Will AI Replace Your Job? The Answer Is About Tasks, Not Jobs

ATMs were supposed to kill bank tellers. The number of tellers went up.

Between 1980 and 2010, the United States went from zero ATMs to around 400,000 of them. Over the same thirty years, the number of bank tellers — the people the machines were supposed to replace — rose. Branches got cheaper to staff, banks opened more of them, and tellers stopped counting cash and started selling mortgages.

That story, popularised by the economist James Bessen, is the cleanest single rebuttal to "AI will replace your job." It is also only half right. The honest version of the answer is more interesting than either the doomers or the cheerleaders will tell you, and it has practical consequences for what you do on Monday morning.

AI does not replace jobs. It replaces tasks.

This single sentence fixes most of the confusion. A job — "lawyer", "graphic designer", "primary-care doctor", "marketing manager" — is a bundle of dozens of distinct tasks. Some of those tasks are routine and predictable and now mechanisable by an LLM — a large language model, the kind of system behind ChatGPT, Claude or Gemini. Most of the others are not.

A 2023 paper from OpenAI and the University of Pennsylvania ("GPTs are GPTs", Eloundou et al.) tried to measure this for the U.S. economy. Their estimate: around 80% of American workers have at least 10% of their tasks exposed to current LLMs, and around 19% have at least half. Read the framing carefully — exposed tasks, not eliminated jobs. That is not a semantic quibble; it is the entire story.

Take radiology, the profession everyone confidently predicted dead in 2016. A radiologist's day contains roughly thirty distinct tasks. AI is genuinely good at three of them: initial scan triage, measurement, and producing a structured first-draft report. It is not good at the other twenty-seven, which include talking to a frightened patient, deciding whether a scan justifies cutting someone open, integrating the image with the lab work, and signing the legal report that goes in the chart. The radiologist of 2030 is not unemployed. She reads three times as many scans, with AI doing the first pass on each one, and probably earns more.

Which tasks AI has already absorbed

Roughly anything that can be described as "produce a competent first draft from a prompt." That category is wider than it sounds.

The clearest cases today: writing routine marketing and SEO copy, summarising long documents, translating between common languages, generating boilerplate code, drafting basic spreadsheets and SQL queries, answering tier-1 customer support tickets, sketching simple graphic designs, transcribing and cleaning up meeting notes, drafting standard contracts. By "absorbed" I mean something specific: a non-expert with a good model can now produce in five minutes what used to take an entry-level professional an hour, at quality acceptable for the use case.

This is where the real, measurable labour-market effect lives today. A 2023 paper by Brynjolfsson, Li and Raymond followed about 5,000 customer-support agents using a generative AI assistant. The average agent resolved 14% more tickets per hour, but the gain was concentrated almost entirely on novice and below-average performers. The bottom of the skill distribution caught up. The top performers gained almost nothing.

This pattern — AI compresses the gap between the median worker and the expert — keeps showing up. It is the most important and least-discussed labour finding of the last three years. It implies the people most at risk in the short run are not the unskilled, but the comfortable middle: the competent-but-not-exceptional knowledge worker whose value was being faster and tidier than the average. The model is now faster and tidier than the average, for free.

Which tasks AI is still bad at, probably for a while

Four broad categories look mostly safe through 2030, in roughly increasing order of safety.

Judgment under genuine uncertainty. Diagnosing the weird patient, pricing the unusual deal, deciding whether to fire the underperforming employee, choosing which of three startup pivots to bet on. LLMs are pattern-completers; they fail when the right answer is not somewhere in the pattern. Anyone whose value is "I am the one who decides when the rule does not apply" is, for now, fine.

Embodied physical work. Plumbing, elderly care, electrical installation, restaurant kitchens, dental hygiene, hairdressing, arborists in harnesses. The robotics gap is real and is widening relative to the software gap. Boston Dynamics' Atlas can do a backflip on YouTube; it cannot install a dishwasher in your kitchen. The cheapest competent pair of hands in your city, for the foreseeable future, is a human pair.

