Speaker Key:
DD Damien Deighan
PD Philipp Diesinger
WM Walid Mehanna
00:00:00
DD: This is the Data Science Conversations podcast, with Damien Deighan and Dr. Philipp Diesinger. We feature cutting edge data science and AI research from the world’s leading academic minds and industry practitioners, so you can expand your knowledge and grow your career. This podcast is sponsored by Data Science Talent, the data science recruitment experts. Welcome to the Data Science Conversations podcast. My name is Damien Deighan, and I’m here with my co-host, Philipp Diesinger. How’s it going, Philip?
PD: Thanks, Damien. Pleasure to be here.
DD: Great. And today we are talking to Walid Mehanna, the Chief Data and AI officer of Merck Group in Germany. Hi, Walid. How you doing?
WM: I’m doing great. Thank you, Damien.
DD: Great. So, Walid is the Chief Data and AI Officer of Merck Group. Where he also chairs the company’s digital ethics advisory panel. At Merck, he leads and helps deliver data and AI strategy, architecture, governance, and engineering across the entire global organization. Walid was born in Egypt and raised in three different states in Germany. And he’s harnessed his
00:01:16
multicultural background to inform his commitment to diversity, equity, and inclusion in the workplace.
Through his many years of experience in startups, in IT, in consulting, and in major corporations such as Mercedes-Benz, where he was also the Chief Data Officer. Walid has developed a strong understanding of the intersection between business and technology. He has a growth mindset and a strength for seeing a tiny bit into the future and translating this into what needs to be done in the present. Walid is a senior corporate executive with the sole of a startup entrepreneur who’s very focused at integrating data and AI into classic company environments. It’s great to have you on the podcast, Walid.
WM: Thank you. I would love to meet that guy you’re talking about.
DD: [Laughter]. Great. Okay. So, we’ll start as always with your background Walid, can you just give us an overview of how you got started in the field of data science and AI?
WM: Sure. So, actually I’m a data guy by training, so my background is in computer science. I studied in Northern Germany at the University of Oldenburg and worked in the internet hype bubble in the 2000s. So, that’s where I got my wings and worked as a database administrator, system administrator, network administrator, programmer, website designer, did every single-entry level job. And afterwards in my career working in management consulting for 13 years as a German consultancy called Hoover.
00:02:55
Always worked in the intersection between business and technology, mostly focused on the data visualization, data analytics, statistics, and then more and more also the data science and AI and machine learning part. After consultancy, I went to Mercedes-Benz as you already mentioned, to the German automaker. Starting as the data officer in the finance organization, and then gradually orchestrating the efforts across all the passenger cars division as the chief data officer. And then three years ago, I joined Merck in Darmstadt, Germany, as the group data officer. And then since last year, extended the role as the chief data and AI officer.
DD: Obviously, you made the decision to go from practitioner and take on the leadership track for your career. What inspired you to do that?
WM: It was skill, to be honest, and leverage. Because early on in my career, I knew that I had a passion for technology, but I also knew that there’s people who are better in programming than I am. And I got curious when I was in university and had a course in strategic management, which was my minor, in business administration. And I said, hmm, combining both dimensions could be very interesting, and that’s essentially what I’ve done. Therefore, my leadership journey, or my interest in leadership started very early on. Understanding that as a single contributor, you can do great things, but as a leader with a solid understanding in technology, you may be able to do more things. That’s how I ended up in this career trajectory.
00:04:38
DD: If we move on to your current role in Merck, you’ve come from a non-life sciences background into a very senior global role. How did you find that transition?
WM: It was interesting, actually, and that’s a fun story. I saw it as an advantage, because we have three different businesses, healthcare, life science, and electronics. Obviously, chemistry goes across, but we have different go-to markets, different business model, different customers, different products in all three businesses. And I always told my hiring manager at the time, “Listen, if you hire me, let me know conflict. Because I’m not a healthcare guy trying to tell the life sciences guy what to do. And I’m not an electronics guy trying to tell the healthcare guy what to do.” And as I was set up globally is very federated. All our three businesses are very flexible with their own CEO, with their own organizations. So, we have our sister organizations, the sector data and AI offices embedded in the business.
And we have an amazing collaboration that we call the hub, hub, spoke with multiple layers where we orchestrate all of our efforts. And I believe to a certain degree it was an advantage not to be out of one of those businesses. And on the other side, it was also for me a learning journey and something that catered to my curiosity that actually I would learn new types of data, new types of businesses. And as you might remember, that was also one of my main motivations to go into consulting because I was curious to understand how the world works and how different businesses operate and how they interact with their customers. So for me, that was not an unnatural transition. Actually, it was pretty intuitive for me.
00:06:27
DD: One observation, I’ve noticed that leaders that come from a non-digital first sector they come from a real-world company sector like automotive, you’re dealing with a physical product in the real world. It’s transporting real people from A to B. I think that that transition is easier than say, someone who has come from finance, which it’s almost a complete digital journey for customers. The product is digital now in terms of financial services. Do you think that there is some crossover there, the mental framework for what can work with data and AI in the real world in more traditional companies? That helped you make the transition?
WM: Most likely. So, but honestly, I believe, ultimately, we are always back in the real world and it’s easier to scale digital platforms, obviously. But I believe the real excitement is the human machine physical integration and interaction. So, when we look at what is going on in robotics, when we look at how automated and how data-driven production also has become, that borders between IT and OT, operational technology, are blurring and integrating more and more. I believe, it is, to me, an artificial distinction between the physical and the digital world. Ultimately, there is one world, and obviously the digital world also talking about digital twins is always some kind of representation and enrichment, or augmentation of the real world.
