00:00:02 Philipp Herzig: You should really focus your own IT resources or your own R&D really on your core product or your services. Like if you are a pharmaceutical company, focus on how AI can help you with scientific discovery, or if you’re an automotive company, how can you build AI into the car? Or if you’re an insurance company, how can you add value to all your clients? This is usually where you should focus on build because that is a domain where you have really the deep domain expertise and also where your customers are expecting from you in the next 10 years that AI is making the experience better, faster, whatever it might be, in order to be able to compete in the market with the service and products that you provide as a company these days.
00:00:46 Chip Kleinheksel: Hello, and welcome to season two of Resilient Edge, a business vitality podcast from Deloitte. I’m Chip Kleinheksel, the chief technology officer of Deloitte’s global SAP Alliance. In this series, we explore some of the most pressing business transformation issues that organizations are contending with today. The question at the heart of our series, how do companies and leaders go from thinking about change to actually making it happen at scale? I’m joined by world-class experts at Deloitte, SAP and beyond to help me answer this question, and today we’re going to focus on the topic at the top of everyone’s minds, artificial intelligence. That’s my specialty and has been way before it was even cool.
Artificial intelligence is everywhere, promising to revolutionize finance, supply chains, compliance and decision-making. We’re at a pivotal moment for AI. It feels like it’s do AI or die for companies and the executives in charge of steering them. The imperative to implement AI wherever possible is stronger than ever, but the challenges in doing so are also in clearer focus and this episode will aim to provide a sort of smart executives guide to thinking through key AI implementation decisions. One of the biggest hurdles businesses are facing in this moment is deciding, or understanding where and how to start with these decisions, potentially having a twenty-year impact and affecting all current ERP systems; they are certainly not easy ones to make.
I have such a wealth of AI expertise on the podcast today. I’m joined by Philipp Herzig, SAP’s Chief Technology Officer and Chief AI Officer, and my colleague Arjun Krishnamurthy, an SAP AI leader in Deloitte’s SAP Alliance. Welcome to the Resilient Edge.
00:02:25 Philipp Herzig: Thank you so much.
00:02:26 Arjun Krishnamurthy: Thank you, Chip.
00:02:27 Chip Kleinheksel: Two gurus on AI, Philipp, Arjun, tell us a bit about yourselves and your roles. Philipp, why don’t you kick us off first?
00:02:35 Philipp Herzig: Well, my role really is to look end-to-end at AI, of course from a product and from an engineering perspective, but then all the way down to sales and services and marketing to really close the loop end-to-end and take a holistic 360 perspective on AI for SAP and of course for our customers.
00:02:53 Chip Kleinheksel: Arjun, how about yourself?
00:02:55 Arjun Krishnamurthy: I’m a partner of Deloitte where I lead our efforts in enabling AI with ERP, specifically SAP for our clients, and my passion is actually helping organizations move beyond the hype cycles, focusing on practical value-driven AI solutions that create real measurable outcomes. Excited to have this conversation here.
00:03:18 Chip Kleinheksel: So this time last year, AI felt like an exciting new development for consumers and companies alike. Now as our thinking around it has grown to take on other dimensions, national security, environmental resources, its impact on the structure of companies and the economy itself, the mood has decidedly shifted a little bit. And I can’t think of two better folks, we obviously all know each other, have worked with each other for years, but I can’t think of two better folks to be a part of the discussion. Philipp, Arjun, how are you seeing companies approach incorporating AI right now?
00:03:47 Philipp Herzig: You mentioned already last year, last year clearly was the year, as we called it, the year of the AI PoC or the GenAI PoC, specifically. Everybody in the C level was priding him or herself on how many hundred use cases they are working on and what they’re all trying and so on and so forth, and sometimes it was really hard to differentiate what adds value versus where it’s just a big number. The reality is also if we look at the end of ’24, beginning of ’25, only a small number of applications in the enterprise space, in the business world, really went into production and really added value right at scale. And there are a variety of reasons on why that is, but what I really see, also, that the whole conversation becomes way more mature also with the customers.
