The biggest engineering challenge at scale is not the AI itself, but teaching the AI to do the right thing at scale — a POC on 10 documents is easy, but at 1000 documents or 20,000 APIs the engineering challenge grows enormously.

(00:00:06) Hey listeners, welcome back to Know Priors. (00:00:08) Today I'm here with Philip Herzig, the CTO of SAP, the enterprise juggernaut. (00:00:13) We talk about their AI strategy, why SAP has endured and thrived through several technology transitions. (00:00:20) Why entrepreneurs are underestimating the challenges of scale. (00:00:24) Why AI is a business model transition, not just a technology transition. (00:00:28) Why he thinks that LLMs are not enough for predictive analytics, and even about the traveling salesman problem in the real world and the Strait of Hermuz. (00:00:37) Welcome, Philip. (00:00:38) Philip, thanks so much for being with us. (00:00:40) Yeah, it's a pleasure to be here. (00:00:41) Thank you. (00:00:41) Everybody knows the name SAP, but I do think that for (00:00:47) lots of engineers or people who aren't close to the system in a larger enterprise, they don't really know like the breadth and function of the platform. (00:00:55) Like, can you just describe what you guys do for customers? (00:00:59) Oh, absolutely. (00:01:00) I mean, look, SAP is the market leader, right, in enterprise of software applications and platforms, right? (00:01:07) So 400,000 enterprise customers and usually I just running their finance, HR and, you know, supply chain, manufacturing, execution, logistics, warehouse management, and then, of course, everything on the customer side, sales, services, commerce, procurement, you name it, right? (00:01:26) End to end, right? (00:01:27) Like SAP, we always say we have the broadest portfolio in terms of end to end, running the business end to end. (00:01:33) This is where SAP started with, right? (00:01:35) Giving real time insight. (00:01:37) And usually I really describe this as it's not just software in itself, it's kind of the operating system, right? (00:01:43) Of a company essentially, in order to get, you know, from everything from order to cash or source to pay, right? (00:01:52) End to end managed for companies around the end. (00:01:55) entire globe. (00:01:57) I definitely want to talk about AI, LLM, some of those stuff that you guys are doing internally and then around predictive models as well. (00:02:07) But just because the macro backdrop is on everyone's mind, both from a technology and an economic perspective. (00:02:14) Oh, sure. (00:02:14) I want to talk about SAP's position in the market a little bit. (00:02:18) SAP has stood the test of time through multiple technology and market cycles. (00:02:24) I, as an early stage venture capitalist, I'm kind of on the other side of this where the narrative is like, well, when you have internet and cloud and mobile and AI and social, like you have an opportunity for new players. (00:02:42) What do you like, SAP, you know, even today is the, I believe the largest like market cap enterprise software vendor versus (00:02:54) sort of the last generation of the new guard, like the sales forces of the world. (00:02:59) How has that happened? (00:03:00) Like how did you do it and what makes it so durable? (00:03:05) what makes it so durable, right, at the end of the day, I mean, if you think about this, and it's happening a little bit the same way also when we talk about the SARS estate narrative or the SARS apocalypse, right? (00:03:14) I mean, anyway, I have the feeling like in this market last year, AI was in a big bubble and everybody was kind of saying, no, it's not. (00:03:22) And this year, SARS estate and so on and so forth. (00:03:26) Look, the reality is (00:03:28) Now, of course, with the costs of building being so low, right, with specifically agentic coding and all these latest powerful models, I mean, something has always prevailed over the years because even when SAP was founded, right, in 1972, right, a long time ago, I mean, why was it started? (00:03:46) Because actually in the 70s, when the founders of SAP worked still at IBM, what did they do? (00:03:52) They went to each customer, right, and they implemented the finance system again and again and again and again. (00:03:58) And then he said like, hey, this makes no sense, right? (00:04:01) Because the economics, it doesn't scale, right? (00:04:04) Because of course you can do this, right? (00:04:06) But you can only add so much value, right? (00:04:08) In any given time. (00:04:09) And by the way, we are basically programming the system very similar. (00:04:14) Of course, there's always a little bit that is specific then to the customer. (00:04:18) And this was the idea where standard, the notion of a standard software was born essentially, right? (00:04:22) And then of course, that stood the test of time, right? (00:04:26) Because there is simply things and companies that need to get managed, right? (00:04:30) End to end. (00:04:31) And that also has transformed throughout the years. (00:04:33) You've mentioned that, right? (00:04:34) First from the mainframe to client server, right? (00:04:37) Then to the internet, then mobile, and now of course, AI. (00:04:40) So of course, the software has changed and evolved all along with these technologies. (00:04:45) What hasn't changed is what customers are seeking for, which is (00:04:48) Outcomes, right? (00:04:49) Outcomes and return on their investment in order to get the things done, right? (00:04:54) And of course, now AI is an