SAP CEO: AI Won't Kill Software, But It Will Change Your Job — With Christian Klein

0:00 Now that this apocalypse is over or paused, software is booming. What's happening? Let's talk about it with the CEO of Europe's largest software company right after this. Welcome to Big Technology Podcast, a show for coolheaded and nuanced conversation of the tech world and beyond. We have a great show for you today. Today, we're joined by the CEO of SAP, Christian Klein, who's here to to talk with us about the state of software amid all these SAS apocalypse worries. and along with some other big decisions that software companies are faced with today. Everything from hiring to the integration of artificial intelligence and more. Great to see you.

0:36 Welcome to the show. >> What a great introduction. Thanks for having me, Alex. >> Okay, so you know we talk a lot about software on the show and AI on the show and we don't really have to you know when we say we think AI is going to do this or or that you know we don't have to make any planning decisions based off of it. We just talk about on the next show. You on the other hand, you do need to make planning decisions. You have to basically decide and this is why I think this is going to be such an interesting conversation. You have to basically decide what you think the trajectory of AI's progress is going to be. And then you plan based on top of it because if AI stops now, it's one set of decisions.

1:11 If AI keeps going a little bit but then hits a wall, it's another set of of decisions. And if we go into recursive self-improvement, it's a third set. each one of these impacts your business dramatically. So actually, you know, you make the choice and then you run a company based off of it. I I think that AI has to hit a wall at some point that we've seen great progress, progress that's gone through many walls that people put up >> or you know, presumed walls and it's continued to build, but it can't keep going on this like this forever.

1:47 >> what do you think? And do you do you think it's we're about to hit like recursive self-improvement or do you think that what I'm saying makes a little sense? >> Yeah, happy to answer that question and let me maybe just describe it how I describe it to my employees. Yeah, when we do an all hands I always show them a mountain. I said look we climbed the cloud mountain. Yeah, we did a cloud transformation and then now we have to climb the AI mountain and on this journey there will be obstacles we have to overcome. AI has to overcome and you know when all of this hype started with generative AI I mean of course everyone was using all these LLM models and start testing it experimenting with it and now it's really about okay but it's nice to experiment with it but you know all of our customers are now saying okay but what is in it yeah show me the value and why does AI understand so much you know all the unstructured documents I can summarize a mail I can summarize a document and so on but what what about the business why does is there always certain limits with the accuracy. Yeah.

2:48 So that is the first obstacle we have to overcome. And the the third one I guess the world is talking about that do we need to regulate AI? Do we need to govern AI? Because obviously when you're running the world's most missionritical businesses I mean obviously governance plays a huge role and this is where I would say these are the kinds of obstacles we have to overcome to further you know climb optimal. Yeah. >> Okay. But Christian this is the kind of the core question here.

3:13 >> Yeah. >> Yeah. All right. AI doesn't understand structured data that well right now I think >> but your your job is basically a big part of your job is assessing whether the technology will be making those leaps in the future. Yeah. >> So, we could talk a lot about what the current state of AI is and we will run the company, right? Let's say AI for instance gets really good at like going into ERP and being able to like make >> serious calculations about things which it's getting better at.

3:46 >> and SAP's business I should say. >> Yes. Yes. >> It's enterprise resource planning. That means that it brings together systems like finance, HR, manufacturing, supply chain and sales into a single system, single software system. >> So your job is basically to to figure out big part of your job is to figure out >> whether AI will be able to overcome those obstacles, what time frame AI is going to overcome those obstacles and and then how you adjust. So do you think it's going to get over the hump is basically the question I'm asking >> 100%. And when and I would say look I mean you know we for example code financial closing agents here and we are sitting here at the wall street and of course they work today I mean AI understands business but you know maybe it understands today 90 with a 93% accuracy but your auditors when you release financial results they say okay sorry this time my numbers were 7% you know too high and they say what I mean you know you need audits you you need it needs to be certified and so on. So it's 100% accuracy is needed and that's where I guess every tech company is now working on and you know every software player is now using especially SAP as we have so much knowledge about industries processes and data I mean we are working this on a on a constant basis and yes of course AI understands business but sometimes you need an accuracy of 100% 90% accuracy is not enough and then when it comes to governance I mean here in the US when we are running the US government when we are running you know many public sector customers you know agents need to understand what is fatwam. Yeah. So that you know which data can I access, what data can I share, where's the data going to be stored and so that is also the governance part and we are working on that. So for me Alex this is more a question of now months. Yeah. Until we are really reaching a level where you can really trust AI running your business and then you can still decide okay the human is in the loop for us.

5:36 That's that's a that's a rule. But then you can really decide on the autonomous level of the agents on how autonomous can the agents run the business. >> Okay. So this is important. so your perspective is within months AI can go from basically working within software and giving us giving your customers like 90 something% accuracy. You you anticipate 100% accuracy within months. >> I I I would say that's absolutely possible. Yeah. So it's for certain tasks and obviously we will then develop the next agents. Yeah. For supply chain optimization and then you know here we are maybe there in 8 to 9 months. So but it's not we are now talking not anymore about years. Yeah. So business AI is real. Business AI is coming and also 100% accuracy is not always needed. I mean for example when I do my customer visits here in New York I'm asking Juel works put me together a customer brief.

6:32 >> My my Exactly. Jul is our, you know, coworker. And when it's 90% accurate, it's totally fine. Yeah. Maybe, you know, when I ask for the the latest earnings and it's not exactly the latest earnings. Maybe it's the earnings before. Okay. I mean, who cares? But, you know, and there a 95% accuracy is totally fine. But I don't need a bunch of people in my office preparing me for these customer meetings and writing all of these preings. Yeah. Because I get the data out of SAP. I mean, the the public content is anyway understood by the LLM. So, you know, here we go. So it really also depends a little bit on the AI use case.

