Showback and chargeback: governing AI costs through visibility, then allocation

Two financial governance mechanisms often confused

Showback and chargeback are two distinct financial governance mechanisms. Showback consists of saying: "Here is what you consumed." It makes consumption visible. Chargeback consists of saying: "Here is what is allocated to your budget." It ties consumption to a responsibility center.

The difference is important: showback informs, chargeback commits. The first can be deployed without changing budgets; the second assumes an allocation rule and identified budget responsibility. The two are often confused, although they do not produce the same effects on behavior.

A starting point: $1.8M in AI consumption per year

Our total AI consumption reached $1.8M per year on Azure, AWS and GCP. This global amount gives the order of magnitude, but it says nothing about the breakdown by team, by product or by use case. Without this granularity, it is difficult to act.

We started with four pilot teams. This choice makes it possible to test the governance mechanism on a limited scope before extending it. It also limits tagging and setup efforts, while producing lessons about available data and behaviors.

The four-step journey

The first step consists of tagging consumption by team, by product and by use case. Tagging is the basic condition: without it, neither showback nor chargeback is possible. It requires defining a convention, applying it and managing exceptions.

The second step is sending a monthly showback. Each team receives a view of its consumption. The goal is not to allocate, but to make visible. The third step consists of identifying uses without value or without an owner. These uses are often the first optimization opportunities: forgotten resources, experiments not stopped, use cases without a clear owner.

The fourth step gradually introduces budgets and caps. We then move from showback toward a logic of allocation, without necessarily switching immediately to full chargeback. This gradual approach makes it possible to test rules, adjust thresholds and limit side effects.

Results on the pilot scope: 18% in four months

On the scope of the four pilot teams, consumption fell by 18% in four months. The annualized saving measured on this pilot scope amounts to $145,000/year. The monthly equivalent is 145,000 ÷ 12 = $12,083/month.

A scope clarification is essential: the 18% applies to the pilot scope, not to all $1.8M in AI spending. This decline should therefore not be directly extrapolated to all consumption. The result measures what was observed on four teams, over a four-month period, with the actions taken at that stage.

Is visibility enough to change behavior?

Visibility is sometimes enough to change behavior even before implementing chargeback. This is one of the lessons of the journey: showback can trigger decisions, for example stopping uses without value or clarifying owners.

But showback can also become additional reporting if it is not linked to decisions. The question asked is therefore: does your showback lead to decisions or only additional reporting? If teams do not change their uses, moving to budgets and caps, then to chargeback, may become necessary.

Chargeback introduces stronger budget responsibility, but it raises practical questions: how to allocate shared costs? How to handle cross-cutting use cases? How to avoid circumvention behaviors? These questions argue for a gradual introduction, starting with indicative budgets before binding caps.

Practical questions before generalizing

As suggestions, several points deserve checking before extending the system beyond the pilot. Is tagging quality sufficient to cover all spending? Are team scopes stable? Are use case owners identified? Is there governance for exceptions and shared costs?

It is also prudent not to confuse savings measured on a pilot with overall savings. The 18% decline on four teams does not prejudge what would be obtained across all $1.8M. Other factors may come into play: team maturity, nature of uses, existence of technical levers, etc.

Finally, the choice between showback and chargeback is not binary. One can start with a monthly showback, then introduce budgets, then caps, and move to full chargeback only when allocation rules are robust and accepted. The goal remains to link consumption to decisions, not to add a layer of reporting with no effect.