SAP's Muhammad Alam: Here’s How You Can Actually Get ROI From AI

The company's head of product and engineering says that AI is most valuable when it comes together with the application layer that produces the data.

This article is presented in partnership with SAP

In the early days of generative AI, some leading tech thinkers believed you’d one day simply dump all your data into LLMs and query it in natural language. Now, three years after ChatGPT’s debut, a more grounded view of the technology is taking shape.

Rather than operating as an amorphous blob, generative AI is starting to deliver the most value when it’s attached to firm data. The closer to where businesses operate the better.

In a fireside chat with me at SAP’s TechEd event in Berlin, Muhammad Alam, SAP’s head of product and engineering and executive board member, made the case that AI’s real ROI can be found in these more rooted uses of the technology.

“AI, when it comes seamlessly together with the application layer that produces the data that needs to run AI, creates the most value,” he told me.

Alam also spoke with me about the evolving roles of engineers and product managers, the importance of data in the AI flywheel, and whether all the AI hype is good. You can read our full conversation below.

AK: Hi Muhammad, let’s start by talking about the ROI on AI — is there an ROI? And what are you doing to ensure that the companies working with SAP are seeing that return on investment?

MA: This question is becoming more and more important for organizations and our customers. They want to know: What’s the ROI? And they go through a pretty long, drawn-out process internally to establish business cases before they can go through the adoption of AI.

My view is I’m very empathetic to all the things that are coming at the customer. Because anybody who’s somebody in the tech stack is going to our customers and saying, “Listen, we can create the most value from AI for you.” And it’s very hard, if you’re a customer, to make sense and make the right decision, if you will.

So we, at least from our point of view for the business application landscape, for our customers, are coming up with a very simple point of view, and that point of view has two parts.

The first part is: We believe that AI, when it comes seamlessly together with the application layer that produces the data that needs to run AI, creates the most value. So if you plumb AI on top as a separate bolt-on, you’re going to have to take on the complexity of making sure you have the data somewhere, and then the data gets access to AI, or AI gets access to data, alongside the right authentication, the right security, the right governance—which is a significant amount of tasks. And even if you’ve done that, where do you make the AI available for the end users to consume?

Which is why, for us, the applications and business processes that produce the data, that give them the most context, that powers the AI, and the AI that’s embedded back into the application—this has the potential to create a level of seamless value and ROI for customers that’s otherwise very hard and cumbersome for customers to go do on their own.

The second part, as it relates to this app-data-AI flywheel: you can argue, “Great, if app, data, and AI seamlessly come together, it creates value. We can understand that. But we have apps from five or six or seven different players. Should we have five of these or six of these app-data-AI flywheels?”

And this is where we believe, as SAP, we also provide unique value in that the business suite of applications we have covers the most significant breadth of business processes, unlike most others that have their specialized areas. Finance, spend, supply chain, HCM, CX—all of these applications with data harmonized across the breadth of those business processes, powering AI, allows you to get to that global maxima, which otherwise would be siloed into local optima if you have five different best-of-breed or five different disparate applications as well.

So the second place where we believe significant ROI gets generated is the breadth of building context that comes out of the box, harmonized in Business Data Cloud with AI, and that data powering AI seamlessly embedded in the application. So both the flywheel itself and then the breadth of it—we believe this creates value, and that’s what we look at as SAP, as our unique value proposition to help create the ROI that we can deliver for our customers. That’s one aspect of why we think we can help solve the ROI equation in a much more unique way than largely anybody else out there.

But the second part of it is—and this is a point of view that we articulated a couple of weeks ago—we do think that because there’s been so much hype in AI, and there’s now realization from customers saying, “Well, I’m not really seeing the value,” we have to sort of simplify the equation. This is where we outline the point of view that, listen, AI is there, at least in this first step, to make people smarter, to make people more efficient, to make people more intelligent and productive, if you will.

And hence, what we’re coming up with is a concept of assistants. You’ve got an organization with a role construct. Today, you have an accounts receivable clerk, you’ve got a billing clerk, you’ve got a planning person, you’ve got a controlling person. We want to make sure that each one of these personas has an AI assistant that can have any number of agents underneath. But this assistant, for me as a human doing the job, makes me smarter, faster, more efficient.

