What should we charge on?

Whatever behaviour you would like more of, because the unit is what people optimise against. Per person encourages hoarding access and says nothing about use, per request encourages batching, and share of total encourages arguing about the allocation rather than changing anything at all.

When is it the wrong time?

During adoption. The same mechanism trims waste once value is established and prevents value from being established in the first place, so introducing it early stops the very experiment that would have told you whether these tools were worth funding at all.

Glossary

Chargeback

Allocating a shared tool's cost back to the teams that use it, so the bill lands where the decisions are made.

In plain terms

Passing the bill for a shared tool on to whoever used it. It works, in the sense that people use less immediately. Whether that is a success depends entirely on whether the use you lost was the use you wanted.

Why it matters

Because it is the fastest lever available and the least selective. A policy asks people to think; a bill makes them stop, and it makes them stop uniformly, including on the work that was producing more value than it cost.

How it works

It works by moving the cost to where the decision is made. A central budget is nobody's concern in particular; a line in your own team's numbers is somebody's, and that difference produces behaviour change within days rather than quarters.

The change is indiscriminate, which is the risk. Usage falls across the board rather than falling on the wasteful part, so an organisation still trying to establish whether these tools help has just suppressed the experiment it was running.

The unit you charge on decides what people optimise. Charging per person encourages hoarding access and says nothing about use; charging per request encourages fewer and larger requests; charging per team by share of total encourages arguing about the allocation rather than changing anything.

Visibility without a charge captures most of the benefit. Showing each team what it spent, with nothing moving between budgets, produces most of the attention and none of the suppression, and it is the right first step almost every time.

Timing matters more than design. Introduced during adoption it stops adoption; introduced once value is established it trims waste, and the same mechanism produces opposite outcomes depending only on when it arrives.

It changes who decides, which is the effect worth wanting. A central budget concentrates the judgement in whoever guards it, and a team carrying its own cost makes that judgement locally, where somebody actually knows which uses were worth having. That is a better decision made in a worse mood.

It creates work of its own. Somebody has to attribute usage, handle the arguments about attribution and maintain the model as tools change, and that overhead is worth weighing against a saving that is often smaller than expected.

Two versions of the same intervention

They do start the same way: somebody has to attribute usage to teams, which is the difficult and useful part, and once that is done both versions are available for the cost of a decision. The temptation is to go straight to the second, because attributing cost without acting on it feels incomplete. What that overlooks is that the first version has already done the thing that matters. A team shown a number it did not expect will investigate it without being asked to, and the investigation is usually more accurate than any policy, because the team knows which of its own uses were worth having. Moving money adds pressure to a group that is already looking, and pressure applied to somebody who is already correcting the problem produces overcorrection. The sequence worth following is to attribute, show, wait a month, and only then ask whether any charge is required. Frequently the number has already come down and the answer is no.

Seen in the wild

A shared assistant subscription split by department, and usage dropping everywhere.
ChatGPT
Automation costs attributed to the team whose workflow generated them.
Make
Model spend broken down by team and shown, without any money moving.
OpenRouter

Common misconceptions

People assume

It reduces waste.

In fact

It reduces usage, which is not the same thing. The reduction lands on valuable and wasteful use alike, and the valuable use is often the part that was quietly discretionary and therefore easiest to stop.

People assume

It is mainly an accounting exercise.

In fact

It is the strongest behavioural lever available in this area, and it works before anybody has read a policy. Treating it as an administrative change is how organisations end up surprised by how fast usage fell.

Questions

What is the safest way to start?: Visibility without a charge: show each team what it spent and move nothing between budgets. That produces most of the attention and none of the suppression, and it lets you see who spends what before deciding whether any of it needs discouraging.
What should we charge on?: Whatever behaviour you would like more of, because the unit is what people optimise against. Per person encourages hoarding access and says nothing about use, per request encourages batching, and share of total encourages arguing about the allocation rather than changing anything at all.
When is it the wrong time?: During adoption. The same mechanism trims waste once value is established and prevents value from being established in the first place, so introducing it early stops the very experiment that would have told you whether these tools were worth funding at all.