Design: Task-level framework analysis synthesising six time-motion studies, ACGME entrustable professional activities and the O*NET task inventory, independently classified by two emergency physicians
Population: 14 emergency physician task categories
Primary outcome: Classification by automation susceptibility and mapping to current AI capability as replace, augment or no application
Effect: Nine tasks (64.3%) non-routine abstract, three (21.4%) routine cognitive, two (14.3%) non-routine manual, none routine manual. AI replacement concentrated in routine cognitive tasks consuming 20-40% of shift time; augmentation dominant in non-routine abstract tasks

Predictions about artificial intelligence replacing clinicians usually operate at the level of the job. This analysis operates at the level of the task, borrowing a framework from labour economics that divides work into routine and non-routine, cognitive and manual — the framework that has correctly predicted which occupations automation reshaped in other sectors.

Fourteen emergency physician task categories were assembled from six time-motion studies, the ACGME entrustable professional activities and the O*NET task inventory, then independently classified by two board-certified emergency physicians and mapped against current AI capability as replace, augment, or no current application.

Nine tasks (64.3%) came out as non-routine abstract — the diagnostic reasoning, the resuscitation decisions, the difficult conversations. Three (21.4%) were routine cognitive: documentation, medical records review, and departmental operations management. Two were non-routine manual: procedures. None were routine manual.

The finding worth carrying is where those three routine cognitive tasks sit in a shift: they consume 20 to 40% of physician time, and they are the ones where AI replacement, not augmentation, already applies. That is the realistic near-term change — not a machine making the diagnosis, but a machine taking back a third of the shift currently spent on the notes.

The near-term target is documentation and records review, not diagnosis
Routine cognitive tasks consume 20 to 40% of shift time in the time-motion studies synthesised
Procedures and resuscitation decisions show minimal automation potential
Evaluate any AI documentation tool on time returned to patient care, not on accuracy alone
This is a framework analysis by two classifiers, not an empirical study of deployed systems

Why it matters

It replaces a vague anxiety about replacement with a specific and checkable claim about which third of the shift is in play.

The statistics, in plain English

These percentages describe how 14 task categories were classified, not how often anything happens: 64.3% means nine of fourteen categories, which is a description of the taxonomy rather than of a shift. The 20 to 40% of shift time figure comes from previously published time-motion studies, not from new measurement here. Two classifiers agreeing on a framework is a reasonable method for this question and a weak one for anything requiring precision.