Now, over time, as the human in that role gets more comfortable with those sets of agents in this assistant that’s there for me, and they build the confidence that the recommendations are the right recommendations and it’s taking the right actions, then you move towards autonomous execution. So you can say, “Hey, now that I have the confidence, I’ve made this role 20% more efficient, 40% more efficient, 60% more efficient. Let me let it run autonomously for a portion of my business process”—not just within a function, but you can do it across functions.

A demand planning assistant can sense the change in demand, send it to a supply planning assistant. Both have multiple agents underneath, multiple tools. Then they can send it to a procurement assistant that can go through a sourcing process and procurement process and a contracting process underneath. But that’s how autonomous execution of doing processes comes into play.

So we believe this ability to create unique use cases that seamlessly bring app, data, and AI together, alongside positioning it first and foremost as there to help the individual, the human, get smarter, more efficient, more productive, leading up to autonomous execution—this can create significantly more value. That’s the strategy we’ve taken that seems to resonate quite a bit now with customers to think about how they get to ROI. It’s a longer answer, but hopefully it makes sense, because I think that’s a very nuanced point of view.

AK: It’s interesting that you bring up the hype. When generative AI first started taking off, some people were saying, “We might not need Excel in the future. You just upload all your data and then query your data and it spits it out.” But what you’re talking about is actually very different. Your argument for the ROI in AI is that you actually need a lot of structure. Am I reading that right?

MA: Exactly. if you look at the tech stack, we believe—I believe—the tech stack hasn’t changed in decades. You’ve got IaaS, PaaS, and then there’s SaaS at the top, and the application layer creates the most value, as I’m sure most people here would know.

I believe, fundamentally, for the first time in decades, there’s a new layer that’s being added to the tech stack: AI. Now that AI layer can’t exist without the stack. Without the stack, you can’t say, “Hey, I’ve got business apps” without a platform or an infrastructure underneath. You get that as a service from a hyperscaler, and you modernize the infrastructure as part of the past decade’s work.

So this AI layer, which is new, needs the data in the app layer underneath. Where else does the data get created if you don’t have the apps? Where else does the action get taken from the AI, from the agentic agents, if it’s not in the applications? You need the level of governance and structure, the localization, the compliance that the applications provide.

AK: Do you like the hype? It does get people in the door. Every time Sam Altman speaks, thousands of headlines are written and awareness increases. On the other hand, you also have, because of this, a lot of companies saying, “Where’s the ROI in my AI?” So what should we think about the function of the hype in this moment?

MA: I think I like the hype to the extent—and by the way, I think I’m also allowed to ask you questions, because I’ve heard you talk to a lot of smart people around the industry as well, so I’d love to sort of hear your thoughts around this too.

My view is the hype has value. Because what the hype is doing, for the first time for AI, is creating this belief that real value can be achieved with AI—and not just that it can be achieved, but it needs to be achieved. It needs to be achieved now and as quickly as possible, because others are doing it, and if not, we’re left behind. And with that comes budget and commitment and organizations that are there to help make that happen, which I think is a very positive development.

But this is where I go back and say: Now that you have that, how do you make sense of the hype? Just because you can go generate an app through a large language model, as some of the large language model providers would say, does it make sense for you to go generate a procurement app? What are you going to do next year when you have to enhance it, when you have to have compliance around it, when you have to make sure it’s in different languages? Is that really your core business—to take a procurement app generated just because you can, and manage it across your global landscape as a core function? It makes no sense.

But of course, large language models and this ability to generate apps would be helpful to replace apps that you’ve custom-developed anyway, because they don’t need this level of rigor. I believe low-code/no-code is going to be completely dead because of you being able to generate much higher fidelity apps, if that makes sense. But hyping it up to say, “I’m going to go rewrite my own application just because I can generate it”—that goes into a more self-serving position from whomever is providing that point of view, as opposed to one that is serving the needs of the customer.

