There is a lot of hype about agentic AI, says SAP CTO Philipp Herzig

SAP Labs India ramps up AI innovation in India with new Bengaluru campus; key work on joule, generative AI, and enterprise solutions now led by 15,000-strong local team

SAP Labs, the research and development (R&D) arm of German software maker SAP, opened its second campus in Bengaluru earlier this month, which will house over 15,000 professionals. The firm says that its centres in India were at par with its headquarters with a mix of product teams, sales, customer experience and marketing. Philipp Herzig, chief technology officer and chief AI officer of SAP SE, in conversation with Avik Das, talks about the critical role of India, future of AI, and why agentic AI is currently more hype than action. Edited excerpts: 

 

How do you demarcate which operations are done in India and what is done in your biggest research laboratory in Waldorf, Germany?

 

Every team is a little bit differently set up. Here, of course, we have an amazing talent pool of both engineering but also product and industry knowhow. It is usually a mixture of various factors because for AI, we are distributed over four locations in Palo Alto, Germany, India, and Singapore. We try to keep certain components to align with the time zone which means a closer proximity here of our AI developers also with the AI colleagues that we have in Singapore and Germany, but then we also try to decouple. So then we give ownership to the various teams and try to distribute it based on where it also fits from a skill perspective, investment, capacity, and talent.

 

Can you also elaborate on what type of AI work is done at SAP Labs in India?

 

We have dedicated data scientists and developers with AI skills, and what we are doing here specifically is a big part of Joule. So our generative AI co-pilot is developed in India. Our main strategy is to provide AI and generative AI (Gen AI) experiences which are out of the box in areas such as HR, finance and supply chain. For example, last year, we had the Joule content program that was done here in India, where a majority of the functionality that customers can experience today within Joule have been contributed to a large extent from the India Labs.

 

When it comes to different areas of AI development, what’s the most promising right now? Is it agentic capabilities?

Agentic AI and agents are on the peak of inflated expectations and there is a lot of hype around it right now. We are following a very consistent strategy because we really are reinventing the entire stack with AI, and agents are kind of the next level of both productivity and engineering. So every layer builds upon each other and if you haven’t solved for the lower layers, you have no chance to build out the agentic layer.

 

While we will have agents, we will also have as our AI foundation, a new kind of operating system that is AI native which will be the next big step forward. Many think of AI as a computing system that can generate some text but what it really does is having a new computing paradigm. Instead of instructions on the CPU that get processed one by one, we now have tokens which is intelligence processed on graphical processing units (GPU). That introduces a new kind of computing paradigm which will be more AI native systems. I believe in a few years, we will see a totally new system that is from the ground up, redesigned with the help of AI.

 

Why do you say that a lot of agentic conversations are inflated right now?

 

It’s always a little bit the same when a new technology comes around. Everybody can create a little proof of concept (PoC) which can be very impressive but if you want to scale that and make sure it works for various use cases in a reliable way, then it becomes hard work. 

 

So there is a curve where you get the first 80 per cent done with 20 per cent of the effort. The remaining 20 per cent actually takes 80 per cent time. That is what happens in the computing industry most of the time. All see potential with the first good results but those are not good enough. AI adoption in the enterprise is therefore at a low scale because PoCs are easy but shipping it to production and adding value is the hard part.

 

Do you see this trend changing for at least agentic AI over the next 24 months?

 

Absolutely as it will get more mature and better. It was the same with Gen AI two years ago. Initially we had 11 use cases and by the end of last year there were 140. Now, we have more than 280 scenarios and will have more than 400 by the end of this year.