Inside Boehringer Ingelheim’s Digital Supply Chain Twin – Christian Petri, Boehringer Ingelheim

00:00:00: The idea is that there's an engine learning from the past, how did your processes behave?

00:00:07: What are you volatility as to forecast accuracy and supply of volatility.

00:00:14: How is the distribution of transportation times?

00:00:18: And then the twin will run simulations with feeding future events like forecasts market developments And making a proposal on how you should set safety stocks and your replenishment quantities to your distribution centers.

00:01:00: I found this approach and initiative super interesting.

00:01:07: It creates transparency end-to-end over the full supply chain, it also supports decision making by simulating but also by optimizing for example inventory level.

00:01:21: with that it unfolds very Also Christian takes us behind the seats.

00:01:34: So what he had to do to build this system up or challenges, he was facing and but also obviously what you learned in this process.

00:01:44: enjoy the episode.

00:01:48: hi Christian welcome to The Data Culture podcast.

00:01:51: Hi Karsten!

00:01:53: Great to have you Christian.

00:01:55: You delivered a great speech presentation at the last data culture summit and your topic

00:02:02: was

00:02:03: the digital twin in supply chain.

00:02:07: And this data culture summit, it's hosted thankfully by Birger Ingelheim and you are at Birger running this topic or you're responsible for these topics?

00:02:18: So my first question would be maybe give us a bit of context.

00:02:22: so what is a digital twin if you think about supply chains And how is that basically?

00:02:31: How did you come to that idea.

00:02:33: So first of all, Kasten thank you for having me here and also for hosting the data card resummit.

00:02:40: It's a pleasure joining you now in your podcast and discussing it topic.

00:02:43: I'm truly passionate about its um...how we can transform the way we make decisions in a complex supply chain environment.

00:02:51: Um..and so when it comes to what is digital twin silos?

00:02:59: that we connect people, but we get real enter into and transparency regardless of systems or organizational boundaries.

00:03:07: And then you can use technology to support us in making decisions by utilizing all the data We have.

00:03:15: excellent I think That's a great overview.

00:03:18: maybe a bit more in practice.

00:03:19: What does it mean?

00:03:20: so obviously?

00:03:21: You move for lot of goods and you move may be a lot of material.

00:03:25: So what do you capture?

00:03:28: Is it like, can I imagine is end-to-end from a supplier to customer?

00:03:34: or what's the scope here?

00:03:37: So in general this scope of a twin is capturing all activities on material and transaction level.

00:03:45: From the vendor where we are sourcing materials... To the point that say its very sellered for first paying customers.

00:03:54: So it's sourcing raw materials, its API is emphasis.

00:03:58: It's manufacturing of bike truck products that finish finishing off goods finished in packaging and then comes to distribution sending the product to distribution centers inventor final find customer.

00:04:12: so I would imagine for pharmaceutical company like Buringer one of their key processes.

00:04:17: two manage

00:04:20: Sure, so at least usually farmer supply chains are rather complex.

00:04:25: It's a highly regulated environment and well it's not far alone.

00:04:31: you also find in the food industry and so on.

00:04:33: of course You have to deal with some boundaries on shelf life transportation And temperature handling materials.

00:04:42: So from that perspective there're enough challenges here To keep an eye on it.

00:04:50: If you would compare it with other industries, these regulatory requirements.

00:04:55: What's the effect of that?

00:04:57: Is it more monitoring or do in general have to fetch more data?

00:05:03: what does this mean?

00:05:06: Well usually is that should take care your products are available on market so patients rely.

00:05:18: Of course, we are also obliged to keep an eye on that.

00:05:22: You can't afford just being out of

00:05:24: stock.".

00:05:26: Then there are other boundaries you may have when it comes to launching a product or changing products because usually you need approval from authorities so your somehow linked to decisions outside the own decision space until this sometimes sets... uncertainty and complexity on how you are managing changes.

00:05:53: Super interesting if You now maybe look into it in a bit more detail, so what's the digital twin here?

00:06:02: What did your build?

00:06:03: How does it look like?

00:06:06: So at the end of few...how Does It Look Like At The End?

00:06:10: It's behaving like an app In A web browser And It's a visualization of all the blight chains, so it totally trim to

00:06:22: how

00:06:23: you humans are consuming data and information.

