SAP has replaced some of the most expensive AI models in its products with cheaper ones where performance remained strong, Chief Financial Officer Dominik Asam told Reuters. SAP says it aims for “healthy margins as scale grows.”
Customers are also asking where those models come from. Asam described a question SAP regularly hears: whether data used by a model could “end up in places where it shouldn’t end up.”
SAP customers who build their own agents can choose the models themselves. At SAP TechEd in Berlin, SAP announced controls in Joule Studio, its agent-building environment, that show developers and administrators agent behavior, model usage, and spend.
Replacing Costly Models Where Performance Holds
SAP provides access to more than 70 AI models and avoids committing to a single one, CEO Christian Klein has said, summing up the approach as “We give our customers choice.” Providers include OpenAI, Anthropic, and Google.
Asam said SAP increasingly picks the cheapest model capable of performing a task to the required standard. Overall AI spending is still expected to rise as adoption grows. In July, Asam told reporters after SAP’s second-quarter results that expecting AI to fix messy legacy data silos comes with “extremely high token costs.”
Reuters reports that SAP charges customers for completed tasks instead of for the computing used to generate responses, so lower operating costs can improve its profitability, and that SAP expects the model changes to help generate more than €2 billion (about $2.2 billion) in AI-driven efficiency gains.
Customers Ask Where AI Models Come From
Export controls and Europe’s sovereignty debate now shape decisions once made on price and performance alone, Reuters reports. Asam said requirements vary: a US customer may not accept components developed with Chinese models, while other back-office work, including SAP’s own, can run on open-weight software regardless of its origin.
Government agencies and companies in regulated industries increasingly exclude models based on origin, and European customers increasingly demand local solutions, according to Asam. A large nation’s defense department will weigh the question differently from a small shop, he added. He pointed to a European debate over a possible kill switch in US cloud infrastructure as another source of demand for sovereign alternatives.
Asam said SAP prefers access to models from every region, including Europe and Canada. The aim is a flexible portfolio for clients such as the governments and defense contractors it has recently signed. He added that working out each client’s precise technology governance requirements across several models has become a craft in itself.
What Cost and Origin Concerns Mean Together
Cost and origin both come down to which model sits behind a given SAP feature. SAP chooses the model inside the features it delivers, so a customer’s rule on model origin has to cover those selections as well as the models it picks itself in Joule Studio.
SAP’s TechEd pitch put context ahead of the model. SAP senior vice president Atul Deo told SAPinsider that “context is more important today than anything else,” arguing that foundation models are becoming commoditized. If models are commoditized, as Deo argues, the choice among them comes down to cost and fit.
SAP is rolling out Joule Studio in phases. Hosting and governance items come later: SAP Sovereign Cloud support for SAP AI Core, Joule, and SAP Datasphere in the UK, Australia, and Canada is planned for the fourth quarter of 2026.
The SAP Enterprise AI Gateway, which SAP says will consolidate policy enforcement, usage tracking, and auditing of AI traffic, is planned for the first quarter of 2027.
What This Means for ERP Insiders
SAP is managing AI costs at the model level. It picks the cheapest model that meets the standard for each task. Its TechEd message puts business context ahead of the model itself.
Government and regulated buyers now weigh which models vendors use. Export controls and Europe’s sovereignty debate now shape decisions once made on price and performance. SAP’s stated answer is access to models from every region.
Governance now has to cover choices the vendor makes. SAP’s planned sovereign hosting addresses where AI services run. The planned SAP Enterprise AI Gateway consolidates policy enforcement, usage tracking, and auditing of AI traffic.