And therefore, for me, the real exciting piece is when you work on the intersection and when you can essentially bring both worlds together. So, I believe it’s also a question of personal mindset and how your perspective is. If you’re a cyber jock and you’re just living in a digital world, obviously it is, will be a hard impact to get into a brick-and-mortar company that has
00:08:29
physical products. And if you believe that physical products will always be sold like a hundred years ago, then also it’s a bit of, let’s say, culture clash, when you meet people who live in the digital world or grew up in the digital world. I see myself, feel also as a walker between both worlds, and an integrator between both worlds. Because I believe ultimately, they will become one.
PD: We’ve worked through the whole internet hype and there have been other hype cycles since then. We talk about distributed computing or maybe blockchain or Hadoop, now it’s GenAI. In your experience, you know, working and being in different phases, working through all of these, do you see similarities, there are differences? What is your take on the latest GenAI hype cycle now?
WM: So, the first difference that I see is that Hadoop definitely had the best mascot with the elephant, and I don’t see anything coming even close for GenAI. But obviously it’s, I guess it is to a certain degree human nature. We get overly excited, or we get overly critical, and there is always a balance, and we always have both. And the truth often enough is somewhere in the middle, or one shade of grey between black and white. So, I believe that is the commonality between it, there is always hypes and there is always also the awakening. And the question is how hard is the awakening and how useful and how timely was it? So, blockchain for example, very exciting technology, I believe it’s time hasn’t come yet. Maybe it’ll be in the future, maybe it will not. Maybe it will be a niche, maybe it will be a mainstream one day.
00:10:17
We’ll eventually find out. And my philosophy for all of those trends, and I went through the internet, I went through the big data, Hadoop Pipe, Firsthand, Blockchain was the one I set out to a certain degree and didn’t get too involved in it. But when we look at GenAI, there will always be a substance that is brought in and there will be waste in terms of attention, money, investment that went into technology but didn’t lead anywhere. And I see the key challenge for me and my teams, first of all, to identify the right inception point. When does a technology become relevant for us, that it could be a competitive differentiator. And at that inception point, we need to be ready to leverage it. And this is something, for example, in the case of quantum computing, where we are definitely observing very much.
Where we are actively being involved in discussions, where we have a solid foundation of understanding, what works, what doesn’t work, and what’s happening in the market where we also somehow have some investments. But we will double down when it becomes relevant for our operations, for our business models. For GenAI, I believe after ChatGPT in November 2022, that was the inception point for us. Where we’d say, we need to have a look at it, or we need now to get active because it is happening, it’s becoming mainstream. And then it’s always a question to find a good balance between the overselling and the over-hyping. And on the other side, on leveraging the potential and the substance that is already there. And that’s, for me, a leadership challenge that is not a technology challenge, that is not a sales challenge, it is a leadership challenge to separate the hype from the substance and put the substance to good use, and ideally ignore the hype.
00:12:10
PD: So, you talked about finding the sweet spot or identifying an inception point, how do you and your team approach this problem? How do you determine whether there is an inception point, or whether it’s not there yet, or maybe it has been there in the past, like how do you approach this?
WM: Ultimately, it’s a people’s business. So, we do have a small team that we call the AI and Quantum Lab that is involved in those things very deeply and essentially is focused as an, let’s say, early-stage research development team on identifying trends and evaluating trends and trying things out firsthand. Because ultimately, you can only say something is substantial or not when you really try it. One thing, for example, we got very excited about a year, year and a half ago, was agents, AI based agents. And we tried them out and we all found out, unfortunately, not yet good idea.
But maybe we need one or two generations further of the model, or we need different architectures or procedures to really be able to use it in a business context. So, those things, so, ultimately, it’s like always in innovation, experimentation is key to those topics. And obviously understanding of the field, understanding of the material, and engaging with the thought leaders and with the practitioners in those fields. You have to get your hands dirty to be able to understand is this something just on a PowerPoint slide that somebody wants to sell me for a lot of money, which sounds like a good idea, but has no chance of flying, or is this something substantial?
PD: And when you talk about experimentation, trying things out firsthand, how far do you go there? Is it building an MVP or a pilot even, or is it less than that?
00:14:06
Do you look for business partners to try use cases, try new approaches? How far do you go there?
WM: It depends. It depends on the, let’s say, necessary investments that you have to do and the complexity of the problem or the potential of the problem. So, I believe we have the full bandwidth from, let’s say, put one person, build a proof of concept, or a minimum viable product for six weeks. Or to strategic collaboration with partners to build something that we scale and implement in one plant and then roll out either in a country or in one business or even globally. So, ultimately, it’s a big bandwidth and depending always on, obviously, you can take more risk and invest more money if there’s even more value and more payback.
But sometimes it’s just human curiosity, so you don’t know where it will end, but you want to do it anyways, which was the case, for example, with OpenAI. So, when we originally logged into GPT for the first time, I was like, okay, we need to talk to these guys, and we need to put an NDA in place. So, I told my team and they just smiled at me and said, “Walid, you signed that NDA nine months ago.” And I said, “Okay. Good job. You guys did well.” So, we started talking to open AI in February 2022, before even ChatGPT was online. And this shows that our AI and quantum lab approach work quite well in this instance.
