From Promise to Practice: Preparing for Agentic Automation in Business

Transcript

Most companies today have some form of AI in place, but very few are actually ready to scale it across their business. In fact, nearly 90% of companies now use AI, but only 1% say they are truly ready to scale it. That's the challenge we are diving into. Agentic automation isn't something you can just switch on. It requires solid data foundations, streamlined processes, and clear governance.

And this is why it becomes so important. Why so? Well, a reason could be the growing wave of shadow AI, where employees use powerful AI tools like chatbots or code generators without IT oversight. With shadow AI now ranking among the top enterprise risk, and many organizations still stuck in pilot mode, we'll explore what true readiness looks like and how keeping humans in the loop can help turn AI ambition into real world impact. Welcome to the Power of Possibility.

HOST: Okay, so today we have a very special guest joining us. From SPS, we have Christian Schierjott. He's with us today. At this point, Christian has definitely become one of our frequent flyers on the podcast, as this is already his third time collaborating with us. He's the Head of Digital Transformation and Innovation at SPS, and we are thrilled to have him back. Welcome, Christian.

CHRISTIAN SCHIERJOTT: Thank very much for having me once again.

HOST: And we also have with us Bertram Weiss. He's the Vice President of Health at Merantix Momentum. He's leading the company's healthcare AI strategy. He brings decades of experience in Pharma R&D, Computational Biology, and Data Science. His work focuses on turning complex biomedical data into AI solutions that could help to provide better diagnostics research and patient care. Welcome, Bertram.

BERTRAM WEISS: Yeah, thanks for letting me join this podcast.

HOST: Okay, so I'm going to start with a question to Christian. There are a lot of topics and a lot of trends around AI right now. And I always prefer to start by setting the scene. What do you think are the most common misconceptions companies face when preparing to implement agentic automation in their core operations?

CHRISTIAN SCHIERJOTT: Yeah, I mean, you touched a little bit on that in your introduction, right? I mean, what we see is the evolution of the current state of AI technology with the large language models in use since now over two years, going in the direction of agentic AI, agentic automation. And then on the other side, you see where we come from, right?

Definitely something that we experience with companies currently when we talk to them is this plug-and-play myth that companies think it's very easy to just implement generative AI applications in the infrastructure. And I mean, who would harm them, when I listen to the promises that some of the technology companies make, saying you just have to have some proper API interfaces and then you just plug it in and it runs.

That was absolutely possible when we think in the era of RPA, for example. If you have a simple process with RPA you could just plug it in and probably with drag and drop build your own process. But if you want to have a meaningful end-to-end automation here based on large language models as we aim to with Agentic AI, then you really need a deep integration in the entire company architecture. You need to somehow get the data, you need to have the interfaces along the complete end-to-end process.

You have to think about the transition of the experience you need to build up a meaningful AI agent. You have to train it. So this shows this is not a surface-level implementation that we do here with this technology. If you really want to have the full potential of AI, you need to do your homework internally. That's one thing.

Agentic automation isn’t plug-and-play. If you want the full potential of AI, you need to do your homework internally. Christian Schierjott

Another thing is, and we talk a lot about that, are use cases. I know a lot of competitors and tech companies or consultancies that claim they have over 100, 300, 500 different use cases. I mean, that's not a race for quantity. You need to have a couple of use cases, but highly quality use cases at scale over your processes where you can really learn from it and where you have a meaningful impact that you also see in your revenues in the end.

So it's not jumping on the next knowledge database and integrating it into existing technology, but really thinking how you can get the best and the most out of your use case.

Thirdly, we look a lot at processes. And yes, it's true with the use of generative AI, it's possible to automate increasingly complex processes, but this also has its boundaries. If you have very fragmented end-to-end processes defined with a lot of different departments, data sources, which are not very well integrated into your infrastructure, it will remain challenging to implement an end-to-end AI application that seamlessly communicates over the different process steps.

I think at this point in time, companies are very much in the trial-and-error phase, which is good. You always have these innovation cycles, but I would say that in a few months, over the next year probably, most companies will see that it's not as easy as plug-and-play or drag-and-drop. They really have to do their housekeeping.

And we see now already that most companies will not get that done themselves. So I would say the fourth myth here is really to try it only themselves and not rely on partners which can scale the entire development of their AI stack.

HOST: Okay, thank you, Christian. From your experience at SPS, what are the essential steps that organizations must consider before deploying Gen AI across business-critical processes?

