THE AI INTERVIEW
“ Each large language model is like a human. They’ re trained to have human-like characteristics”
WALTER SUN, SVP, GLOBAL HEAD OF AI, SAP interoperability, which was a higher layer of interoperability than within the sub-agent layer of technology.”
The partnership extends beyond Google to include Microsoft and AWS. Walter describes the approach as creating“ orchestration and connectivity at a higher level” rather than traditional API integrations, allowing SAP’ s agents to access capabilities from Google’ s Agentspace platform and vice versa.
SAP has built an agent builder capability within Joule that creates agents across the company’ s business applications. The platform includes pre-built agents for common processes, while Joule Studio enables custom agent development for specific customer requirements.
Walter suggests this approach addresses customer concerns about vendor lock-in.“ By having agent to agent as an open standard and having other partners join, we actually at this conference also mentioned that Microsoft and AWS are working with us as well. We can actually make it easier for our customers to have a wider ecosystem without anything being closed.”
Addressing AI performance challenges SAP supports more than 30 large language models through its Generative AI Hub, creating an optimisation challenge that the company addresses through partnerships with Not Diamond and AI search startup Perplexity. The company has also announced partnerships with Palantir Technologies for data connectivity and Adobe for demand forecasting solutions. The Not Diamond collaboration focuses on prompt optimisation across different AI models.
“ Each large language model is like a human. They’ re trained to have human-like characteristics, and by the same token, each LLM has its own preference of how it’ s spoken to,” Walter explains.
The technical issue stems from how different AI models respond to identical prompts. Instructions optimised for OpenAI’ s models may produce suboptimal results when used with Google’ s or Anthropic’ s systems, not because of model quality differences but due to different training approaches.
“ If you have a prompt that’ s optimised for one language model and then you say, hey, I have a new language model,
aimagazine. com 47