Skip to content
Contact us
Insights

Token chargeback per cost centre in four steps

  • DateNovember 17, 2026
  • CategoryCost & Operating Model

Model providers bill per contract. Large organisations budget per cost centre, business unit and legal entity. Between the two sits a gap that grows with every new use case. When the monthly invoice arrives as one consolidated number, it usually lands on the IT cost centre. Business units consume, IT pays, and nobody can say which use case is responsible for the increase since last quarter.

If you cannot allocate AI costs, you cannot steer them. Budget owners have no signal that would lead them to choose a smaller model, cache repeated context or retire an unused agent. Finance cannot compare the cost of a use case with its benefit. And the platform team ends up defending a growing line item it does not control. Chargeback is therefore not an accounting detail. It is the mechanism that connects consumption with responsibility.

We recommend a method in four steps that finance and the platform team own together.

Step one, every request carries metadata: each call to a model passes through the gateway with mandatory attributes for cost centre, use case, tenant and environment. In practice these are request headers that the gateway validates, for example a cost centre code, a use-case name such as claims triage, the business unit acting as tenant and a marker for production or test. The decisive rule is enforcement. Requests without valid tags are rejected, not estimated afterwards. Allocation that relies on reconstruction after the fact produces disputes and unallocated remainders.

Step two, measure tokens instead of estimating: the platform captures input, output and cached tokens for every call, together with retrieval operations and tool calls that create their own cost. This metering lives in the platform, not in each application. If every team implements its own counting, the numbers will not add up to the invoice, and finance will lose confidence in all of them.

Step three, set prices and allocation: direct cost follows from measured usage multiplied by the applicable price. A worked example makes the logic tangible. Forty million tokens at two euros per million tokens result in eighty euros of direct cost. On top of that comes a share of the shared platform cost, distributed by a fixed key that finance agrees with the business, such as share of usage or number of active use cases. The key should be simple, stable for a budget period and documented.

Step four, showback before chargeback: run one quarter in which teams see their costs without being billed. This gives them time to understand the numbers, question the allocation and change their behaviour before budgets are affected. After that period, the same data becomes real internal bookings.

The main trade-off is granularity. Allocation per prompt or per user creates booking overhead that nobody wants to reconcile. Allocation per business unit is too coarse and leads to disputes about who caused which share. In most organisations, cost centre combined with use case is the workable level. It is fine enough to steer and coarse enough to book. A second trade-off concerns shared costs. Platform operations, evaluation runs and idle capacity do not map neatly to individual requests, and any key for distributing them is a convention rather than a measurement. Being transparent about that convention matters more than finding a perfect one.

There are also limits to what chargeback can achieve. It makes costs visible and attributable, but it does not by itself make them lower. Budget owners need complementary instruments such as quotas per use case, model allowlists that steer simple tasks to smaller models, and regular reviews of the most expensive use cases.

For an architecture team, the next steps are practical. Agree with finance on the mandatory metadata fields and on the source of truth for cost centre codes. Configure the gateway to reject untagged requests, starting in non-production environments to find gaps early. Define the metering record per call and confirm that its totals reconcile with the provider invoice. Choose and document the allocation key for shared costs. Then schedule the showback quarter and agree in advance when real bookings begin.

Once tags are enforced at the gateway and metering is consistent, chargeback stops being a project. It becomes a query.

cloudstrata builds and operates AI platforms for enterprises. In AI Foundry, our managed runtime for AI agents, usage and cost metering is a built-in building block alongside multi-tenant workspaces, routing of models and tools, knowledge retrieval, evaluation hooks, audit trails and central governance. If you want to discuss a chargeback model for your organisation, you can request a technical introductory conversation via cloudstrata.io/contact.

CONTACT

Let's talk about your project

Describe your use case. You get a concrete next step — not a sales pitch.

We reply within one business day.

Follow Cloudstrata on LinkedIn and Instagram to stay up to date with our work and openings.

Details used only to respond. Data privacy

Token chargeback per cost centre in four steps | cloudstrata