PD: You mentioned that your approach to the pilots to exploitation is very business focused, so, you look at investment versus return. Let’s say you implement something, it goes into MVP phase, and then you approach the
00:15:46
point where you need to make a decision whether to scale up or whether to wait or make changes or so, how do you approach this rollout decision?
WM: Well, rolling out anything in the physical world always depends on the people in the physical world. And as we are an internal technology provider, we don’t own the plants, we don’t own the processes, the sales, the customer relationship, the supply chain, the warehouses, and so on. And therefore, it is always a collaboration. So, you have to have a convincing case and you need to convince the people who essentially will need to do it, ultimately. And in my previous experience also in consulting, I coined the term popcorn customers. And for me, popcorn customers were those experiences where we did a proof of concept, where we did an MVP, and we were excited about the first results, and everybody was also excited to do the fancy AI machine learning stuff. And then we got in the meeting and the people relaxed, laid back, brought out their popcorn and wanted to enjoy the show.
And I had always to tell them, I’m sorry, but there’s a big misunderstanding. Because if we’re successful with this teeny tiny MVP and proof of concept, then the blood, sweat and tears start. And the blood, sweat and tears will be in your business, with your people. Because we need to change the way we operate. We need to change processes, we need to train people, we need to harden the code, we need to scale the infrastructure, and, and, and. So, essentially an MVP as promising and as exciting it could look is the easy part. The hard part is the scaling, the hard part is the rollout. And no data and AI organization in the world is able to do it on their own, they will always
00:17:30
need to rely on the relationship, the openness, the willingness, and the ability of their organization, their partners on the ground to do it.
PD: Makes a lot of sense. Yeah. So, I think for scaling and helping the organization, that seems to be a central role for you and your team. Merck is very special in the way that it has these different sectors. I think last time when we previously spoke, you mentioned that you have embraced the hub-hub-spoke model to make that happen. Could you talk us through that approach and why you chose it, what the advantages are, how it makes utilizing AI and data at Merck becoming a reality?
WM: So, I strongly believe in two things. Number one is you cannot separate data and AI. If you want to be successful with AI, you need quality data that is relevant to your business, that is specific to your business and to your processes. And quality is the key word here. Data everybody has, but if your data is a waste product or a side product, will not help you with the AI. So, if you’re really interested in leveraging AI as an organization, as a competitive advantage, you have to take good care of your data. That’s the first thing. The second thing, I’m very of is, data and AI cannot be an ivory tower capability. Obviously, you need central organizations, global organizations like mine, as a catalyst, as a supplier, as a governance entity, as a strategic entity, as a cultural change entity. But it cannot be a delegation of accountability.
And the accountability for data and for artificial intelligence needs to be embedded in everything we do as an organization. And that’s the underlying
00:19:21
philosophy for the hub-hub-spoke model, because the hub-hub-spoke model essentially means we want to have our understanding and capabilities as embedded as possible in the organization. But we also want economies of scale, and we also want to be very disciplined in our usage of technologies that can easily become very expensive. And therefore, we have the central global hub, which is my organization, the Merck data and AI organization, as a global orchestrator across all three businesses in our enabling functions. In a second instance, every sector has their own data and AI office that is either embedded with the IT organization, or separately from the IT organization, this is up to the sectors to decide.
But what they, in either setup take as a responsibility, they are on the one side a multiplier of our global standards and our global strategy. And the other side, an aggregator for the local needs of the organization. And the spokes are essentially our embedded teams on, let’s say, the customer side, on the operations side, on the R&D side, that are closest to our core business, to our value chain, and that are directly embedded there. Obviously, there is specialist roles that you don’t need decentralising in those organization, and there are business roles you may don’t need in a global organization. But we need to talk to each other, and we need to have enough overlap to understand each other and be effective and efficient in the way we apply data and AI to our business.
PD: Going into that a little bit deeper, how can one understand how this model works practically? Like what would be responsibilities or value that comes
00:21:08
from the central hub? What would be responsibilities that lie more in the sector, and then how does the spokes level contribute?
WM: So, the global level is all about scale and efficiency. So, what we have is, and direction to a certain degree. So, we are not an entity that tells everybody how they should do stuff, but what we are trying to do is we listen, we understand, and we align collectively for the joint way forward. So, for that, my organization has the data and AI strategy, again, not in ivory tower, but in collaboration with the sector data offices. We do have the data and AI governance, which is also important for us because of the global regulation that is more and more upcoming. Also with the EU AI Act, and the Biden executive order, and much more coming down the way. Then we do have the mindset and skillset of people, which is our way of culture.
So, this is about the literacy and the right approach to leveraging the technology. And last but not least, it’s the technology itself. So, the global platforms also to make sure that we are cost efficient and that we are scalable in our usage. And that we are monitoring the market for a proactive approach to develop our landscape. So, everything that is reusable, everything where we believe that we as one Merck can benefit from, that we try to bundle and make available centrally in our organization. And we’re also now establishing more and more capabilities like Agile, AI machine learning, data engineering. That we can make available to the organization that are well trained in our practices, in our standards, in our technologies.
