Bridge Technology for AI Agents in SAP On-Premises

The DSAG Investment Report 2026 shows that many companies continue to run SAP on-premises. The transition to the cloud is proceeding much more slowly than SAP had planned. SAP’s AI strategy relies on Joule as a generative AI assistant. However, as a cloud service, Joule is still not a reality in many companies.

In a highly dynamic environment, users face the dilemma of whether to wait and see or to risk ending up at a dead end by implementing AI services within their tried-and-true systems. AI solutions still rarely use SAP data. However, the DSAG study also shows that 43 percent of companies have already implemented their first AI use cases. These primarily involve processing information from documents. The direction is set by the use of AI in business processes: Business AI applications have the potential to significantly increase productivity. However, the use of SAP data is essential for this.

 

AI-driven workflows

The euphoria surrounding AI in the SAP community has now given way to questions about its added value. While queries and image generation are already standard, the focus in the future will be on autonomous AI-agent-based workflows. At the same time, the view has taken hold in the SAP community that AI applications can only be used in the cloud. In fact, however, SAP users do not have to wait for S/4 HANA or the cloud. The Milliarum AI Construction Kit bridges this gap. It enables AI applications and agents to be implemented on-premises using ABAP-based technology and to leverage SAP data.

SAP customers and SAP partners are provided with a tool to make their solutions AI-ready. AI functions run outside the core ERP system and securely access SAP data via defined interfaces, subject to appropriate authorizations. Business objects and process data are utilized via standardized APIs, allowing AI applications to operate independently of the ERP system. They are explicitly designed to comply with SAP’s restrictive API policy and can also be seamlessly integrated into Joule, migrated to the cloud, or used in a hybrid environment via an OData service.

 

Cost-Benefit Analysis Slows Down AI

The biggest hurdles on the path to AI implementation lie less in the architecture and more in scaling to complex business processes. The token consumption of large language models can quickly lead to high costs while delivering limited added value. SAP processes are complex and data-intensive. Milliarum offers the bridge technology needed here: By extracting Customizing data and utilizing standard SAP functions, even complex, chained business use cases can be mapped. Batch functions enable the efficient processing of large volumes of data and lay the groundwork for AI-driven process automation. The open Model Context Protocol enables integration with other AI platforms, such as ChatGPT or Claude AI. The tool offers over 1,000 predefined functions for core SAP modules such as FI, SD, MM, CO, PS, and CS, which can be used directly to implement automated use cases. This gives even users without in-depth SAP knowledge quick and easy access to SAP data and processes—for example, in call centers or customer service when handling inquiries about delivery orders or prices. Even complex workflows can be automated, such as generating an invoice as a PDF and then automatically sending it via email.

In this way, the software lays the foundation for AI-based agent systems that will increasingly execute processes autonomously in the future. A key driver of efficiency lies in reducing token consumption: Companies can freely choose the appropriate LLM provider or operate LLMs themselves and deploy the most cost-effective model depending on the use case. At the same time, on-demand function calls ensure that only the data relevant to the respective use case is passed to the model. In this way, companies can generate tangible added value even within existing on-premises environments while simultaneously laying the technological foundation for the next generation of AI-powered agent systems and agent-to-agent communication.