I think we are past the hurdle of where we discuss data privacy and security issues as a major concern or legal topics like IP and indemnification topics for example. And of course we see really more and more the question on what is really adding return on investment, so the whole business conversation became also more mature. So it’s not just about, oh, another great PoC that somebody did over a week or so, but what’s really moving the needle for the business and how does this, in terms of the return, but also the investment that’s required in order to reap that value, so to speak. And so that conversation became pretty mature on what I call GenAI the last two years, which even felt, meanwhile, it feels like a decade, but it’s kind of already surpassed from a hype perspective, with, and I think no podcast these days can go without agents. Agents are really on the peak of inflated expectations right now where again, we are going through that curve of understanding what the technology really can and cannot do. So we are living almost in this kind of bimodal world where on the one side we are in this trough of disillusionment for understanding much better what the technology can do and where it adds value while the agents, it’s so much buzz in that topic, it feels like two years ago when everybody was trying to figure out what the tech can and cannot do.
00:05:53 Chip Kleinheksel: It’s the hype cycle of the excitement of what agents can do to then balance with the practicality of now where does it apply to business? And I love what you’re saying.
00:06:02 Philipp Herzig: Oh, absolutely.
00:06:03 Chip Kleinheksel: Less around the preliminary type questions around legalese and data privacy and more now focused into ROI and value, which is a good thing. That’s where we should be. Arjun, what about yourself?
00:06:13 Arjun Krishnamurthy: Yeah, I totally agree with Philipp, especially companies rapidly moving away from just experimenting with AI to, how can I start embedding it into my core business? Our data shows that there are about 70, 80% of organizations are already moving in that direction. But to Philipp’s point, most of this understanding is based off the consumer experience of generative AI. It convinced a large majority of how AI can solve enterprise problems, but most of the organizations are still looking at AI powered productivity tools, improving daily ways of working, document management, email management. Those are the areas we are seeing significant push.
When it comes to challenges that clients are facing in order to incorporate, it’s not essentially the AI and the technology wireless hype cycle, but the big critical factors surrounding it. To Philipp’s point, what value do I choose? It’s not about AI for AI, but AI for ROI. Which function or scenario or the use case do I start with that gives me the most value? How do I make sure that my data is good and the AI has access to that good clean data. And when I say data, we’re talking about all enterprise data, it’s no longer just the data in your systems and how do I make sure it is secure and trustworthy and I can stand behind it. And then all of this comes at a cost. How does the economics play? How do I make sure that the architecture… So there’s all of these questions surrounding… Talent is another one, what happens to our people and the operating model? Yes, the big answer is always around transition to higher elevated value work, but how do I drive that shift? Those are all the questions that clients and organizations are looking at in order to make this beyond just POCs and proof of concepts to make it real impactful. But those baby steps are what we are seeing, I think that’s what 2025 has been for us so far.
00:08:05 Chip Kleinheksel: So both of you talked about how we’re transitioning out of the world of POC and into a world where we’re actually seeing real use cases. And I love where you went with this, Arjun, not AI for AI, but actually AI for ROI. Where are businesses actually seeing value with AI today? Where are some examples where we’re seeing it?
00:08:22 Arjun Krishnamurthy: I would say there are broadly two categories. We have seen organizations having success when they tie AI to what was traditionally called process automation. So improving transactional efficiency, just imagine invoice management or processing receivables, but even supporting financial flows as an example. What we have seen in these organizations, is that they have started to see significant success when you’re going beyond logic and code-based automation to having, even traditional AI, I’m not even talking about latest breakthroughs of generative AI, We’re starting to see those as low-hanging fruits giving significant efficiencies.
The second category, which is more recent is with generative AI, organizations are seeing benefits where AI can perform basic data retrieval tasks in a simple conversational style. You can actually ask an agent in a very conversational way, and specifically what data, what time period, what are you looking for, and it brings it for you. These are some low-hanging capabilities that you can get started with and you can actually demonstrate that AI is real and it can actually move the needle.
00:09:31 Philipp Herzig: I would say when I specifically look at generative AI, where most of the value today really is in the unstructured world, specifically when you look at large language models, maybe image models for document understanding and so on, and specifically in areas like coding very clearly, so that’s more and more really a domain where a lot of benefit is. And developers love it, and the same is true if you look into support understanding tickets and deflecting tickets and customer support, for example, while helping sales reps to make sense of vast amounts of documents and unstructured data.