00:22:58
But then can be used in single solutions, in single use cases, some people like to call it sometimes, and then essentially deliberate. The sector data and AI offices are very focused, first of all, they are obviously our collaboration partners. They have a say in everything, all of the topics that I just mentioned. In technology, in culture, in strategy, in governance, and so on. But in terms of delivery, they are more focused on the applications, how can we generate impact in the business? How can we build the data products? How can we evolve the ontologies, the semantic models for our different data domains? How do we implement our data governance, specifically in healthcare, in life science, in electronics and the enabling functions?
So, they are essentially the wardens of the data and AI strategy and also the champions in the organization. And the spoke is hyper focused, hyper focused on their business. Spoke is hyper focused on their customers, on their supply chain, on their plant, whatever it is. But they have the highest proximity, those can be low involved, people using the self-service that we make available, or it can be small teams that use our technology. And sometimes even in a very conscious approach, teams that use different data technologies that were not of global standards. But we know, and we endorse, and we still support on an overarching when there’s good reason for those things as well. So that’s the rough outline, how the hub-hub-spoke collaborates.
DD: So, it sounds like the strategy rather than top-down. While it is a real co-creation between the central organization, your team, and the rest of the business, that’s quite unusual, I would suggest [laughter].
00:24:47
WM: There are times where we have, or we want to be top-down to a certain degree. Obviously, disagree, decide and deliver is one of our high impact culture values. We talk, and it’s okay to disagree and it’s okay to have different priorities, and obviously I remember my time in consulting, every organization tells you we are different. Which is absolutely okay, it’s okay to feel different and to be different. Ultimately, there is some kind of abstraction that we can agree on. I always say, what can we put in front of the equation? And ideally that is a solid compromise. And when we have that, we need to be disciplined about it. So, that is then the piece where we are flexible, we are collaborative, we are very open to hear each other out, understand each other. And then think about what’s the best way forward as one Merck, and then implement accordingly.
PD: You mentioned before that when you approach a data product, you look in, or you think in three different dimensions, platform, data, and then also analytics. Can you talk us through that approach? Like how do you evaluate that the data product is ripe for scaling, or it is a high-quality product, it’s potentially a high impact product when you assess those three dimensions?
WM: So, in the old times, we would have use cases. And those use cases, it’s either a new system we’re building or something, an application somebody’s building, or sometimes even just a dashboard. What people are doing is they look into the source systems and identify which data do I need? Then they find out they need data from different DAP systems and that data doesn’t directly connect to each other. So, they start building mappings and clean the data. Alternatively, there is somewhere a data set they somehow get access
00:26:42
to in data warehouse, in a data leak, or in other application, so, that’s the shortcut. But then they find out that that data is not complete, and they need additional data. So, they start to build what I like to call Frankenstein data sets. So, because essentially it’s a second-hand data set, that is then enriched by firsthand data, nobody has a lineage of that data.
And essentially it works somehow, but yeah, to me as a data guy, it feels like a Frankenstein. So, a data product is a more strategic approach to us. So, we are going top-down with different data domains and what we are trying to do in the future is to push responsibility to the data as close as possible to the point of creation. We haven’t implemented data contracts, we have limited data lineage for, let’s say, a lot of systems, for some we do have good lineage. And establishing that transparency, and most importantly, the right accountability and responsibility across the organization. And then also supporting it with the platforms available to deliver good quality, to monitor data, to monitor pipelines. And one differentiation we like to do is there’s always a quick fix, and it’s okay to apply a quick fix. We like to call it stop the bleeding.
Obviously, that is a priority and that’s okay. But after you stop the bleeding, there is a tendency to stop even more bleeding, put more patches on it, because it works so brilliantly, but it’s not a sustainable, healthy solution. So, the second part of the journey and the more strategic piece should be to heal the patient. So, how do we get to healthy behaviours? How do we get to healthy setups? And as I said, in my mind, this ultimately is that those who create the data, have awareness that data is an asset to the organization.
00:28:48
And that we need to take good care of that, that is part of their daily job. It should be part of the daily job, and they ideally also incentivized as such. And ultimately, that they also have transparency about what happens with their data.
And this is why I’m also a big fan of data marketplace approaches, because ultimately then you see, oh, that data is used. People are generating business value with it, customers are delighted because my sales reps have that information that I procured, that I made available to him. And at the moment, we don’t have a real economy of data, we just have people trying to make gold out of coal, which is not an easy task. But again, it’s a journey and it’s an organizational transformation at scale. They bring everybody on board and all hands-on deck. And as such, it is technology, obviously, one dimension. But even more, and this is also when we talk about our data and AI strategy, we always talk about, well, I like to call the holy trinity, which is people, ways of working, and technology. And you have to start with the people.
PD: How about these other two dimensions that go with the data product like a platform and then also obviously the analytics, KPI, analysis aspect?
WM: Exactly. To a certain degree, the data generation is the first mile of our journey. So, getting there is not typically something data folks have in mind. So, when we talk about data, people always start with the ETL, the extraction of the data, but they don’t put any attention in the creation of the data. And I believe that’s a fundamental issue that we have. So, pushing down the
00:30:41
quality mindset into the point in the process where the data is created, where people put something into systems, or where interfaces are built, or where data models are designed, or when a company is acquired, for example, in M&A process. So, that’s the first mile. And when we talk about the second, third, fifth tenth mile, that’s essentially how data is made available for secondary use in the organization or shared between systems.
