AI, Automation and Work: What Task-Level Analysis Actually Shows

A task-level, not job-level, look at what AI automation actually displaces and creates, with historical parallels from past automation waves.

▶ Open the simulation

Why job-level analysis gets the question wrong

The most common way this topic gets discussed — "will AI take my job" — asks the question at the wrong resolution. Almost no job is a single homogeneous activity; a paralegal's job includes drafting routine contract clauses, reviewing discovery documents for relevant keywords, managing client communication, appearing for procedural filings, and exercising judgement about which precedents actually apply to an unusual case. Each of these is a distinct task with a different susceptibility to automation, and economists studying this (most notably Acemoglu and Autor's task-based framework) treat a job as a bundle of tasks rather than a single unit, because that's the resolution at which automation actually operates: a technology automates specific tasks, not entire job titles, and the net effect on a given occupation depends on what fraction of its task bundle is automatable and what happens to the remaining, non-automated tasks.

This distinction produces genuinely different predictions than job-level analysis. A job-level framing asks "is this occupation at risk," which tends to produce alarming binary headlines. A task-level framing asks "which specific tasks within this occupation have costs falling below the value of doing them, and what does the person previously doing those tasks now spend their freed-up time on," which is a harder but more accurate question, and the empirical record of past automation waves gives a fairly consistent, non-catastrophic answer to it.

The routine/non-routine, cognitive/manual grid

The task-based framework typically sorts work along two axes: routine versus non-routine, and cognitive versus manual. Routine tasks are ones that can be fully specified as an explicit procedure — bookkeeping entries following fixed rules, assembly-line welding following a fixed sequence, basic document sorting — and these have been the most consistently automatable category across every wave of technological change going back to mechanised looms, precisely because "can be written down as an explicit procedure" and "can be automated" have always been closely related properties. Non-routine tasks resist this: non-routine manual tasks (a plumber diagnosing an unfamiliar leak in an oddly-configured older house, a care worker adjusting how they physically assist a specific patient) require sensorimotor flexibility that remains genuinely hard to automate even with modern robotics, while non-routine cognitive tasks (a strategy consultant weighing ambiguous, conflicting signals about a client's market position, a novelist making an aesthetic judgement) require exactly the kind of contextual, ambiguous judgement that both earlier rule-based automation and current large language models still handle unreliably compared to a skilled human.

What changed with the recent wave of large language models is that a meaningful slice of what used to be classified as non-routine cognitive work — first-draft writing, code scaffolding, summarising a pile of documents, answering a fairly specific factual question — turns out to be more routine in the relevant sense than it looked, because it has enough recognisable statistical structure that a model trained on a huge corpus of similar past work can produce a competent first pass. This is the genuinely new part of the current wave compared to earlier waves of industrial and computing automation, which concentrated almost entirely on the manual-routine and cognitive-routine quadrants and left non-routine cognitive work largely untouched.

Displacement and creation are not the same event, or the same people

Every serious historical study of a major automation wave finds the same two-part structure: specific tasks get displaced, and specific new tasks or entire new occupations get created, but the crucial complication is that displacement and creation are rarely the same event happening to the same people in the same place at the same time. When ATMs were introduced starting in the 1970s, the routine cash-handling task within a bank teller's job was substantially automated, and the number of tellers per branch did fall — but total teller employment in the US did not collapse the way naive automation forecasts predicted, because banks responded to lower per-branch staffing costs by opening more branches, and the surviving teller role shifted towards the non-routine tasks (relationship-building, cross-selling, handling exceptions) that ATMs couldn't do. That's a genuine net story of task reallocation within roughly the same occupation and workforce.

Textile mechanisation in early 19th-century Britain, by contrast, is the harder, more disruptive historical case: it displaced skilled hand-loom weavers directly and rapidly, and the new jobs created (mill operative roles, and eventually a much larger set of downstream manufacturing and commercial jobs as textiles became cheap enough to massively expand demand) went largely to different people, often in different towns, frequently the next generation rather than the displaced workers themselves, producing real, sustained hardship for the specific cohort caught in the transition even though the long-run aggregate employment and living-standard trends were clearly positive. The honest historical lesson is that both patterns are real and the outcome for any specific automation wave depends heavily on transition speed, geographic concentration of the affected occupation, and how portable the displaced workers' remaining skills are to the newly created roles — none of which can be predicted from the technology alone.

What this suggests about the current AI wave specifically

Applying this framework to large language models and generative AI rather than speculating at the job-title level, the more defensible current-state observation is that within many white-collar occupations, a meaningful chunk of the routine-cognitive task bundle (first drafts, boilerplate, summarisation, initial research passes) is becoming cheaper and faster to produce with AI assistance, while the judgement-heavy, ambiguous, accountability-bearing tasks within those same occupations (deciding what to actually recommend to a client, taking legal or medical responsibility for a decision, navigating a genuinely novel situation with incomplete information) remain resistant, at least with current model reliability. This points towards task reallocation within occupations, similar to the teller pattern, being the more likely near-term dynamic for many roles — the job title survives with a changed task mix — rather than wholesale occupational elimination, though this is not true uniformly, and occupations whose task bundle is unusually concentrated in the routine-cognitive quadrant (some categories of transcription, translation, and basic content production, for instance) look structurally more exposed to the textile-weaver pattern of direct, rapid displacement than the ATM-teller pattern of gradual task reallocation.

The genuinely uncertain part, which historical parallels can only partially inform, is the speed of the current transition relative to earlier waves — language model capability has improved on a timescale of a few years rather than the decades over which mechanisation or computerisation previously diffused through an economy, and the historical record consistently shows that transition speed, more than the ultimate direction of change, is what determines how much real hardship falls on the specific workers caught in a given wave versus being smoothed out across a longer working-life horizon.

Frequently Asked Questions

Why do economists analyse automation at the task level instead of the job level?

Because automation technologies act on specific tasks, not entire occupations, and a job is a bundle of tasks with different levels of automatability; task-level analysis asks which parts of a job's task bundle are actually affected rather than treating the whole occupation as a single at-risk or safe unit.

Did ATMs actually reduce the number of bank teller jobs?

Per-branch staffing fell, but total US teller employment did not collapse as naively predicted, because lower staffing costs led banks to open more branches, and the surviving teller role shifted toward non-routine relationship and exception-handling tasks that ATMs could not perform.

What makes large language models different from earlier waves of automation?

Earlier automation (industrial machinery, early computing) mainly displaced routine manual and routine cognitive tasks; language models are notable for automating a meaningful slice of what was previously classified as non-routine cognitive work, such as first-draft writing and summarisation, because that work turns out to have more learnable statistical structure than expected.

Does history suggest automation always creates enough new jobs to offset losses?

In aggregate and over the long run, yes historically, but the new jobs frequently go to different people, in different places, sometimes a different generation than those displaced, as seen with early textile mechanisation, so the aggregate story can coexist with real, concentrated hardship during the transition.