On the other side where we haven’t yet fully cracked the nut yet is in terms of really applying generative AI models to the structured world. When we look in the structured world, like the tabular world and so on, or predictive tasks, when we look in demand forecasting for planning purposes for example, or just if you want to match open items with payments that you receive as a company, and still these non-generative models, classical machine learning models are actually very, very good, but we haven’t cracked yet the nut on making them as scalable as these generative models are today in the unstructured world. So there is still a tremendous amount of research that is required to bring them to the same level of, I would call it democratization and accessibility in that space as generative models brought to us in the unstructured data world.
00:10:57 Chip Kleinheksel: I think of, in the unstructured data world, one area where I’m seeing it is I think towards the future especially is we spend so much time with clients creating reports, we’ll create hundreds of reports when we do an SAP implementation, and I do believe that there’s a world we’re going to be in and not too far off where we may not have to create all of these reports, we have access to the data, we leverage AI and the CFO or CIO can say, “I need to see a report with XY, or Z,” and it just creates it for you. It’s real and it’s coming.
As we work through AI and the landscape, what are the greatest challenges and maybe common misconceptions you’re seeing in the market when it comes to implementing AI? Arjun, let’s start with you.
00:11:41 Arjun Krishnamurthy: Yeah, this is a hard one because there are a lot of challenges, especially if you want to make a move and make it production and scale, but if I were to put my finger on one thing, it’s data and trust, those are the most immediate challenges. There is a struggle to ensure that the data that the models are using are accurate, timely, complete, because now you’re giving the agency of insight generation to the agent. So now you are completely reliant on the insights based off your accuracy and completeness of the data. How do I ensure that I can trust and stand behind the data and the insights I get? This is a huge challenge. And this is not limited to structured tabular data, to Phillip’s point, imagine all data of an enterprise. If I’m talking about emails, shared folders, how do I know that if I’m pulling a policy or a contract, that’s the latest version that actually the model is reading and responding to you. So that to me is sort of the biggest challenge. From a misconception perspective, I don’t know if it’s a misconception, but I would say it’s more of a mismatch of expectation.
Phillip touched on it earlier, there is an expectation of plug and play to deploy AI. That obviously the consumer world has given you that, you download the ChatGPT app or Perplexity app and it just starts to work. But when it comes to functional domain and industry specific data and the language that you use in your organizations, there is time required to teach the models, how do you use the organization’s specific context so that it can actually solve for the specific business problems? Especially in the world of agentic AI, you have to take the time to train the model similar to how you would actually train a campus recruit if you may. Deloitte is actually working on solving this specific problem by pre-training with functional domain and industry specific nuances so that we can solve real world problems that are business and enterprise oriented in order to achieve that plug and play expectations that is there in the market.
00:13:40 Chip Kleinheksel: Phillip, how about yourself?
00:13:41 Philipp Herzig: Well, I can just echo what Arjun said, there’s specifically in the senior levels of management, there is this misconception on, how hard can it be? Because today everybody can build a little RAG, Retrieval Augmented Generation, as we sometimes call it, app, or let’s say 10 documents in your enterprise, something that the AI model has never seen because it’s private data. A hundred documents is a little bit harder, a thousand documents already becomes an interesting engineering challenge. And then there was this misconception, if you can do it on 10 documents over an afternoon, which gives you the 70% version of your app, it becomes more of a logarithmic function step-by-step. It gets increasingly harder, the more we move [inaudible] you need to put way more effort in until it really gives you then at scale.
And it’s not just then for all these documents, it’s then also how do you serve this then with enterprise authorizations? Because if you, for example, build an enterprise chatbot on your, let’s say your HR or your procurement, your travel policies, now you want to know, ”Oh, that employee is actually in Brazil.“ Or, ”This employee is in the US.“ Or, ”That employee is in Germany.“ Of course different rules and different answers should be sent back depending on, it’s still the same company, but it still needs to send contextualized answers back based on where that user is located, which authorizations and privileges does this user see, what is he or she also allowed to see? So all of a sudden it becomes a huge complex system where you have to take a lot of things into account.