7:04 >> Yeah. Okay. So 100%. That's that's really interesting because I think the big thing that has been holding back a lot of enterprise roll out and the ability of enterprises to integrate this technology has been the fact that yeah if you're doing something like tax supply chain >> projecting your sales pipeline and you're at like 92%. >> Yes. >> Might as well not use it. >> Yeah. I mean, you could check the work, but you don't know where the problems are.

7:30 >> Exactly. >> Can you ask another chat about where are the problems? If it finds 90% of those problems, you now have a compounding issue. >> Yes. >> But things change when it gets to 100. So, I think the way that you're framing it is we're about to see a real >> explosion of enterprise. It depends on the use case. It also depends. just you know an hour ago I visited a big customer here in New York and they said oh Christian you know I love your financial closing assistant but it's only 93% accurate but okay there are 100 finance systems warning hanging somewhere around in this company some are non SAP so okay we need to clean up the house I mean you know your data is a mess so it's also not only the agent per se so it's also the data yeah the data quality the data silos in a company where we then need to match data so that the agent can really understand okay where does all you know the financial data sit how do I depreciate a certain asset you know and so on so and the more systems you have and the more you lack data quality I mean of course you know this also then reduces the accuracy of your AI agent there >> okay but I thought smarter and smarter AI was supposed to solve that problem right that basically like the big one of the big reasons that enterprise point at for like not being able to roll out AI.

8:48 Like there are all these stats about like one in 20 pilots or one in 10 pilots actually make it to production. Yes. Is because the systems are working on sort of faulty data foundations. But you know is the AI at the point where it's getting good enough to figure that out on its own? Like we companies are putting teams and teams on this try to figure it out. But why is the AI sort of not able to do that on its own?

9:12 >> Yeah. I mean AI will also solve that challenge. I mean look at when I look at our data platform I mean we are partnering you know with other data lake providers like data bricks or snowflake bigquery and so on but obviously I mean it's easy to do zero copy so that you know we can share data without moving the data but then of course the semantics you know matching data fixing data quality issues of course AI now you know in our data platform AI will help us to match data from a SAP to a Salesforce system from a workday system to an SAP system so that also that that the cleanup of the house is getting easier and easier and also that the agents can really access a semantical data layer which is also coming out of the box so that you don't need an army of data scientists anymore always you know from every month and you need to clean up your financial data and you match it to your HR data and to your employe data and to your payroll data. I mean you know all of this work will also get automated because AI is very good in in matching those data points together and that is of course also part what is happening underneath the agents in the data. Business today moves very fast >> but it moves fast despite you know it moves fast despite the fact that processes are a mess. It moves fast despite the fact that culture can be a problem. so is basically the net of this that just business, you know, I guess like we've heard stories of companies that used to be on this annual plan of planning. You plan, you release at your big event and then you plan again. Now they're on quarterly or monthly plans.

10:46 >> So is that what we're going to see is basically a speed up of the ability of businesses to release products, to serve customers, to to grow is just is everything going to get faster. the speed of execution is definitely getting faster and that's also I mean you you need to also then redesign your own planning cycles inside the company to react to that and no matter if you do supply chain planning financial planning workforce planning I mean we at SAP we look at our virtual agents and then we look our employees and I said hey this now needs to be looked at together we can't plan you know here the agents and here we plan you know do our employee plan and so and we need to do this faster because the software the AI is moving now much faster faster. So absolutely and then the second point what I would say is you mentioned one important point and that's culture because you know when I look at some of the ACP projects I mean obviously when you technically migrate to a new software to system but you ch don't change anything on the business side because also the business has why to change and and so on it just creates you know uncertainty I mean obviously you are not seeing the business benefit now with AI the mindset in in many customers I'm seeing has completely changed they say wow Wow, if we miss that boat, I mean it's really disruptive, you know, for our company. So, we have to change.

12:03 So, the openness for change is also now very very different. So, speed of execution getting much faster, planning cycles getting shorter, but also the business is now very clear. Hey, we we have to be on we have to ride this wave otherwise it could be too late. >> Okay. So, I have a lot of culture and hiring questions, talent questions to ask you throughout our discussion today, but I'm just going to start with the first one.

12:27 >> Yes. >> Are you noticing people getting exhausted? Because if you you know there was a story in the Wall Street Journal I've referenced a couple time on this show >> that I initially hated and now I'm like starting to see the wisdom where it was like bosses were mourning the loss of busy work because you know even though you know we the the example is doing expenses people are like I'd rather my expenses be automated so I can focus on the bigger thing and that makes sense but like >> if AI handles let's say it h you know it handles all this rope work >> so all you're doing is like very highly intense cognitive work.

13:03 >> and everything is so fast. >> Mhm. >> Seems I I don't know if it's a recipe for burnout or not. What do you think? >> I I I actually it could be actually. I mean, you you have to see. I mean, many jobs will change really dramatically. I mean, you know, to look at all the the controls in finance, in HR and so on, they will get all automated now with with the AI agents we are delivering other players. what's going to get automated >> the the controls, the compliance checks, everything. Yeah. So also the the typical stuff. Yeah. What many people are doing in their business life and then of course >> full-time job for thousands of people.

13:40 >> That's a full-time job for thousands of people and now you're even in in planning and steering a company and putting the pricing list together and so on. I mean AI will you know be infuse a lot of intelligence into that and then of course people have more time yeah to spend on we would say the the value adding task yeah to to spend more on okay when the earnings is already prepared including my remarks and they even you know AI even tells me how to when to raise my voice or lower my voice I mean obviously you know it's such a different preparation it's also then the way on okay now I have more time to spend of really on delivering you know my earning speech and so I guess that is really a a different change of working and definitely needs a lot of change management. Absolutely. Yeah.

