Every afternoon the scheduler undoes what the software booked that morning. It sends follow-ups to clinics that closed, to times the patient already said they could not make, to specialists who stopped taking that insurance two years ago. She calls, rebooks, apologizes, and catches most of them.
Upstairs the dashboard reports 94% of discharged patients leaving with a follow-up booked, up from 71% back when three schedulers worked the queue by hand. Two of those three have since been reassigned. Every one of those numbers is accurate, and so is the record: the appointments exist. Most of them were booked twice.
Her afternoon appears nowhere. There is no field for it. The metric she is repairing is the metric that says she is not needed.(note 1)
What a good number hides
We know how to read a bad number. A queue backing up, a complaint rate climbing, an error log filling — these are legible, and institutions have machinery for them.
We are much worse at reading a good number that someone is holding up.
She is not hiding anything. Nobody asked her to conceal the rebooking; nobody would have stopped her from mentioning it. She does not mention it because it is not a distinct thing in her mind. It is the job. It has always included fixing what arrives broken, and the fact that what arrives broken now arrives from software rather than from a fax machine does not change how it feels to fix it.
Asking her will not find it. A survey asking her to report unlogged corrective work would return nothing useful, because she has not filed it under that. She would report that things are busy.
And her competence is what removes the evidence. If she were worse at her job, the follow-up rate would fall, and someone would look. Because she is good at it, the number holds, and the software's failure rate is stored nowhere except in her afternoon.
The experiment already run by accident
One question finds it, and she never has to notice anything.
Ask what happened the last time she was out.
Before the software, when a scheduler took leave, the queue moved slower. Bookings dropped a few points and recovered the following week. The operation bent.
After the software, the week she had flu, the follow-up rate stayed at 94% — and the clinic spent the next month absorbing patients who arrived for appointments that did not exist. The number held because the number was never measuring the thing. What collapsed was downstream and delayed, so nothing connected it back.
Ask one question:
Before this system, what happened when these people took a week off? What happens now?
Not does it degrade — everything staffed by humans degrades when the humans are away. Watch for divergence: the reported performance improving while the tolerance for absence collapses. A system that has genuinely reduced work should be more robust to a missing person, not less. When those two lines separate, the gap between them is the work that is not being counted.(note 2)
The useful thing about this question is that it needs no new instrumentation. Every organization already holds the two halves: staffing records know who was out and when; operational records know what happened. Nobody has cross-referenced them against the date the software went in. The experiment has been run many times already, by accident, and never read.
Two kinds of work that look identical
A sharper version separates a healthy system from a concealing one.
When the software routes something to a human because it cannot handle it, that is the system working as designed. Exceptions exist. Someone should catch them. Automation is supposed to look like this: the machine takes the volume, the person takes the hard part.
When the software marks something complete and a human quietly fixes it afterward, that is different in kind. The dashboard has already counted it. Correcting it afterward does not amend the count. She is not handling an exception — she is manufacturing the number.
Same effort. Same person. Opposite meaning. And the only way to tell them apart is to ask, of the work she actually does, how much of it is on items the system already reported as done.
Most institutions cannot answer that, and the reason is duller than concealment. Nothing was ever built to tell the two apart.
What gaming this looks like
Any measure becomes a target, and this one is no exception. But it has an unusual property: the cheapest ways to cheat it are the fix.
An operator who wants to pass the absence test can staff up, so a missing person is survivable. Or they can find the informal repairs and make them logged, resourced duties — which puts the labor on the books and ends the false efficiency. Or they can fix the software so there is less to repair.
All three are the outcome. There is one real evasion — push the absorption somewhere the telemetry cannot see, onto contractors, onto the patients themselves in the form of a second phone call — and the metric will drive exactly that if the rest is closed off.(note 3)
But the ordinary escape routes here run toward the destination, which is rarer than it sounds.
The claim
An institution that cannot say who its systems depend on does not know how well those systems work. It knows how well they appear to work, which is a different fact, and one that people are producing on its behalf without being asked and without being counted.
The scheduler is not a failure of the software. She is part of it. Does the institution know that?
Note 1.
The scheduler is a composite. The double-booking, the reassignment and the flu week each come from a different organization; no single one of them is a case I can name. Read her as the shape the pattern takes, not as reportage.
The term for what she has become is not mine. See Lineage for where it comes from, and Everyone is a Crumple Zone Now for the argument this essay is the test for.
Note 2.
Organized labor has known this for over a century. Work-to-rule — following the written procedure exactly and withholding the informal judgment that normally fills its gaps — is not a slowdown tactic in the ordinary sense. It is a demonstration that the specification was never sufficient, and it works because the gap it exposes was always there.
Note 3.
On what happens when cost is moved rather than removed, see Optimizing the User.