And this is where the platforms come in. Because I don’t believe in one size fits all, and I don’t believe in the Highlander principle when it comes to technology, because that would be too much of a lock in. I’m a big fan of global standards, I’m a big fan of open source. I’m super excited about Apache Iceberg, to see that really become a relevant piece of our architecture and our global standard. But the platforms for me are fundamental parts to be effective and efficient in scale across the company. And because of that, we decided to take our key platforms and integrate them into what we call the data and AI ecosystem. We even gave it a name, it’s called Optimize. And Optimize is essentially our brand promise to our organization, that when we called it Optimize, it is well integrated and it’s safe, secure and compliant.
And the different components work as seamlessly as possible, together. It is a journey, also that ecosystem is not perfect, but it is improving significantly. Every product increment, every three months that we do, and we are growing it from here. And then we have a multitude of what we call analytics and AI products that are built on top of that ecosystem. That can be full-fledged applications, that can be simple dashboards, that can be AI, API services
00:32:32
that we expose and make available to our organization. But that is then exactly what our community of people, our practitioners, our sector hubs, and also our spokes, but also individual contributors, leverage and build.
PD: Makes a lot of sense. Yeah. We talked earlier about the term use case MVP or pilot. Can you explain a little bit more the process of thinking, shifting away from that concept of the use case?
WM: We are still, so, the central technology where we document is still called use case portal, but I’m not a big fan. Because when looking at it from a business perspective, a use case, or the term use case feels to me like the old man with the hammer. I just saw an interview yesterday with Tim Cook, the Apple CEO, and he was asked by Marques Brownlee, Apple has never used the term AI, why? And he said, “As a company, we focus on not the mean to an end, but we focus on the end, we focus on the fall detection. We don’t need to tell our customer that it’s AI, and that’s why they’re not actively using it.”
And for me, similarly, it’s also that a use case is an application of technology. But I’m not interested in, let’s say, show casing the technology, that is the sales mindset. I’m interested in the business impact and the value for the people we serve, for our patients, for our customers. And therefore, I like much more, again, use case feels to me like an old man with a hammer looking for a nail. And I like much more to have the mindset of a solution builder, someone who understands a pain, an opportunity, and builds a solution to cater to it. And not apply a technology in a use case to showcase the technology.
00:34:31
DD: Makes sense. So, what do you actually call it then? You avoid the term altogether and you call it solution?
WM: First of all, we try to establish the product mindset across all categories. So, we talk about platform products, things like AWS, like Snowflake, like Foundry. We talk about data products to make sure that a data set is really of high quality, actively managed and governed. And then we also talk about analytics and AI products to make clear that this is something that is the last mile essentially. So, delivering the insight, delivering the functionality needed to solve a business problem or realize an opportunity.
DD: You’ve touched on a couple of things there and I’ve heard you talk in the past about AI as a first mile and a last mile problem. Can you just bring that together for us and tell us what you mean by that?
WM: So, the first mile problem is the creation of data. Because at the moment we talk about data, but we, unfortunately, is the data people [laughter]. So, the rest of us in the business always look at us and say, yeah, yeah, that’s nice, but that’s not my reality. I still have to write stuff down and somebody else maybe types it up if he can read my handwriting. Or yeah, great, but I have to put in the same information in four different systems and I don’t know where they ever end up. So, that’s our first mile problem. Our reality is if you start the data journey on the extraction of data from operational systems, you are already at a disadvantage.
Maybe you’re even playing a losing game altogether. So essentially, if you are serious as an organization about it, you have to start at the creation. And
00:36:15
that is the people on the ground, and that is the systems we give them. And that has never been an opportunity or have never been a priority to our organization. So, discipline and good systems and ideally automated data capture, those are the things that are our first mile problem and could also ease a lot of pain further down in the value chain of data analytics AI. So, that’s not a small one, that’s a pretty big one to be honest.
DD: Well, it’s the first principle, isn’t it? It’s the foundation, if it’s not there, then how can you do anything downstream?
WM: And I haven’t seen yet. So, maximum a handful of companies, you maybe have been born in that environment that really think about data this way, expose data, every team has to do it. You know, the API mandate by Jeff Bezos at Amazon, for example. Those are very, very few examples that we see out there that have that first principle in place and that really can reap the benefits of such data mindset literacy, and then also quality available at scale.
DD: So, how does a large complex company tackle that problem?
WM: Step by step. You have to preach, you have to talk, you have to collaborate, you have to put some Band-Aid on it. Obviously, when somebody has to put in data in four different systems, you can build a wrapper powered by AI to make sure that the data is pushed into four different interfaces but is consistent. But that stops the bleeding, if you want to heal the patient, obviously you have to talk about enterprise application architecture, and you
00:38:00
have to start talking about interfaces. You have to start talking about what we call data languages [inaudible 00:38:06] that people agree on data models.
So, if you have three, five, ten thousand applications globally, which large corporations in our size normally have, then how do you, and none of them have ever been built with a target state in mind. They always have been built, okay, what’s the new data model for this system? Oh, we take either what’s coming out of the box or alternatively we do a project, and we start greenfield to think about how a data model should look like. But not strategically, there wasn’t a data domain, there wasn’t an ontology in place and there was no mapping that was available. Or as we did, you do an acquisition.
Congratulations, there’s another 500 systems that you are inheriting where you don’t have a single fighting chance to integrate them. So, ultimately what you want to do is you need to define your target state, you need to define your strategy, you need to define your semantics ontologies, your data domains, and then you need to start to get disciplined to map to them. But you also need to continuously evolve them because your business might change. As I said, you might acquire companies who have a different business or that know more about you, about single domain, and therefore it is a journey. And the question is, do you know where you want to go, and do you have a plan how to get there? Or are you just firefighting and putting Band-Aids on everything to stop the bleeding?