And the second one really is of course there is a lot of legacy in companies and the legacy of course, specifically where you have not brought the data together, we always say that AI is only as good as the data that feeds it, because data is the lifeblood of AI and that is of course also true for generative AI. And I think those companies who really created a comprehensive data, harmonized data layer with a harmonized semantical data model that also explains to AI the various relations and the semantics of the data and so on and so are far better off and have a head start versus those who still have these siloed data boxes that are non-integrated and so on and so forth. Specifically when you look from an SAP perspective, for things, what we do with the business data cloud and so on is exactly helping customers to clean that now up because those who haven’t done that yet, it’s now just more urgent than ever to bring this really together to run off a harmonized data model to fuel your AI to get the insights and the predictions you need.
00:16:04 Chip Kleinheksel: I think you’re right, and I chuckle as you both hit on data and trust. I think we’ve all been in an executive meeting where one executive brought a report and said, ”This is my version or my analysis of the truth.“ Another executive brings their version and analysis of the truth. Now there’s a third, which is, what’s the analysis that AI drove?
00:16:22 Philipp Herzig: Correct.
00:16:22 Chip Kleinheksel: And so this harmonized data is going to be key for everybody to get behind that. The other thing you didn’t mention, and I was a little bit surprised on, is the challenge I’m also seeing is just purely habit. We build these capabilities, and what I’m seeing quite frankly is if you build some type of AI solution that sits outside of an end-to-end process, less adoption, less use. If you force it into the end-to-end, then people don’t have an option, it just becomes part of the process, and I think we’re hopefully transitioning there, but that’s been a gap I’ve been seeing.
00:16:53 Philipp Herzig: It’s interesting that you’re saying this, and I’m actually the wrong person to ask this question to because that’s the whole hypothesis that we from an SAP standpoint started with because if that’s a disconnected experience, that is not part of the workflow and the software and user experience you use today, the change management, I mean you still have with AI, even if it’s embedded in the business process, if it’s embedded in the apps you use today, even then you have change management, but it’s far easier from an adoption perspective than when it just sits outside and now you need to learn how you combine these things and so on and so forth. But it’s just interesting. I don’t observe that, Chip, that often because that is part of the SAP strategy to just embed it everywhere. So usually that is not of a major concern, at least from what we see.
00:17:39 Chip Kleinheksel: Your strategy is working well, and Arjun and I will both say, we see it when we’re out with client’s. Let’s transition a little bit. So given that there’s so much potential with AI and so many different possible approaches to bringing it in, I want to talk a little about taking those first steps. How do you think companies should get started with AI? What should they be thinking about? What should they focus their efforts on? Arjun, let’s start with you.
00:18:01 Arjun Krishnamurthy: Sure, this has always been the million-dollar question, where do I start? How do I go about it? One of the things that we utilize is a value to effort framework. ROI value has become way more important because AI is not cheap, so where do I actually see the most value where the least effort, where it’s mostly plug and play? That helps us to prioritize scenarios that offer medium to high value, requiring low to medium effort to implement instead of solely relying on rule-based automation, we leverage machine learning and drive automation there. Those are examples around accounts payable, accounts receivables, financial close.
With generative AI, we also recommend strategy by enabling more areas where giving better user experience, user querying data, easy conversation based prompts, and these things actually set the stage so that there’s a good base where you can start to scale these to become agents that can functionally be proficient and even do deep business problem solving. By getting these baby steps in place, what you’ve also done is you’ve made sure that you have the right data pipeline, security, trust, even talent model, all of those things are also being achieved and you start to kick the tires by getting some value and proving to your cohorts that you can start to implement these kind of things and genuinely improve user experience and start to see more easy activation.
