What Automation Does

When people say AI will automate a job, they often picture a robot walking in on Monday morning and doing the whole thing while the human packs their desk. The reality is almost always more nuanced and more gradual than that. To understand what automation actually does, you need to think at the level of tasks — not jobs.

Jobs Are Bundles of Tasks

A job is not one thing. A radiologist's job includes looking at medical images to spot abnormalities, writing reports, consulting with other doctors, explaining findings to patients, teaching residents, attending staff meetings, and keeping up with new research. These are very different activities bundled together under one job title. Automation rarely eliminates all of those activities at once. It targets the tasks that are most repetitive, most rule-governed, and most dependent on processing large amounts of structured information quickly. In the radiologist example, AI can now scan images for certain abnormalities with impressive accuracy — but the consultation, explanation, teaching, and judgment calls in ambiguous cases remain deeply human work.

This task-level view was developed by economists David Autor, Frank Levy, and Richard Murnane. Their research categorized tasks by type: Routine cognitive tasks: following defined rules to process information — think data entry, basic accounting calculations, scheduling. Routine physical tasks: repetitive, precisely defined physical movements — think assembly line work. Non-routine cognitive tasks: judgment, analysis, creativity, complex communication. Non-routine physical tasks: physically adapting to unpredictable environments — think plumbing in old buildings. Automation has historically been strongest on routine tasks of both types and weakest on non-routine tasks. AI is now beginning to make inroads on some non-routine cognitive tasks, which is what makes this moment different from earlier automation waves.

Three Outcomes When a Task Gets Automated

When automation takes over a specific task within a job, one of three things typically happens to the worker: Labor substitution: the worker is replaced because the task was nearly all of their work. A toll booth collector whose entire job was taking cash and raising a barrier is fully substituted when that booth becomes automated. Task recomposition: the worker stays, but their job changes. An accountant who used to spend 60 percent of their time on manual data entry now spends that time on analysis, client advising, and exception handling — tasks the software cannot do. The job became richer, not extinct. Job augmentation: the worker is empowered to do more. A warehouse picker using an AI-assisted routing system can fulfill twice as many orders per shift. The task of figuring out routes is automated; the task of picking and handling items is still human — and the combined output is higher.

Which Tasks Are Most Automatable?

Researchers have developed ways to estimate how automatable different tasks are. The most automatable tasks share these characteristics: they follow well-defined rules; they involve processing structured data; they produce outputs that can be evaluated objectively; and they require little physical adaptability. The least automatable tasks tend to require: understanding context and nuance; physical dexterity in unpredictable environments; emotional perception and genuine empathy; original creative judgment; and ethical reasoning about complex, novel situations. A practical way to think about it: if you could write a very detailed instruction manual that a careful person with no experience could follow to do the task, it is probably automatable. If doing the task well requires human judgment that is hard to articulate — knowing when to break the rules, sensing how someone really feels, responding to something completely unexpected — it is much harder to automate.