00:06:27: It is not this on our tables with our crafts dynamic class who can move?

00:06:32: You can click on your can dive in.

00:06:34: so we Their maps.

00:06:38: there are asset that crafts.

00:06:40: you see diagrams.

00:06:43: And if you go there, at the end of your really start that you are depicting the area for supply chain.

00:06:48: You're interested in.

00:06:49: this can be a plant it can be reach and you want to analyze but off course we can also see.

00:06:56: please provide me with entire supply chain?

00:07:00: And then you start very high level UC for example.

00:07:03: okay these our hours apply chain nodes from renders chemical plants pharmaceutical plans distribution centers customers.

00:07:12: Are they connected by product flows?

00:07:15: So, at least all those visuals support you a lot understanding what is going on.

00:07:23: Who suppliers are contributing to which final product in the market.

00:07:29: and The second thing we build this but don't only focus how it's moving through the nodes But can also tell your story Which resources or machines are utilized who process those goods.

00:07:42: So it's already linked to the more operational perspective on resource management and machine management.

00:07:50: Interesting, let's dive a little bit deeper into that.

00:07:54: so first That sounds pretty easy.

00:07:57: you see the whole truth supply chain.

00:07:59: for me?

00:08:00: That sounds like a tremendous challenge.

00:08:03: There must be hundreds of systems involved in terms of delivering data maybe even IoT sensors and so on.

00:08:15: So it's my perception, right?

00:08:18: That this is the first tremendous effort to build... This actually data integration.

00:08:23: Well with last but not least one of the biggest challenge you have Of course at very beginning that you know where your data are what kind of data do want use sorting out need for your vision what is available?

00:08:40: sometimes you meet some.

00:08:42: Some grounding and rethinking, but I think before this it's even more important than the book was a very intense discussion when we started building at What is?

00:08:52: The first use case in water support of scope because depending on what do want to model And off course It will have huge impact on on that model You are creating or how you create the data model and then also which kind of data you need.

00:09:08: And so for example, we decided that we do not want to make any visuals where you see how product is physically moving through a plant with storage locations at the buildings.

00:09:19: The other thing we did was that if current state does not want achieve real-time it means they really get the current state as now their vision will be on a daily update.

00:09:34: So then you can see okay It's lacking one day.

00:09:38: But for the decision-making processes we want to support, we figured out it's efficient.

00:09:43: And this ends up.

00:09:44: then finally that, for example... We were able to carve our internet of things and don't need sensors but can rely on systems.

00:09:53: those sensors are already providing data too.

00:09:56: So from a system set I never counted them.

00:10:03: To be honest they're a lot.

00:10:06: So, of course there are ERP systems.

00:10:09: We're running for managing transactions There.

00:10:12: our system we use for special planning purposes Network Planning.

00:10:18: their systems we used to manage our risk and resilience processes Quality System.

00:10:23: so it's a its the usual bunch of systems you have.

00:10:27: And of course the challenge accessing them getting the data knowing which data to use then finally connecting them So that you know, the material in system A is same than material in System B. But it's possible if we have right people at a desk who can do this?

00:10:46: Yeah, let's talk about people!

00:10:49: You made an interesting distinction between operation steering and you said there are indifference systems And your more on decision side as decisions supports Who?

00:11:00: other users of their digits when here?

00:11:03: Who's using it?

00:11:04: So we define the digital twin on use cases.

00:11:09: It is not that if you want to make a digital twin and now have one, think of who may be able to do this.

00:11:15: With the use case in scope so far let us turn back way.

00:11:19: This might change for the future as we extend them but really heading towards the global supply chain community during Ingeheim.

00:11:27: The colleagues working at the Global Arena As well as the colleagues working as demand planner supply planners.

00:11:34: In we in the local organizations and supply sites on commercial side,

00:11:39: okay?

00:11:40: And you also mentioned that teams.

00:11:44: so what team do You need to build and run this?

00:11:50: We set up the development of a screw the twin in an agile way.

00:11:55: It's not but we are running pure scrum.

00:11:58: So it's more HIR, Kanban systems.

00:12:01: At the end there is a core team of ten members created from IT people being responsible for the core system From data scientists and engineers.

00:12:15: There are people around focusing really on statistics And front-end design and user experience as a product owner, being the one somehow coordinating prioritizing and bringing with value case.