DD: Very eloquently put, Walid. So, how would you then describe the last mile problem, Walid?
00:39:38
WM: Well, the last mile is essentially how do we translate it into our daily life? So, we’ve seen so many dashboards, reports who are fancy, shiny, look enticing, but then are not used regularly. That become outdated at a certain point in time. So, the last mile for me is essentially how do we really get to scale? How do we embed this in our routines, in our daily doings? And for that, obviously, we need to talk about user experience, about user interfaces, but we also need to talk about things like responsiveness, data availability, and processes. And this is where I also believe that this intersection of, “Oh, great. I have a new fancy analytics in AI product, an application that I can use.”
To how do I need to work differently? And this is sometimes a bit the differentiation between, what do you mean by digitization? Digitization is just we do everything the way we did it previously, but just with a tool that maybe gives us five or 10%. Or do we rethink the way we do business because of the possibilities that we have? And this is the one thing where I, because of our federated setup, I need to rely on my partners on the ground in the sector hubs, but also in the spokes. Because our remit is so broad, we cannot be an expert in material science and electronics and also with key opinion leaders, and doctors in neurology, in healthcare.
So, that is exactly where our federated approach comes in. But ultimately also, this is a team’s sports, and also the senior leaders and the middle management must understand digitization as an opportunity that can help us to up our game. And obviously, we can be an early follower, we can be a late follower, we can be a laggard, but ultimately things will change and it’s up to
00:41:54
us to find out how they can change. And we’re coming back to the experimentation and how we really make a difference for those we serve.
PD: Yeah. We talked a lot about strategic approaches, how to make things work, you know, governance ,orchestration, collaboration, and so on. Do you have any specific applications at Merck that you are free to talk through where you actually implemented it and maybe to make it a little bit more tangible?
WM: Sure. So, one of the biggest areas for us in our healthcare business, obviously, is the drug discovery piece. So, this is something identifying molecules targets. This is of high interest, and this is something where we are broadly on the one size doing collaborations and partnerships with companies like Benevolent AI, and Accenture on the other side, also developing our own technology stack in-house. And also, through our life science business, building it into an AI and drug discovery product that we’re selling out in the market. And even our life science, even our electronics business is on board.
Because material science is also identification of molecules, completely different molecules, obviously. So, we’re talking about liquid crystals, we’re talking about other specialty chemistry that is needed in the semiconductor industry. But the silico process of going through a search space and identifying a fitting model with properties is very similar. So, that’s one of the high value targets that we see in the AI space that we’re very much invested in. In both sides, through partnerships, in own development, and across all of our three businesses of healthcare, life science, and electronics. Not GenAI,
00:43:53
classical AI. It does have some, let’s say also GenAI pieces around it, but it’s a combination.
PD: And how did you solve the first mile, last mile problem there?
WM: So, you have to work with what you have from historical things. So, when you look for example at clinical trial data, or when you look at experiments that you have done in the past, obviously, you cannot reduce 300,000 experiments, you wouldn’t want to do that. So, you have to work with the data that you have. The first mile is mostly about building a foundation for being better off in the future. So, if you have data, sorry, you have to work with the data that you have at the first start, but obviously, especially in ongoing business processes that generate a high volume of data, you want to make sure that the future setup that you have is more sustainable in line with your strategy. And the last mile problem is convincing people that they do the experimentation.
In this case, it is pretty straightforward because the drug discovery space and also the material science base is normally quite a long cycle. So, a new drug needs about seven years until it gets to the markets, and there’s a high failure rate. So, what you want to do is obviously, you have to have the different test balloons. And this is something that will pay off and show off if it was successful or not in one or two years time. So, there’s still, but when you have the results and more and more of those, let’s say, AI identified drug candidates are coming now into the early and mid-stages of clinical trials, when they are successful in less time with significantly less investment in the
00:45:41
research domain, this will convince people. But again, this is a long game, but also a high risk, high value game.
DD: I would like to take a brief moment to tell you about our quarterly industry magazine called The Data Scientist and how you can get a complimentary subscription. My co-host on the podcast, Philipp Diesinger, is a regular contributor and the magazine is packed full of features from some of the industry’s leading data and AI practitioners. We have articles spanning deep technical topics from across the data science and machine learning spectrum. Plus, there’s careers advice and industry case studies from many of the world’s leading companies. So, go to datasciencetalent.co.uk/media to get your complimentary magazine subscription. And now we head back to the conversation.
Circling back to GenAI, Walid, where have Merck found the most value in terms of implementing generative AI?
WM: So, generative AI is an exciting topic, and it’s so mainstream. Obviously, the main modality at the moment is text. So, we see it most valuable in all domains where you have to, where you already have a very extensive corpus of knowledge codified in text, in semistructured, unstructured text. And where you need to consume a lot of information that is codified as text and where you need to generate a lot of information codified in text. So, everything in legal, where you have a lot of regulation, everything in regulated environments, so, EMA, FDA. It is a helper, don’t get me wrong, it
00:47:37
doesn’t mean that any of our submissions that we do to these are generated by Biogen AI.
