독서리서치레이더투자 프레임워크
로그인 / 회원가입
← 읽기로 돌아가기
뉴스주제 · SAP表格型AI중대2 분 소요DisconfirmAI 편집부

German Software Giant SAP Says LLMs Struggle with Complex Enterprise Predictions, Advocates for 'Table AI' to Fill the Gap

SAP highlights limitations of LLMs in enterprise data analysis and forecasting, proposing 'Table AI' for structured data combined with LLMs for natural language understanding to improve ROI.

2026-10-20 08:22 GMT+8

German software company SAP recently stated via its official blog that relying solely on Large Language Models (LLMs) like ChatGPT is insufficient for solving complex business problems, particularly in data analysis and future predictions involving numbers and tables. SAP proposed a technology called 'Table AI' (Table Foundation Model) as a complementary solution aimed at enhancing prediction accuracy in areas such as supply chain and finance.

SAP explained that while LLMs excel in document writing, information retrieval, and programming, enterprises heavily rely on structured data like sales figures, inventory levels, and order records. Accurate analysis of this data requires not just identifying numbers but also understanding the complex relationships between them. In contrast, LLMs often produce inconsistent answers when table formats or row orders change. This limitation was also confirmed in a paper presented by researchers from IBM Research and Stanford University at an ACL workshop in July.

As a representative case study, SAP introduced 'SAP-RPT-1.6', recently provided to customers. This AI model is specifically designed to analyze numbers and tables to predict future events. For example, it can forecast customer payment timeliness, product popularity trends, or stockout times based on past transaction records and sales volumes. Traditionally, companies had to develop separate AI models for each prediction task, resulting in high costs and maintenance difficulties. SAP's proposed 'Table Foundation Model' allows a single general-purpose model to handle multiple prediction tasks, thereby reducing development and operational costs.

To strengthen its capabilities in this area, SAP acquired Pryor Labs, a company specializing in Table AI, in July. It is combining Pryor Labs' TabPFN model with SAP's own AI models. SAP emphasized that while generative AI has improved individual employee productivity, its impact on significantly increasing corporate profits remains limited. According to McKinsey survey data cited by SAP, 80% of AI users reported increased personal efficiency, but only 6% of companies believed AI substantially boosted operating profits.

로그인 / 회원가입
독서리서치레이더투자 프레임워크

SAP recommends a hybrid architecture: using LLMs to understand human queries and explain information, while employing Table AI to analyze internal structured data for predictions. SAP stated that the key is not choosing one technology over the other, but combining both to leverage their respective strengths, addressing the real-world challenge of high investment and low output in enterprise AI adoption.

출처

  • zdnet.co.kr↗
주제

SAP表格型AI

德国软件巨头SAP正试图重塑企业级AI的技术范式,主张在大型语言模型(LLM)之外,引入专门处理结构化数据的“表格型AI”以解决商业预测难题。【背景】尽管LLM在文本生成上表现优异,但SAP指出其在分析财务、库存等数字关系时存在局限,导致企…

5 단계1 건업데이트됨 2026-10-21 01:51

현재까지의 과정

  1. 2025-11-04SAP首席技术官Philipp Herzig发布博客,首次提出利用“上下文学习”处理关系型数据的新范式,宣布SAP-RPT-1模型概念。community.sap.com ↗
  2. 2026-02-01微软研究员Amit Lal发表预印本论文,对SAP-RPT-1进行独立基准测试,指出其在小样本分类中具竞争力,但在回归任务上不及XGBoost等传统模型。arxiv.org ↗
  3. 2026-07-01SAP完成对德国表格型AI初创公司Prior Labs的收购,承诺投入超过10亿欧元将其打造为全球领先的结构性数据前沿AI实验室。news.sap.com ↗

아직 해결되지 않은 문제

  • SAP-RPT-1与TabPFN-3.5在真实生产环境中的长期稳定性与维护成本是否真的低于传统机器学习流水线?
  • 微软指出的表格型AI在大规模回归任务上的性能劣势,是否会阻碍其在核心财务预测场景的大规模普及?
  • SAP提出的LLM与表格型AI混合架构,在实际部署中如何平衡两者的调用延迟与计算成本?
전체 주제 읽기 →
리서치 계속하기 →