Coordination cost

How it shows up

This arrives in three stages, months apart.

Each stage looks like good news or like a temporary problem. The third one is where the money is, and by then it has stopped feeling like an AI question.

1. First, within weeks People start finishing more, and they are right

   The drafting, the coding, the first version of anything. Real, visible, and worth having. Nobody is exaggerating at this stage.

2. Then, within a quarter The queue in front of the reviewer gets longer

   Twice as much arrives at the person who has to check it, and that person did not get any faster. Everyone treats this as a temporary backlog. It is the new normal shape of the process.

3. Much later, when somebody asks The delivery date has not moved

   Request to delivered still takes what it took last year. The gain is real and it is sitting in a queue, and now you are explaining to a board why the invoices went up and the calendar did not change.

What it is

AI made production cheaper. Review stayed exactly as expensive.

Producing something went from hard to easy. Judging whether it is any good takes a person, the same person as before, reading at the same speed as before. So the volume doubled at the point where it was cheap and stayed flat at the point where it was expensive.

A queue forms wherever those two meet. That is not an AI phenomenon, it is what happens in any system where one station speeds up and the next one does not, and it has been true of factories for a century. What is new is how fast the first station got faster.

Everybody optimised the part that was already the easy part. The waiting was always where the time went, and nobody bought anything for that. Paul Musters, founder of emaho

And there is a second effect that is easy to miss. When there is more to review, the reviewing gets worse, because the same person is now working through twice the pile with the same attention. So the queue gets longer and what comes out of it gets less reliable at the same time.

Why nobody notices

Three reasons the waiting never got attention.

Waiting is invisible in every system you own

Your tools count what people do. Tickets closed, documents produced, messages sent. The hours a piece of work spends sitting still belong to nobody, so they appear in no report, and a cost that appears in no report does not get a budget. Asana's long-running finding is that knowledge workers spend about 60% of their time on coordination work rather than the work itself.3

The bottleneck moved and nobody redrew the map

Two years ago production was the constraint, so that is where the money went and that is where the AI went. Production stopped being the constraint some time in the last eighteen months. Almost nobody has moved the attention to where the constraint went, because the plan was written before it moved.

Everyone is measured on their own step

The writer is judged on output, the reviewer on thoroughness, the manager on their team's throughput. Nobody in that chain is judged on how long the thing took to reach the customer, so nobody in that chain is wrong. The whole is slower and every part is doing exactly what it was asked to do, which is the most expensive kind of problem to have because there is nobody to correct.

Where it sits

Level 2 buys speed. Level 3 buys flow.

In the five AI Culture Levels we use with clients, Level 2 is daily individual use with no shared standard. Everybody gets faster at their own step, which is exactly the pattern on this page. About a quarter of organisations are there.

Level 3 is where the way of working gets written down, and the reason that matters here is dull and specific. When everyone drafts the same way, review takes less time, because the reviewer stops rebuilding their expectations for every piece of work. That is where the queue starts to shrink, and it costs nothing except the decision to agree on a method.

What a measurement adds to a stopwatch

An Operating Profile in use. Personality type and AI level in one profile, with the agents that fit it.

The stopwatch tells you where the work waits. It does not tell you why it waits there, and that is usually a person question. Somebody two levels ahead of the rest is producing in a shape nobody else can check quickly. Or the reviewer is the only one who can judge this kind of work, and that has been true for three years and nobody said it out loud.

The measurement is two things per person: the way somebody works, and their AI level. Put that next to your timestamps and the queue gets a name. In most companies it turns out to be one handover between two people, which is a much smaller thing to fix than the process redesign somebody is about to propose.

Fifteen minutes per person. No credit card, no strings.

What it costs

Everything finishes sooner and arrives at the same time.

You are paying for the tools and getting the benefit in a queue. That is the cleanest way to say it. The invoices are real, the individual gains are real, and the thing your customer buys has not changed, so on paper you have bought nothing.

Then quality starts to slip in a way that is hard to trace. Reviewers working through twice the volume catch less. Faros saw bugs rise 9% in 2025 and 54% per developer across the longer window, alongside a large rise in incidents.12 That lands on the same people who are already the bottleneck, which makes the queue worse again.

And there is what it does to the people producing. Finishing work that then sits for a week is demoralising in a specific way: it teaches them that their speed does not matter. The ones who care most notice first, and they are the ones who were making the gains in the first place.

Nearby

The queue has neighbours.

These three sit around the same problem, from a different angle each time.