But people working in those areas have a very powerful tool at their hand combined with RAG, Retrieval-Augmented Generation, to consume and also very targeted find information in those documents and they had in the past. And especially when you generate information in marketing, in sales, in communications, but also in scientific writing, it is a useful technology. So, those are the baseline, let’s say, usefulness that we see, so, personal productivity, most definitely. And our approach has been maybe slightly different again than to other organizations. Maybe quickly describing our journey, so, as I said, when ChatGPT came up in November 2022, we saw the opportunities and the risks.
Obviously, the risk was people would be using it and OpenAI would use our people’s information and our people’s data, our company’s data, our customer’s data to train on that model. That’s what happened to other companies, and we wanted to make sure that that didn’t happen. So, the first thing that we did was essentially communicate to our organization to tell them the good, the bad, and the ugly about ChatGPT. So, we give them guidance, we told them guys, that’s it’s great. What we don’t want to do is essentially to kill it and to block it through our firewall so that they’re out of our network, could access it. But we wanted to make sure that people have the possibility to try it out, but that they are aware and cognizant about the limits and the risks that are associated with it.
00:49:22
So, that was guidance that we gave out very early on, also in the training, to all of our employees. The second thing that we did was we wanted to have a safe, secure, and compliant ChatGPT option. So, we already started in late 2022, early 2023 to talk to OpenAI how we can access their service to build essentially a safe, secure copycat of ChatGPT. Then Microsoft made these services available through OpenAI Azure, Azure OpenAI services. And that was for us the inflection point and the sweet spot where we said, okay, now we have an opportunity. And what we did, we built a wrapper around the API with the front end, no database in the background, storing all conversations and the cookies on the browser of the user. And we made it available to our digital community across the company. And we started in May 2023 with the early access to our digital community, and we put up a waiting list for everyone in the company who’s interested into it. And three thousand colleagues signed up on that waiting list.
And we gave them access every week to make sure that they have this possibility to try out and to learn new things and the productivity gains that they can get out of it. We then in October made it generally available to all of our 63,000 employees. And then did a pivot earlier this year with a
00:51:09
partnership with Langdock German startup, that we essentially sunk down and retired our own application and got into the partnership with Langdock and used their application in our infrastructure. So, it’s still safe, secure, and compliant with our requirements, but added us the possibility to have a broad set of models. So, we’re using OpenAI’s and Tropic’s and Mistral’s models in the meantime for our people.
And we are connecting it more and more also to our different company databases. So, that’s our, let’s say, baseline journey. And I have one slide that I always share where I say we have three layers of AI that will go along in the company. The one is the base layer, that is our everyday AI, the GenAI at scale layer, that’s my GPT. Our custom application, and that’s our APIs that we make available through optimize. Then the middle layer is essentially all the embedded AI functionality or standalone AI functionality that you can buy on the market. That might be a useful addition, but also is, sometimes an open question, is that already value for money?
Because sometimes it’s just essentially, oh, instead of copy and pasting something, I can do it in line in the application. That’s not something I would like to pay $30 per person for 63,000 people. So, we always need to evaluate, and we do that together with our IT business partners, our procurement colleagues. Is that value for money? Is this something the base layer with a little bit less of comfort can also supply or not? And the tip of the pyramid is essentially the game changing options. If you’re into history, you’d know that the original pyramid in Egypt where I was born, was golden, and that’s also the tip of this pyramid.
00:53:02
So, when we talk about training models, when we talk about fine tuning models, obviously we want to make sure that there is a clear business case behind it, and we limit it to those cases where it is necessary to do that. But also, we see the cost there going down, so there is a certain appetite also to do more and maybe invest a bit more also on that domain. But that’s our general go-to on GenAI. So, we make it broadly available, very cost efficiently to all of the organization. We give technical possibilities to integrate it into analytics and AI products, into solutions. And we are very deliberate about the proliferation and the usage of AI in our applications or bringing in new AI functionality. Because we want to make sure that they are really value for money and that we’re not essentially over buying in the hype.
DD: You mentioned fine tuning there, because I think it’s clear that RAG implementations definitely there’s value there and it can be implemented in achievable way. A lot of organizations seem to be not getting great results from fine tuning, it’s almost like no one really knows how to do that in an enterprise environment. How have you been finding that at Merck compared to RAG implementations?
WM: A similar experience. But honestly, I believe the reason, but again, that’s just one person’s opinion, I believe the reason is the quality of the data. I believe fine tuning is relevant when you have high quality data for a very specific case. And if you just throw what you historically have at the fine tuning, then you might be better off with the pure play model.
00:54:49
DD: And where do you see the value of GenAI in the next two to three years? What’s realistic for the technology given where it is on its evolution?
WM: So, I still haven’t given up on agents, it was too early. Sometimes it’s just you have to wait a year, 2, 3, 4, 5. We will eventually find out, but I still believe that the idea of agents is very enticing and that it’s extremely powerful. So, that’s something I would like to keep a keen eye on, and still have my money on it without a timeline. The other thing that I’m pretty excited about is the integration in the operating system of the business. So, when I look at the Apple keynote and we’ve seen a lot of things happening, first in the consumer space, and then moving to the business space. But when I see how powerful very small models or how useful very small models can be when they are integrated with personalized or relevant information, like for example, in the very highly governed environment like the apple ecosystem and the operating systems of Apple.