amazing technology that, again, helps to get more things done in the enterprise, right? (00:05:03) And then that is actually what SAP is standing for, right? (00:05:06) And so what we are really doing is in, given, of course, also the breadth of the portfolio and the customers, (00:05:12) is of course to help customers to achieve more by deeply embedding AI, AI agents, and of course transforming now the user interface and so on and so forth to help them get more done in whichever industry that they are working in. (00:05:29) And we believe that still will continue, right? (00:05:33) Because this is exactly what we're also seeing right now with, of course, there's still, of course, there's tremendous progress, (00:05:40) But we also see that the AI adoption in the enterprise is still not where we want to see it, right? (00:05:45) 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? (00:05:51) And the gap almost... (00:05:53) increases, right, versus getting narrow. (00:05:56) And that is what we are really focused on, right, that customers get the outcome from AI to achieve more, given, of course, the foundation we have. (00:06:06) But simultaneously, of course, the system, we are kind of re-engineering the entire system, right, with the help of AI in a totally new way. (00:06:15) Yeah. (00:06:17) You now, as CTO of SAP, have like a (00:06:22) very broad purview, and it also includes the AI strategy piece of it, internal and for your customers. (00:06:30) What do you think of as your own top priorities for the organization, and where is SAP on this re-engineering or re-imagination journey? (00:06:39) No, look, I mean, we are, in the meantime, all in on AI, right? (00:06:45) So, I mean, everybody in the company is using agentic coding, right? (00:06:48) Because that's, of course, an amazing productivity boost, right? (00:06:54) That our developers have, no matter in which programming language they are building the software for the customers. (00:07:00) But of course, it's also really (00:07:03) again, focusing on customer outcomes, right? (00:07:05) And we've seen this, for example, early in the early days now with consulting, for example, right? (00:07:10) We built this thing called Joule for Consulting, which is phenomenal because it was one of our fastest growing AI products because what this actually helps is to build the (00:07:20) to help the consultants, right, in an SAP project or in a complex landscape if they're, I mean, again, right, we are serving some of the largest customers. (00:07:28) They have a lot of heritage, they have a lot of complex landscape to help them actually to move into the cloud to adopt the latest AI capabilities and so on. (00:07:36) And with true for consultant, they can reduce like 30% of their efforts, right, to get to the outcome faster, which of course then directly reduces the cost, not just the time, but also the costs that are necessary in order to get to (00:07:50) to get to the latest software. (00:07:52) And we've seen this, of course, with Conqueror, for example, where now our travel booking agents, our expense agents are alive. (00:07:59) And so there's many of these outcomes that we are designing. (00:08:02) But when you look from a CTO perspective, really, in my mind, it's three things that are really getting not disrupted, but are massively changing. (00:08:11) To me, the metaphor is a little bit like, (00:08:14) When we moved from on-prem to the cloud, originally everybody thought, Hey, we just take the on-prem software we already built, we put it on the internet, call it cloud, right? (00:08:23) But then, only over a certain period of time, people realized, Oh, what does actually CI/CD mean? (00:08:30) deploy daily or multiple times per day. (00:08:32) And then you realize, oh, hmm, we always had multi-tenancy in the on-prem software already. (00:08:37) But then, of course, in the cloud, you have to learn how to scale it up and down, right? (00:08:41) And all of a sudden, the software kind of got re-engineered, right, to really build real SaaS software that (00:08:51) with all the concepts in cloud computing. (00:08:53) And with AI, the same is happening. (00:08:55) And then it's happening on three levels. (00:08:57) It happens, of course, on the UI side. (00:08:59) The time is clearly over where you design software, where the dump software that requires the intelligence to sit in front of the computer. (00:09:08) So I mean, if you look at classical software, what did you do? (00:09:12) You decide a user interface. (00:09:14) trying to, hopefully you did some user research, try to figure out in the easiest way or the most intuitive way to teach a human how to get their tasks done by clicking through the UI, essentially. (00:09:25) And this is over, right? (00:09:26) It's now (00:09:27) We call this generative UI, so the UIs get dynamically generated. (00:09:31) If you have analytical questions, for example, or if you want to do your deep research, not just the deep research you find on Perplexity or the usual chatbots, but deeply rooted, let's say tariffs are being introduced or new taxes or the strata form was, what does this mean for my supply chain? (00:09:45) And then you can analyze this in conjunction with your SAP data. (00:09:48) So there's a lot of exciting opportunities and new things you can build that we only dreamed of in the last, I don't know, (00:09:57) 20 years, at least since I'm a developer, where now the system becomes much more multimodal, much more proactive, because it can run overnight. (00:10:06) And