The Cognitive ROI: Why Time Saved Is the Wrong Metric

By four in the afternoon, the senior on call is a different person than she was at eight. Not less capable — less available. The morning went on passwords that expired overnight, billing codes that needed validating, a renal score she calculated by hand because the system would not, and a queue of alerts that mostly concerned someone else's patient. Now a family is in the relatives' room waiting to talk about whether to keep going. She has done the difficult part of this job a thousand times. The question is how much of her is left to do it well.

When a tool is brought to a hospital board, the case for it almost always rests on one number: minutes saved per encounter. It is a tidy figure and an honest one, and it measures the wrong thing. It treats a clinician's day as a factory line, where every minute is worth the same as every other. A day in medicine is not built that way.

Attention is a budget, not a tap

We talk about clinical attention as though it were a utility — always on, drawn down without consequence. It behaves more like a budget. Every clumsy interface, every code to confirm, every score worked out by hand spends a little of it, and the account does not refill on demand over the course of a shift.

So we have arranged things almost perfectly badly. We make some of the most expensively trained people in the building spend the first half of their day on administrative trivia, and then ask them to make the day's hardest calls on what is left. The reflexive antibiotic prescription, the consent conversation rushed because the clock is against it, the subtle finding skimmed past at the end of a long list — these are rarely failures of knowledge. They are what depletion looks like.

What a machine should and should not carry

It helps to be blunt about which work is which. A good deal of clinical labour is rule-bound and predictable: assembling a timeline from five years of notes, checking a drug against renal function, coding an episode for reimbursement. Machines do this kind of thing tirelessly and well; people do it slowly and resent it, with good reason.

The rest is the work no machine can take. Negotiating a treatment plan that fits a particular life. Noticing the fear behind a composed face. Weighing a year of quality against a statistic about survival. This is the part of medicine that justifies the training, and it is precisely the part that suffers when the practitioner arrives at it already spent.

The useful image here is not the robot doctor but the exoskeleton. A frame on a warehouse worker does not lift the load for them or decide what to lift; it lets them carry what they already carry without wearing out their back. An AI that turns five years of scattered notes into a coherent timeline has not done the doctor's job. It has cleared the rubble so the doctor can begin it.

A better question for the strategy review

All of which suggests a different question to put to any tool, and to the people who own your systems. Not only “does it save five minutes?”, but “does it give attention back?” Does it reduce the constant switching between tasks? Does it leave the practitioner with more of themselves at four in the afternoon than they would otherwise have had?

By that test, a system that shaves minutes off a task but demands hard concentration to operate is a poor bargain: it pays you in time and charges you in judgement. A system that quietly absorbs the drudgery, and asks nothing of the clinician's attention in return, is doing something closer to a safety function.

We do not need machines so that doctors can work less. We need them so that when a clinician sits down with a frightened family late in the day, there is still enough of her there to do it properly.