A Queue Is Not a Shelf

Underneath every staffing model sits a small claim about physics. Work goes up, waits, comes down weighing what it weighed. Service targets, the business case for one more reviewer, the whole apparatus of capacity planning is priced on the assumption that the ticket you open at hour six is the ticket that arrived at hour one.

Sometimes that holds. Often the shelf is warmer than a shelf, and things grow on it.

Start at the far end of the wait, with the person who inherits it. A support worker opening a conversation that has already gone bad. An engineer opening a proposed change whose author has moved on and can no longer quite reconstruct their own reasoning.

The chat clock is the one you can watch happening. In a study of live-chat logs, stretching the gap between messages from 21 seconds to 80 slowed the customer's own next reply by 20.8%, and the drag accumulated down the length of the exchange. Before the agent had said anything at all, though, a long wait made customers move faster once contact began. Scarcity concentrates people. Waiting to be seen is one experience; waiting while someone is ostensibly already attending to you is a different one. Only the second spoils.

Spoiling is the entire mechanism. Waiting turns a broken order into a grievance, and grievances get handled differently, run longer, end differently. Same customer, same defect underneath, new problem.

A randomized deployment at Taobao covered 647 workers and 680,676 conversations, routing the 5.8% it judged eligible, 39,432 chats, to an AI agent with human supervisors watching. When the system handed off a technical problem, the humans took over and carried roughly 65% of what followed. When the handoff was emotional, frustration or dissatisfaction rather than a broken order, those same workers carried 43%. Fewer messages. Less information gathered. Fewer solutions offered. The emotional escalations also tended to arrive later.

Disengagement is one reading. Triage is another, and triage strikes me as the closer one. Effort poured into a relationship that has already ended is effort gone, and the workers appear to have priced that correctly. The data cannot adjudicate between the two readings, and that inability is more interesting than either verdict would be. What it says without ambiguity: emotional handoffs stretched conversations by 40.8% and cost close to a full point of customer rating.

The second clock has no feelings in it and rots anyway, on a different principle. In the first, the work itself becomes harder. In the second, the conditions for attending to it get thinner. This publication asked a miniature version of the same question when it wondered whether the page was still fresh by the time an agent got around to clicking it.

In a study of 196,212 code changes at an anonymized industry site, output per developer roughly doubled. The reviewer pool grew by half. Review activity per reviewer doubled. The review itself went the other way:

The authors are careful to say the automation came "in place of, not merely alongside" human review. And whoever finally opens a change that has gone cold is reassembling intent from someone who has already gone somewhere else.

More reviewers. Less review.

The cost sits in the latency numbers, which is where this sort of cost usually hides. Organization-wide cycle time spiked and then came back down, a shape that reads like a system absorbing a shock and recovering. But among the changes that drew an actual human comment, the slow tail peaked near 114 hours and was still sitting around 73 in 2026, against 63 before the ramp. The average improved partly because the population being averaged had quietly changed underneath it. Fewer things received real attention, so attention looked faster.

Which makes capacity neither a headcount nor a throughput figure. It is the volume of attention that arrives before the work has turned into something else. The moving edge deserves a name: call it the recovery frontier, the point past which the good outcome can no longer be bought at any staffing level, because the thing you could have repaired has stopped existing.

Neither clock shows up in a staffing model. They run at different speeds. And acceleration does not manufacture the decay. It moves the deadline forward while the spreadsheet stays exactly where it always was.