CHRISTIAN SCHIERJOTT: So, beside what I already said, to find the right use case and prioritize first, which takes some time. It's not a one-time effort to just select the right use case and also to look into process standardization, process improvement, etc., because you just don't want to have a non-functional process and put it into the AI. They won't fix it. So, really have to define your business rules. You have to make process descriptions available from someone. The AI needs to learn your specifics of your process, your individual business rules that you have. And only then you will get the full potential out of it.

AI will not fix a broken process. You have to standardize, clean, and define your business rules first.Christian Schierjott

But besides those two, we see a third thing or topic which is very important: the data availability. It's nothing new, right? If you look back in history, I mean, since we have technology, technology is as good as the data you feed into. And this is the same case with AI or generative AI technology. See what happens if you use Google or ChatGPT, you see it still after two years of heavy development, so many cases of hallucination and where we cannot rely really on the output to 100% and it gets even more relevant if you use it in a business context or in an industrialized operational context. And this mainly results from controversial or contradicting data, bad data quality, not unique data.

Many companies will learn very fast — or did that already in the past — with knowledge databases where you just feed your knowledge that you have, actually your data into those engines that you won't get the results you would expect because you didn't clean your data or you have not all data available. In this case, even more important, if you have end-to-end processes where you have interconnection between different process steps, different business rules, you need to interpret data. You need to make data available across all the different systems that you run or databases that you run within your company. And if that is not given, you will have not the outcome that you expect, right, where you can really benefit from.

And I would say that the last point here that I'd like to mention, it's a little bit focused not on the technology itself. I think it was Boston Consulting Group who said there's a rule that only 10% is algorithm, 20% is technology, and 70% is the change management or the organization. And we as a people company, we know that. We're dealing since two years very intensively with generative AI and the development of our own processes, but also on client side and to really get the organization behind this. I would call it really a shift in technology and how to handle processes. To get the people behind it is very difficult for various reasons. In our experience, the most time that you will spend changing towards agentic automation, you will spend with change management and dealing with skill shifts, with understanding, with rule sets, etc.

HOST: You mentioned already some of the reasons why this is so important, but why do you think data readiness is such a critical enabler for generative AI, especially when scaling across multiple departments or service lines?

CHRISTIAN SCHIERJOTT: Well, I think the biggest challenge that we still face is the limited explainability of the algorithms. So if I don't start with the data that I feed into the system, I make it even worse to interpret the outcome or to interpret the data and then seeing maybe the wrong outcome. I won't get the system audited because I'm losing a little bit here the database. This is maybe one thing, this is tied to the models.

The other thing, and I touched on it before already, really that especially when you have business processes, you have for example from the data entry out of an insurance claim or a client request in banking until then really it's solved to the outcome, that you touch several business departments, several different process areas and all these processes or process steps need to rely on each other and this is mainly based on the data that are provided within the different business rules and if partially you have only one of these segments not in order, you won't get the outcome. So this makes the challenge really over the entire end-to-end process with the right or with the correct data availability and quality.

HOST: Getting into that specific point, it's great to also focus on what are the stops or the risks that not having proper governance could take. From your perspective, what are the biggest risks of deploying autonomous workflows without proper governance? And how can companies mitigate that effectively?

CHRISTIAN SCHIERJOTT: Governance topic is a very important topic. Let me refer that to this technology cycle and where we started two years ago. I think it's normal that you have a lot of trial and error and then the government and politics will just run behind it and I think that they're catching up pretty fast. We have since a couple of months the European AI Act in force. We have in Switzerland the FINMA circular which governs the use of AI and we also have in the US already a couple of states that released governance around the usage of AI.

And if you now look at — and I'm saying this is good, right? — because when we see that especially in the BPO area for example, we're with very sensitive and highly regulated services. Just an example here is the entire payments area where SPS is processing most of the payment slip relevant business here. We have a lot of regulation already, even without the newest technology. And knowing about the limitations of AI and what you can do with AI when you don't use it under proper governance is really system critical. So for Switzerland, for example, for the entire payments area, but also highly regulated and also complex processes such as payroll processes, where we're dealing really with people, where little mistakes lead to bigger failures, there is an absolute must for regulating this to a certain extent. Because if you have compliance violations, which mostly will lead to reputational damage for your company and end up in fines and financial loss.

HOST: Okay, and how can human-in-the-loop frameworks help organizations to maintain trust, compliance and control in this kind of AI-enhanced workflows?