That’s where, honestly, we try to balance at SAP to say, “Listen, are we participating in the hype?” Now, because I will tell you, Alex, there was somebody who actually wrote it out in the public domain and said, “The number of agents SAP announced at SAP Sapphire, we could have sent an email. We didn’t need a conference to be able to go talk about this.” So we actually end up taking a much more pragmatic view that’s grounded in delivering the ROI, as opposed to hundreds of agents that nobody really knows how they fit into your organization’s ecosystem.

But again, now I’m curious as to what your point of view is on ROI adoption, the hype, and how that’s translating in terms of how customers understand it.

Conflicting Studies on AI ROI

AK: Well, let me start this way. There have been multiple studies about whether enterprises are seeing an ROI from AI. MIT had this study that said 95% of companies are not seeing ROI from their AI projects. Wharton, last week, came out with a study that said 74% of companies are seeing positive ROI from their AI projects. I don’t think you need to be a math major to say one of those has to be wrong, right?

The problem is there’s not good data yet, because this is so early. You can even go to different parts of the corporation and you’ll get completely different answers about whether there’s an ROI from AI. In fact, one of the most interesting things about this Wharton study was VPs and above are much more likely to say that their companies are getting an ROI from AI than managers. And if you’re on the manager level, you’re probably the one that’s implementing it. So clearly, the hype is impacting the highest level of companies and not the lower level who’s seeing that this is actually tough.

The studies also have interesting data points where you can really start to see exactly what’s happening. Yes, there’s a difference in seniority, but there’s also a difference in company size. In fact, the Wharton study found that 74% of companies say they’re seeing an ROI from AI, at least those that measure it. But among big companies, it’s only 57%.

And some people have looked at this and they’ve said, “Well, it really comes down to a company’s ability to adapt.” If you try to fit AI into your current business processes, you’re probably not going to see the same results as if you rethink your culture, rethink the way that you work. So I’m actually curious to hear your perspective on this—How much is adaptability key to being able to see a return on investment on AI?

Trust and Adaptability

MA: I think trust, in my mind, before adaptability, is critical. And this is why I believe that if customers, if people, take the approach to AI which is, “For AI, you need to completely go rethink your business processes and then approach deploying AI,” I feel like that’s going to be a much more elusive end state than being able to say, “Let’s start with making the roles that run our company, that understand our company, faster, more efficient, more productive, if you will.”

As you then build trust—because every company, largely, can say, “We want to change” and “We want autonomous this or autonomous that,” but they still, I would guarantee you, nine out of ten would have an AI ethics policy that says, “We want a human in the loop,” right? And then who’s that human going to be if you’re going to completely change every organization?

So the first step, we believe, needs to be: you make the roles and the construct smarter, more efficient, deliver the value, if you will. And as you build that trust and that set of capabilities that allow you to autonomously execute all the different functions of, let’s say, an AR clerk or billing assistant, then you can say, “Hey, listen, maybe I’m getting to autonomous billing, or I can get to autonomous financial close.”

Or you take an example that’s much closer to being autonomous today, but that’s how it started, which is customer service agents, where a case comes in. First, what you need is to be able to get a recommendation based on your knowledge base to say, “Because of this case, because of this knowledge base, this is probably the recommendation we should send to the customer.” Are you going to go do that without at least having somebody first in the loop to validate whether that’s the right recommendation or not? Hopefully not, probably not.

Once you build that confidence, you can say, “This percentage of cases I have high confidence in, I can just let the AI agent work through actually crafting the response, sending it to the customer, waiting for the feedback from the customer to say, ‘Did it resolve my case?’ and then closing the case, and maybe even doing a survey afterwards.” And there you have the touchless service experience. But it starts with the roles you have, with an ability to build trust step by step towards that autonomous execution.

And then there’s a third step as well in that journey—in that value journey, if you will, which is more like deep research—but maybe we’ll get to that separately. But to me, I believe that we have to start with the humans, the people that are in the roles, to be able to build the trust, to be able to think about how to then go either fully autonomous or think about much higher value work that organizations are ready for.

AK: So starting with humans before you get the technology? That’s kind of a lost art.