14:27 >> And how about the part about people getting tired. I mean I'll give you an example from my life. I'm blessed. I mean I I really I really am you know it's nice to be able to run you know this company. I don't won't say on my own because I have a lot of help but like the you know running you know a media company with one person mostly wouldn't have been possible to the level that we're doing it >> but like >> I'll have claude co cloud claude code is doing my invoicing claude co-work you know soon it it will do these projections it will keep on top of my inbox it's going to do you know all these other things like expenses and things like that >> and so my time is like really spent on >> Mhm.

15:09 >> interview prep, interviewing, like doing these higher higher low cognitive load tasks. >> I love it, but I can't tell if I'm more tired because of it where like >> rarely do I have a moment to like downshift. >> Mhm. >> And that's it's it's great. It's great, but it's also like I'm starting to questioning whether that's kind of whether it's what I want. So, you're obvious you're your company's 100,000 people in the neighborhood. >> Yes. Yes. Yes. you know what do you hear from people about this? You know, I I see this also in my daily life. You know, all the all the prep work, you know, is suddenly getting automated, the customer brief of that. And so I can really focus more on the content, but do what you know, and then it's actually much better time, you know, where you have with your team talking about the strategy you you're discussing, you know, certain things on are we are here on the right track or need do we need to course correct about what about these leadership decisions I have to take or what are about other things. So I would say yes there is a change of the type of work what you're doing but I would say will it get more boring or more exhausting? I mean I I don't believe so but you have to structure indeed your dayto-day in a different way.

16:19 Absolutely. Yeah. >> Yeah. Okay. So no increase in fatigue >> I so far. No no okay that's good. And I also honestly I to the customers I talked to many customers and also end users of AI and so on. I I didn't hear that now. they are all happy if okay let's get rid of this all of these workflow approvals let's get rid of all these compliance checks and controls and and then let me focus on the stuff what really matters so yeah >> yeah I mean I'm framing it in the most negative way possible but I also think that like yeah I would never go back to like wanting to do all this all the invoicing and expenses on my own software takes >> what I'm of course getting in ably all hands is okay is my job still secure yeah and what about restructuring what about that and so I get that yeah so not about So necessarily I'm getting exhausted or it's more >> because AI can do so much of the work they might be afraid to tell you. Yeah.

17:08 Yeah. But also honestly I I really appreciate I mean you know we are running at at full speed these days. The transformation happens at full speed. So and the workload is is not is not getting lower for our people. Yeah. So we AI helps them to get more productive but it's not like that we are running out of work in in development or in finance or in HR or other parts of the company. Well, that's why like when I think I think I spoke about this with Muhammad Alam, one of your colleagues.

17:33 Like when I think about at least with today's technology, like is it going to create mass job loss? >> My answer is always like I don't think so. It used to be I'm sure no. Now it's I'm I don't think so. And the reason is is because companies >> they have a massive road map. They have so many things they want to do. They have things their competitors are doing. And so if your choice is, you know, do the same with what you're doing with fewer people but with technology or keep people but do even more because they have this technological boost.

18:04 >> The companies that survive will do more. >> Yes. >> With people and tech. >> Yeah. And you know I believe that and I tell this my people honestly. I said hey I mean our workforce you know the way all our skills and you know the people working here at SAP I don't believe that in 12 months from now you know we will you know there will be a mix a different mix of job profiles yeah we will need other you know jobs will we we will hire people data scientists full stack developers etc but you know we will need less people in other jobs now it's a question about how much can you reskill how much can you you know train people on a new job but also sometimes you know in these moments A company also needs you know you know new employees to come in with new skills with a new mindset and pair them with the experienced colleagues we have as well because you needed to the domain knowh how you need to they understand the software which is underneath the AI so I would say it's it's there will be a workforce transformation but it's not necessarily that it's only about restructuring it's really about you know at the end of the day you need to have the right skills and the right mix of people >> yeah people often hear the term reskilling and they don't believe in it.

19:14 >> They think it's a nice thing that academics think of, but when you go to somebody and say, "We're going to reskill you. You've been doing job A forever. You're going to do job B." it doesn't work in practice. >> what's your experience been like? Have you seen successful reskilling programs? >> Oh, yes. I mean, inside our company. So, talk about it. I mean, for example, we had a lot of developers coding on premise software. Yeah. And now coding cloud software is different. Yeah, it's it's more devops. You know what you code need to be tested automatically. You need to then ship it and it's in production and it's then on us. Yeah.

19:48 That it's bug free and that we deliver it and that the systems run stable. So it's really it's really a different way of coding and it's a it's a different mindset. Yeah. And then the question is also it's not only about a reskilling from skills, functional skills. It's also about the mind mindset. Does someone really need to go into this new field into this new world? And so and and that is so absolutely it's possible or taking out when we are rolling out our coworker to work 80,000 people are using it but at the beginning it was also there certain resistance yeah you know what does it do to my job how can I practically use it going across this barrier of you know does it really help me or do I need to still cost correct certain things what jewel is producing for me but once you you know you train the people you coach the people and you see hey you can get you stuff done faster and you can focus on higher value adding tasks that is the moment where they say okay let's use this tool in legal in HR in finance in sales and and then you see yeah that people are you know able to adapt to that and learn and also acquire new skills >> you know pre-hat GPT I wrote a book talking about how the tech giants were already using AI to like minimize wrote work and make room for more inventive work >> and you know I it came out in April 2020 not not a great time to release a book, but I'm glad I did it. And the main push back that I got was or basically what I what I said was about work was that you take people off of these wrote tasks and you put them on more inventive tasks and it's more fulfilling and it helps a company grow and move faster and be more inventive.

