What actually happened
SAP CFO Dominik Asam told Reuters on October 28 that the company has replaced some expensive AI models in its products with lower-cost alternatives when quality remains strong. SAP uses more than 70 models from suppliers including OpenAI, Anthropic and Google, choosing increasingly by the requirements of each task rather than defaulting to the most powerful option.
Reuters reported that the approach will contribute to more than €2 billion in wider AI-driven efficiency gains. The figure is not a claim that model switching alone will save €2 billion. SAP also expects its total AI spending to rise as more work uses the technology.
Why a small team cares
Take a small SaaS product handling thousands of support tickets each month. It may need to classify a ticket, extract a customer identifier, draft a response and decide whether a human should intervene. Those jobs differ in complexity. Paying for the strongest model to perform straightforward classification can consume margin without improving the result.
A cheaper model can take the routine step while complex or uncertain cases go to a more capable model. That only works if you can recognize an unacceptable answer. A small benchmark should include noisy inputs, ambiguous requests and examples where the expected answer is known.
The metric is cost per successful task, not the published price per million tokens. A low-cost model that needs three retries, produces incorrect labels or creates human rework may cost more in practice. Routing also adds monitoring and failure handling, which must earn their place in a small codebase.
Hype vs useful
SAP's scale does not translate into a guaranteed percentage saving for a small application. The company has extensive model choice and operational data. Your workload may be too small or unpredictable to justify maintaining multiple providers.
But the underlying discipline is sound. Define a quality threshold, compare candidates on real tasks, and send only suitable work to a cheaper model. Include latency, regional or contractual restrictions, fallback costs and correction time. If the numbers fail, keep the existing model. That is still a useful test result.