then only if you wake up in the morning, tell you, hey, Sarah, have you looked at this? (00:10:10) Here's a problem on the sales side. (00:10:11) Maybe the order entry is going down. (00:10:13) You should do something here. (00:10:14) And here are some recommendations already for this. (00:10:16) Or here, there's a problem in the supply chain. (00:10:19) Because now if you're an oil and gas customer, obviously you want to know what are my options you need to replan. (00:10:24) So all these things become super important (00:10:27) for customers and that changes the UI. (00:10:29) Then the second one is, of course, the business processes like an order to cash in the past. (00:10:33) Of course, it has variances and so on and so forth, but it was a rather rigid process. (00:10:39) like the standard operating procedure of a company. (00:10:42) But now, of course, with these agents, right, we can blend the structured and the unstructured world more seemingly, so to get actually more work done. (00:10:51) So this whole move from software as a service to some call it service as a software outcome as a service, that is, of course, what these agents are building for us. (00:10:58) And then, of course, below that, you have the whole data layer, right? (00:11:02) The whole data layer of bringing, of course, SAP has a lot of super valuable data for a company, like all your general ledger and your invoices and your warehouse and inventory information, et cetera. (00:11:14) But of course, you now want to combine this, right, with the plethora of other data, right, in order to kind of build this one harmonized semantical view, because only, we always say AI is only as powerful as the data is, right? (00:11:27) So, and that is exactly what we are then also doing and transforming on the data side to help our customers to benefit from a globally harmonized data model to fuel the AI. (00:11:38) What is the biggest engineering or technical challenge for (00:11:42) for you guys when you look at these three bodies of work or anything else that you're doing at SAP? (00:11:48) Well, the biggest challenge, quite frankly, is, of course, when you look at this is how do you, it's actually not the AI so much, right? (00:11:57) But it's actually teaching the AI to do the right thing at scale, right? (00:12:01) Because, I mean, you can, look, you can build two years ago, right? (00:12:06) Everybody build a rack service, right? (00:12:07) And you could easily with a POC blew off everybody's, (00:12:11) Like the CEO socks and they're like, look how easy it is to build a chatbot on 10 documents, right? (00:12:16) But SAP and these large customers, right, they always have a problem of scale. (00:12:21) Okay, what do you now with 100 documents? (00:12:23) What becomes a little harder? (00:12:24) 1,000 documents becomes a big engineering challenge. (00:12:27) And now if you go into Joule, you are Sarah, you're maybe an SAP US employee, right? (00:12:33) Of course, if you ask a question, (00:12:35) Of course, for travel policy, for example, of course you expect a very different answer than me as a German employee would get. (00:12:41) So you now need to connect this actually with your master data, like where are you located, in which country are you, under which payroll are you actually, which taxes apply to you, and so on. (00:12:52) So all of a sudden it becomes a very, very tricky problem. (00:12:55) Same with MCP. (00:12:56) Last year, oh, everybody could build an MCP server. (00:12:59) It was so super simple to hook up your MCP server and do amazing things with it. (00:13:04) But that's becomes very, like for 10 APIs, not an issue, 100, you'll get already context bloat and all these challenges, but we have 20,000 APIs. (00:13:15) Oh my goodness. (00:13:15) So it becomes this like, because it's so huge, right? (00:13:19) There's so much things. (00:13:20) So it becomes this problem of scale, right? (00:13:23) And doing this really end-to-end for the customer, because what we also build is really... (00:13:27) an integrated experience across so you can ask your finance question, your HR question, your supply chain, you can correlate that. (00:13:33) Like this is the biggest challenge to bring that, so to speak, together, and design it then really for the right outcome. (00:13:42) And you said this also, it's what's interesting. (00:13:44) And another interesting thing is, from my perspective, is you had recently this other podcast, I think with Andre, you had it on (00:13:56) But the most important thing from a development perspective is actually that people start writing their evals. (00:14:01) That is like I was on this tour for a very long time because the problem and why does agentic coding work so well, Sarah, is of course you can verify the outcome, right? (00:14:14) You can either say, hey, is the program compiling or are your unit tests, right? (00:14:19) Does it work, et cetera. (00:14:20) And of course, combined with a little bit of taste, (00:14:23) and a lot of hard engineering work and Tropic and OpenAI built these phenomenal code generation models. (00:14:29) The problem is if you now want to build a reliable outcome in finance and so on, you need the data that say, hey, with this input, that's the output, right? (00:14:38) In order so that the coding agent can validate that and assert that against the reliable outcome. (00:14:44) And that's something where there's a mindset shift in terms of how you (00:14:49) describe the right boundary conditions to your coding agent, the harness. (00:14:55) All