Business AI with little training effort: SAP announces SAP-RPT-1

At SAP TechEd, which is taking place both in Berlin and virtually this year, the software company has introduced three novel AI models for businesses. The SAP-RPT-1 family consists of two proprietary models and one open-source model, which is available on GitHub and Hugging Face. Unlike language models like ChatGPT, these are based not on General-Purpose Transformers, but on Relational-Pretrained Transformers.

While the former are optimized for linear text sequences, the latter are particularly adept at handling tabular data. For this reason, they are said to be particularly suitable for business data. A research group at SAP had already presented this approach at the NeurIPS conference in June.

Unlike previous machine learning models, which had to be trained separately for each business scenario, the new model family also uses a method called "In-Context-Learning" (ICL). Dr. Philipp Herzig, SAP CTO, compares it in the blog post to a consultant who only needs to be presented with a few sample data points for correct predictions. The data basis is the "Tremendous TabLib Trawl" collection, or T4 for short, which was published last year. Three million tables from the dataset, which comprises four million tables, were used for finetuning.

Error-free and significantly more efficient than LLMs?

In its announcement, SAP highlights the new model family for its efficiency and, compared to large language models, low error rate. The Walldorf-based company used test scenarios for evaluation that are intended to correspond to customer challenges. The new models are said to particularly stand out in terms of efficiency. Unlike LLMs, they would use one hundred thousand times fewer Floating-Point Operations (FLOPs) for the same task and still be 50 times faster.

Interested parties can try out various scenarios online. The company provides an environment where either predefined examples can be tried out or predictions for one's own datasets can be made. File uploads in the portal are limited to CSV files with a size of five megabytes, or tables with 2073 rows and 50 columns. The models are expected to be available to end customers in the current quarter via the Generative AI Hub of SAP AI Foundation.

New AI Agents and Extended AI Capabilities

The company also announced a series of new AI agents for SAP Joule. In addition to some supply chain and SAP Ariba agents, these include further helpers for SAP Concur. For example, there is now a service that can examine receipts for authenticity. It was only recently announced that companies are increasingly finding fake receipts in expense reports.

Other cloud applications are also receiving updates. The Agent Builder, with which end customers can develop their own AI agents, was originally announced at SAP Sapphire in early 2025. At TechEd, the company now announced general availability by the end of this year. The Agent-to-Agent Protocol is also to be supported then. New LLMs for ABAP development are also to be available in the AI Foundation by the end of the quarter.

Local model context protocol servers are already available, enabling access to SAP-specific frameworks through code assistants like Cursor or Windsurf. Interested parties can find a complete overview of all announcements in SAP's "Innovation Guide".

(mack)