out of tasks. Survival belongs not to whoever has a \"safe\" job, but to whoever decomposes their work into tasks faster than their manager does.",
  "author": "Дністер",
  "language": "en-US",
  "published": "2026-08-12T03:01:21.000Z",
  "updated": null,
  "tags": [
    "ai",
    "automation",
    "future of work",
    "task decomposition"
  ],
  "translationOf": "https://neurodrift.org/blog/dumai-zadachamy-ne-profesiyamy/",
  "sourceUrl": null,
  "body": "Imagine that one day your name in the work chat stops appearing without a slash. That is exactly what happened to Olivia Lipkin — a 25-year-old copywriter from San Francisco. At first, tasks in Slack were addressed simply to \"Olivia.\" Then to \"Olivia/ChatGPT.\" Soon the slash disappeared: just \"ChatGPT\" remained. In April 2023 she was let go without a single explanation. She found out the reason only after the fact, scrolling through the chat history: the managers had been quietly calculating that paying the bot was cheaper than paying her ([Futurism](https://futurism.com/the-byte/copywriter-fired-chatgpt)).\r\n\r\nLipkin wasn't a bad copywriter. Her mistake was subtler: she thought she was selling a **profession**. What people were actually buying from her was a set of **tasks**. And the moment all those tasks fit into someone else's prompt, her name became that prompt.\r\n\r\nKeep a second number in mind, from a completely different era. In 2013, two Oxford researchers, Carl Frey and Michael Osborne, calculated that 47% of U.S. jobs were at risk of computerisation within two decades ([Frey & Osborne, 2013](https://oms-www.files.svdcdn.com/production/downloads/academic/The_Future_of_Employment.pdf)). The figure went viral, landed in hundreds of \"end of work\" presentations, and still lives on in keynote speeches. Ten years later the forecast aged not just badly — it aged exactly backwards. Insurance underwriters, whom Frey and Osborne placed among the highest automation risks, grew employment by 16.4% through 2021; recreational therapists, logged as the safest, lost 8.9% ([ITIF reassessment, 2022](https://itif.org/publications/2022/09/30/oops-the-predicted-47-percent-of-job-loss-from-ai-didnt-happen/)). What was supposed to die first grew; what was supposed to outlast everything shrank. A compass that points due south when you need to go north isn't broken — it just measures the wrong thing. And what Frey and Osborne measured were entire **professions**, nouns. The profession level is precisely the level at which nothing can be predicted, because professions don't get automated; their individual tasks do, and whatever remains gets reshuffled into something new.\r\n\r\nThis scene isn't about a fired copywriter, nor about \"AI arrived and took the bread.\" It is about a mechanism we deliberately conceal behind a convenient word. We say \"AI replaces **professions**\" because a profession is a position, a headcount, a line in the org chart that someone else is responsible for. In reality, automation bites off **tasks** — small, concrete, one at a time. And whoever confuses the two systematically looks in the wrong direction: fears the wrong risks, prepares the wrong skills, defends the wrong positions. I'll dissect this through three lenses — economist, manager, and freelancer — and make one uncomfortable argument: **survival belongs not to whoever has a \"safe\" profession, but to whoever decomposes their work into tasks faster than their manager does.**\r\n\r\n![A 1970s lecturer breaks profession nouns into task fragments on a blackboard; half the audience panics, half leans forward with notebooks](./images/inline-1-task-displacement.png)\r\n\r\n## Let's call it by its name: task-based displacement\r\n\r\nAutomation researchers have a precise apparatus for this. Daron Acemoglu and Pascual Restrepo described technology, even before the wave of generative AI, not as a \"human substitute\" but as a force operating at the task level: capital automates a continuous segment of tasks in production, **displacing** humans where they are no longer needed (displacement effect) and **reinstating** them where new tasks arise with a human advantage (reinstatement effect) ([Acemoglu & Restrepo, 2019](https://shapingwork.mit.edu/wp-content/uploads/2023/10/acemoglu-restrepo-2019-automation-and-new-tasks-how-technology-displaces-and-reinstates-labor.pdf)). Let's call this framework by its name — **task-based displacement**. A profession doesn't sink or float as a whole; its contents sink and float.