the boundary conditions need to be true from a security perspective and from a data privacy perspective, all the code qualities, because you also still want to maintain that code on day two and day three and day four, not just get the first version vibe coded. (00:15:08) And then, of course, these evals that then tell you, Hey, this agent is actually doing what it's supposed to be doing in a variety of ways. (00:15:18) And sometimes (00:15:19) you to laugh because do you still remember when I was a computer science student, where the Google guys came in in a lecture, and they said like, hey, I can go home at 5 p.m. (00:15:29) because I wrote my tests. (00:15:31) And of course, this was non, you still remember that test first or test driven development? (00:15:35) Of course, yeah, it's coming back. (00:15:38) It's coming back. (00:15:40) The reality is nobody did it. (00:15:41) At least I never did it because it was so much more fun. (00:15:45) It was not very popular at the end. (00:15:46) Why was it? (00:15:47) Because (00:15:48) A, it was A, so much more fun to write the code first, right? (00:15:51) And then B, of course, usually the product manager gave you a very messy requirement. (00:15:56) It was very hard for you to write the test actually first. (00:15:59) So while you wrote the code, you kind of iteratively discovered what the system, how the system would actually behave, right? (00:16:07) Now the behavior and the writing the code is so much automated, right? (00:16:10) Because now you can write almost software like completely on its own. (00:16:15) But of course, now you need to describe the right outcome, what you want from this thing. (00:16:20) And so that changes very much how the developers, of course, also now need to work specifically now that the models have the steps changed since December last year. (00:16:30) It's really hard, or I think it's not obvious how to picture if there's a version of agents and (00:16:41) models powering those agents in enterprise systems like SAP, getting better in a compounding way, the way they have in generic code generation. (00:16:53) Do you think it's possible in terms of verifiability or the ability to go understand and evaluate against that intent? (00:17:01) Because it is much, I don't know if I would say it's more diverse than code, but it's not obviously verifiable, as you pointed out. (00:17:10) Like, do you think it can be? (00:17:12) That's exactly the point. (00:17:13) That is where the starting condition is great, right? (00:17:15) I think in terms of like 2 lanes. (00:17:18) The first lane is, of course, you have the system of record today, right? (00:17:20) You know exactly. (00:17:22) in the system, hey, given this or that instruction, right, what is the outcome, right? (00:17:28) Because you can see it in the database, right? (00:17:30) And then you can construct, hey, if the order to cash process runs like this, then you need to expect, right, that the cash, like the accounts receivable, needs to come in this way, right, with the following taxes and so on and so forth. (00:17:43) So that gives you verifiability. (00:17:45) Now, the challenge, of course, is, rather, this is never enough, right? (00:17:50) Because (00:17:51) If you just look into the system of record today, that data is insufficient for this grand vision that everybody has, that it becomes this autonomous enterprise, or the agency of these agents is increasing over time. (00:18:05) So at the beginning, the agents, of course, are coming back to you. (00:18:08) Some people call this human in the loop or whatever, right? (00:18:11) So they need to come back to you, like also still with Cloud Code or Codex, and still ask you some clarifying questions. (00:18:18) Hey, I could now go this way, I could do that way. (00:18:21) And with that, what you want to design for is that you start to capture more of that context. (00:18:27) I always call this the tribal knowledge, the stuff that is not in the system, stored somewhere, that just lives in people's heads or maybe in Slack channels, maybe in Teams channels, maybe it was just a discussion on the phone. (00:18:38) So it's not stored anywhere. (00:18:40) So how can you drive a decision from that? (00:18:42) And then so the question is, the agent needs to come back, ask you for input, now you wanna store that. (00:18:48) And now what we do, in the past we called this process mining, now we call it agent mining, because you record all these decision traces, these contexts, what the users are entering into the system. (00:18:58) And then you can either use it to say like, Hey, wait a minute, this is actually an anomaly. (00:19:03) The folks in, I don't know, in UK from our company or the folks in Australia shouldn't do this because the standard operating procedure is this. (00:19:13) Or you say like, that's actually a very good improvement. (00:19:15) And then you can elevate this to be the new standard operating procedure, maybe not just for Australia, but maybe for the rest of the world or more countries to run your company more efficiently, because now you'll learn something how the organization behaves, because it can go two ways, right? (00:19:28) It could be either good. (00:19:30) Or it could be a bad thing, and then you maybe want to streamline the process, how people then actually conduct the process in a different way. (00:19:38) And that then leads to this kind of, I call this then this data flywheel, so to speak. (00:19:42) So because with every trace, every input a user gives you with all