Executive Summary
- According to the article, several independent benchmarks from 2026 show that LLMs are markedly weaker at table-reasoning than at pure language tasks.
- Tabular foundation models are intended to detect patterns across columns, tables and business processes, enabling predictions from structured enterprise data.
- The article states that SAP has made the model SAP-RPT-1.6 available to customers and that Prior Labs was acquired in July; Prior Labs’ TabPFN-3 is mentioned in the article as a capable open-source model for tabular predictions.
- The article names concrete use cases such as detection of payment defaults, demand forecasting and risk detection in supply chains, and estimates a macroeconomic potential (according to the article: roughly €440 billion additional gross value in Germany by 2034).
What happened?
Facts: In the published piece SAP argues that the first wave of generative AI mainly increased individual productivity, while significant profit increases were realized by only a few companies (according to the article: a McKinsey-reported figure of 80 percent user benefit vs. six percent of companies with materially higher operating profit). The article states that several independent 2026 benchmarks reportedly show that even powerful LLMs perform worse on table- and relationship-logic tasks than on language tasks.
Facts: As a technical response the article describes so-called tabular foundation models (“tabular AI”) that are meant to recognise patterns across columns, rows, tables and temporal relationships. Specifically, the article says SAP has made the SAP-RPT-1.6 model available to customers and that Prior Labs was acquired in July. The article further states that Prior Labs’ TabPFN-3 is one of the capable open-source models for tabular prediction. It also notes a trend toward additional foundation-model types, for example for time series.
Why does it matter?
Context: The article links the relevance of tabular models to the structure of the data many mid-sized companies use: invoices, orders, inventory levels, payment behaviour and measurement series are typically organised in tables and depend on relational links. According to the article, the value lies in such models’ ability to capture relationships across different data sources and thus signal risks or opportunities earlier than isolated, domain-specific models.
Context: The article also raises a scaling argument: instead of many individual, specially developed ML models, tabular foundation models should be more universally applicable and thus reduce development and operational effort. This perspective is particularly relevant for companies that already run or plan multiple isolated analytics projects.
What does this mean for mid-sized companies?
Context: For small and mid-sized companies this can mean, according to the article, that investments in data preparation, data quality and integration pay off more if they use models that recognise relationships across data silos. Concrete use cases mentioned in the article include early detection of payment defaults, improved demand and inventory forecasts, and detection of supply-chain risks.
Context: The article is product- and vendor-specific: SAP presents SAP-RPT-1.6 as an example and points to Prior Labs and TabPFN-3. For users this means a technological option is available that is intended to integrate into existing SAP environments (according to the article). At the same time, it should be noted that the benchmarks and technical assessments presented in the article are interpreted from the provider’s perspective; independent validation is not shown in detail.
Related: AI Readiness Check
Recommendation from Future AI Solutions
Short-term recommendation: Identify which business questions in your company are based on structured data (e.g., payment flows, orders, inventory, measurement data). Prioritise use cases with clear economic levers (liquidity, ability to fulfil orders, avoidance of losses). Start small, time-boxed pilots to test tabular models with your own data and compare them to existing ML approaches.
Mid-term recommendation: If you use SAP systems, clarify available integration paths to SAP-RPT-1.6 and to components that use Prior Labs technology (TabPFN-3), according to the article. Validate vendor claims with proof-of-value projects using your real data. At the same time, invest in data quality, metadata and governance, because the performance of tabular models strongly depends on consistent, linked datasets.
Operational recommendation: Combine, as the article suggests, the strengths of LLMs (language, interaction, knowledge work) with tabular models (predictions from structured data). Establish monitoring and explainability processes for tabular predictions to build trust among users and managers. Also review legal and data-protection requirements when using company-specific data in foundation-model workflows.
Finally: The article describes a possible technological shift toward tabular foundation models as a complement to LLMs. Decision-makers should use this depiction as a prompt to carry out their own tests and feasibility checks rather than accepting vendor claims uncritically.