00:19:30 Philipp Herzig: That’s absolutely right. Everybody should do that. Maybe just to add, it’s more to consider the introduction of AI and the company not as a big bang. Not all the things need to be in place, but you can start incrementally, but what you need to do, I call these, sorry, I get now a little bit technological and architectural, is to not focus so much on the things that may change. The large language models will change, the use cases will change. What shouldn’t change are these, I call them the architectural invariance in your landscape, like establishing these standardized set of services, like for example, what we do with business data cloud where you can bring all the data together and then it doesn’t matter whether it’s other tools to bring your data together. You can do this under the hood incrementally. Or when you talk about your natural language interface, what we provide with Joule, establish that. And maybe at the beginning it’s only able to do HR tasks, but then you can bring the finance tasks underneath that. Then you can bring the analytics tasks and the reports that you’ve mentioned earlier underneath that and step by step, you make the experience with every integration as you clean up the things you make your enterprise architecture data ready, as you make everything come together step by step, you make the experience richer.
Give yourself under the hood the technological flexibility and breathing room to be able to swap technologies and so on, because we are always living in this discrepancy between, on the one hand side, you want to provide enterprise readiness and stability to the business and not bother them with changes all day long, whereas simultaneously, it’s hard to make a bet these days because the technology is just moving so fast. That’s a little bit of a principle on what I advise to customers around the world to get started, not to wait, but to establish these common set of principles and services that allow you to incrementally make it better, but also switch components whenever it is necessary.
00:21:17 Chip Kleinheksel: Yeah, a flexible architecture that allows you to evolve as the technology evolves and then-
00:21:22 Philipp Herzig: Exactly.
00:21:22 Chip Kleinheksel: You have to think about it like LEGO blocks. If you try to think about the end-built thing, sometimes you’re going to get just paralysis-by-analysis versus if you start with Lego blocks and you can build, well, now you can start stitching the end-to-end together, which that’s quite frankly where I’m seeing clients be successful.
Where would you say companies should generally build or where should they buy, or are there hybrid approaches that are likely going to be the best bet? What are you guys thinking around that?
00:21:49 Philipp Herzig: Well, I mean that’s an easy to answer question, I would always say what comes out of the box right from already the applications you use today, just use them. There’s no need to reinvent the wheel. Why spend your precious IT or R& D resources on building something that anyway comes out of the box? But then instead, because AI is so pervasive, you should really focus your own IT resources or your own R& D really on your core product or your services. If you are, I don’t know, a pharmaceutical company, focus on how AI can help you with scientific discovery, discovering new drugs, for example.
Or if you’re an automotive company, how can you build AI into the car, for example, and provide a better experience to the car driver, for example, and so on and so forth. Or if you’re an insurance company, how can you add value to all your clients? This is usually where you should focus on build because that is the domain where you have really the deep domain expertise, and also where your customers are expecting from you in the next 10 years that AI is making the experience better, faster, whatever it might be in order to be able to compete in the market with the service and products that you provide as a company these days.
00:23:01 Chip Kleinheksel: Philipp, what you said where R&D is precious, so spend your R&D time on the things that make you unique and that are going to solve the problems specifically for your customer base. Leverage out of the box for as much as you possibly can. Arjun, from your experience, could you give us some examples of real AI-driven transformation stories? What does a successful enterprise AI strategy look like and what are some key signals that a company is getting it right?
00:23:27 Arjun Krishnamurthy: Sure. Let me share two examples. The first actually involves a large healthcare organization that’s committed to making their entire procure-to-pay operations lean and highly effective. And in this AI world, they want to achieve this, a fully touchless ordering to payables process, significantly reducing their cost to serve, and that’s the aspiration. What they did was initially they focused on automating the downstream activities. There are things around invoicing and payables and things that are more on the downstream, and then now, once you accomplish that, and most of these come out of the box, most of these are extremely easy to deploy, and you start to get those wins. And then once they’ve got those wins, they start to work themselves towards their more upstream activities in procure to pay, to optimize contract management, vendor onboarding.
So this sort of illustrates how you start with a targeted improvement and you can start to pave that from an end-to-end process transformation, and you do that in baby steps. The second example I want to give is of a manufacturing organization. They are looking for seek deeper insights into their manufacturing costs to enable better decision making. Especially with generative AI and the capability of generative AI, we are able to solve this hard problem because this requires a fine-tuned model that is advanced reasoning capabilities for strong understanding of business problem solving rather than taking a broad but superficial approach, what this company did was chose a narrow area, but went deep, and then they actually saw significant value in their decision making and actually saw tangible dollar savings. Now they’re gradually expanding that scope over time. By doing so, they’re able to address the complex challenges, effectively, demonstrate the tangible value and systematically broaden their impact.