the observability that an agent writes you, you have new data sources that can then lead to new evals where somebody says, Yes, that's a verified output, so to speak, that I want. (00:19:58) And then of course you can optimize (00:20:00) the system more towards that outcome, depending on which data you gathered. (00:20:06) Do you have a strong point of view today as to whether agents operating against these business processes within SAP or otherwise an enterprise software, do you think it's going to be computer use? (00:20:24) Do you think it is all, you know, code and tool use on APIs? (00:20:29) It's an interesting question. (00:20:30) I have not a very finite answer yet to this. (00:20:36) So I think given, of course, also how clunky UIs are and so on and so forth, and knowing the challenges also from UI automation from the past, I mean, it's phenomenal what they can do already today, quite frankly. (00:20:48) I mean, it's still a little bit slow, right? (00:20:49) And so on and so forth. (00:20:50) But I still believe for the most part, it will (00:20:55) The majority will live with tool calling and agents running in the background and so on, right? (00:21:02) Because you also don't maybe want to have the browser open all the time. (00:21:09) Okay, you can do this with headless browsers and so on. (00:21:11) But I mean, if you can do this with a more structured approach from an integration point of view, I think that will be the preferred method. (00:21:18) But then of course, there will be always kind of things where an API is maybe not available or you have a legacy (00:21:25) system for a time being and so on. (00:21:26) And then, of course, these computer use approaches and so on will nicely tie in, so to speak, as well. (00:21:35) If we zoom out a little bit and just think about agents and automated business processes, in what domains do you hope customers will see that be most effective first? (00:21:50) I mean, we need to be clear, right? (00:21:52) I mean, it has been, for the most part, very productive in what I call the unstructured world, right? (00:21:58) Because let's face it, I mean, large language models are very good in the unstructured world, right? (00:22:02) Text and... (00:22:03) the images and stuff like that. (00:22:05) And so, of course, everything where unstructured data is concerned for the most part, like in services and in support and maybe sales, right? (00:22:14) And then, of course, in anything related to knowledge work, right, that deals a lot with documents. (00:22:20) Of course, this is where we see, like, just do it for consultant product I've mentioned, right? (00:22:23) This is a lot of unstructured information. (00:22:26) This is, of course, where, you know, (00:22:28) it was the easiest to get quickly to the return on investment. (00:22:33) It was harder now to kind of combine this. (00:22:35) Also, you mentioned tool use, for example, right? (00:22:38) I mean, the models had to learn, of course, first of all, and need to get better, right, on how to use the tools. (00:22:43) And then you need to build orchestrators, right, and disambiguate, oh, what does an order actually mean? (00:22:47) You mean a maintenance order, sales order, purchase order? (00:22:51) Order is a very overloaded term. (00:22:53) It's very ambiguity. (00:22:57) So, and that's, of course, this orchestration logic, that is a hard thing to build. (00:23:02) Yeah. (00:23:03) And, and so I think overall, but now that it's become gotten better, right? (00:23:10) Now you can do things like chat with your data, right? (00:23:13) And instead of going to the data analyst, business analyst that curates you some dashboards, and in 80% of the cases, that might be a good enough dashboard. (00:23:21) But for all the other 20% of the question, you always need to go back to your IT department. (00:23:25) No, now you can just converse and (00:23:27) natural language with the system, it pulls the data, natural language to SQL or whatever have you, pulls that data, you converse it until you have that point of view of the data that you want to have. (00:23:37) And then you just pin it and it's like, okay, that's actually my problem. (00:23:41) Now I want to manage that problem for the next, I don't know, two, three weeks until the problem maybe has disappeared. (00:23:47) And then, of course, you move on, maybe then you delete that tile and so on and so forth. (00:23:52) So this kind of combination of the structured-unstructured world (00:23:57) which is required, right? (00:23:58) If you want to go into the tabular world, right? (00:24:00) Because lots of data in finance is stored in tables and sales and the supply chain and so on and so forth, right? (00:24:08) Unlocking that took a little bit of time, but now it's actually, we are seeing through, for example, the knowledge graph, the SAP knowledge graph that we've built, which is kind of the glue between natural language and the structured data in the system to really bring this together. (00:24:22) That actually leads to one of your, I think, like, I don't know if it's unconventional, (00:24:27) but it's certainly not the dominant narrative in AI right now, which is your interest in like predictive and tabular models. (00:24:38) Can you talk about like, you know, why LLMs aren't the be all, end all here, or why we can't just use tools and calculation external to the model in combination with LLMs to achieve what you want to achieve? (00:24:57) Yeah. (00:24:58) Now, first of all, from a business motivation, it's a great