And when I translate and project that into our business environment with ERPs, with Salesforce, ServiceNow, and things like these, start to get creative on how useful actually our business applications and how efficient business applications can become when that is embedded. And at the moment, AI is in the business applications, starting to get available, but it’s mostly both. They’re not embedded in the fabric, they’re not embedded in the knowledge of the background. So, essentially, it’s just, oh, here’s a functionality that has a text field, so I can put GenAI on it so people can generate the text field. That’s great, maybe useful, but that’s not a game changer, that has nothing to do with logic. And that’s what I’m excited about
00:56:57
in the business context, the combination of agents and embedded in, let’s say, in the fabric of our operating systems, that can be extremely powerful.
DD: So, it’s really interesting because the, for the agents to work, I think effectively what we’re asking them to do is automate the amount of decision making that a human does to perform even intermediate tasks. And I think that challenge is going to be very interesting to see how the industry tackles that. My sense is we might be four or five years away yet, may maybe longer, I don’t know. But yeah, it’s interesting.
WM: I would tend to agree, definitely. But, I’m sorry if I sound like a broken record, but I’ll pitch the data quality again here. Because I believe decision making, we as a human need to make are because of lacking data or lacking quality of data. So, essentially if there’s clear rules and the quality data available to those questions, then I believe also agents could make decisions more stringently and more in the favour of the human. So, also this one, it’s not purely a data problem, don’t get me wrong. But I believe that also higher data quality could help agents to be more useful as well.
DD: For Sure.
PD: Walid, I wanted to ask you from your diverse set of experiences across so many different sectors now, you have the whole Merck spectrum, then you have automotive and so on. The question was a little bit, you know, how did it help to have these diverse experiences? Where do you see similarities? Yeah. But where is it also very different, like truly different? You mentioned, you know, everybody thinks their sector or industry is very special, their
00:58:45
company is super special and so on. And oftentimes of course it’s not true, but sometimes it is. So, wanted to have your perspective on that.
WM: Well, I believe it actually is always true. So, if you believe it, you are, to a certain degree. But it is also question of the culture of the openness. Are you ready or are you open to, let’s say, reflect on what made you successful until here, is also what makes you successful in the future. And honestly, if an organization says yes, we don’t need digital, we don’t need data, we don’t need an AI, then that’s absolutely fine. But you won’t find me working in that company. So, obviously this is a legitimate discussion and a legitimate perspective also to have. Coming back, is it a, I guess it’s like always, it’s a mix. Is it an advantage, is it a disadvantage to come or to be in different industries? On the one side, I believe you need to have domain depth to a certain degree. And the leap of faith is always, can I get into that domain deep enough to be, let’s say, connected and not disconnected.
And also, on the other side, how much knowledge is already in the organization about my expertise about data and AI, so that there is also not a gap. So, you need an overlay about those things. Obviously, the overlay, when you come from the same industry as a given, because you’ve been there, you’ve done that, you’ve applied that. So essentially this is, this automatically, that’s the advantage. The disadvantage is you might be missing out on fresh thinking. Yeah. You’ve got another schema. Maybe it’s not, let’s say your schema and your company, but it’s a competitor’s schema or your supplier’s schema or your customer’s schema. But someone who’s
01:00:41
very, very close to your business, to the way you’ve been doing things forever.
Or you have someone from an outside industry that has a fresh perspective on things and can tell you how things are done in different industries. So again, it’s a combination of both, it has advantages, has its disadvantages. And ultimately, I believe it is a personal fit if you say, okay, you are coming from another industry, but we’re excited about you and you’re excited about us and we want to work it out. And ultimately, at my level, and Damien you mentioned already the leadership, this is a leadership challenge, this is a global leadership. So, data and AI, it’s not a technology challenge, it is a global leadership challenge for all of the organization. And obviously, you can try to set it out, you can try to delegate it as a commodity to the CIO, but it will not go away by itself. And the question is how you approach it and how you tackle it.
DD: Great. So, Walid, as we wrap the conversation up, are there any sort of final thoughts that you would like to leave with current data and AI leaders or aspiring leaders that are trying to make data and AI work at scale and complex organizations?
WM: So, my recommendation would always be, don’t worry too much about scale, worry about the impact, worry about the value, worry about the difference that you can make in the lives of those that we serve, our patients, our customers. And if you make sure that you are delivering on the promise, either directly or indirectly depending on your responsibility, then you will be
01:02:21
on the right trajectory. That is, I believe, the number one thing. The second thing is, as you’re doing this, also try to think about the future and try to minimize the proliferation and the diversity of too much technological depth, and too much technological complexity, or too much process complexity, and too much process depth. Because ultimately those will hinder you to scale at a later point in time. But I believe scale is not a singular priority, scale is only useful when it is preceded and built on top of two impact and value that you can show to the organization. Otherwise, scale will be unachievable.
DD: Sadly. That concludes today’s episode. Before we leave you, I just want to quickly mention our magazine, The Data Scientist. It’s packed full of insight into what’s happening in the world of enterprise, data and AI. And you can subscribe for the magazine free at datasciencetalent.co.uk/media. We will be featuring an article on this conversation in issue eight of our magazine, which is out in September 24. So, Walid, thank you so much for joining us today. It was an absolute pleasure talking to you. Full of insight and some absolute gems for any leaders or aspiring leaders in large organizations. Thank you so much.
WM: Thanks for having me. It was great to chat.
DD: And thank you also to my co-host, Philipp Diesinger, for his great questions, and of course to you for listening. Do check out our other episodes at datascienceconversations.com and we look forward to having you with us on the next show.
