AI costs: token consumption becomes a budget line of its own

What happened

In a German-language SAP News Center article from 1 October 2026, SAP describes how the company handles spending on AI tokens internally: costs are allocated to the business areas, and as causes of disproportionately high consumption SAP names concentration on a few heavy users and automated agents, an unsuitable choice of model and uncontrolled use of tools.

Who is affected

The CFO organisation and programme leadership in companies bringing AI functions into their processes. As long as AI is an experiment with a central budget, consumption hardly stands out. Once assistants and agents run in day-to-day purchasing, sales or logistics, a type of cost arises that fluctuates with usage and that in many business cases so far appears as a fixed line or not at all.

What to do

SAP draws two conclusions from its own practice that can be transferred. First: transparency comes before precision – to begin with, it is enough to know which areas and use cases drive consumption; the model can be refined once patterns become visible. Second: whoever consumes should also see the costs, and the question is not only “What are we spending?” but “What do we get for it?”. For a running SAP programme this means, in our view: carry consumption-based AI functions as a separate, variable line in the business case, name an owner per use case and review the consumption assumptions after go-live instead of carrying them forward from the pilot phase.

Our assessment

The article comes from SAP and describes its own experience; the stated increase in reviewing and implementing code changes, in the mid double-digit percentage range, is SAP’s figure without further breakdown. We think the direction is right: anyone looking only at the cost per token may throttle exactly the usage that pays off. For companies using the software there is something the article does not cover: as a rule they do not manage token consumption itself, but what their contract makes of it – quotas, tiers, billing units. Before AI functions go into a process, the question of how consumption is measured and billed, and who in the company looks at that figure every month, therefore belongs on the table.