SAP is bringing AI token spending into its regular finance processes after finding that rising use among employees and automated agents could produce disproportionate costs.
SAP Chief Controlling Officer Lukas Deutsch and SAP Financial Management Chief Marketing Officer David Imbert outlined how the company has been testing and refining its own approach. They say token caps, model routing, tool rationalization, and greater cost ownership helped contain what they describe as a “triple-digit-million-dollar financial risk.”
Deutsch and Imbert present the work as an evolving internal framework. Their central lesson is that managing AI spending requires better visibility into consumption, clearer ownership, and a way to weigh cost against business value.
AI Usage Exposed Gaps in Cost Visibility
Deutsch and Imbert write that an overall spending figure did not show which teams, workloads, or usage patterns were driving AI consumption. Finance worked with commercial, engineering, and product teams to add operational detail, then refined its forecasts over repeated planning cycles as usage patterns became clearer.
The company also reconsidered how those costs were assigned. Centralized funding had helped teams experiment with AI, but Deutsch and Imbert say it weakened accountability as usage spread into regular business processes. SAP began allocating token costs to individual business areas so managers could see what their teams were consuming and assess that spending against expected outcomes.
The authors caution against treating lower consumption as the goal. A workload may justify higher token use if it produces enough business value to offset the cost.
They point to SAP’s AI development tools as one example. The company says the tools increased the rate at which developers got code changes approved and incorporated by a mid-double-digit percentage.
The exercise shifted the focus from how much AI SAP was consuming to where that consumption was occurring and what the company was getting in return.
Token Caps and Model Routing Target Waste
Deutsch and Imbert identify heavy users and automated agents as one source of disproportionate consumption. They also point to workloads using models that were more expensive or capable than the task required, along with overlapping AI tools.
SAP responded with token caps, model routing, and tool rationalization. The authors say those controls are intended to reduce avoidable consumption without restricting AI use that is producing useful results.
The company now plans to bring those practices into regular planning and reporting, improve cost allocation, and develop forecasts that give finance teams a better view of where consumption is heading. Deutsch and Imbert describe the framework as a work in progress.
That work is becoming more relevant as SAP pushes its Autonomous Enterprise strategy. Introduced at Sapphire in May, the vision calls for AI assistants and agents to take on more work across areas including finance, supply chain, spend management, human resources, and customer experience.
SAPinsider Vice President and Research Director Robert Holland recently reported that customers who initially responded positively to that vision are increasingly asking questions about cost, security, governance, and implementation timing. Many of the underlying agent capabilities are still being developed with selected customers.
SAP’s internal experience does not provide a finished model for those customers. It does show how the financial side of AI adoption can become more important as usage spreads beyond individual employees and into automated business processes.
What This Means for ERP Insiders
AI spending may need business-level ownership. SAP’s experience suggests centralized AI budgets can become less useful as adoption spreads. Finance teams may need to assign responsibility closer to the business areas creating the consumption.
Successful AI can make budgets harder to predict. SAP refined its forecasts as usage patterns developed, suggesting adoption itself can change expected consumption. Finance may need to update AI budgets more frequently as successful workloads expand.
Useful AI can still become expensive AI. SAP’s developer example shows why higher consumption is not automatically waste. Finance teams may need to distinguish productive growth in usage from spending that rises without a clear operational return.
This article was first published by SAPinsider.