CHRISTIAN SCHIERJOTT: So, I hope you enjoy. I think at the current state that we are in with language models and the development of AI automation, the human loop is a very critical part. Especially, I mean, we have today when we look at the payments processing, we have four or six, even six principles with critical payments that we process. And this will not go away, especially not if you have quality SLAs for your process, which are beyond 99.9%. You need to have it 100% correct what you deliver. And again, given the limitations that we still see out there, it is not possible to really have an autonomous process design and technology, like really a co-worker, as it's promised by the tech companies, we really need the humans in the loop to ensure the data quality that we have on the output side.

Although we see that work and the skills will shift towards from a simple operator doing data entries to a more quality assurance kind of work. It's not that we automate the people component. We will see over the next months and years, we will see a shift in the skills that people are doing different type of work, but will remain very important in this whole end-to-end automation strategy.

HOST: OK, thank you, Christian. I also wanted to have the opinion of Bertram Weiss on the current situation. Many companies in Europe, for example, feel pressure to implement AI, but struggle with legacy systems or even strict regulations or risk-averse cultures. From your perspective, what are the biggest structural and cultural barriers preventing successful AI adoption today?

BERTRAM WEISS: Miguel just mentioned already some of them, right? I mean, first of all, of course, especially in Europe, we have a lot of champions in the economy who have under-invested IT departments and under-invested IT infrastructure and landscape, which makes it very difficult if you have these systems designed decades ago. They were more or less prepared for security and on-premise. Now you need very agile systems that can have a very fast acting interchange of data between agents. This is not going to take place with these very old IT systems, so I think we are paying here really a credit to having spent many years with IT departments spending like 2% in healthcare and pharma, for instance, especially in healthcare, 2% for a hospital on IT. When they now really want to go digital and use AI, that is not going to work.

So this is one of the first hurdles they see. They are sitting on old infrastructure and now they need to invest a lot to make it actually AI-ready. Let's call it like this. But this is not the only problem. This is the tech side of it on one hand. But you also have the cultural topics where we see that in Europe especially we have this very compliance-first attitude and the fear of something could go wrong, and Christian just mentioned it, that we are 99.99% control or correctness that we want to have and we have a six-eye principle, and that's not going to go away.

Now, of course we need an attitude that is more, let's say, open to exploration, experimentation, also to handle errors from a probabilistic perspective. That means maybe it's not a good idea to start with a process like payslips, where you have a 99.99 certainty, and take out the people first, right? There, I would not start with an AI transformation in a company. I really would look for processes that are a pain in the neck of many people, that are very redundant and repetitive in what they actually accomplish and that's really a routine task. And then go from there and say, okay, is this a process that even if an invoice goes, for instance, wrong, yeah, this still can be fixed somehow? It cannot go maybe wrong with a payslip or it cannot go wrong with a therapeutic decision making, right? There we still need the human in the loop and we all understand, I think, that if you go for an AI implementation, of course, you need to very thoroughly check these kind of questions.

Is this where I need a human in the loop at the end and take just a routine load from employees, but not, as Christian said, the control function and the governance function to really make sure we take the right decision here? Or can I go really to an agentic AI where I think I can manage the errors that will occur, because with human beings they also occur. But can I manage them from a probabilistic point of view and have a staff that is also, let's say, eager to experiment and deal with errors?

The biggest risk for European companies is the temptation to wait and see. While others move fast, we focus too much on avoiding failure instead of creating value. Bertram Weiss

We have in Germany especially, for instance, a very much attitude of no errors may occur. And this is not going to work for AI as it is not going to work for people. Because otherwise we wouldn't have a six-eye principle. Because also humans do the error, right? But people expect AI always to be perfect. They do not expect this from their staff. And I think there we need also a cultural shift. And I think the biggest problem actually I see also in the culture is the temptation to wait and see. It's the biggest risk for European companies. They think they can wait another year or so until to really move forward. They want to be perfect before experimenting and this mindset is what's going to make a huge problem for them.

Because in the meantime, we see that in the US and in China, people are moving forward with really amazing speed, thinking rather about the value it can contribute and less about the risks and failures it can produce. Somehow they will manage it with the risks and failures, but they focus on the value. And we are very much compliance-first, avoid failure first and then we think about the actual value. And these are big hurdles for us in Europe to really come forward. While at the same time we all appreciate we have governance and control and checks in place. But we may end up as a museum of economy, of the old economy, while the US and China are moving forward. And here we have a very nice lifestyle and rich culture. But technology-wise, we are really becoming a customer and client to the other big economies.

HOST: Okay, and what happens when the fail arrives? Where do companies most commonly fail when introducing AI? Is it the technology itself? Is it the leadership mindset? Or maybe what we've been talking about, the data readiness, or maybe just the organization itself? What do you think?

