Your backlog is a queue problem, not a headcount problem

The request almost always arrives as a headcount question. The backlog is four days deep, the team is drowning, how many people do we need to hire. It is rarely the right question, and it is never the first one.

Where the time actually goes

In the inboxes we have measured, between a half and two thirds of the queue is made up of questions that have a documented answer and no judgement in them at all. Those messages are not slow because they are hard. They are slow because they are sitting behind something else.

A first-in-first-out queue treats a request for a tracking link and a complaint about a damaged order as the same object. One of them takes eleven seconds to answer and the other takes a considered reply from someone senior. Handled in arrival order, the eleven-second messages spend hours waiting, and the considered reply gets written at the end of a long day.

Why adding people does not fix it

Adding people to a queue that sorts badly gets you a faster mess. The same order-status questions get answered sooner, the same hard conversations still wait until Thursday, and the backlog reappears at the next seasonal peak with a larger payroll attached to it.

There is a second cost that shows up three months later. A team hired to clear volume gets measured on volume, and a team measured on volume stops writing documentation, because documentation is the one activity that does not close a ticket today.

Sort first

Sorting by urgency and tone before anything else moves the angry and the time-critical to the front, and moves the routine to somewhere it can be answered without a person opening it at all. The backlog does not get worked harder. It gets shorter at the bottom.

This is also the cheap experiment. Sorting is reversible in an afternoon; a hire is not reversible for a year. If you are going to be wrong about your queue, be wrong in the direction you can undo.

What sorting actually looks like

It is less sophisticated than it sounds. Three signals do most of the work: what the message is about, how urgent it reads, and how upset the person writing it sounds. None of those requires a model to be clever. They require the queue to stop pretending that arrival time is the only thing it knows.

The effect compounds in a way that surprises people. Pulling the two-minute messages out of the line does not just answer them faster; it shortens the wait for everything behind them. In the teams we have watched do this and nothing else, median first response fell by more than half while the hardest decile barely moved — which is exactly the right shape, because the hardest decile was never the problem.

The part that needs care is the upset detection, and it needs care in one direction only. A false positive costs you a person reading a message that did not need them. A false negative costs you an angry customer waiting behind a tracking enquiry. We tune it to be wrong generously.

What to do this week

We have yet to see a team get to step four and still need the hire they asked for at step zero. That is not an argument against hiring. It is an argument for knowing which problem you are buying your way out of.

What to expect

Two numbers are worth holding on to. The first is that most teams find between forty and sixty percent of their queue is answerable from something already written down, which is higher than anyone guesses beforehand. The second is that the backlog, once sorted, tends to stop being a backlog at all — it becomes a small standing set of genuinely hard conversations, and those were always going to take the time they take.

That is the outcome worth aiming at. Not an empty inbox, which nobody has ever had for long, but an inbox where the things still waiting are the things that deserve to be.