21:26 And the main push back I got was when AI automates like back office work like data moving data from one place to another or like you know one example that we gave in the book was like AI is going to eventually be able to write like new hire letters and you know benefits letters and stuff like that >> which like was felt crazy at the time but it's it's been doing it for years now. the push back was people who are used to doing this like data let's say moving data from one to another >> are not going to want to do the more inventive work.

22:00 >> It's a different type of thinking. It requires different culture, different set of permissions. >> and they're not going to be able to adapt to the new type of working. >> What's your read on that? Is that is that a fair criticism? >> Yeah, it also depends on you know on each individual. I I give you an example in our world. you know for 50 years we coded software and for 50 years we were never running out of customer requirements on features new features for our software. But over time obviously yeah you know the business you know the business which you are running but the product managers didn't need to go out always to the customer and to completely reinvent how businesses is running because it was one feature more and the software was running the business in a very stable way. Now with AI, we tell our product managers, no no don't sit too much in-house. Go to the customer and reinvent how supply chain works, how we do inventory optimization, how payroll will work in the future so that not an army of people need to make sure that the payroll is working correctly and people are getting paid in in a in a in a in a correct way. And now it's really about this excitement of oh now I'm going there and with with this technology I can really completely reinvent how businesses how companies will work in the future. I really reinvent the future of work and I would say with the vast majority of our people that creates excite excitement are there maybe a few people who say I'm not sure if I'm into that I'm used to a certain type of working. Yeah, it could be. Yeah, but this is, you know, really about, you know, how do you bring your people with you and how you do the change management? I'm not saying that every every single employee will make that move. But I guess a lot of people get it. Oh my god, this is exciting.

23:47 This is new and I love to learn something new in my working life. >> Okay. And you've segueed perfectly to the ces apocalypse because you know there's there was this belief and remains remains a belief among many that >> >> software like yours >> was developed for you you develop it for the masses right so you build software like SAP >> people have to be use able to use it in different functions different companies and so when you get a seat to it you're going to get this software that's kind of built for everything there features you're going to want to features you're never going to click on.

24:23 >> Yes. And this idea once this idea h like came through became popularized that once people could build software on their own yes you know because you could prompt it and they can show up for you >> then the old way of building software >> old way of building software wasn't going to work anymore and we would see new software vcoded replace >> you know the old software u and that that narrative has plagued software companies until recently for for you know about the better part of a year.

24:54 >> Mhm. >> At this point. >> Yeah. >> >> let's just high level, >> you know, what what was what's your reaction to that? What was what's what was your feeling as that narrative became popular? >> Okay. Yeah. Let's go back in time, especially when this narrative came up. I mean, of course, the first moment you think to yourself, okay, I see development productivity is going up like hell. why can't people just reproduce what we built over 50 years?

25:22 And then I ask my product managers and my portfolio team and go through our portfolio and which solution do you believe can you know can you wipe code easily and just reproduce it and replace it maybe and the answer was very quickly that they said oh my god I mean you know we are building ERPs yeah so we are building supply chain payroll finance it's not only that you need you know you have 7 million data fields in in in such an ERP which you need to correlate so there are hundred hundreds of millions data correlations what you need to understand to run a business but there's also of course deep process knowledge by industry you know by country local requirements I mean we are investing hundreds of millions every year to keep our software compliant and then you know the answer was okay there are maybe you know one or two solutions small solutions where there's not so much process and data context inside which you might you know just can replicate but the west I mean definitely ely you know you need to have a lot of domain knowledge it's definitely not easily you can definitely not easily white code it and so yeah and that gave me a good feeling but still Alex I mean obviously there's not not the time to lean back because clearly now I guess in the next phase you know the value creation is now moving up from the system of record into the into the agenda layer so absolutely our AI needs to be world class to still also defend our existence in the system of record Okay. But now now let's go back to something you said earlier. Yes.

26:55 where you said that AI is getting good enough to be like 100% accurate on some of these tasks. You said that AI is getting good enough to sometimes reconcile data. >> Yeah. >> >> the the just to I'll sit on the side of the SAS podcast and you can be software for this example. >> Okay. SAS pacos people would say >> man the AI is getting good enough that that those hurdles >> domain knowledge process knowledge you know being able to put data together >> it's going to do it >> and therefore the mode while it still exists today cuz the AI wasn't good enough when the AI gets good enough that mode's gone.

27:36 >> Yeah. >> What do you think? >> Yeah. I mean well when we talked about the beginning about all of this accuracy I mean that was the accuracy on our AI platform. I mean we tag an LLM and then we put you know our AI foundation next to it where we train business data business process knowledge where we build knowledge graphs semantical ontology layers and and and so and that's why we are giving the LLMs now the understanding yeah to deliver you know agents who can actually live up to all of these accuracy targets. We talked earlier on about it. If you just use an LLM alone, there's no way that you can run a warehouse with it, that you can do a financial close with it, etc., etc.

28:13 But it's an LLM plus our AI foundation plus the governance. Yeah, let's it's not only the business data and the business process. It's also that you need to understand all the legal requirements. I mean, the agents cannot just go wild and do a financial close without understanding the tax requirements in a certain country. So there's more to it and that's why you know you know I'm so confident that you know these LLMs are super powerful without any doubt and they're getting better but they need the the process context the data context and the governance and this is you know I would say where SAP I mean this is what we are doing for a living since 50 years. Yeah.

28:49 >> Yeah. Okay. Well I am not done with my questions here and I have some other things to speak with you about regarding this conversation. So let's do that when we come back right after this. Yes. >> And we're back here on Big Technology Podcast with SAP CEO Christian Klein. Great to see you. Thank you for coming in. I said before the break I wasn't done with my questions and I'm sure not. All right. So, okay. This idea that like AI is going to get better and people can vibe code you know their own SAP is going to be hard because even if the AI gets better, all this domain specific knowledge is is tough to include. I have this theory that what you might see instead of people vibe coding software >> is that the AI companies which currently make like the foundational labs open AAI and anthropic which currently make their >> best models available via API >> will close those models off. So for instance, let's say GPC GPT6 Astra right now it's pretty smart but it has those limitations.

