The Jevons Paradox in Healthcare: Why Faster Doctors Are Not Better Doctors

A board meeting, a slide, a single line that sounds like common sense: if artificial intelligence cuts documentation time by a fifth, the same doctors can see a fifth more patients. The saved minutes appear on the projection as empty inventory, slots on a conveyor belt waiting to be filled. Someone does the arithmetic out loud, and the room nods. It is the most reasonable-sounding mistake in healthcare management.

In 1865 the economist William Stanley Jevons noticed something that ought to have been impossible. As steam engines grew more efficient and wrung more work out of every ton of coal, Britain did not burn less coal. It burned far more. Cheaper to use, the resource was used everywhere, and total consumption climbed. Make a thing more efficient and you do not necessarily save it; you may simply invite more of it.

Hospitals are now poised to reproduce this paradox with clinical time, and the consequences land on patients rather than on a coal ledger.

The bottleneck was never time

The linear logic — minutes saved equal patients gained — rests on an assumption that does not survive contact with a ward. It assumes the scarce resource in medicine is time. It is not. The scarce resource is the capacity to think clearly about one hard case after another. You can use AI to speed up the typing. You cannot use it to speed up the reasoning.

There is a physiological ceiling on how many high-stakes, non-routine decisions a clinician can make in a day before judgement starts to fray. Automation can lower the load that surrounds those decisions — the dictation, the letter, the search through the record. It cannot raise the ceiling on the decisions themselves. If management harvests every freed minute and feeds it back into the appointment grid, it does not buy more good decisions. It packs more difficult ones into the same tired hour.

The result is a hospital that moves patients through more quickly and treats them slightly worse. Throughput rises; so do the quiet failures. The discharge that comes apart because nobody had the headspace to ask whether the patient could actually manage at home. The contraindication that was there to be seen, in a record no longer read with full attention. These do not show up in the efficiency report. They show up later, as a readmission, a complaint, a harm.

The cognitive dividend

There is a better way to treat the time that AI returns, and it begins by refusing to call it spare capacity. Treat it instead as a dividend to be reinvested in the quality of the work rather than its quantity.

Say a tool gives a cardiologist back ten minutes on a routine consult. The tempting move is to slot in another patient, and it produces a clean short-term number: one more case, one more code, today. The more durable move is to let those ten minutes go where the judgement is — into actually examining the raw angiogram rather than skimming the report, or into the unhurried conversation with a family that decides whether a treatment plan will be followed at all once everyone goes home. The first choice books revenue this afternoon. The second avoids costs that would have arrived, larger and later, on someone else's desk.

This is where the usual return-on-investment case for AI quietly misleads. The return is not the additional consultation squeezed into the day. It is the readmission that never happened, the litigation that never began, the deterioration caught early because the clinician still had the attention to catch it. These returns are real and substantial, but they are invisible to a dashboard built to count volume, and so the institution optimises for the number it can see and erodes the value it cannot.

Ring-fence the surplus

The decision is not really about technology. It is about what the organisation does with the surplus the technology creates, and that is a choice made in scheduling rules and capacity targets, not in the software. If the freed time is left unprotected, the scheduling system will consume it automatically, because that is what scheduling systems are built to do. The minutes will be gone before anyone decides they should be.

So the surplus has to be ring-fenced on purpose, defended as deliberately as any clinical resource. Used to run the existing treadmill faster, AI buys a little revenue now and a more exhausted, more error-prone workforce soon after. Used to step off the treadmill, it buys back the one thing medicine has been losing for two decades: the room to think before deciding. The technology is neutral. The paradox is a management choice.