00:12:34: And for business story to these of course more technical teams.

00:12:39: um.

00:12:40: and then there is another part of the endeavor we started very early onboarding people from businesses.

00:12:50: so even before we start coding trust line We had a very intense scoping phase running through organization making interviews getting pain points of today's processes, wishes constraints we should follow.

00:13:04: Sometimes which is we can fulfill.

00:13:06: sometimes maybe it will take some more time to do so because you have to manage our data first.

00:13:13: and so at least there are people being a subject matter expert out the day already entering test systems and developer system testing from very beginning to guide us what is working, and not already validating if the connections we made are correct or not.

00:13:32: So it's ten developers more than fifteen to twenty subject matter experts around the globe at the moment joining others in this endeavor creating something new

00:13:44: Interesting.

00:13:45: Let's talk about the use case a little bit more, maybe first from a functional perspective.

00:13:50: so you already mentioned that obviously transparency across supply chain was one of your major goals and what do achieve?

00:13:59: You mention end-to-end visibility into this supply chain.

00:14:05: So that seems to be like a primary target.

00:14:07: But then I'm sure you also support some analytical tasks.

00:14:11: so before they take the decision, they get more insights into their data?

00:14:17: Is it correct assumption?

00:14:19: Yeah!

00:14:20: The first target was designing between as addition-to-port system As we defined in the First Decision Support Use case.

00:14:28: and of course if you create those kind models You first need the model to optimize and simulate.

00:14:34: And of course, then getting the transparency is a kind of side effect Of course bringing some value off its own but it was not primary target.

00:14:44: It was more than be The yeah matter of incidents that it is additional there

00:14:50: Mm-hmm

00:14:51: of course helping that if you are still working on simulation That they're some other value already.

00:14:57: so you can Still Create value to our organization.

00:15:02: so they get the fast feedback on what it's possible.

00:15:04: What not.

00:15:05: and Yeah, I'm the value case itself.

00:15:08: It's a decision support system on The parameterization of all inventory targets.

00:15:15: Okay, so interesting.

00:15:16: So you say simulation and optimization was actually the primary target.

00:15:21: transparency is a side product that's super interesting?

00:15:24: So maybe it can give us a bit more flavor into that.

00:15:27: How does they look like?

00:15:28: how do I simulate what?

00:15:30: What kind of optimized?

00:15:32: so in day In the first step.

00:15:34: we said okay B We want to somehow optimize.

00:15:41: or let's see optimum.

00:15:44: It's not optimized that we didn't do it so far.

00:15:47: We want to change the way, we are setting for parameters sort of that way make it faster closer to real data and much more convenient.

00:15:57: or community The decision itself is not new.

00:16:01: this why I say we changed your way?

00:16:03: We made our decisions and not we remake new decisions.

00:16:06: So um be the idea said there is a an engine learning from the past, how did your processes behave?

00:16:15: What are you?

00:16:15: volatility is how it's forecast accuracy.

00:16:18: How does this apply?

00:16:19: volatility and how is the distribution of transportation times?

00:16:27: so sometimes its half a day if we go by track but with ocean freight can be far more.

00:16:34: yeah.

00:16:35: So all these parts of the input layer.

00:16:43: And then this, the twin will run simulations with feeding future events like forecast market developments and making a proposal on how you should set safety stocks and your replenishment quantities to your distribution samples?

00:16:59: Okay super interesting!

00:17:01: You already mentioned value case.

00:17:05: did you measure benefits or values get feedback, how this approach improved the situation or maybe even some measurable outcomes?

00:17:18: Yeah.

00:17:19: At the moment to be honest we are still thinking on that.

00:17:21: now of course you can go the usual way but try to measure okay... How did my inventory levels behave?

00:17:31: it's not cramped and will always go down.

00:17:33: You see they also stay the same in certain areas.

00:17:36: increase there is no obvious direction.

00:17:41: what a success.

00:17:43: So how often do people or colleagues really adopt the proposal of the Twin and its engine to propose solution?

00:17:53: How close do they stay through those solutions, I think that's better because then you can also understand okay why didn't we follow?

00:18:01: so it was something missing.

00:18:02: are their other constraints known phrase upfront?

00:18:06: Is there still some options in getting better simulation results?

00:18:13: Do we have to change

00:18:13: something?".