So let me ask, before we move on to our next segment, I cited two studies: MIT saying 95% of companies are not getting an ROI from AI, Wharton saying 74% who measure it are. What do you think is closer to the truth?

MA: I feel like the truth is in the middle. It’s probably the most objective answer I can give. I do think a lot of customers have started getting what I’ll call initial value from AI, but it’s not to the extent that the hype is talking about, to be able to say you can just have autonomous this or autonomous that. So I do believe there’s a level of value creation that’s happening.

I also believe it differs by the function and the industry. So there are certain functions that lend themselves well, particularly for a large language model, to be able to create value—functions that rely more on text. So if you look at customer service we talked about, if you look at HR being able to do your performance or your goals development, if you look at marketing, those functions benefit first, before you get into more tabular stuff, which we’ll also talk about later, just as a teaser for the keynote, when we’ll make announcements particularly around tabular data, which is critical for business AI to be able to function.

You also see more progress in what I would call software development, because the industry that arguably is the most impacted by large language models is software development. So if you ask me, from an SAP perspective, our 35,000 to 40,000 colleagues—are they using AI today to do things in a much more productive way? The answer is absolutely yes. But do I feel like we’re done? Absolutely not, because I feel like the productivity throughput that is left to be gained is far more than what we’ve gained. But have we gained significantly? Absolutely.

From that perspective, I feel like it depends and it varies. I do think the answer is probably in the middle, but we certainly have more runway left to cover in terms of creating and realizing value from AI than the distance we’ve covered so far.

Impact on Software Developers

AK: So software developers, the ones that you mentioned, are the ones that are feeling the most impact here. And oftentimes, when we look at tech, we can see what’s going to happen for the rest of the economy. So what type of changes are you seeing in the role of software developer? And then what is SAP doing to help developers evolve as their roles change?

MA: It’s a good question. In terms of the change that we’re seeing from a software development perspective, I feel like there’s mixed sentiments out there on this topic. You wake up one day, you’ll read, “AI will effectively kill the role of a developer—you won’t need to write code anymore.” So what should you have your kids or your nephews or your nieces studying in school?

And on the other side, there’s so much software still left to be written. None of us are short on backlogs. My backlog, even though I have 35,000 to 40,000 people, is probably worth 200,000 or 300,000 people’s worth of backlog to be able to go execute on.

AK: Can you just say that again? So you have 40,000 people, and you think that you have enough work to hand to 200,000?

MA: Yeah, our backlogs are never-ending.

AK: So I just think that’s so important, because when people talk about, “Will AI automate developers?” they just think it’s enough to just take away the work that’s done today. And the response that I think of always is: you’re assuming that people are happy with what they’ve built and they’re done. But clearly there’s so much more to do.

MA: There’s so much more to do. Not just for me—I believe largely in development, it’s going to be about acceleration. Now, that does mean, stepping back, that in aggregate we didn’t reduce from an SAP perspective this year. That’s not our plans for next year, obviously, keeping all other external factors constant. What we’re focusing on is taking the staff we have, even growing it, but making sure that they have the right tools to be able to deliver innovation faster.

So a product manager—you’re talking about how jobs could change or should change—a product manager at SAP now should be able to generate an app that he or she takes to a colleague, an engineering development manager or development lead, to be able to then complete, if you will. So this gives acceleration. A software engineer should be able to leverage all the tools that are out there to be able to accelerate doing unit tests, writing code, generating code on top.

Think about design. Design is a significant one as well. So all fundamental roles have now an ability to rethink the value and acceleration they can gain. And that’s what we are doing already within SAP with a bunch of front-runner teams, as we call them, to evaluate it by various dimensions.

And some of the productivity examples we’re getting are pretty phenomenal. One team would come in and say, “In this last sprint that we did, we changed our processes, we changed some of our role definitions in this or that way, and we actually produced more throughput than we did the entire previous quarter,” which is pretty phenomenal. One team would come in and say, “We did A, B, and C, and this sprint, our throughput was 7x what we were predicting, and we were just learning this. Frankly, the next one we think is going to be 12x.”