29:54 >> But we've seen like unreleased versions of OpenAI's technology. >> Yes. >> Be able to do things like team up in a swarm >> work for 88 hours and solve like >> you know the hardest math problems ever. So is there a world where they say we're not going to keep we're we're not going to release these models maybe for safety >> and but we are going to put them on some of these problems that SAP has been working on for a long time and then decide that in order to make money because AI models aren't going to commoditize they have to go upstream which means >> take that knowledge and do a chatbot version of your business.

30:35 >> Yeah. Yeah. Fair question and look the the truth is when you come to our business AI platform as a citizen developer as a business user or you know as a IT guy a developer you can actually develop agents on top of our platform. So we are delivering hundreds of agents, standard agents for finance, supply chain, HR but also the business and the the the pro code the developers can also build agents on our platform and what you find is all these models where you just talking about. So you know but what you also find then on top of it is that we connect you know those models you know to the business process and the data knowledge the context of what our ERP actually you know knows about your business and we are reproducing that in our ontology in our data layer. So you know you have all of these models >> sorry what's ontology explain that >> the ontology is actually when you have you know you have SAP data but you also have non SAP data what the agents need to understand when you do asset management. Yeah, the agent needs to understand the sensor data. are the signals of a machine to understand when does the machine needs maintenance and then it needs to pair this with the SAP data about okay let's create a maintenance order who is the worker who can fix this machine etc etc >> so it's the logic >> so and that is the ontology what we are building now what we are bringing together and we also tell the agent okay machine has a problem where do I find the spare part for this machine who's maybe not working anymore where where can I find the worker who is fixing it etc and everything on time So that there is no downtime to the machine and that is the ontology. Yeah, that is the semantic layer and again why would you go to you know a third party platform when you find all of these great models on our platform plus you get all of the semantics that the ontology plus we are going to run these agents to make sure they adhere to the governance of your company of a country of an industry. So you're skeptical that there's a even if they hold those. So right now OpenAI and Antropic are making their models available on your platform, but you're skeptical that if they were to withhold >> their latest models from your platform and try to do it themselves, they'd be able to do it.

32:44 >> No, I'm I'm you know the same question I got years ago about the hyperscalers. Yeah, I mean we are partnering with >> this is a different technology. >> Yeah, that's a different technology, but we are partnering and you know they are of course they are super powerful. I mean they are running you know the most you know the most systems of the world they are owning you know a lot of our systems and they could also say yeah we are not providing this infrastructure anymore so and this is you know and this is similar but I see you know there are so so many models now coming to the market that I would say there's a healthy competition also going on and we are not seeing this one model who you know solves all of the problems actually we are seeing a lot of open source a lot of weight models who also doing a very very good job when it comes to also looking at the price outcome ratio of an agent. Yeah, it's not only important to that the agent does a great job, but it's also at what cost does the agent does a good job. And so I feel there is so much healthy competition that I'm not so much worried about okay there is now a provider saying oh we are not offering you this model anymore because again there's competition there are many models and there is not this one model who actually will solve all of the problems. Let me ask you this.

33:53 Would you accept or I don't know if you do this already. Would you accept OpenAI or Anthropic as like a front end to your software? And by that I mean like you have all this great data. could let's say a user of SAP who's used to like logging in and going through the graphical user interface or using your agents. Yes. >> Are you open to them like having a connector through Chat GPT for instance and being able to query that data and do that through chat GPT for business? I mean the way how we solve for that is there's jewel work you know our cowork it's our interface >> but you won't integrate >> we actually embed those models into jewel work yeah so you can also switch the models yeah you can do certain task with topic you can also do certain tasks if you're in Europe and soft is important you can also use mistral you can use their latest model and and then but then it's really about SAP's interface that's our tool to work it's our coworker and and then of course there will be also scenarios where you will build an agent you know let's say Salesforce, Adobe and so on and they need to also then access you know SEP data they and and that's totally fine as long as people are going through our API gateway and the agent gateway and so it but I would say predominantly you know what we are seeing now with our customers that users working with SAP software they will use tool work but there will be definitely also third party agents accessing SAP who will have access to SAP systems via our agent gateway >> including opening Ian and Anthropic or >> including including >> so there will be a world where >> yes yes yes but yeah but it it needs to be controlled yeah and customers will also rely on that that it go through our agent gateway because again governance is important back is important you you don't want to go having an AI agent doing uncontrolled things in your ERP system yeah so that's why I'm very very confident And that first of all a lot of the SAP users who do work in finance in HR and supply chain will come via jewel work because that is the place to be.

35:58 You can do everything what you do with an LLM model plus you get the SAP data the business process the context. We also have a data platform who infuses non SAP data and if you build a third party agent for example for your next marketing campaign and you want to weed out some order data I mean yes you can do it but it will happen while our agent gateway and the marketing users are anyway is not you know our prime users of the SAP software.

36:23 >> Okay let me ask the question one more time in a maybe different way because I don't know if I'm getting the right answer or full fully understanding the answer or getting the question out the right way. All right. There's an example recently where Meta has this Muse. Yes. >> Personal assistant. >> And you could get Muse to shop for you. >> Yes. >> There's two approaches to this. >> Yes. >> There's Amazon's approach where Meta is we want Muse to be able to use Amazon. And people can say, "Buy me a leaf blower." And then the Muse will go into Amazon and buy them a leaf blower.