BERTRAM WEISS: Yeah, so I think there are definitely several breakpoints and we already talked about the dimensions: technology, culture, organizations, and you have breakpoints in all of those dimensions and they are different ones. And typically when things fail, we see a concatenation of that you had errors or mistakes or failures in several of them. And in combination, then they really make stall the project and the idea, right? So we have developed Merantix Momentum with the University of St. Gallen strategic framework, we call it the AI Canvas, where we look at AI use cases through all these dimensions.

So first about the value, where we say, OK, first of all, we need to make sure the AI use case actually pays into the corporate's protocol, corporate strategy, and that secondly, there is a clear KPI we can measure to see if we are successful. And then, of course, we move on to the tech part and look, is the data actually ready, available, accessible sometimes even? Is it also, from a quality point of view, ready for AI? Or do we need first to invest in that area? And also, do we want to have a system that is as agile as Christian also mentioned? You need very agile systems and not legacy. If you build it on legacy IT, it can be very slow and then people get very frustrated or customers and clients can get very frustrated.

So we look also on the technology side: where are there the typical errors? And then you have the whole point of cultural topics, and I would also call societal topics. Like we already talked about the governance. But of course, also the cultural changes. You mentioned shadow AI, where people actually already use AI without organizational oversight. And there also, we see, of course, big risks to fail if we miss to find the right framework from a governance perspective. And then, for instance, a model could thrive away in terms of bias, or it could become obsolete because new data has come in and it's not managing the new data properly and well, right? So there are a lot of breakpoints there as well.

And I also would add the management tier. We nearly always see also that AI use cases fail because of a lack of support from the right level of management. This is critically important and missed by many leaders. They think this world will come bottom up and they fail to lead for their AI. And I'm not talking about governance so much, but really support the people in getting this big transformation somehow done.

This Canvas helps to find all these breakpoints and discuss them early on before you actually start off, right? And that reduces enormously then the risks later in the project. And that's why this AI Canvas has been so successful for us, and we have delivered now over 200 projects in the last seven years for many corporates. And we always have seen these kind of things being important and, if valid, rest being rather a success factor than a failure breakpoint.

HOST: Okay, if you could provide a list of the internal steps that any company should transform internally to make AI truly effective rather than just a surface-level innovation, what would it be?

BERTRAM WEISS: So I think typically the first thing you should do is educate your people. And that starts with the leadership team, but also of course staff because innovation must come from all leaves of the tree. So people need to understand what this technology actually is so they can ideate about how to make use of it in their daily lives. So education comes first, then you can do ideation. And from there you come to the many use cases. And Chris had mentioned some companies are sitting on hundreds of them. I can really only confirm this. That is what we see already. They have done this ideation.

But then what we also see is they very much lack a structured process to make out of these many ideas a portfolio that is going to work for the company. And here we introduce mostly the proof of value and the proof of technical feasibility. Then you very early on check if you actually can do it and if it brings in return on AI. And that's very important because that helps you enormously to select those use cases to really start with out of the 200 or 300 you have ideated and really pick the ones that then bring you on the right path. And then manage your AI use cases as a portfolio. I think that is critically important. You need to balance some quick win projects with other foundational projects because you can leverage some stuff easily, but also on the other hand, for the data we talked about it, you need to get the data understanding in the company that your data becomes a product.

So you need to also invest in your data more than you have been doing this in the past. And the past data was kind of registered or stored somewhere for more legacy reasons and reasons for compliance than for really being reutilized. And that is going to change now enormously because now you want to use this data to train your models to do the tasks. And that is, of course, conceiving data as a product and making them available through APIs is something that is probably new to many of the departments to break up those data silos and make it really available and let's say create a flow of data throughout the company. And then you find all these interoperability issues that let's say across any kind of unit that you use, maybe very different in a sales department use than in a financial department. And you will need to resolve these kind of semantic issues as well to make data really flow.

And then once you have built up your portfolio, you have a structured approach to it, you have a governance about it, you probably introduce something like an AI hub or a council that somehow surveys and manage the budget and also the synergies between the different use cases so that you really can make, let's say, also a learning effort out of it. And I think with such a structured approach, we help many of the corporates and companies to actually get out of this kind of questions like, where should I start? What is this all about?, into a very structured approach to leverage and manage return on invest.

HOST: Merantix is working with companies to effectively transform them to make them ready for AI. What role does Merantix play in enabling that?

BERTRAM WEISS: Merantix is actually a group of three main different things. One is that we are, on one hand, investing as Merantix Capital, which is investing in startups that are all AI-first or AI-native. We have on this campus here like 80 different startups being represented and more than a thousand AI workers coming together on the campus.

