SAP’s AI Acquisitions Point to a Post-LLM Enterprise Future

SAP’s planned acquisitions of Prior Labs and Dremio suggest the company is pursuing a different vision for enterprise AI than the large language model (LLM)-centric systems that have dominated the market since ChatGPT’s release.

SAP CTO Philipp HerzigIn May, SAP announced plans to acquire German AI startup Prior Labs and invest more than €1 billion over four years to expand the company into a frontier AI research lab focused on structured enterprise data. Around the same time, SAP also announced plans to acquire Dremio, a data lakehouse company focused on interoperability and AI-ready enterprise data infrastructure.

SAP CTO Philipp Herzig framed the company’s strategy around structured business data rather than general-purpose language models “The greatest untapped opportunity in enterprise AI wasn’t large language models; it was AI built for the structured data that runs the world’s businesses,” Herzig said in SAP’s announcement.

The acquisitions point to a broader shift in enterprise AI architecture. SAP appears to be betting that enterprise systems will increasingly rely on specialized foundation models optimized for ERP tables, financial records, supply chain systems, procurement workflows, and operational forecasting.

Why SAP Thinks LLMs Struggle With Enterprise Data

Large language models are optimized primarily for unstructured internet text. SAP’s systems operate on a different type of information entirely.

Enterprise software environments generate invoices, procurement tables, inventories, ledgers, supplier records, and transactional databases that require deterministic outputs, statistical reasoning, forecasting, and auditability. Those requirements differ significantly from the conversational tasks most LLMs are designed to handle.

SAP said current LLMs have only a “rudimentary understanding” of tables and statistics.

That limitation is central to SAP’s interest in tabular foundation models (TFMs), a category of AI systems designed specifically for structured datasets.

Prior Labs, founded by researchers connected to the Max Planck Institute, developed TabPFN, a transformer-based model built for tabular enterprise data. According to the company, its models can perform classification and prediction tasks directly on structured datasets without retraining for each task.

SAP had already been working on similar systems before the acquisition. The company previously developed SAP-RPT-1, a relational pretrained transformer designed for enterprise datasets such as ledgers, invoices, and inventories.

An independent evaluation paper found SAP-RPT-1 achieved between 91% and 96% of the performance of tuned gradient-boosted models without task-specific training. The paper also found the model performed particularly well in low-data environments.

The models are designed for operational tasks such as payment delay forecasting, supplier risk analysis, customer churn prediction, and inventory forecasting. Those systems require statistical prediction and deterministic outputs instead of conversational fluency.

SAP appears to view reasoning over structured business data as the larger technical challenge in enterprise AI deployment.

SAP Is Building a Different Enterprise AI Stack

The acquisitions also reveal a broader architectural strategy.

Many enterprise AI systems today are centered around copilots and external LLM APIs. SAP’s emerging stack separates data infrastructure, predictive modeling, orchestration, and conversational interfaces into different layers.

Dremio addresses the data layer. The company’s platform is designed to unify fragmented enterprise datasets across cloud environments and legacy systems. SAP said the acquisition would strengthen its Business Data Cloud platform and improve AI-ready interoperability.

Prior Labs addresses predictive modeling over enterprise datasets. Its models are designed to perform statistical prediction directly on tabular enterprise data rather than generating text responses alone.

SAP’s Joule assistant operates as the orchestration layer. The company has increasingly positioned Joule around “agentic AI” workflows that can retrieve enterprise information, automate processes, and execute tasks across SAP systems.

Herzig described the direction in a LinkedIn post following the acquisitions.

“Tables on the fly + predictions on the fly = decision intelligence like never before,” he wrote.

The structure differs from monolithic LLM-centric systems where a single conversational model handles retrieval, reasoning, and interaction simultaneously.

SAP’s architecture suggests enterprise AI systems may become increasingly specialized, with separate systems handling language, prediction, enterprise context, and workflow execution instead of relying on one general-purpose model.

The Bigger Risk for SAP Is Losing Control of Enterprise Intelligence

The acquisitions also address a broader competitive risk for enterprise software vendors.

As AI assistants become more capable, the interface layer increasingly shifts toward external AI systems such as copilots and enterprise agents. That creates a strategic problem for ERP vendors. If AI agents become the primary interface to enterprise systems, companies like SAP risk becoming underlying infrastructure rather than the central intelligence layer inside organizations.

Microsoft has embedded OpenAI models throughout Copilot and Dynamics. Salesforce continues expanding Einstein AI across CRM workflows. Oracle has also integrated generative AI capabilities across Fusion Cloud applications.

SAP’s response appears focused on controlling the structured reasoning layer underneath enterprise workflows.

Rather than competing directly with OpenAI as a frontier LLM developer, SAP is building systems optimized for enterprise operational data. SAP’s strategy indicates structured-data intelligence may become more strategically important than conversational interfaces inside large businesses.

The Dremio acquisition helps SAP unify fragmented enterprise data. Prior Labs provides models designed to reason over that data. Joule orchestrates workflows and automation across SAP environments.

Together, the acquisitions suggest SAP is building an enterprise AI stack where large language models remain useful but are no longer the central component.

The strategy points toward a broader shift in enterprise AI architecture, where specialized foundation models, workflow-native systems, and enterprise data orchestration become increasingly important alongside conversational AI.

SAP’s AI strategy suggests the next competition in enterprise software may center less on chatbot interfaces and more on who controls the data, reasoning, and execution layers underneath them.