36:53 And you won't ever have to touch Amazon. Yes. >> And then there's the Okay. So, Amazon blocked that. >> Mhm. Yes. Shopify said that that's fine. >> So we Shopify said we want a third party agent to come in and use our technology and Amazon said no if you want to buy on Amazon you can use you can use Amazon.com or use our agent. >> Yes. >> What is your philosophy the Amazon or the Shopify? >> Actually we offer both. Okay. So that's and and ideal model the ideal world is that you know because our AI platform tool work is so great. I mean in my eyes customers end users will understand why would I use an LLM model standalone if I'm not getting you know the same business context and not the same governance so that we are connection >> we are yeah but the API connection is not giving you all the business context is not giving you the semantics it's not giving you this deep process mode what a >> some of the stuff that >> yeah yeah a pure API connection pure MCP server is not the same like the ontology this the context we having on our platform so and and the governance.

37:58 Yeah, when it's SEP managed, we also make sure that your when you're doing a financial analysis that you know when we govern it, why it should work, it is ensure that only the people who are allowed to see the numbers are also getting access to this report. So the agents understand that. So why would you then go into a third party platform where you have yes API access via our API gateway? That's possible. You can access from a third party platform, you know, into the SAP system via the API gateway. But there's so many reasons yeah for for SEP users to use tool work because better context better governance that we believe yes the main route will ch to will will goes via tools yours.

38:38 Yeah. >> it is interesting like when we sit down with someone like yourself >> and hear it it's like oh this is actually you know not as simple as a one-s sentence tweet. >> Yeah. >> And that's what the market seems to have been responding to. I mean SAP this year is it's down 21% but it's been up 43% over the past 2 months. >> It's a roller coaster for you, isn't it? >> Oh yes. Oh yes.

39:01 >> But I'm used to it. I mean when I became co when I'm used you know I I did the cloud transformation >> right >> the market was say oh are they going make it or not and you know the share price was down then we reached the all-time high. we we proved you know that we can transform the company and that's another transformation and and it's a similar pattern now the good piece this time we are not alone in this the whole software industry you know went down but I guess more and more and investors analysts and of course also customers realize hey an LLM alone will not cut it so we will need the software the app the rights the context the governance and then the LLMs are of course the key you know to unlock this value but it needs to be really brought together with the with the business process context and the governance and I guess that is something what in the last two months many people realized in the market as well >> software is it's made a dramatic comeback.

39:56 >> Oh yes. Oh yes. Yes. >> I want to hear a little bit about you mentioned open source. I want to hear a little bit about the way that you think about spending with these models and which models you want to use. Yes. >> So we recently on the show cited some data from RAMP. They have an economics lab which obviously is like >> they're getting credit card data from lots of startups but they think it's indicative of what the rest of the economy is going to do.

40:20 >> Yes. >> in August 53% of spend on AI was on the frontier models >> models like Opus and Fable and so on. >> In September that number was 40 or at the end of August the number was 45%. >> Which is quite a decline. from 53% to 45% decline in terms of the spend on the frontier. >> Yes. >> What that suggests is that companies are discovering they don't need to use the best AI >> to succeed.

40:53 >> Yes. >> Have you found that at SAP? >> Similar patterns. Absolutely. And I mean again that's also maybe to the discussion you know to the point you brought up earlier on. Would customers go via third party platform directly to the ACP system or will they use you know our tool work our platform I mean in our platform you're also not logged in to one model because you see yeah with every model release you know it can change again you know the one model gets better then there's a new open source model who is really producing great results and we also see agent by agent not every model you know performs the same yeah so this multimodel being agnostic is really also a key differentiator what customers are loving. They're saying, "Okay, I'm not locked into any frontier model. I can and SEP is even doing the switching from one model to another to always optimize the token span with regard to the outcome." Yeah. In comparison to the outcome. And now to your question, yes indeed. I mean, we of course also started you know coding with a lot of frontier models. Still coding is something what we do a lot with the frontier models also for softronty reasons etc etc. But then you know for many agents we are running you know and for also for tasks like okay create me a headcount report for my manager or do me a financial report we actually see that some of the open source model performs so good that it and then compare it to the cost of those models. We we are going to switch and many agents we are building in the meantime have seen the the fifth model because we are always you know optimizing yeah outcome versus the tokens and the cost of a the cost we pay for such a model. Yeah. And has that emphasis on cost grown recently?

42:30 >> Oh yes of yes. Yeah. Because the overall token spend is up which is good but you you know it doesn't help you if some of your employees is getting 20% more productive if at the same time you know the cost going 30% up. Yeah. And of course and by the way this is also for our product manager a cultural change. Yeah. We told them test every agent which we are releasing to the market needs to be tested with different models. And they need to do this all the time. And because we see new model comes, oh maybe we can switch the model.

43:00 Yeah, because it's a better price outcome ratio at the end of the day. And so yeah, that will continue. And we see this also as a clear differentiation that there's a software company like SAP who does this for you so that your AI tokens are not running away while you are celebrating your maybe your productivity gains and then at the end you look at your P&L sets, oh shoot, you know, my my my profit is actually not hitting the mark anymore. Yeah.

43:24 >> Yeah. it's it's interesting, right? It it says that it says a couple things actually. First of all, it says the AI has gotten good enough >> that the frontiers is only going to be applicable to certain use cases. Yes. >> Which is wild. >> Yes. >> it also says the standard models are are, you know, are really working. And an interesting thing about the standard models is they don't tend to be at like standard versus frontier.

