SAP Expands Tabular AI Lineup With RPT-2 and New Forecasting and Decision Models

For nearly a year, SAP has relied on a single general-purpose AI model to make predictions from business data rows and columns. At its annual TechEd conference this week, the company expanded that lineup with SAP-RPT-2, a next-generation model, and two specialized models for forecasting and decision-making.

First introduced in November 2025, SAP’s Relational Pre-Trained Transformer (RPT) model is designed for relational and structured business data. RPT-1 was available in several variations: SAP-RPT-1-small, described as “super-fast”; SAP-RPT-1-large, described as “extra-strong”; and SAP-RPT-1-OSS, the open-source model available on GitHub and Hugging Face. Unlike large language models, which learn from text, RPT-1 learns from rows and columns stored in CRMs, enterprise resource planning systems, and financial databases—areas where SAP is a major player. Customers can use the RPT model for predictive analytics, financial modeling, customer analytics, and scenario simulations.

SAP has upgraded the model since then. At its Sapphire conference in May, the company unveiled SAP-RPT-1.5, which improved accuracy on complex data and could explain predictions by pointing to comparable examples. SAP also previewed Tabular Orchestration, a managed workflow featuring retrieval-augmented prediction that pulls records from a customer’s data sets and feeds them to the model as context. And last month, SAP released SAP-RPT-1.6 to handle larger data sets and made Tabular Orchestration available to customers.

SAP-RPT-2 builds on that foundation. Previous RPT versions handled two kinds of tasks: regression, which predicts a number, such as how many days late a payment will arrive, and classification, which sorts a record into a category, such as which shipping point a sales order should use. SAP-RPT-2 adds anomaly and outlier detection, flagging records that deviate from usual patterns or transactions. 

RPT-2 also creates what SAP calls structured data embeddings, converting a customer, order, or invoice into a numerical fingerprint that lets software and AI agents compare and spot similar records without needing a custom rule written. In addition, the new model can read data spread across related tables, so teams don’t have to merge everything into a single table before the model analyzes it. Though announced this week, SAP-RPT-2 won’t be generally available until Q1 2027.

SAP also introduced two specialized models. SAP-RPT-TS forecasts business metrics that change over time, such as product demand within SAP’s Integrated Business Planning software, without requiring customers having to build their own machine learning models. Customers can use RPT-TS in SAP HANA Cloud or the SAP AI Core foundation, available now.

SAP-RPT-Decisions automates decisions inside apps and agent workflows. Customers can use it to decide whether to approve a customer order, route a supply chain exception, or choose the right approval path for a purchase requisition. SAP expects it to be generally available through SAP AI Core in Q4 2026.

RPT-Decisions unveiling comes as more companies embrace decision models, with TypeSafe’s Jev model putting them in the spotlight. OpenAI, Cloudflare, Perplexity, Amazon Web Services, Snowflake, Microsoft, and Together AI are some companies that have released or announced decision models. But while most of these models evaluate text, SAP is applying the idea to structured data in ERP systems, where many of a company’s most consequential decisions already live.

But that’s not all: To show SAP-RPT’s appeal, the company revealed it has deployed the model on the HPE Spaceborne Computer-2—an advanced edge computing platform designed for use in space—aboard the International Space Station. However, SAP didn’t disclose what the model will predict, whose data it uses, or whether anyone aboard the space station interacts with it.