\r\n\r\nThis is what explains the forecast gap worth memorising forever. When the OECD recalculated the same risk not by professions but by tasks inside professions, instead of Oxford's 47% it got **9%** on average across countries ([OECD, 2016](https://www.oecd.org/content/dam/oecd/en/publications/reports/2016/05/the-risk-of-automation-for-jobs-in-oecd-countries_g17a27d8/5jlz9h56dvq7-en.pdf)). The reason is simple and sobering: professions written off as \"doomed\" almost always contain a substantial share of tasks that are hard to automate. An accountant isn't just reconciling numbers; they're also sitting across from a terrified business owner who doesn't want to hear the truth about their balance sheet. The algorithm takes the first and leaves the second alone.\r\n\r\nThe variation within countries is equally telling. On the same task-based OECD count, the share of vulnerable jobs ranged from **6% in Korea to 12% in Austria** ([OECD, 2016](https://www.oecd.org/content/dam/oecd/en/publications/reports/2016/05/the-risk-of-automation-for-jobs-in-oecd-countries_g17a27d8/5jlz9h56dvq7-en.pdf)) — a factor of two, and not because Koreans and Austrians work in different professions, but because the same professions are sliced into tasks differently there. The internal task structure of a position matters more than the position's name. This deserves a second reading, because it's the entire point: your vulnerability is determined not by **what** you are, but by **what tasks** make up your day.\r\n\r\nThis doesn't mean \"relax, nothing will happen.\" A later OECD estimate raised the share of jobs at high risk to **27%** ([OECD, 2023, via CESI](https://www.cesi.org/posts/oecd-27-of-jobs-at-high-risk-from-ai)). Anthropic, looking not at theory but at real conversations with Claude, puts it more sharply: roughly **49% of professions already have at least a quarter of their tasks that people occasionally offload to the model** ([Anthropic Economic Index](https://www.anthropic.com/research/anthropic-economic-index-january-2026-report)). Note the phrasing — \"a quarter of tasks,\" not \"half of professions.\" That is task-based displacement in action: not a door shutting in front of a profession, but a thin line inside it that shifts a little further every month.\r\n\r\nThe task breakdown itself is not a publicist's metaphor — it is infrastructure. The government database O*NET, on which nearly all these forecasts are built, describes thousands of professions not as holistic \"callings\" but as sums of elements: over **19,000 task formulations** aggregated into more than 2,000 detailed work activities ([O*NET / Pew methodology](https://www.pewresearch.org/social-trends/2023/07/26/2023-ai-and-jobs-methodology-for-onet-analysis/)). Which means that even when governments want to talk seriously about labour, they stopped thinking in \"professions\" long ago — they think in tasks. Strangely, the last people to internalise this remain the workers themselves.\r\n\r\n![Two adjacent desks: the junior celebrates a stack of completed task cards, the senior holds a hand-drawn decision map — two different kinds of power](./images/inline-2-junior-senior.png)\r\n\r\n## Why juniors win and seniors don't (and why that terrifies the wrong people)\r\n\r\nHere is where things get most interesting — and most counterintuitive. Common sense suggests a smart machine should above all amplify the smartest people. The data say the opposite.\r\n\r\nThe largest and cleanest study on this point tracked 5,179 customer-support agents who were progressively given an AI assistant. Average productivity gain: **+14%** calls per hour. But the average lies here. Novices and low-skilled agents added **+34%**, while experienced and high-skilled agents gained almost nothing ([Brynjolfsson, Li & Raymond, QJE 2025](https://academic.oup.com/qje/article/140/2/889/7990658)). The model, in effect, extracted the best practices of the strong operators and distributed them to the weak. The same pattern holds in software development: in a field experiment across three companies (including Microsoft and Accenture), juniors raised output by **27–39%**, seniors by a modest **8–13%** ([MIT Sloan](https://mitsloan.mit.edu/ideas-made-to-matter/how-generative-ai-affects-highly-skilled-workers)).\r\n\r\nAt first glance: wonderful news — AI levels the playing field, lifts the weak. But read that same statistic through the task-based lens and it stops being reassuring.\r\n\r\nWhy does AI give a senior almost nothing? Because a senior already **knows how** to do the tasks the model does well. Their edge didn't lie in the tasks — it lay in the experience of **decomposing** them: looking at a chaotic client request and dissecting it into ten clean steps. And that operation — decomposition — is exactly what generative AI does **not** perform on its own. It brilliantly closes a step that has already been formulated for it, and is helpless until the step is spelled out.