00:18:16: Because at the end, between will follow the obvious targets on supplying with patient.

00:18:22: Of course he would avoid destruction.

00:18:24: so you'll take care of our capacities.

00:18:27: So... We must take care that both results are really when used and if not then why and maybe tweak between and at further aspects of decision-making.

00:18:40: Okay, so you would say that this constant thinking and monitoring of adoption is one of the key parts to make this a success?

00:18:51: I think one important aspect when it comes to adoption is that even though there's transparency maybe inside effect having this transparency may be the key element talking a lot about with our supply chain meta experts at the moment when they give us feedback, testing for front ends and trying to understand the recommendations of the twin.

00:19:19: So in usually if you run formulas like... You may find them in textbooks on economic order carriages And there's actually stocks where you enter some data and then get your number.

00:19:31: But it is very hard to understand what are the effects of my decisions If I go up or down, what is the impact on the supply side?

00:19:42: For example with all of these costs and efforts there.

00:19:46: And on the other hand if i am more keen in getting a higher lot size for example because maybe I need it where I want to be.

00:19:56: What are the effects in the commercial site so that when we'll support this hopefully then also get adoptions and more understanding across the supply chain.

00:20:07: If we really showcase with every recommendation, what are the economic effects if you follow this recommendations on the crucial targets in aspects of local organizations which can be a supply site?

00:20:21: With managing down times and setups is not productive time so I want to have as little possible off them while commercial sites of course they look for shelf life that can provide it to customer.

00:20:37: i don't want to scrap product issues but you want to manage the working inventory.

00:20:44: So we hope and this is feedback we get between create a lot.

00:20:49: transparency on helps understanding the text of decisions and then also, accepting maybe the result, even if it may not be preferred optimum for local organization.

00:21:04: Super interesting!

00:21:05: So can you say that you produce better decisions because of digital transfer?

00:21:11: This is what we are heading to.

00:21:14: at least they're more data-driven.

00:21:19: If there's a better... Maybe.

00:21:23: but this I think still has been proved by our ambition

00:21:28: Now, I think we've got a pretty good understanding now.

00:21:32: Why is it important?

00:21:34: What can do?

00:21:36: maybe share little bit more about the technology side of things.

00:21:40: so how did you build?

00:21:41: what are the importance components?

00:21:44: So...I don't know them all by heart.

00:21:49: give your ruck oversight at the end.

00:21:53: if you work with data science have to leave me standard reporting tools to usually use.

00:21:59: So we are cloud-based with a cloud provider where we also host our data catalogs, We create and provide enterprise wide to share data make data available for being for cooperation.

00:22:14: Because I think you all have to understand that if we are working this between It's not that we're the only one who are trying too To make use of data and get insights.

00:22:26: try to get turning data through transparency, better decisions.

00:22:30: It's an enterprise-wide endeavor with all areas working in this company.

00:22:36: so the twin is one contribution and from that perspective of course we also have a target where there will be marketplace for data from various systems managed by data owners you can use later on for these cases.

00:22:49: So first off we started in our discussion on providing data, of course.

00:22:53: We had to take care that the data we need are staged through this data counterpart if not there already.

00:22:58: That's tremendous work.

00:23:01: so meanwhile you can use with two datasets just make your model and somehow validate it.

00:23:07: but to make a final call you need the data there.

00:23:11: And one central aspect is having this data platform at this data foundation place in an acceptable state.

00:23:21: That's very bad thing, right?

00:23:24: Then of course you get the entire stack architecture stack and cloud service with various database technologies as no flake so on.

00:23:37: And finally when there is one piece of software we are creating which is carrying data model into simulation engine optimization engine of a digital twin Maybe within there it is that you can think of, we are having various agents.

00:24:01: It's a multi-agent approach but we're simulating for example a Rails operation and an RAP.

00:24:08: We are simulating our customer or forecaster in the machine.

00:24:12: so from that perspective if you kind of think about this simulation as working like real multi agent approach very really is limolade roles and elements of a supply chain which interact with each other.

00:24:29: So just had to smile, so you are working with agents before even AI agents were basically invented?

00:24:40: In that sense yes I think the idea of having both those elements it's not really new.

00:24:46: we knew for decades in IT work multi-agent approach, even if you just think of real coding with.