question, right, Sarah? (00:25:00) I mean, first from the business motivation point of view, right? (00:25:04) Again, LLMs, unstructured world, that's all good, right? (00:25:07) But most of the time, if you want to plan forward, right? (00:25:11) If you want to make good decisions in a company, you need (00:25:15) predictions, right? (00:25:16) You need predictions in terms of, oh, what's my demand, right? (00:25:19) For, oh, is this depending on the seasonality effects and so on? (00:25:22) What's my demand forecast maybe, right, for my products in the retail store? (00:25:26) Or what's my demand, right, for my products so I can plan accordingly my manufacturing, right, if I'm a manufacturing customer. (00:25:34) Or you want to predict your cash flow, right? (00:25:36) You want to predict, and that has a bunch of input variables like, oh, what are actually my day of sales outstanding, right? (00:25:43) And that is determined based on (00:25:45) Are customers paying yes or no? (00:25:46) That's a classification question. (00:25:48) And if you then say, okay, if a customer is not paying within the payment terms, what's the payment delay? (00:25:54) A classical regression question. (00:25:56) And so on and so forth. (00:25:58) The problem is, of course, still today, if we look at these predictive questions, right? (00:26:03) And then you wanna maybe do a what if analysis from it, right? (00:26:06) Now, if you want to do these predictions, quite frankly, then the challenge is large language models are not made for this, right? (00:26:12) The way how they generate just one token after another, essentially, in a sequence-to-sequence modeling. (00:26:18) I mean, they're language models, right? (00:26:20) And they do this phenomenally well. (00:26:22) But if you still want to do these predictors, you have to go back to these classical machine learning approaches. (00:26:29) You use XGBoost or AutoGluon and many of these AutoML approaches that are still out there. (00:26:38) The problem is just it doesn't scale. (00:26:40) So we haven't seen in the predictive space the same level of democratization. (00:26:45) You still need to hire a very good talented data scientist, right? (00:26:48) And then if you, for example, if you're a large company, we did this, for example, at a pharmaceutical company, if you just want to solve the payment delay prediction problem I've mentioned, right? (00:26:59) They are running in 90 countries around the world and they need these two models. (00:27:04) So you end up with 180 models you need to train. (00:27:06) You need to curate the data, you need to train the models, figure out, right, what the right model is, feature engineering, like the classical, (00:27:15) machine learning kind of approach, right, that was used in the past. (00:27:21) And what we said all the time is, okay, look, we have all this data stored in these tables, right? (00:27:26) Thousands of tables, right, where all this information is stored. (00:27:30) Can we not apply the same idea that large language models or multimodal models did for the unstructured world, actually for the structured in order to start predicting things? (00:27:38) So you can just basically provide a little bit of context, a small amount of data, not a large amount of data, because that will (00:27:44) always a problem, small amount of data, and then starting making high accurate predictions, so to speak, in that domain. (00:27:50) And that led, actually, it was two years of research. (00:27:53) We published that also at NeurIPS and a bunch of other conferences. (00:27:57) We call this RPT1, so RAPID1 stands for relational pre-trained transformers. (00:28:01) It's still based on the transformer architecture, but with a very different architecture. (00:28:07) We released this and we see some, meanwhile, some very, very good results from that in various domains where, as I said, classification and regression, sometimes time series. (00:28:17) and so on are concerned. (00:28:18) And we believe this will be huge because it obviously will allow way more people from a business impact to make these predictions, which large language models have a really hard time with. (00:28:30) When you think about the gap in, I think, (00:28:38) I don't know, I think you described it as like hype versus adoption within the enterprise customers. (00:28:43) The innovation race versus the outcome race. (00:28:46) You're right. (00:28:47) Innovation race versus outcome race. (00:28:50) It's a good framing. (00:28:54) The change is happening very quickly. (00:28:55) That's hard for companies to absorb. (00:28:58) Where do you see challenges for the enterprise and adoption today, and where are customers making the most progress with you, or where are they most excited? (00:29:08) Yeah, that's a good question. (00:29:10) I mean, usually I say, (00:29:14) The primary problem, as I said, is the problem of a data, right? (00:29:19) Because most of the time, the data is, of course, very disaggregated in a company, right? (00:29:25) I mean, for a variety of reasons, right? (00:29:28) Either because you made certain decisions, how you purchased solutions in the past, (00:29:34) Or you did an M&A, so you acquired a company naturally. (00:29:36) Of course, they bring a very different IT system landscape as well, and so on and so forth. (00:29:41) So you have to segregate that information. (00:29:43) And the problem is, of course, that limits the potential of what you can actually do with AI. (00:29:48) And then the question next is, how do you integrate