\r\n\r\nNow for the paradox that chills you. If AI lifts a junior nearly to a senior's level on the **tasks** themselves, then the only thing that now distinguishes a senior is the ability to set tasks. Not to execute them — to set them. That \"post-tutorial paradox,\" where after a short onboarding a newcomer suddenly nearly catches a veteran, doesn't mean the veteran is unnecessary. It means the veteran is now valued **solely** for what remains after the tasks are subtracted: the ability to see work as a set of tasks and to distribute them — to people and machines — faster than anyone else.\r\n\r\n![A lone manager at night in a boardroom, ringed by glowing sticky-note tasks; behind the glass wall sixty ghostly former team members stand in rows](./images/inline-3-management-audit.png)\r\n\r\n## Weak management and its favourite alibi\r\n\r\nNow for the second, most uncomfortable thesis — and for the antagonist of this piece. He doesn't look like a villain. He looks like a tired department head, call him Andriy, who just proudly reported to the board of directors that he \"optimised\" the content team.\r\n\r\nThe loudest case of this kind: a tech company reduced a team of more than 60 writers and editors to **one** person managing ChatGPT. Then fired them too ([Vulcan Post / BBC](https://vulcanpost.com/862972/bbc-tech-company-cuts-60-people-leaving-one-managing-chatgpt-then-fires-him/)). It sounds like a triumph of automation. In reality, it is a confession about management. Because when 60 people collapse to a single prompt operator — and then even they become redundant — the question isn't \"how powerful AI is,\" but \"what exactly were those 60 people doing that a single prompt replicated?\"\r\n\r\nThe answer is usually buried. If a team is entirely replaceable by a handful of prompts, it doesn't mean the team was weak. Most often it means the opposite: no one in management **ever managed to decompose the team's work into tasks**. As long as the work remained a vague \"they write articles over there,\" it couldn't be measured, optimised, or automated — and it held together by inertia. The moment a tool appeared that forced explicit task formulation, it turned out that half the tasks were ritual, and the other half were genuinely valuable — but nobody had ever thought about them separately, because they lived inside people's heads.\r\n\r\nHere is the uncomfortable mirror. Imagine you are that Andriy. The owner comes to you and asks exactly one question: \"Describe what your team did this week as a list of concrete tasks.\" If you can do this, you probably won't be firing anyone, because you'll see which tasks are alive and which are ritual. If you **cannot** do this and reach for the word \"optimisation\" instead — that is the moment poor management disguises itself as technological progress. If your team is \"entirely\" replaced by a few prompts, that is a diagnosis not of the team but of whoever spent years unable to say what exactly it was doing. AI here is not the killer. It is finally an honest auditor of poor task definition. It doesn't destroy value — it makes the **absence** of value visible.\r\n\r\nAnd this cuts both ways. The tracks of hasty replacements are already showing in the data. A peer-reviewed study of a freelance marketplace (Hui, Reshef, Zhou) recorded that after ChatGPT's launch, order volumes and earnings for freelance writers dropped — on average minus 2% of contracts and minus 5.2% of monthly earnings in text-heavy professions ([WashU Olin / *Organization Science*, 2024](https://olin.washu.edu/about/news-and-media/news/2023/08/study-ai-tools-cause-a-decline-in-freelance-work-and-incomeat-least-in-the-short-run.php)). But the average here is the text's biggest trap. Because the blow landed not evenly but **inversely**: the hardest hit were not the weakest but the best. In the direct words of Xiang Hui, \"for every 1% increase in past earnings, a freelancer experiences an additional 0.5% drop in opportunities and a 1.7% decrease in income\"; his formulation worth pinning above the desk — \"top freelancers suffer the greatest losses\" ([phys.org / INFORMS](https://phys.org/news/2025-03-generative-ai-upending-freelance-safe.html)). This is an inversion of Darwinism: in nature the most adapted survive; in the AI market it's the mediocre, because the machine closed exactly the delta for which professionals used to command a premium, and the skill premium evaporated first. What should concern more than any percentage: the machine wasn't taking \"the weak\" — it was taking entire types of tasks, and it cut hardest those who had spent years building an advantage in precisely those tasks.