SAP is swapping some of the most expensive artificial intelligence models inside its products for cheaper alternatives, a move its chief financial officer says should generate more than €2 billion in efficiency gains over the next two years. The German software giant, whose systems run core operations for many of the world’s largest companies, told Reuters on October 28, 2026 that it had already removed costly models from products including its internal consulting assistant, replacing them where performance remained strong. Overall spending on AI is still expected to rise as adoption grows, SAP said, but the goal is healthier margins as that scale grows, not an absolute cut in AI expenditure.
What SAP Changed
The change is deliberately practical rather than ideological. SAP uses more than 70 AI models from providers including OpenAI, Anthropic and Google, and chief financial officer Dominik Asam said the company increasingly chooses the cheapest model capable of performing a given task to the required standard. That is a routing decision, not a rejection of frontier models: where a top-tier model genuinely earns its price, it stays, but where a smaller or cheaper model delivers the same result, the expensive option is removed.
The internal consulting assistant is a telling first example. Such assistants answer questions, summarise documents and draft analysis, and for much of that work the difference between the most expensive model and a well-matched cheaper one is invisible to the user. Cutting the cost of each answer, multiplied across thousands of employees and millions of interactions, is where the €2 billion figure comes from. SAP framed the gains as efficiency gains over two years, not as a one-off saving, which matters because it implies the benefit compounds as usage grows rather than shrinking once a single contract is renegotiated.
The Cheapest Model That Works
SAP’s approach reflects a wider shift in enterprise AI from capability at any cost to fitness for purpose. In the first wave of adoption, companies reached for the most powerful model available, partly because benchmarks made that choice easy to defend and partly because the cost of getting a task wrong felt higher than the cost of the compute. As usage has scaled, the arithmetic has flipped. Inference is now a recurring operating cost, and a model that is ten times more expensive but only marginally better on a routine task is a bad deal at enterprise volume.
The discipline SAP describes, picking the cheapest model capable of the task, is the same logic behind model routing, cascading and distillation that has spread across the industry this year. A simple classification or summarisation job goes to a small, fast model; a genuinely hard reasoning task goes to a frontier model. The hard part is not the idea but the measurement: knowing, task by task, where cheaper stops being good enough.
Where Models Come From Matters Too
Asam also pointed to a second pressure shaping SAP’s choices: customers are paying closer attention to where AI models come from. Government agencies and companies in regulated industries increasingly set specific requirements and exclude models based on origin, while European customers increasingly demand local solutions. That is a procurement reality as much as a political one, and it favours a multi-model strategy over dependence on a single provider.
What To Watch Next
Providers will also be watching. A public statement from one of the world’s largest enterprise software companies that it is actively trading down on price, where it can, sends a signal to OpenAI, Anthropic, Google and others that enterprise buyers are no longer price-insensitive. That pressure may accelerate the tiering of models, with cheaper, faster options marketed specifically for high-volume enterprise work, and frontier models reserved for tasks that truly need them. In that sense, SAP’s move is not just a cost exercise for one company; it is a vote for a maturing market in which price and performance are traded off openly, task by task.
Why It Matters
This matters because it marks the moment when enterprise AI stopped being a capability showcase and became a cost discipline. For years, the debate was about which model was smartest; now, for the buyers who actually pay the bills, the question is which model is smart enough for this job at this price. SAP’s answer, more than €2 billion of efficiency gains from choosing cheaper models where performance holds, gives that discipline a number, and other large buyers are likely to run the same calculation. The sceptical view is that the figure is a forecast, not a result, and that quality is harder to measure than cost; the optimistic view is that a market that prices models by the task is a market that can finally scale. Either way, the era of reaching for the most expensive model by default is ending, much as industrial AI is being judged on promised returns and professional platforms are being judged on completed work, not on the model behind it.
Background
SAP is a German enterprise software company whose products are used by many of the world’s largest companies, and it uses more than 70 AI models from providers including OpenAI, Anthropic and Google across its product range.
Source: Reuters report
0 Comments