this safely? (00:29:51) And what I see is clearly customers who did that kind of homework. (00:29:55) Now, of course, it's not a new topic. (00:29:58) We're discussing this for 10, 15, maybe more years. (00:30:01) The ones that did, (00:30:03) their homework. (00:30:04) They, of course, have a much easier life to then also reap the benefits, for example, of AI. (00:30:10) The second one, as I mentioned, is already is the problem of scale. (00:30:12) The bigger, the more complex the landscape and so on. (00:30:15) And of course, also then bringing this together in a unified experience is a challenge. (00:30:19) And then finally, of course, everything around then security and so on and so forth. (00:30:24) Because then there's always then this gap between, oh, there's an amazing innovation. (00:30:28) Take OpenClaw, for example. (00:30:30) I mean, amazing. (00:30:32) what this has brought to the world in terms of further ideas. (00:30:36) And of course, I mean, from a security perspective. (00:30:39) And so that's a problem. (00:30:40) You don't want to run this just like it is there on GitHub and deploy this in your organization to do something. (00:30:47) I mean, nobody would ever do this. (00:30:49) So then, of course, you need to bring this, make it secure. (00:30:52) I mean, we have seen with LightLLM, how long is this now ago, two weeks or something? (00:30:56) You probably saw it, right? (00:30:57) Like with this vulnerability that still all of a sudden steals all your keys and credentials and so on and so forth. (00:31:04) And you don't want, right? (00:31:07) If you were to chief information and the security officer in the company, you don't have a job anymore, right? (00:31:13) And that's, of course, another big challenge from an adoption perspective as well. (00:31:17) What do you think the function of a finance or an HR or a supply chain team that would've been operating out of SAP in their day-to-day work a year ago? (00:31:35) What do you think that looks like a few years from now if they're successful, if you and your customers are successful with the AI transformation? (00:31:45) Yeah, I mean, first of all, it's very simple. (00:31:47) They will get rid of a lot of the mundane work, right? (00:31:49) Like collecting information and preparing PowerPoints for decision making and so on. (00:31:54) So what we're going to see is... (00:31:56) a much, much faster way of making decisions, making better decisions. (00:32:01) And then, of course, automating the mundane work. (00:32:03) So what the people will do is they will run more scenarios. (00:32:06) They will run, get better, deeper insights in a much faster way in order to then really think about, we always call it this more strategic thinking. (00:32:15) And kind of in a way, Sarah, if you will, for me, this is the same way, like everybody who works today, maybe in the finance shared service center, it's for me the equivalent of a junior developer today with (00:32:26) So now they actually become, they get one level higher, right? (00:32:30) They're now not so much anymore tasked with then writing a lot of the code, right? (00:32:36) With Codex or with Codecode, but they actually... (00:32:39) then start supervising the code, give feedback, and capture, of course, the essence of what the code should look like, and then do much more review, and then rather think about what to build next. (00:32:51) Think about the next requirement and how is that actually differentiating. (00:32:55) So every role, every level will kind of get up-leveled, so to speak, because the work that's being done today will be pushed down to these agents. (00:33:07) And therefore, I think I believe (00:33:09) in general, what we will see is that people will just achieve so much more because there is a lot of intelligence baked into the system that gets rid of many of the things that we're doing today and that are actually, well, at least in many cases, a lot of fun. (00:33:26) I must admit my ignorance here. (00:33:29) I'm thinking about this and I want to talk a little bit about the impact on the business, if you're right as well. (00:33:35) I don't actually know how SAP prices broadly today, but the question would be like, how do you price? (00:33:40) And if you are, you know, delivering more outcomes for customers or serving them, (00:33:46) you know, services software in a different way. (00:33:48) Do you think that changes the business model for SAP? (00:33:51) It does, absolutely. (00:33:52) I mean, there's no question, and we have prepared for this already. (00:33:55) So for me, it was always very clear. (00:33:58) I mean, for the most part, SAP software is seed-based, licensed today, with a few exceptions, like Concur or Fieldglass, for example, or the business network. (00:34:10) But, you know, (00:34:11) very clearly with AI, it was very clear for us that, you know, step by step, it will go towards this consumptive world, right? (00:34:19) At first consumptive and then maybe in the next step, once we have more verifiability in the system, then also towards maybe an outcome-based license model to, for example, what Sierra is doing and so on and so forth. (00:34:35) But the reality is also, it is today for us, it's a hybrid model. (00:34:39) It's consumptive, but it still has a certain element of seats in there and so on and so forth, because also it's a joint journey with the customer. (00:34:45) Because the customer is saying they are not yet ready. (00:34:49) in many cases, for a purely consumptive model, right? (00:34:53) Because