\r\n\r\nThen the rollback began. The loudest example: Swedish fintech Klarna proudly cut its headcount from ~5,500 to ~3,400 and announced that AI was covering the work of hundreds of support agents. Within a year CSAT and NPS crept downward, complaints multiplied — and by mid-2025 the company quietly returned to a \"human plus model\" hybrid, with the CEO himself acknowledging the cuts \"went too far\" ([CNBC](https://www.cnbc.com/2025/05/14/klarna-ceo-says-ai-helped-company-shrink-workforce-by-40percent.html)). Klarna is not an exception: by industry estimates, **around 29% of companies that cut staff for AI have already rehired for the same positions** ([Washington Times](https://www.washingtontimes.com/news/2026/mar/10/ai-layoff-reversal-companies-rehire-customer-roles-eliminated/)). This is displacement without reinstatement — a culling executed by a manager who confused cost reduction with understanding work.\r\n\r\n\"Weak management hides poor task definition behind the word 'freelancers'\" — this isn't a metaphor. It's the literal mechanism. Four freelancers are easier to write into a \"external contractors\" budget line and one day replace with \"four prompts\" than to admit you never knew which four tasks they were closing and which one was irreplaceable.\r\n\r\n## Pushback: maybe this is just another panic?\r\n\r\nHere I'll stop honestly and strike at my own thesis — because otherwise this is a sermon, not an argument.\r\n\r\nThe strongest objection: \"You're describing exactly what people described with every automation wave. ATMs were supposed to kill tellers — teller headcount grew. Excel was supposed to kill accountants — there are more of them now. Maybe 'think in tasks' is simply fresh alibi for the same layoffs that would have happened anyway, for the quarterly report.\"\r\n\r\nThe objection is strong, and partly fair. The ATM story here is not rhetoric — it is arithmetic: from the 1980s to 2010, the U.S. installed roughly 400,000 ATMs, and over that same period teller headcount didn't fall but rose — from approximately 500,000 to 600,000 ([Bessen, IMF, 2015](https://www.imf.org/external/pubs/ft/fandd/2015/03/bessen.htm)). The machine consumed the dull task of \"counting bills,\" branches got cheaper to run, banks opened more of them — and the teller's task migrated up the stack, from \"count\" to \"sell the loan, retain the customer, be the face.\" The teller didn't die. The teller moved up a floor. Acemoglu and Restrepo themselves underscore: alongside displacement, historically **reinstatement** has always operated — the creation of new tasks where humans have an advantage — and in prior waves it offset displacement substantially, if not fully ([Acemoglu & Restrepo, 2019](https://www.aeaweb.org/articles?id=10.1257%2Fjep.33.2.3)). Moreover, fresh Anthropic data show that **augmentation** — model-assisted human work — 52% of interactions, is again outpacing direct **automation** at 45% ([Anthropic Economic Index, November 2025](https://www.anthropic.com/research/economic-index-primitives)). The dominant mode, in other words, is not \"machine instead of you\" but \"machine alongside you.\" A correlation between AI's arrival and layoffs is not yet proof that AI caused the layoffs rather than Andriy's quarterly plan.\r\n\r\nThere is a second objection, even more uncomfortable, because it strikes at the framework itself. The entire story — \"give the task to the machine and move up a floor\" — silently assumes that the floor above exists. For the teller it did. For a 55-year-old miner from Dnipropetrovsk Oblast or a warehouse packer in Bangladesh, \"the floor above\" is a class of cognitive work whose access requires education, language, network, and capital — none of the four available in the required quantity (OECD data show that after 50, only a small share successfully migrates into a fundamentally new field). The ATM displaced the teller's task — and the teller moved to the adjacent desk; the warehouse robot displaced the packer — and their \"floor above\" turned out to be in another country, another language, and with another diploma. This is not one situation at different scales — it is structurally distinct situations, and the \"think in tasks\" framework is not a recipe of equal opportunity but only a description of the mechanism: it shows exactly where the wall stands. For some people it opens a door; for others it explains why the door is shut and that change must happen at the system level, not the résumé level.\r\n\r\nWhat would **disprove** my thesis? An honest test: if in the coming years productivity gains from AI were distributed evenly across all experience levels; if companies that replaced teams with prompts did not come back for people; if what was valued turned out to be something entirely other than the ability to decompose work — then \"think in tasks\" would be a beautiful but empty formula. So far the data lean the other way: the gain is uneven (juniors > seniors), some replacements are already reversing, and the line between those growing with AI and those stuck runs exactly along the decomposition divide.