00:24:55: I have a class for example planner and he has method like events.

00:24:59: Releasing an order, scheduling in the order or executing an order is okay, I'm on machine.

00:25:05: it's something very basic that's turned back way.

00:25:09: so be the Machine Learning and AI skills.

00:25:14: they come later out.

00:25:18: Yeah, that was the question I wanted to ask anyway.

00:25:22: Does AI play a role?

00:25:24: Because supply chain obviously is very sophisticated and established exercise in a way.

00:25:31: Optimization also serves lots of good purposes but it's very bad understood how this is done.

00:25:38: Are there any new influences now from the discussion or do you pretty stable on building already?

00:25:47: To be honest, I think if you want to.

00:25:50: You can do it without AI and machine learning but of course having both technologies in place make things much more easier or more convenient for users And i don't know they're really If you ever prove that They have better results.

00:26:07: But um They just provide very good results.

00:26:11: so first of all the foundation.

00:26:14: It's plain data science.

00:26:17: It's a moderated model where we teach the computer, okay these numbers map this are in those datas.

00:26:24: and then first of all Of course you need statistics on distributions.

00:26:27: This is pure statistics.

00:26:29: You don't think machine learning for that.

00:26:32: When it comes to guiding the search or good parameters We really use machine learning algorithms.

00:26:39: So there two of them embedded And Then One use of machine learning being in the front end.

00:26:49: When it comes to data literacy, if you work with those data and make this recommendation... Of course we are confronted by highly sophisticated statistics And We provide results on all these statistics for users.

00:27:07: But we cannot expect that every user can really read now these detailed boardships make his own gasses on that and then try to follow the recommendation.

00:27:20: So what we really do is they use large language models too, but screen those statistics and translate back the insights of both statistics.

00:27:30: two key messages to be users so he does not have to consume the statistics or forecast quality on its own But it's translated somehow back to sentence and statements.

00:27:42: what is the forecast quality, supply volatility.

00:27:46: What is driving volatility?

00:27:48: Is it transporters or supply volatility demand

00:27:52: volatility?".

00:27:53: To avoid that he really has to go deep into all these nitty-gritty details we of course provide and use them more convenient for him.

00:28:01: just consume in a kind of summary being readable for humans with the herd and to get all of these statistics.

00:28:14: And maybe it's totally lost is, off course reduced.

00:28:19: So this what we expect at least

00:28:22: interesting.

00:28:23: so a good mix or combination?

00:28:26: We're heading towards the end our episode so far super-interesting.

00:28:30: thank you Christian but still have few minutes now.

00:28:35: I think you highlighted all the good things, other successes here and the benefits.

00:28:41: But I'm sure there were also lots of challenges we had to overcome a little bit looking into it already.

00:28:47: but what are your lessons learned?

00:28:50: What were challenges that you have to overcome?

00:28:53: maybe how did you do

00:28:54: that?".

00:28:55: So i think these are the obvious ones.

00:28:58: yeah.

00:28:58: so whenever you work with data or even if they have an idea about what their data will tell them and you're testing your profiles of data, what's in there?

00:29:09: What not.

00:29:10: Of course every week to get some surprise at the end.

00:29:15: for us as a team this is fun part because just doing it straightforward would be somehow boring.

00:29:21: so but most interesting things even when I started learning about our company was that the data look different than we assume they are.

00:29:28: um i think the agile approach with good network people always enabled us solve this final solution.

00:29:38: And interestingly, it's not always that we have to remove features.

00:29:43: But very often is also the case where you say okay its in there.

00:29:47: this enables a feature which has never been thought of.

00:29:51: So It really nice endeavor and I don't want miss it.

00:29:57: The second thing when you are scoping your product We started with Yeah, it's catching the dashboards and profound ends of a file in users to avoid that.

00:30:10: They have two discuss techniques with them And now they are filled for real data.

00:30:15: but Of course you have to revisit them because It's the same effect that you see effects.

00:30:21: You can't explain what you requested at very beginning.

00:30:26: until Now we're entering a loop too to bring off course all those feedbacks and missing links there, but more on visuals.

00:30:35: And how we maybe name something or how we sort something that's easier to read?

00:30:40: It's worthwhile because if you do it... We talked about adoption!

00:30:45: ...it must be really comfortable for people to use without being frustrated by technical stuff so-on.

