Every efficiency claim in healthcare software has the same shape. A task took twelve minutes; now it takes four. The vendor reports the difference as a saving, and the arithmetic is correct.
What the arithmetic leaves out is a decision. Eight minutes do not disappear. They become available to someone, and someone allocates them.
Three destinations
Inside an organization, a recovered minute lands in one of three places.
It becomes throughput. The schedule fills, the panel grows, the same staff sees more patients. The saving is real, and the organization keeps it.
It becomes new work the tool created. Someone reviews the generated note, corrects the extraction, checks the summary against the record. The typing shrank; the checking did not.
Or it comes back. Shorter hours, fewer patients, time for the part of the work that carries no billing code.
The first two are easy to measure and easy to bank. The third requires someone with authority to decide that the saved time belongs to the person who spent it. That decision is not technical, and many deployments never make it.
Returning time is also the only destination that asks an organization to give up capacity it has already paid for. A filled schedule is a plan. An empty slot is a decision to leave the plan unfilled, and someone has to own it.
The accounting boundary
An automation can report 94% straight-through processing while pharmacy technicians reconcile mismatched drug codes in a shadow spreadsheet. The number is not a lie. It is an accounting boundary: work outside the tool’s runtime is booked at zero.
The same boundary explains where saved minutes go. When an organization measures only the tool’s throughput, it will spend the minutes on more throughput. The metric does not have to intend that outcome. It only has to be the number leadership sees.
A clinician who finishes notes faster and finds two more visits on the schedule has not recovered time. A clinician whose notes a model drafts, and who then has to verify every line, has traded typing for checking, which is not obviously cheaper.
Who receives the difference
The buyer and the user are usually different people. An employer or a payer buys the tool; a clinician spends the time. When the saving is captured by the buyer, the user experiences the tool as a new obligation with a better interface.
That is not an argument against the tool. It is an argument about where to look for its effect. A deployment can improve the organization’s cost per visit and leave the clinician’s evening unchanged. Both facts can be true, and the second one rarely appears in the business case.
What to measure
I would measure three things: time to produce the artifact, time to check it, and time to repair it. Then one more: who receives the difference between the first and the other two.
A tool that shortens production and lengthens checking has moved work. A tool that shortens production, leaves checking flat, and shortens the schedule has returned time. The difference is visible only if someone measures all three.
There is a fourth measurement that is harder and more useful: what the clinician does with the time if it is returned. If the answer is “more of the same work, just later,” the saving was real and the benefit was not.
The decision
None of this argues that AI cannot return time. Some tools do, and the clinicians who use them can say so. The point is narrower: a time saving is a surplus, and a surplus is allocated by whoever holds authority over the schedule.
Ask where the minutes went before accepting that they were saved. The answer is an organizational decision, made whether or not anyone makes it deliberately.