they need one predictability, right? (00:34:56) And then, of course, they are not yet fully also everywhere trusting the outcome, right? (00:35:00) Or node and also, of course, is the value already there, but then they are afraid of that the costs may explode from a consumptive perspective, et cetera, et cetera. (00:35:09) So at the end of the day, what we have designed is a hybrid that is basically ready for this consumptive world, but actually meets the customers where they are today (00:35:19) knowing that they demand still a lot of predictability in the enterprise space in order to cost control the whole thing for themselves as well. (00:35:28) That makes sense. (00:35:29) It's unclear how... (00:35:30) I also believe that transition is going to happen. (00:35:32) It's unclear how quickly it will. (00:35:35) Exactly. (00:35:35) No, absolutely. (00:35:36) I agree. (00:35:36) I mean, nobody knows this. (00:35:37) And at first you see (00:35:39) customers that are more, you have a wide range, right? (00:35:43) They're also of opinions, right? (00:35:44) And of course, some customers are a bit more forward leaning already, and then others are more still asking or demanding a classical model, so to speak. (00:35:55) And so therefore, it's a journey. (00:35:58) What do you... (00:36:01) Let me rephrase that for a second. (00:36:04) When you look forward and think about SAP's position five years from now, and you compare it to the broad market pivot away from SaaS and software in terms of just how investors are valuing these businesses and their enthusiasm about their durability, (00:36:27) My own opinion is like the challenge is real and yet it will affect the incumbent software companies like very differently, right? (00:36:36) There will be winners and losers versus like universally, like everybody's market cap should come down. (00:36:41) What do you think is going to be characteristic of a winner or why does SAP get to endure again? (00:36:48) I think at the end of the day, it's all about adoption and the outcome you bring to the customer, right? (00:36:53) I mean, the technology, (00:36:55) Look, the reality is for most companies, the technology doesn't matter. (00:36:59) I always tell to my developers all the time, our job at SAP is to make the technology disappear. (00:37:05) We need to get the outcome in front of the customer. (00:37:08) And of course, not just the value itself. (00:37:10) Of course, you'll also be able to produce and price it in a way. (00:37:12) So it's a win-win situation for the customer, and of course, the vendor at the end of the day. (00:37:19) So what we are really trying to do, and this is also from an architecture, we are so flexible. (00:37:25) We don't over index on a specific element. (00:37:27) We have partnerships with all of them, right? (00:37:30) And really only invest in the things that are actually differentiating for our customers versus the things that anyway will likely get commoditized in the tech stack. (00:37:42) And then try to make sure that we, of course, bake the enterprise qualities in and the integration is there and the customers can turn these capabilities on almost instantaneously in order to benefit from it. (00:37:52) Why is this important? (00:37:53) Because if you take a lot of time (00:37:55) time to reap the value, then your return on investment is essentially gone, right? (00:38:00) Or the business case becomes harder, right? (00:38:04) And therefore, what we are really focusing on is to deliver these outcomes to the customers. (00:38:09) And I think that will differentiate the winners from the losers at the end of the day to really focus on the business outcomes for the customer at the end of the day. (00:38:16) As we wrap up, I want to ask you a few quickfire questions, including a little bit more personal one. (00:38:23) Our listeners always want to know, like, what do you do all day as like a CTO of SAP? (00:38:28) Like, can you just describe how you spend your time? (00:38:34) Well, I spent most of the time reviewing the progress with the teams, right? (00:38:39) And we're thinking along, you know, from all the layers with the teams, from the database to the models, right, to the UI, review the progress, give guidance, feedback, learn something new, study, of course, what happens outside, do a lot of prototypes, right? (00:38:54) While we speak here, I have a bunch of command line interface instances running here, prototyping a bunch of things, right, trying things out, see what works, what does (00:39:04) doesn't work, right? (00:39:05) And then give this also as kind of inspiration back to the team. (00:39:10) And then of course, you know, yeah, work a lot with the teams. (00:39:14) Think about, of course, how we connect the vision then to the execution. (00:39:18) And of course, work a lot with customers as well. (00:39:20) So I love working with customers because they always, you know, give you (00:39:26) They keep you honest in terms of what you can do already and where you can still get better. (00:39:32) Can you give us a flavor of just because you have such an amazing like global customer base, like something challenging or interesting that a customer is asking you to do right now? (00:39:43) there are plenty of things, right? (00:39:45) As I said, right? (00:39:46) The predictive awareness just immediately comes to my mind, right? (00:39:49) Like if you can improve the accuracy of the demand forecast by 3, 4%, right? (00:39:54) That is huge for them, right? (00:39:55) This is multi-million value.