00:30:52: This is also something where we don't write a user handbook.

00:30:57: So the entire twin is working like an app.

00:31:00: So if you go there, You get this in-app features like and he's guiding you through with a tutorial where we can click.

00:31:07: so We don't want users anymore to read first of all manuals Guiding you Like you are used today with apps on your cell phone or somewhere else To also get their very fast positive user experience.

00:31:29: I think these are the obvious challenges you have.

00:31:34: The remaining part was then whatever you do in data science, two steps forward one step back but of course it's working and fine.

00:31:47: Christian thanks so much for all this insights.

00:31:49: let's maybe sum it up.

00:31:51: Maybe You can think about your top tips That is starting on such an endeavor, trying to bring more transparency simulation into supply chains or maybe even any other business department.

00:32:08: So I think one important piece was the big start small.

00:32:15: don't start with a story that tells you i will have be entire driven supply chain and no human touch in whatever.

00:32:25: Maybe this is a vision, but it's not what you can achieve within one year.

00:32:30: And I think it's very important that you have road map where we create value in at least yearly buckets maybe faster if you scale products like these to do not lose the momentum and really prove all those investments are worth.

00:32:52: This also brings us We didn't start with, we want to create a twin when you have it.

00:32:56: Now what do they did?

00:32:58: Really started off the use case and maybe one business decision that you want improve or take data and start there but keep in mind where are your goals.

00:33:09: so don't method things by design decisions.

00:33:15: then if you do this its not built-by technology.

00:33:21: build mobile with people and very good cross-functional teams.

00:33:26: And obviously, it's sometimes also necessary that they are discussing hard about their interests and focuses to make a strong solution where everyone can adopt it and accept it.

00:33:39: So you keep in our case supply chain expertise unit data scientists engineering guys Good product management end of course business stakeholder who are convinced As in every project support you when necessary.

00:33:54: And as a product on my role it's not the coding, its more bringing those people together and managing this community.

00:34:03: Yeah thinking capability is not in projects.

00:34:06: so of course that transparency was society effect.

00:34:11: but if we go to our roadmap The first thing was create the model and for transparency because otherwise without a model in front of parents, you can go up to simulate or optimize that.

00:34:22: I'll start with transparency.

00:34:25: Harvest the value in their start with simulation then goto optimization at.

00:34:29: finally talk about decision support.

00:34:31: um Because off course it does start with decision support if your start scoping delivering its way around And always keep in mind but most important KPI for the entire product is that people are trusting you.

00:34:47: We all discussing, but should I trust AI and why not my own brain?

00:34:52: And so creates trust.

00:34:56: But also keep in mind that you are providing a decision to port system an auto-decision systems.

00:35:01: So it's been pensions who support the position on what tool replace the decision owner.

00:35:07: And if we want to create a trust be talked about That you have need early engagement that you must enable them to test.

00:35:15: but own also two have a good data lineage to show.

00:35:18: The underlying data.

00:35:20: so if we go through some fields in the pop-up really see the SAP data field which is populating this field because people ask for, though it's appeal.

00:35:29: I also use another thing?

00:35:32: You will explain logic and take time for even if it takes half of day yeah If u say its complicated you'll understand People will say, okay but then I can't trust this black box.

00:35:46: In both subject matter experts earlier i already talked about and body data results together.

00:35:52: so what we really did is when we had the first version where you put that data send them to the supply site.

00:35:57: ask him Is this your world?

00:35:59: And we totally accept it or were expecting?

00:36:02: of course there are some elements in which they're not reflecting What's Really.

00:36:06: We told Them That It Not The Final Version But We Need Their Knowledge.

00:36:11: And by the way, with more subject matter experts you have less work.

00:36:14: it is for individual.

00:36:17: Now and if you really had a good story and to invite them usually You get A lot of support from those colleagues To go with you this journey in making this endeavor and something totally new.

00:36:32: So these are I think the best tips.

00:36:37: What should do?

00:36:37: It's connecting the community and building trust.

00:36:41: Excellent, I think that's super helpful advice.

00:36:43: Christian thanks so much for all this and i wish you are the best for your future endeavors here.

00:36:50: And i can say a lot today is bye-bye and see us soon.

00:36:54: Thanks for invitation.

00:36:55: cast member talk ends.