00:25:24 Chip Kleinheksel: What I like though is both of those, two different scenarios, two slightly different approaches, but at the end of the day goes back to what we were saying earlier, which it’s still LEGO blocks. It’s finding something, picking a specific thing you want to go after, leveraging AI to make it more efficient, to make it more insightful, and then continuing to build upon that. Philipp, in your role as CTO and chief AI officer at SAP, you’re in a pretty unique and powerful position. As you well know, data is what fuels AI really in over 80% of business data generated in the world runs through an SAP system. What are you focused on and what are you most excited about?
00:25:59 Philipp Herzig: Oh, I think we have not enough time to enumerate all the things that I’m excited about, so I’ll give you just maybe three, four things. So I’m really super excited about the way how we are rethinking the user experience for the enterprise. For me, the days where we are sitting in front of GUI transaction screens or Fiori apps where people are typing in data or retrieving data from, they’re just over. If I see what we are doing with an iPhone or Android phones and so on, with real-time voice interactions right now, you can converse just freely with the system, ask it anything you want, and whether you’re on the car or on the go, that’s super exciting. So this whole idea that you can basically have the enterprise in your pocket and can ask it anything or just note down anything you want in the system wherever you are.
Then let me say the second thing really, of course is agents, because that will help enterprises of course, and not just our own agents that we are building again and finance procurement, but also with respect to this agent to agent protocol that we’re going together with Google and with Microsoft and a few others that we announced already to really create this ecosystem, this multi-agent system that is to me like the new dawning of the internet, if you will, that we are creating. They’re super exciting from a technology perspective.
And then thirdly, what I’m super stoked about is robotics. This work that we are doing also with Neura and with Nvidia, for example, to help basically humanoid robots that will come in next, because it’s just so obvious, on how we actually, with the integration, what we do with Joule and agents, how to help these robots, not just to move phenomenally on the shop floor, on the warehouse or wherever, but actually then interact with AI and the agents we are building with the SAP system because as much as of course, a human can interact with the system, a robot can do as well through AI models. So that’s super exciting work as well that we are doing. I’m really excited about this.
00:27:54 Chip Kleinheksel: It cracks me up a little bit because we spent as much time as we’ve spent talking and we haven’t even got into robotics and physical AI yet. That might be our next episode, who knows? All great things. To our kicker question, Arjun, Philipp, what’s one problem in your life you would love AI to solve for you? Arjun?
00:28:16 Arjun Krishnamurthy: I would love an AI agent that could attend all my no-value add meetings on my behalf, and then report back to me confirming exactly why that meeting could have been a simple email.
00:28:29 Chip Kleinheksel: I like it, but if you don’t show to a meeting, you’re going to have some people questioning why they’re not valuable to you. But I think it’s great. Philipp, what about yourself?
00:28:38 Philipp Herzig: Yeah, no, it goes along the same lines actually. So I have also not figured that one entirely out yet. I have some basic versions of that working, but it’s like how can I make myself more obsolete? And I would not say the meeting, I still think all these mundane meetings add value for somebody just not necessarily for oneself, because how often Chip or Arjun are you in meetings where just repeating the same stuff over and over again? You have your knowledge in your brain and you’re just playing like this FAQ master to somebody else answering all these questions. Why not just create a digital twin of yourself that has all this knowledge that you shouldn’t just go and you are really busy teaching your own twin and growing that twin by growing yourself and doing something novel in the world. So that is still what I’m looking for.
00:29:27 Chip Kleinheksel: Well said, well said. Thank you for a truly fascinating discussion. My guests today were Philipp Herzig, SAP’s Chief Technology Officer, and Chief AI Officer and Arjun Krishnamurthy, an SAP AI leader in Deloitte’s SAP Alliance.
00:29:42 Chip Kleinheksel: I’m Chip Kleinheksel, host of Resilient Edge, a Business Vitality podcast, paid and presented by Deloitte, and produced for Deloitte by BBC StoryWorks Commercial Productions. Thanks so much for listening. We’d appreciate a review on your podcast app. Goodbye for now.
