SAP was founded in 1972.
“I mean, something has always prevailed over the years, because even when SAP was founded, right in 1972, right, long time ago.”
Person
Philipp Herzig
1 episodes · 23 quoted claims
Herzig says SAP's founders were still working at IBM in the 1970s when they went customer to customer.
“actually in the 70s, when the founders of SAP worked still at IBM, what did they do? They went to each customer, right?”
Herzig says the SAP founders repeatedly implemented a finance system for individual customers, which led them to the idea of standard software.
“And they implemented the finance system again and again and again and again.”
Herzig says the notion of standard enterprise software was born from SAP founders' realisation that re-implementing finance per customer does not scale.
“And this was the idea where standard, the the notion of a standard software was born, essentially right.”
Herzig says everybody at SAP now uses agentic coding.
“We are in the meantime all in on AI, right? Uh, so I mean, everybody in the company is using agentic coding, right?”
Herzig says SAP Concur now has live travel booking agents and expense agents.
“with Concur, for example, right, where now our travel booking agents or expense agents are are alive.”
Herzig says AI is changing enterprise software on three levels: the user interface, business processes and the data layer.
“And then it's happening on three levels. It happens, of course, on the UI side, right?”
Herzig cites an analyst adoption curve that separates an AI innovation race from an AI outcome race and says the gap is widening.
“There's like, there's this Gardner curve, right? Where say, like, there's this AI innovation race, and then there's this AI outcome race, right?”
Herzig says an enterprise assistant must tailor travel policy answers to employee master data such as country, so a US employee and a German employee get different answers.
“of course, for a travel policy, for example, of course, you expect a very different answer that me as a German employee, right, would would get.”
Herzig says MCP servers work easily for ten APIs, but with hundreds of APIs the model's context overflows, and SAP has far more.
“for ten API is not an issue, hundred, you get already, uh, context, blowed and all these these challenges”
Herzig says agentic coding works well because outcomes can be verified, for example by checking whether a program compiles and passes unit tests.
“of course, you can verify the outcome, right? Uh, you can either say, Hey, is the program compiling?”
Herzig calls context that lives in people's heads or chat channels rather than in business systems tribal knowledge.
“I always call this the tribal knowledge, the stuff that is not in the system, stored somewhere, that just lives in people's heads.”
Herzig says SAP now calls the recording of decision traces from agents and users agent mining, where it formerly used the term process mining.
“In the past, we called this process mining, now we call it agent mining.”
Herzig names XGBoost and AutoGluon as classical machine learning tools used for business predictions that do not scale well across many use cases.
“You use XGBoost or AutoGluon and and many of these, uh, AutoML approaches, right?”
Herzig says a pharmaceutical customer operating in 90 countries would need 180 separately trained models to predict payment delays and non-payment.
“They are running in 90 countries around the world, and they need these two models. So you end up with 180 models.”
Herzig says SAP's RPT-1 stands for relational pretrained transformer and is a transformer-based model for structured table data.
“we call this RPT1, so rapid ones and for relational pre trained transformers.”
Herzig describes the SAP knowledge graph as the link between natural language questions and structured data in SAP systems.
“the SAP knowledge graph that we've built, which is kind of the glue between natural language and the structured data in the system.”
Herzig cited a recent LiteLLM vulnerability that steals keys and credentials as an example of why enterprises cannot deploy fast-moving AI tools unvetted.
“Like with this vulnerability that still all of a sudden steals all your keys and credentials and so on and so forth.”
Herzig says most SAP software is licensed per seat, with exceptions such as Concur and Fieldglass.
“for the most part, SAP software is seat based, licensed, uh, uh, uh, today, with a few exceptions, like a Concur or a Field Glass, for example”
Herzig says SAP's current AI pricing is a hybrid of consumption-based charges and an element of per-seat licensing.
“it's consumptive, but it still has a certain element of seats in there”
Herzig says enterprise customers still demand cost predictability, which is why SAP keeps a hybrid pricing model.
“Knowing that they demand still a lot of predictability in the enterprise space in order to cost control.”
Herzig says improving a customer's demand forecast accuracy by three or four points can be worth millions of dollars.
“if you can improve the accuracy of the demand forecast by three, four, right, that is huge for them, right? This is this is multi million value”
Herzig says SAP's quantum computing research will be hardware agnostic, much as SAP software has always run on different kinds of computers.
“we will be very hardware agnostic in that sense, right? Uh, like, like, SAP always, always ran on different computers”