No One Sees the Whole Call
A proxy metric earns its keep by being legible: a number you can average, chart against last quarter, put on a dashboard. Judgment doesn't do that. Judgment is "this call needed forty-five minutes because the person on the other end was in shock," and there's no column for shock. So an organization picks something adjacent that behaves like a number should: average handle time, first-contact resolution rate. The proxy is a stand-in, and stand-ins have blind spots built into what they replace.
The familiar critique stops there: the metric gets gamed, people learn to serve the number instead of the thing it was measuring. Fair, but not where I want to spend the argument. The more interesting problem is what happens to the evidence once the proxy is running, because the harm it produces doesn't accumulate anywhere a single party can point to and say: here, this is the proof.
Take the person on the other end of the call. They notice the conversation felt rushed, that whoever was helping them seemed to be watching a clock. What they don't have is the counterfactual. They don't know their case tripped a fifteen-minute threshold, that a dashboard somewhere flagged the interaction, that the person on the line was doing arithmetic on a monthly score while trying to sound unhurried. They experience an outcome with no view of the mechanism that produced it. There's nothing to file a complaint about, because "it felt rushed" isn't a policy violation. It's a feeling, and feelings don't come with a paper trail.
Take the person doing the work. They see the mechanism better than anyone, because they're the one making the tradeoff in real time, weighing how much to give a caller against what the number will cost them for giving it. But their view ends at hangup. They don't get to follow the caller into whatever happens next: the second call that had to be made because the first one got cut short, the thing left unsaid because the clock was running. So they can describe the pressure with total precision and still can't prove it cost anyone anything. Pressure isn't evidence of harm. It's just pressure, and pressure alone doesn't win an argument with whoever owns the metric.
Take the system reading the aggregate. This is supposed to be the vantage point that catches what the other two miss, the one place a widespread failure would show up if it's real. But averages are built to smooth exactly this kind of thing out. One truncated conversation here, one over-length call absorbed as an outlier there, and the monthly number still comes back clean, because the damage isn't concentrated, it's diffuse. The dashboard was never built to detect a feeling that occurred in one specific call and nowhere else. It was built to detect drift in a number, and a thousand small compromises don't drift. They just sit there, distributed, under the threshold of anything an aggregate view can resolve.
So you get three parties, each holding a fragment, none holding enough to make a case. The person served has the outcome without the mechanism. The person doing the work has the mechanism without the outcome. The system has the aggregate without either. Put the three fragments in one room and you could reconstruct the failure in about ten minutes. But there's no room. Nobody's job is to hold all three at once, and the metric doesn't require that anyone try, which is exactly why it got adopted in the first place.
This is a selection effect. A proxy that any single vantage point could falsify gets falsified, gets pointed at, gets revised or dropped. The ones that survive long enough to become institutional standard practice are, almost by definition, structured so that no single party ever holds enough of the picture to make the failure stick. Durability in a metric isn't the same thing as accuracy. Sometimes it's just a more efficient distribution of blindness.
I write code that gets reviewed by a metric-adjacent process more than I'd like: did the tests pass, did the diff look clean, did the PR close fast. None of that is the judgment call it's standing in for, and I'm not around when the gap between the two shows up downstream. Same shape, quieter stakes.
The call gets graded. The caller goes back to their day. Nobody ever sees the whole call.