So even now, going back to the question before, where would you put this? Would you put this in the bucket of “We’ve realized the value,” or in the bucket of “Listen, we think phenomenal value is still left to be gained”? The runway that’s left is still far more. And again, for me, being in the tech space, in the business application space, this is exciting—not because I can take the 40,000 colleagues and say, “Great, now I only need 10,000,” but because now I can take the 40,000 colleagues and say, “Now we’re producing at the pace of 200,000 colleagues with the help of AI agents that are making our developers much smarter and faster.”

Now, closing with the question as to how are we helping developers: if you think about SAP Build, we’re going to make some significant announcements that we’re bringing AI agents and tools to SAP Build in a way where it meets developers where they are, so they can leverage the best of the tooling out there to accelerate their development as well. And not just that, because we’re SAP and we bring that seamlessness of AI, data, and app layer together, the out-of-the-box context, process context, data context, and understanding is already there in Build. So it allows you to produce applications and extensions at a much faster pace than you would be able to do on your own or with a third-party platform, if you will.

So those are a few examples. Again, there are a lot of AI agents in HANA Cloud. There’s a Joule Studio agent builder experience that we’re launching as well, where our customers can build agents in both a low-code/no-code fashion as well as in a pro-code fashion. So the list there is pretty long, but these were some of the highlights.

The Future of Product and Engineering Collaboration

AK: You have such a fascinating position because you run not just engineering but product at SAP. I wonder what you think about how product and engineering will be able to work together in the future.

My view of an engineer is often someone who’s like, “Let me work. I’m going to build it for you. I’ll come back, you tell me, and I’ll adjust it.” Now they have pesky product people being like, “I built it.” So how are these two divisions going to be able to sync up? Is it going to cause tension or closer working relationships?

MA: A lot of it, I think, we’ll find out in the near future, at least from an SAP perspective. I talked about how we’re doing a bunch of different front-runner projects here that will help us shape a point of view as to what is the evolving role of a product manager, an engineer. What should a scrum team look like, frankly—QA, unit tests, design, if you will.

I do think that because there’s value on all sides—I mean, you could generate an app, but particularly in the class of applications that we work in here, that app is not something you’d want to go put in production. You still need that finalization of the last mile that you would need to take that app to. So it’s still going to be a partnership, a very deep partnership.

But should the ratios of what we know of a scrum team in terms of a PM to engineers or devs change? Probably. Should the ratio of how many designers you need on a scrum team change? Probably. But I think these are the things we’re going to figure out along the way, if you will.

Should the role of a product manager evolve from being a product manager to a product builder, and a developer just one that takes that and makes it production-ready? So I think there’s a lot to be figured out, and I agree—I think it’s exciting, and that’s the exciting part of my job, a lot left to be figured out here.

But I also believe that there are different types of applications, right, and different types of work you need to do. There isn’t going to be a one-size-fits-all, and there isn’t a one-size-fits-all even just within SAP, in the sense that there are teams that are building net-new products and innovation where this concept could work really well. There are teams that are building capabilities on existing products and driving innovation. There are teams that are saying, “Hey, on this product or this semester, we just want to do lights-on work, or we want to do fundamental, foundational work, but not innovate.”

All of those things will use different tools and different ways of approaching it as well. So I do think we’re going to see a view of what should the roles look like based on the kind of work that you’re doing. Not everything is going to be one-size-fits-all, even in the eventual state, if you will.

But again, similar to the other question, Alex, I’m kind of curious as to what are you hearing now broadly, because you also speak with a lot of not just tech companies, but companies that now—every company is now a development shop. They have developers in-house to be able to build applications and extensions as well. What are you seeing?

AK: A couple of months ago I wrote a profile of Dario Amodei, the CEO of Anthropic. And for the profile, one of the things I had to do was go speak with people who were using Claude Code. And there was this fashionable thing to do then: you’d run as many Claude Code agents as you possibly can, and then you see how much value you get from it.

People were spending $200 a month with Anthropic and getting thousands and thousands of dollars of tokens out of it. And I spoke with this one developer who said, “Yeah, I’m running like seven or eight instances of Claude Code at the same time.” So you end up having somebody who previously would be limited strictly by their time—now they’re effectively running a team of AI agents coding things up.