Relational and accountable work. Therapy, primary-school teaching, end-of-life care, complex enterprise sales, leading human teams. Not because an LLM cannot produce empathic-sounding text — it obviously can — but because the social contract requires a human counterparty. People do not want their seven-year-old taught by an algorithm, even if the algorithm scores better on the test. They also do not want a model on the legal hook when something goes wrong, and right now nothing else can be.

Genuinely novel research and creative work at the frontier. Designing the next category-defining product, writing the next great novel, finding the next bit of physics. LLMs interpolate inside their training distribution; they extrapolate badly. This is why the labs that build them still hire human researchers.

What these four categories have in common is exactly what makes them hard: they are the parts of knowledge work that nobody can fully write down in a job description. If you cannot specify the task, you cannot prompt the model.

The honest part of the analogy, and the part that breaks

The ATM story works because the technology automated a narrow task (cash handling) and complemented the rest (advising, selling). Excel did the same to accountants: VisiCalc shipped in 1979, the U.S. had roughly a million accountants then and has about 1.5 million now. Photoshop did not destroy graphic designers; it expanded the market for design by an order of magnitude. The pattern is consistent. Boring part to the machine, human moves up the value chain, total amount of the activity grows because it is now cheap.

What is genuinely different this time is the breadth. Previous waves of automation hit one or two tasks per profession. The Eloundou paper finds LLMs touch a long list of tasks across almost every white-collar occupation simultaneously. ATMs hit cash dispensing. Excel hit ledgering. LLMs hit drafting, summarising, coding, designing, translating, scheduling, analysing and answering — at the same time, in the same quarter, across the same workforce.

The likely consequence is not mass unemployment. U.S. unemployment in early 2026 is still hovering around 4%, three years into the LLM rollout, and Goldman Sachs' famous 2023 estimate of "300 million jobs at risk" has aged in ways its authors might prefer to forget. The realistic consequence is the compression of middle-skill knowledge work. The junior lawyer who used to spend two years redlining contracts now does it in two months and either gets promoted faster or never gets hired in the first place. The middle of the org chart gets thinner. That is the part nobody at the World Economic Forum wants to put on a slide.

The interesting question, then, is not "will my job exist in 2035?" — it almost certainly will. The interesting question is: in the version of my job that exists in 2035, what am I doing that the model is not?

What this means for you

If you do knowledge work, the right move is not to learn to code or to "pivot into AI." It is to become unusually good at the parts of your own job that involve judgment, taste, relationships and accountability — and to use AI ruthlessly for everything else. The premium will accrue to people who can supervise a model well: prompt it precisely, check its work, catch its lies, integrate its output into something a human will sign. That is a learnable skill, mostly by doing it badly for a few months.

If you manage people, start measuring output rather than hours. The 10x engineer of 2027 is the one running three AI agents in parallel; the 10x marketer is the one whose campaigns ship in a day instead of a month. Promotion ladders built around "years of experience" are going to age badly, because the apprentice now arrives with an apprentice of her own.

If you are early in your career, the pessimistic reading is real: the bottom rung of many ladders — paralegal, junior copywriter, first-year analyst, tier-1 support, junior front-end developer — is being eaten. The optimistic reading is that the best 22-year-olds are now doing genuinely senior work in their first job, because the model handles the grunt work that used to be the apprenticeship. Both readings are true at once. Behave as if the optimistic one applies to you.

If you are a citizen and a voter — the hardest part — the policy debate worth having is not "ban AI" or "UBI for everyone." It is whether the productivity gains get captured by the small number of firms that own the models, or get shared through wages, prices and shorter working weeks. That is a political choice, not a technological one, and it will be made by people who vote.

Ask Gradient

If you want to go deeper on a specific occupation — yours, your child's, your team's — try Gradient's Ask Gradient tool, which searches the underlying research papers rather than guessing. A useful place to start: "Which tasks of a [your job title] are most exposed to current LLMs according to Eloundou et al. 2023, and which are least exposed?" The answer will be more useful than another think-piece, including this one. If you want to understand what the models actually do under the hood before you trust them with your work, our What is an LLM? guide is the right next stop.