43:48 >> They don't tend to be like you know, 80% of the price for 80% of the performance. Yes. Right. They tend to be like 10% of the price for 80% of the performance. >> Yes. And and look, I mean, we also acquired an AI model, one a table AI model who does predictions and I mean it's remarkable. Yeah. We took all the data of retailer SAP non SAP and without a data scientist touching it and building data pipelines, doing the semantics, matching the data so that it makes sense. the table AI model actually did it with the same accuracy than for the customer than what a team of 10 data scientists did before. So you see you know it's not only the frontier models it's also the tap AI models and there are other models coming up for the structured data the all the models we are building yeah for the business data.

44:34 So I would say there is not this frontier versus you know opensource standard models there's also no tap AI models coming because for predictions these models are becoming better and better as well. as a business person what do you think it says about the business of these frontier AI labs if the frontier is >> you know it's harder to charge a premium for the frontier >> yeah keep on innovating keep keep on you know making the model better and better better >> no they are making the model better and better but they people don't need the cutting edge >> but others but others are on it as well yeah so that's why I'm saying I'm not I'm not afraid of all that you know one model is or one provider says oh you are not going to allow to use our model anymore. There are so many models in the meantime and they are making all good progress and but again it's the same like SAP when Hasslatner founded this company there was not not many competitors now we have hundreds of competitors I mean that's the name of the game we have seen that others have seen that the only way is keep on innovating to justify the price and if you don't if you're not much better than the rest of course at a certain point it's then hard to justify the price but this is a game we all playing >> I'm going to I'm going to answer my own question I I think that if you're spending billions and billions on training frontier models and the standard models are doing just as good or or good enough job for many of your customers to the point where we're seeing things like 8% declines not year-over-year but month over month that might be a problem for your business.

46:02 All right, I won't ask you to >> agree with me on that one or disagree but I have to put that in for the record. >> Yeah. Yeah. I mean I mean I guess this is also the question what everyone at the frontier model is also asking themselves and you know and it's also interesting and maybe one addition to that I mean you know there's of course a financial close there is inventory this is when you ship when you build accuracy needs to be record high yeah we have a rule 95% or better otherwise it's not worth to ship the age and the customers will just not use it but then there are many other agenda use cases yeah take you know a customer brie leaving a head country but what I mentioned I mean this is not doesn't need to always be 98%.

46:43 Yeah it's also okay if it's 94%. because I still you know all the people who needed to prepare that stuff you know we need less time on that we can focus more time on the value adding stuff prepare myself for the meeting etc etc so and that's why I would also say all of these frontier models and you know we don't need to always use the best best model from an outcome perspective you always needed to match it to the price and to the nature of the model what is the accuracy what I need yeah to have the acceptance of the business >> yeah Very interesting. We we have IPO, so we'll learn a little bit more about this. yes, >> how about how about you know, you've sort of talked a little bit about this token maxing reckoning or the token reckoning.

47:27 >> RAMP also has interesting data about the spend for the top 1% of employees. >> Oh yes. >> Right. So top 1% of employees, they're not just consuming a little bit more tokens, they're cons consuming way more. So knowing smile for those on audio. So we're gonna ask this. All right. So, in a month, >> the spend from the top 1% of users fell from $7,976, according to Ramp Economics Lab, to $7,25, which is a 9.7% decline.

47:58 >> >> with the users, the the top token gobblers within SAP, have you encouraged a similar pullback? Have you seen a similar pullback? what's that knowing smile all about? >> Yeah, we also have this 1%. And my CFO says, "Hey, where is this going to?" And honestly, it took some time until we actually introduce certain token limits in SAP because I told my CFO and my CEO, I said, "Hey, I mean, now we tell everyone use AI to you know, reskill yourself to change and become more productive." And when I'm now the first one who actually introduces this token limits, it's also the message what I don't like. But at a certain point, we really saw, oh my god, now the token spend is really coming to a point where we definitely need to set certain limits. Yeah. For certain jobs. But you know, for the top 1%, I mean the ones who are doing the model training and so on, we didn't really set a limit, let them run because we see this is a very important task, very important job to make the accuracy of the agents better. So I said hey we can't compromise on that because all you know the success of our AI really depends on the accuracy of our agents.

49:14 let them do the job. But then of course you know for you know all the developers the product managers the designers I mean there we also introduce limits also for people in finance and HR and of course there are certain limits where we can't can say hey you can't just experiment around spend the tokens when we are not seeing you know the the respective productivity outcome but we set very healthy and still I would say fair thresholds where everyone says okay I can definitely do my job with AI and I'm not reaching this limit you know every month and Then of course people are allowed to say but I have this certain task where I feel AI can really help me. It's this time it's very special so I need a higher tone spend.

49:53 And then we have a approval process where a manager can also say okay for this month or for this task or for this project we give you more tone spend. And I feel the people accepted it well and it was also the right message yeah to not send the message of we forbid you know the use of AI or you're reaching the limit so fast that it's hard you know to really optimize your day job. but vice versa also making sure that tone spend you know stays under control. I guess what was even more important than these tone limits is the model switching. I mean yeah of course I mean we are testing as I said different models and when you switch from one model to another that can cost can cut your cost by 10 by a factor 10 and that is of course even more effective than you know introducing all of these token limits for the certain job profiles in a company.

50:37 >> When did this all go into place? Oh, the I would say the the token limits, the budget by job profile we did four months ago, three months ago and and the model switching we built this into the platform from day one on but in all fairness and we were just now building the agents getting them out celebrating success getting references and now I would say in the last three months we also intensifying our product manager intensifying the more the work on switching the models testing the models and you know optimizing them and really also look into the TCO war model because you know at the end it's it's it's not about only about delivering great agents. I mean at the end it's also needs to be from an economic standpoint it needs to be reasonable. Yeah.