The campus is actually the second pillar where we create the community where cross-fertilization really can take place on a daily basis. There's a single coffee machine in the center of 6,000 square meters of space for those many companies. And amidst these startups is then the third pillar: Merantix Momentum. Many of the corporate investors who have invested in the startups through Merantix Capital realize that they need help in-house and that waiting for a product coming out of a startup may last too long and also not address all the problems they have.

That's why Merantix Momentum was founded: to help corporates get the AI journey really up in a structured fashion and make it a success story, not a “pilotitis” disease where you run a lot of pilots and do not gain a real return on investment. These are the three pillars, and we believe that this community thing is extremely important because the AI transformation is just too big to be handled alone.

We see a lot of especially mid-caps who think they can just do it on their own, and Christian alluded to it already. This is probably also one of the big failures we see very often. They swim in their own soup and think they can reinvent the world out of their own thinking. Sometimes it's so beneficial to team up, to come together and bring stuff together that has not been played together in the past, but now can really be beneficial.

We have a fourth part in the Merantix ecosystem: the AI House in Davos during the World Economic Forum, where we try to shape the dialogue on the international stage to make sure that European values like data privacy and ethics are also being heard globally. We do not want to leave this discussion alone to the US or China, where they have a very different view on those values. Shaping this framework in which we can deploy AI across our economy here in Europe is also a very important role that we actively play with the AI House in Davos, which now takes place in January again.

HOST: I'm afraid that we don't have too much time left. We've arrived at the last question and feel free both of you to answer it. If you could give one piece of advice to a CEO who is unsure about starting their AI journey, what would be the most critical first step that they should take? Go ahead, Bertram.

BERTRAM WEISS: If I may, the first thing is really the mind shift: go away from building an AI strategy to building a business strategy with AI. That is absolutely the first thing they should do. And that is something that only happens in your mind, right? But then, of course, what you need to do is to get out of the temptation to wait and see and start experimenting.

Stop building an AI strategy and start building a business strategy with AI. Bertram Weiss

Watch out for a single use case, even if it's only one that you can afford to do at that moment. Look for this repetitive thing that is a pain in the neck to most workers in the company. Every employee would be happy if you can get it done by an AI agent. Then build all the dimensions we talked about in the AI Canvas: address all these five dimensions, think hard through them, and then implement it to create a success story that helps you move on to the next thing and get buy-in from employees and unions by getting it right in the first one or two attempts.

Also, be prepared to do it in a portfolio approach. Maybe the first one may not fly, but if you do three or four, your chances that a few of them succeed are higher. Think a little bit like an investor: do three or four things instead of one, and you are probably there. And then be prepared for the foundational changes that may come from those AI use cases. You will learn about how you need to change your data landscape, your under-invested IT. These things require significant transformations in the mind of people, need a lot of break-up of past behaviors, and this is what I would do in the first place. Then scale as you learn. Do not boil the ocean on the first day; rather build it up gradually.

HOST: Okay, Christian, what do you think?

CHRISTIAN SCHIERJOTT: I can absolutely agree with what Bertram just said, and this reflects our experience too. A couple of years ago, we just said, look, this is incredible what's going on there. We just dig into it and not wait or overthink it too much. Just really get a use case ready, fail early, learn from your mistakes, jump to the next more structured use case. And just try, right? Don't shy away from all the regulation, from the complexity that you might see out there.

When we learned it as a BPO company, building up an ecosystem of the right partners, then I would say that everybody else can do that too. You just have to start with this and be open-minded about it and see where it takes you. This is one thing. And the other thing that we learned early on is don't try it on your own. Bertram, you said it before too: build up your ecosystem, not only with technology providers like AWS and Microsoft with their shiny models and cloud infrastructure. Look on the market, what is there, with smaller technology providers, with BPO providers like us, with other knowledge companies out there where you can learn from and also see what their experience is.

This is a huge opportunity for the entire market. It affects nearly all different industries. I think it's a huge thing, and this hype is not over yet.

BERTRAM WEISS: And I think that's also what we see with the Merantix ecosystem: with the startups, you have a pulse on what's possible; with Merantix Momentum, we then see what is able to scale into bigger corporates; and with the AI House, you have this general approach to framing the discussion and building the framework we need to have the right political environment and legal environment so that AI can flourish. But I understand also that it's hard for CEOs to make this fly just on their own. So again, team up. If you want to get far, team up. If you want to go quick, you may go alone, but you won't get far.

HOST: Okay, so that was the last question. Thanks a lot, Christian and Bertram, for collaborating with us. It was a really great conversation. Thank you.