SAP Joule Operating System

SAP Joule is more specialized than universal

The effectiveness of LLMs (large language models) lies in their universality. This means they contradict the zeitgeist of specialization. While the ERP software market has seen a proliferation of interfaces and data silos in recent decades due to an influx of specialist programs, AI systems such as ChatGPT and the associated LLMs are exactly the opposite.

LLMs are often prone to hallucinations—but who could blame a "polymath" for such lapses and digressions? In the tough day-to-day business of ERP, however, precision and specialization are paramount. This apparent contradiction can also be seen as an opportunity. In the future, AI agents could guide and discipline LLMs. These AI agents would then act as specialists based on universal LLMs.

An operating system for AI agents and LLMs

The concept of a meta-level for large language models as an AI operating system for ChatGPT and similar systems is not new. Jonas Andrulis, co-founder of the German AI start-up Aleph Alpha, had the idea of developing a management network for multiple LLMs.

He made a virtue out of necessity: as his own resources were insufficient for the world's best LLM, he came up with the idea of configuring the world's best LLMs into a meta-system with the help of an operating system. A meta-level for LLMs and potentially AI agents could evolve into an operating system for AI applications such as SAP Joule.

SAP BTP versus AI Operating System

SAP BTP (Business Technology Platform) offers an interesting approach to managing and orchestrating LLMs. The BTP's AI gene hub enables the use and administration of different LLMs, allowing users to select the appropriate AI system for their use case.

Further developing the BTP AI Hub into an AI operating system would be worthwhile and could simplify LLM orchestration and, in the future, SAP Joule.

AI orchestration instead of platform chaos

If the hype surrounding AI agents becomes a reality in ERP, SAP's customers will require an IT tool for managing AI agents. A platform will be necessary to consolidate all AI tasks so that AI agents do not interfere with each other or work against each other. This meta-platform for all AI agent platforms will essentially become an operating system for AI environments.

Too many AI agents for end-to-end processes in the ERP environment would interfere with each other. Without administration tools for AI agents, there is a risk of uncontrolled growth and chaos. One solution would be AI agent management. Ultimately, there will also be a meta-level for AI agent management. Those who have read the computer culture bible, Gödel, Escher, Bach by Douglas R. Hofstadter, know about the power of meta-levels (not to be confused with the IT company Meta). Often, problems cannot be solved directly. Transforming to an abstract meta-level can provide a new perspective on the problem and often reveal a solution.