51:21 >> how long does it take between when we see some crazy frontier AI behavior and when it gets diffused into business? Let me just give you one example >> like >> the hugging face attack from open AI. hundreds of bots, you know, attacking a problem and well, that's I mean that's like well known, but there's also the math the attacking the millennium problem >> problem and succeeding. >> Mhm. >> That seems to me to be, you know, these sort of swarms of agents working to solve business problems. I don't know. It's it's it hasn't I haven't really heard about it much from from businesses like even those that have said, you know, we have like a multi- aent system. It's like six or seven, not hundreds or thousands working together.

52:09 Yes. >> How long does it take from, you know, from when we hear about something like that at the frontier to when you get it? And where do you think we are on this agent form thing? >> I would say look on cyber security obviously it's something Yeah. where you always have to be ahead of the game. Yeah. So we are testing different models all the time. >> But it's not cyber it's like these type of uses anywhere.

52:29 >> Yeah. Okay. On the uses everywhere. I mean you know whenever a new model comes we are testing it right away and model switch can be done in one week two weeks we need to change you know when there's a new model we need to change APIs we need to you know adjust the MCP servers but we can do this very fast yeah and so it doesn't take a long time to really test out a new model and how does it perform on the on the business side >> but like swarms of swarms of agents for business use cases are you testing that now or >> yeah yeah absolutely yeah we're doing this And when we are building all of these A2A use cases, I mean obviously we we need to test it. We also need to test you know the collaboration of these agents, how do these models also perform and so yeah we are doing all of that.

53:12 Yeah. >> Yeah. I will say if I was sitting in your seat Yes. >> the thing I'd be most afraid of is what's going on with cyber security right now because when you think about your customers, they've got finance, supply chain, sales, HR. These are the most sensitive records a business could have. And if we see, you know, agents going zero day >> and people being able to direct them. >> Yes. >> that's that would scare me.

53:38 >> Exactly. And that's why guess what we are doing. Yeah. In in our deaf operating model inside the company. Yeah. We are always now working on you know how can we use AI to detect vulnerabilities earlier. How can we put more proactive measures in play to test the firewalls of the company and the product and what do we do on patching the system and so on. Yeah. So that's definitely now something what I would say for every technology company this needs to be at the very top of the agenda.

54:05 >> Your cyber security budgets going up. >> Yeah. >> Substantially >> actually I mean on the one hand side it's going up because you're investing into AI but you know finding vulnerabilities patching I mean of course that also gives you productivity gains. Yeah. So, it's going up, but it's going up in a reasonable way. >> Okay? You know, we're we're a podcast where we don't like to do the theatrics or the we try to really understand.

54:28 >> So, I'm really going to resist being like, "What's wrong with Europe?" Okay. I'm going to resist that question. >> I actually want to ask it to you this way. >> Yeah. >> Europe has put so many regulations on its tech >> companies and people would say business in general. Mhm. >> to the point where if you speak with people here in the US >> Mhm. >> they'll they will be like it's not even worth doing business there.

54:52 >> we have a lot of European listeners. as I told you before, I'm from a this mixed marriage here. I'm American. My wife is European. >> >> I I think that like it would be good if Europe became a good place to do business. >> Yes. >> What So this is I'm not going to herang. I want to know from your perspective cuz you're running the largest software company in in Europe. >> What is the thought process that the European regulators are going through >> to get to them where they are >> like what is what are their intentions and do you believe that those are good intentions?

55:30 >> Yeah, I mean first I really believe they have all good intentions. I mean who in Brussels does something to disadvantage Europe? Yeah. But of course what people sometimes do especially the ones who are not so close to the technology they tend to regulate the technology itself. I guess what we are now learning and we talked about also the concerns and the risk we are seeing coming with AI. I guess you still you need a regulation but you should regulate more the business outcome the impact on a society and not the technology per se not the use of data because otherwise you know it's so hard to do business and get a startup going in in in Europe and also for us I mean we are a global company and we are doing our research and our development also in Palo Aldo we're doing it in Bangalore we're doing it in you know everywhere in in our labs but for a startup it's really it's really hard and I guess you know Brussels and Europe I mean it's definitely they realize that probably they stretch it too much. That's why they call it omnibus. This is where you can where they now collect all the simplification measures, the deregulation measures. So, and I know they're dealing with it right now and I hope there is stuff coming which definitely helps us. Yeah. To put the regulation to the right level and again I can I can really emphasize not enough that the need to make sure that you regulate the impact on societies but not regulate the technology per se because how can you be competitive if you are the only part of the world who is really regulating the technology?

57:02 >> Yeah. So you think there's going to be some form of of roll back? I'm I'm I'm confident I mean the message is loud and clear not only by SAP I mean many startups there are other companies like Seammens and so on I mean we are building a lot of industry AI in in Europe so there is AI I mean more applied AI industry AI and all of these companies you know say message to to Europe we have over regulation let's scale it back >> yeah and it's mostly GD sorry I come from like more of like the ad perspective GD I mean GDPR is Yeah, speaking of the intentions, like I get it. You want to protect people's data.

57:39 >> Yes. >> Is it mostly that or it's just >> it's actually not so much about GDPR. I mean, what Europe does is the AI act, the data act, >> which then puts another layer of regulation on top. >> And, you know, it's it's not only another layer. Sometimes, you know, these layers are also overlapping. So that even the legal people or the your data protection officer is not even knowing anymore. Okay, there's so many quones, so many overlaps. So it also takes forever until you understand how can I use node data to build AI and to do research and to apply AI in the customer's business. And so that is you know these all of these layers of regulations and then you know the the recent ones with the AU data hack and the AI act.

58:23 >> Mhm. All right Christian feeling dunk. >> Yeah. Thanks for coming in. Thanks a lot for your time. >> Should very good German actually. >> I'm learning. All right. Well, thank you for being here. Thanks to the New York Stock Exchange, for hosting us today. Good to be here. Great to speak with you. Thanks everybody for listening and watching. and we'll see you next time on Big Technology Podcast.