Concept

Task vs Job

conceptaijobseconomics-of-aiautomation

Task vs Job

Benedict Evans‘s frame for thinking about which work AI automates: ask whether the hard part of a role is the task — the visible output, the code, the SKU, the slide deck — or the job — knowing which code, which SKU, why the politics block it, what the customer actually needs. Automation eats tasks; it rarely eats jobs.

The distinction

  • Sometimes the task is the job. A 1950s elevator attendant pulled a lever; automate the lever and the role disappears. These cases are rarer than they look.
  • Usually the task is a fraction of the job. Amazon gets you the SKU if you already know which SKU you want — choosing it is another job. Claude Code writes the code if you know what code you want — and “what features, for which customer, taken to market how?” is the job. McKinsey’s deliverable is a 75-slide deck, but you don’t pay them for the deck; you pay them to walk your company, read the politics, talk to your customers, and work out what to do.

The same lens explains why frontier AI labs are buying consultancies and hiring forward-deployed engineers: reimagining a company’s workflows is a project needing people, not a prompt.

Why tasks-automated rarely means jobs-destroyed

Evans grounds this in the Jevons paradox (applied price elasticity) and the lump-of-labour fallacy: make a task cheaper and you often do far more of it. Before Excel, junior bankers worked brutal hours; after Excel, they still do — the output expanded. Before IDEs and libraries, developers wrote every line; Apple now writes ~90% of an iPhone app’s code, and there are more developers, not fewer. The count of accountants rose across a century of adding machines, mainframes, ERP, and spreadsheets.

The corollary: don’t trust ”% automatable”

Two fallacies follow:

  • The expert-systems fallacy. Scoring a profession by breaking it into automatable percentages (“17% of a senior partner’s work could be automated”) repeats the doomed attempt to recognise a cat with hand-built edge/ear/eye detectors. A job isn’t a sum of separable steps.
  • Unpredictable exposure. Which jobs get transformed is genuinely hard to foresee. In 1997 “taxi drivers are safe from the internet” looked obvious — then came Uber. Evans’s running examples: personal trainers, video editors, coding itself (the role thought least automatable became the most transformed).

Where mainstream views differ

The frame is a deliberate rebuke to the “job apocalypse” narrative (e.g. Dario Amodei’s warning that entry-level white-collar jobs vanish). Evans’s counter: such forecasts misread labour economics, ignore 200 years of automation creating new work, and underrate diffusion lag (18-month enterprise sales cycles; SAP isn’t ripped out overnight). The opposing camp argues this time is different because AI automates cognition itself, not just a tool — so the new jobs may not materialise fast enough, or at all. Evans’s honest position is radical uncertainty: “it depends”, but “it’ll probably be okay” — while conceding that for someone entering an exposed profession now, the averages are cold comfort.

See also