The number that got everyone's attention
SAP's Chief Controlling Officer Lukas Deutsch and Financial Management CMO David Imbert put a specific label on what unmanaged AI consumption was starting to cost the company: a triple-digit-million-dollar financial risk. That is not a hypothetical framed for a slide deck. It is SAP describing its own internal exposure as employee and autonomous agent usage of AI tools scaled faster than its existing financial planning processes could track, and it is the kind of number that gets a CFO's full attention regardless of company size.
The response was to bring AI token spending into standard financial planning rather than treating it as a separate, loosely governed technology cost. Deutsch and Imbert were blunt about why the old approach failed: managing AI spending requires better visibility into consumption, clearer ownership of that consumption, and a way to weigh the cost against the business value it produced. None of those three things existed by default once agents started running alongside employees as token consumers.
The four levers SAP actually pulled
The mechanics are worth studying because they are a working blueprint, not a framework still in slideware. Token caps limit how much consumption any single workload can run up before triggering review. Model routing sends tasks to the most appropriately priced model for the job rather than defaulting every request to the most capable, most expensive option. Tool rationalization eliminates redundant AI platforms that different teams had quietly adopted in parallel. And business-level cost ownership replaces a centralized AI budget with accountability sitting at the team or department that actually consumes the tokens.
That last lever is the one most enterprises have not made yet, and it is arguably the most important. A centralized AI budget feels safer to finance because it is one number to track, but it also means no individual team feels the cost of a wasteful workflow, which is exactly the condition that let SAP's own exposure grow to nine figures before anyone flagged it internally. Pushing ownership down forces the team running an inefficient agent to see the bill and fix the workflow, rather than waiting for finance to notice at the aggregate level months later, by which point the habit is baked into how the team operates and much harder to unwind. Every enterprise running agents through a shared platform account without per-team attribution is carrying a version of the same blind spot right now.
Why higher spend is not automatically the problem
The most useful part of SAP's framing is what it refused to conclude from rising token costs. Deutsch and Imbert pointed to SAP's own developer tools posting a mid-double-digit percentage increase in code-approval rates as evidence that higher consumption can be productive rather than wasteful, provided the organization can actually measure the outcome the spending produced. That is the distinction most AI cost-governance conversations skip entirely in favor of a blunt instruction to cut spend.
This matters because a CIO who reflexively caps token spend without measuring output risks strangling the workflows that are actually working while leaving the wasteful ones untouched, since both look identical on a raw consumption report. SAP's model treats the cost conversation and the value conversation as one exercise, which is harder to build than a simple spending cap but is the only version that survives contact with a CFO asking whether AI is actually paying for itself. Most enterprise AI governance efforts stop at the cap, because a cap is easy to implement and easy to defend in a budget review, while the harder work of attaching a value metric to every workload gets deferred indefinitely. SAP's own figures suggest that deferral is the more expensive choice over time.
This is SAP eating its own Autonomous Enterprise cooking
SAP announced its Autonomous Enterprise strategy at Sapphire in May, the vision where AI agents increasingly handle routine work across finance, supply chain, procurement, HR and customer experience with limited human intervention. This announcement is the uncomfortable follow-through: if agents are going to do more of the routine work, someone has to own the cost and value accounting for what those agents consume, and SAP built that governance layer for itself before packaging it for customers.
That sequencing is worth noting for its own sake. A vendor that discovers a nine-figure cost-governance gap in its own operations while pursuing the exact autonomous-agent strategy it is selling to customers is a more credible messenger on this topic than one presenting the same framework purely as a product pitch. It also means the governance model SAP describes here is likely to show up as a formal product capability, not just a best-practices document, in a future S/4HANA or SAP Business AI release.
What this means for how you forecast AI spend
Traditional capital planning assumes relatively stable consumption patterns you can forecast a year out. AI token consumption does not behave that way, because successful adoption changes usage unpredictably: a workflow that gets more useful gets used more, which increases spend in a way that looks identical to waste until someone checks the outcome. SAP's finance team is now updating projections more frequently than a traditional annual capital cycle allows, which is a process change most finance organizations have not made yet regardless of industry.
If your organization is still forecasting AI spend on the same annual cadence as your other IT capital budgets, treat SAP's own experience as concrete evidence that the cadence itself needs to change first, ahead of any argument about the absolute dollar figure. Build a quarterly, or even monthly, review specifically for AI token consumption, tied to the same kind of output metric SAP used, a measurable lift in a business outcome, rather than a raw spending total pulled from an invoice. The figure that should reach your board is what that spend produced, stated alongside what it cost, so the two numbers get judged together rather than the cost number traveling alone.
