"If you have some hallucinations in the process, the errors will actually compound statistically over many steps."
— Dominik Asam, Chief Financial Officer of SAP, July 2026
When the CFO of the world's largest enterprise software company says AI token spending is "going through the roof," it is worth pausing to understand exactly what he means, because it is not the headline most people assume. Speaking around SAP's second-quarter 2026 earnings, Dominik Asam did not say AI is a bubble or a waste. His argument is more precise and far more useful: today's soaring token bills are being spent on the easy wins, and the genuinely valuable use cases, the ones inside finance and supply chain, are dramatically harder and more expensive to do safely. Here is what that means for anyone budgeting for enterprise AI.
What "Tokens" Are, and Why the Bill Is Climbing
A quick definition, because the whole argument rests on it. A token is the unit AI models use to process text; roughly speaking, a token is a chunk of a word, and every request you send to a model and every answer it returns is billed by the number of tokens consumed. When Asam says token spending is going through the roof, he means the raw volume of AI usage across enterprises is rising fast, and the meter runs on every single interaction.
The important nuance is where that spending is currently going. Asam noted that the "lion's share" of AI token consumption today is spent on what he called "low-hanging fruits": coding assistants and chatbots. These are the natural first use cases precisely because they are forgiving. If a coding assistant suggests a flawed snippet or a chatbot gives an imperfect answer, a human catches it and the cost of the error is small. The output carries limited risk, so you can tolerate the occasional mistake.
The Reframe That Matters
"Token spending is going through the roof" is not a complaint that AI is overhyped. It is an observation that enterprises are spending heavily on the safe, low-value use cases, and that the expensive part, applying AI to core business processes where it actually moves the needle, is still ahead of them. The rising bill and the missing ROI are two sides of the same coin.
Why Finance and Supply Chain AI Is So Much Harder
The heart of Asam's argument is about compounding error. In a chatbot, a single response either helps or it does not, and that is the end of it. In a core business process, one AI step feeds the next, which feeds the next. As Asam put it, "if you have some hallucinations in the process, the errors will actually compound statistically over many steps." A small error rate that is harmless in a single chatbot reply becomes a serious reliability problem when it is chained across a ten-step financial close or a multi-stage supply chain plan.
This is why he says these use cases require "much more excruciating assurance levels." To run AI safely inside a process that touches compliance, auditability, and money, you cannot tolerate the error rate that is perfectly acceptable in a coding assistant. Getting to that level of assurance, more validation, more guardrails, more retries, more human checkpoints, is exactly what drives what Asam called "extremely high token costs." The expensive AI is not expensive because the model is fancier. It is expensive because doing it safely requires far more computation per unit of real work.
We saw the same structural point from the opposite direction when we wrote about why you cannot simply vibe-code an ERP: the core of a business system is the red zone where compounding errors and compliance make casual AI dangerous. Asam is making the CFO's version of that argument, and putting a price tag on it.
The Line Everyone Should Copy Down
The single most important sentence in Asam's remarks is this: "The idea that AI will solve all these problems if they are messy, legacy data silos is not true." This is the quiet demolition of the most common enterprise AI fantasy, the belief that a powerful enough model can be pointed at a mess of disconnected systems and simply make sense of it.
It cannot. A model acting on bad, ungoverned data produces confident, expensive, wrong answers, and in a multi-step process those wrong answers compound. This is why the real work of enterprise AI is not model selection but data foundation. It is the same conclusion we reached in our analysis of why AI stalls at scale, where 83% of organizations named data quality as their single biggest AI blocker, and it is the premise behind SAP's own push to let customers ground Joule in their own governed knowledge rather than relying on a generic model. Clean, governed data is not a nice-to-have that precedes the AI project. It is most of the AI project.
Use the Cheapest Reliable Tool, Not the Fanciest
Asam's guidance on model choice is refreshingly unglamorous. Companies, he argued, will use the cheapest reliable tool that can deliver the required outcome safely, whether that is simple traditional software, an open-source model, or an expensive frontier model. The most advanced model is not automatically the right one; the right one is the least expensive option that clears the safety and reliability bar for the specific task.
This is precisely the procurement logic we documented when Shopify, Airbnb, Coinbase and others switched to cheaper AI models, cutting costs by factors of 9 to 75 by matching the model to the task instead of routing everything through a premium frontier model. Asam is confirming, from inside the largest enterprise software vendor, that this right-sizing discipline is not a fringe cost hack. It is becoming the default enterprise stance. For a lot of steps in a real workflow, the correct tool is not AI at all, it is deterministic software that never hallucinates.
A Practical Budgeting Rule
For each step in a process, ask two questions in order. First, does this step even need a language model, or would deterministic software do it more cheaply and reliably? Second, if it does need a model, what is the cheapest model that clears the assurance bar for this specific step? Answering these deliberately, step by step, is how you keep token spending from going through the roof while still capturing the value.
Why SAP, of All Companies, Is Saying This
It is worth asking why SAP's CFO is the one making this case publicly. SAP sits at the center of exactly the high-value, high-risk processes Asam is describing, finance, supply chain, procurement, the systems of record that run the world's largest companies. SAP's entire AI strategy, from Joule to its embedded agents, is a bet that it can deliver AI inside these processes with the assurance levels they demand. Asam is, in effect, explaining why that is hard, why it is expensive, and why SAP's position, owning the governed data and the process logic, is the thing that makes it possible at all.
There is also a candid financial subtext. SAP has been managing its own AI cost pressure carefully; earlier in 2026 the company froze hiring and cut travel explicitly to help fund its AI investments. A CFO who has felt the token meter run inside his own company is a credible narrator on why the rest of the enterprise world should expect the same, and plan for it.
What This Means for Your AI Budget
Strip Asam's remarks down to their practical core and three planning implications stand out. First, expect your AI bill to rise before your AI ROI does, because the cheap, safe use cases come first and the expensive, valuable ones come later; budget for that sequence rather than being surprised by it. Second, invest in data governance now, because no amount of token spend compensates for messy legacy silos, and the model cannot fix what the data foundation gets wrong. Third, build the discipline to match each task to the cheapest reliable tool, because in a world where token spending genuinely can go through the roof, cost control is not the enemy of AI value; it is the precondition for it.
The most reassuring thing about Asam's comments is what they are not. They are not the words of a skeptic writing off enterprise AI. They are the words of a CFO who believes the valuable use cases are real and coming, and who is telling the market clearly what it will take to reach them: governed data, right-sized tools, and the assurance levels that core business processes have always demanded. The token bill going through the roof is not the failure of enterprise AI. It is the sound of the industry paying tuition on the way to the use cases that actually matter.