So from my perspective, a developer goes being somebody who’s sitting at a keyboard coding things up to orchestrating teams and actually managing these groups of AI coding bots. Now we’re seeing that these bots are starting to be able to code autonomously for hours on end.

I think it’s also going to be the age of individual empowerment, where somebody who would effectively be limited by their own capacity, now they’re limited by their imagination and maybe the budget they have. Because Anthropic did put in some rate limits, but it’s still extremely flexible if you’re on the pro tier. So that’s my perspective.

MA: Again, this is where the math of 40,000 to a throughput of 200,000—it’s not because we’re going to go to 200,000 people. It’s because we do believe that, in effect, AI is going to be a team for everyone, and not just one. You might have multiple agents working for you. Everybody, then, in effect becomes a manager of a team of AI agents, if you will. So as an engineering manager, I should be able to say, “Hey, how many AI agents do I want to do the things that I was originally doing myself?”

Data Strategy and Partnerships

AK: Okay, let’s talk data. SAP made a big splash in February with this Databricks partnership. So what’s the latest on your data system?

MA: Data is critical for us. As we talked about, the app-data-AI flywheel is what makes both the value proposition that we have and the value we can create for our customers. But also, it’s one that I believe needs to be done seamlessly for value to be created and ROI to be realized. The data layer is obviously one of the most critical ones because that’s what powers AI.

And for us, we launched Business Data Cloud in February. But it was important because we were launching it in 2025, and a lot of data work and data platform strategies have already been set for a lot of our customers. So we needed to have a very open strategy around embracing the ecosystem perspective.

So we launched it with Databricks, with both aspects of embedded capability of Databricks within SAP BDC. So if you are looking for an enterprise-wide data platform, you can run it within BDC. But in case you’ve already made a platform decision, a data platform decision, at least the value that you get from your SAP data being governed, managed as data products, and maintained over time—which otherwise you would have to do with people, humans, or other tools that you would spend a lot of money on without as much value—these governed and managed SAP data products that we produce in Business Data Cloud, you can then share them out to whatever platform you have. In this case, Databricks and Google Cloud BigQuery is what we announced, and you can zero-copy share them out to them.

We’re also going to make an announcement today with Snowflake. With Snowflake, again, we’re taking a twofold approach. One, if you have a data platform in Snowflake, you can zero-copy share your SAP data within BDC to Snowflake. But if you want to use Snowflake as your data platform strategy and you don’t already have it, you can do that within SAP Business Data Cloud as well. So we’re also making Snowflake available as a solution extension within Business Data Cloud as well.

So we’re not done with our partnerships, but we’re excited. Now with Databricks, having the embedded version in Business Data Cloud and the zero-copy share; with Google BigQuery, we have the zero-copy share; now with Snowflake, we have both the zero-copy share as well as a solution extension capability that can run within Business Data Cloud. So we’re excited about providing a solution to our customers in an open, flexible manner on the data layer, if you will, that allows them to govern and manage the SAP data but then also bring in their non-SAP data as well.

Quantum Computing

AK: Okay, let’s end on a fun one. Muhammad, do you have a prediction for the year that we’re going to see quantum roll out?

MA: I think we probably need a separate session on AI versus quantum. I mean, we need to get AI rolled out. That’s near and here, top of mind for customers. We do believe, from an SAP perspective, that quantum has a lot of potential to offer—a little bit different use case than AI. And we’re working with some of the early players to be able to make sure that, from a business application perspective, the right things that can run in quantum do run in quantum, so our customers can benefit from the performance and the speed and making possible what was impossible before, and being able to solve complex problems.

But in terms of, are we close enough to shipping something or making something generally available—and not because just because, but because of where quantum is—I think it’s probably not this year. But there’s a lot of early-stage work that’s happening. But I’m also curious what you think?

AK: I believe after 2030—that’s my guess. If you gave me over/under 2030, I’d say after.

MA: I’d say after as well.

AK: All right, thank you, Muhammad.

MA: Thank you, everybody.