\r\n\r\n**Why now?** Because until 2018, automation decomposed primarily physical and routine office work — things easy to describe as instructions. With the arrival of generative models (2022–2026), for the first time **cognitive, linguistic, \"creative\"** work fell within the task-based frame: write, summarise, sketch code, draft a letter. Pew finds that the highest AI exposure is not among blue-collar workers but among people with higher education and analytical professions; roughly one in five workers may see AI affect half or more of their tasks ([Pew Research, 2023](https://www.pewresearch.org/social-trends/2023/07/26/which-u-s-workers-are-more-exposed-to-ai-on-their-jobs/)). The caste that spent decades watching automation move top-down felt it themselves for the first time. Hence the temperature of the debate — and hence the temptation to hide behind a soothing \"it's a profession, you can't automate that.\"\r\n\r\n## What this looks like at your desk\r\n\r\nEnough theory. Take one ordinary job title and decompose it the way automation does — silently and mercilessly. Let's say it's \"marketer at a small business.\"\r\n\r\n| Task inside the role | Who covers it better right now | What does NOT fit in a prompt (and stays with the human) |\r\n|---|---|---|\r\n| Generate 20 headline variants | AI (seconds vs. an hour) | Choose the one that hits the client's pain |\r\n| Rewrite copy in the brand voice | AI with a good prompt | Know which voice is right in the first place |\r\n| Compile a weekly metrics report | AI / a script | Notice that one figure is lying |\r\n| Decide what we are advertising at all | human | Formulate the task for all three rows above |\r\n| Reassure the owner that \"everything's under control\" | human | Hold someone's gaze across the table |\r\n\r\nLook at the right-hand column. That is the part of the work that **cannot** be passed to a prompt — because it consists of **setting** the prompts, not executing them. And look at the bottom two rows: they don't split between human and machine at all; they are entirely human, because these are tasks about **decision and trust**, not about product.\r\n\r\nNow, an honest question to yourself. If you run this exercise with your own role — how many rows land in the left column, where \"AI is now better\"? And, more importantly, can you even compose such a table about yourself — or does your work still live in your head as an undifferentiated fog of \"well, I do lots of things\"? Because that fog is exactly what you will be fired for, without explanation, with your name written after a slash.\r\n\r\n<mark>Survival belongs not to whoever's profession \"can't be automated,\" but to whoever decomposes their own work into tasks faster than their manager, competitor, or model does it for them.</mark>\r\n\r\n## What to do tomorrow\r\n\r\nNo coaching mantras. Three simple moves.\r\n\r\nFirst: **make the table about yourself.** The same five-row breakdown above, but for your job. Without an honest decomposition you don't know where you stand — and therefore don't know what to protect and what to hand over.\r\n\r\nSecond: **hunt the right column, not the left.** Instinct says grab the tasks AI does well — because they're visible, easy to show to management. That is suicide in reverse: you're competing with the machine on its home turf. Value has migrated to setting, deciding, and trusting. Invest there.\r\n\r\nThird, for those who manage: **stop buying \"professions\" and \"headcount.\"** Buy closed tasks. Before you fire a team because \"there's AI now,\" try writing out exactly what it was doing. If you can't — the problem was never the team.\r\n\r\nThere's good news here too, easy to miss in the anxiety. Decomposition is a skill, not a talent. That same experiment where novices gained 34% also showed this: the model **distributed** best practices to those who lacked them. Decomposition can be learned — faster than it seems. You just have to do it **yourself**, rather than waiting for someone else to decompose you first.\r\n\r\n---\r\n\r\nOlivia never found out, that day in Slack, at exactly which moment \"Olivia\" became \"Olivia/ChatGPT\" and then simply \"ChatGPT.\" That moment didn't come when the model learned to write better than she did. It came when someone — for the very first time — finally managed to write down exactly which tasks Olivia was doing. The slash in her name was not a verdict from the model. It was the handwriting of a manager who, for the first time in his life, had decomposed someone else's work into tasks — and never thought to do the same with his own."
}