00:00:08 Decision-first approach in supply chain optimization.
00:02:25 The importance of generating decisions to understand data semantics.
00:04:55 Decision generation as a validation mechanism for understanding data.
00:06:10 The challenges of optimizing metrics and identifying errors in the decision-making process.
00:07:50 The relationship between metrics, business strategy, and decision-making.
00:09:08 The end goal of capitalistic numerical recipes for decision automation.
00:11:37 The necessity of generating decisions to understand if previous steps were effective.
00:12:40 The role of forecasting in supply chain decision-making and the problem of not being decision-first.
00:14:45 Alternative approaches in the industry and splitting up decisions.
00:16:01 Reconciliation of forecasts and atomic decision-making.
00:18:16 Learning from past failures and embracing a new perspective.
00:20:59 Learning by doing and the philosophy of praxis.
00:21:57 Industry readiness for a new approach and previous attempts.
00:24:10 Acknowledging failures and the limits of cartesian approaches.


In the interview, Kieran Chandler and Joannes Vermorel discuss Lokad’s decision-first approach to supply chain optimization. Traditional methods rely on forecasting, but Lokad focuses on tangible decisions to improve understanding of data and enhance optimization. Vermorel shares the company’s shift from a Cartesian perspective to a decision-first approach after repeated failures. He believes organizations can only learn by doing and emphasizes the need to generate decisions and iterate on them. Convincing companies to change their approach is challenging, but Vermorel thinks most are open to it once they acknowledge past failures and limitations of traditional methods.

Extended Summary

In this interview, Kieran Chandler, the host, speaks with Joannes Vermorel, the founder of Lokad, a software company specializing in supply chain optimization. They discuss Lokad’s unique “decision-first” approach to supply chain optimization, which differs from traditional forecasting methods.

Historically, supply chains have relied on forecasting methods, leaving the actual decision-making to senior staff members who rely on their intuition and expertise. Vermorel explains that Lokad has developed an alternative approach, focusing on decision-making first. This approach emerged after years of working in predictive supply chain optimization.

Initially, Lokad operated with a Cartesian perspective: gather data, clarify its semantics, apply numerical recipes for forecasting and optimization, and then produce decisions based on specific metrics. However, Vermorel realized that this approach did not work as expected. He found that focusing on the decision-making process first was not only more effective but also counterintuitive.

By “decision-first,” Vermorel refers to the process of making tangible supply chain decisions, such as purchasing an additional unit from a supplier, producing one more unit in a production line, moving stock from one location to another, or adjusting product prices. These decisions have real economic impacts on supply chains.

Vermorel shares his experience at Lokad, where the company initially focused on data preparation. They collected historical data, such as sales and stock movements, and documented the data to ensure its proper understanding. However, they often found that they were misunderstanding the data, only realizing their mistakes when they generated decisions based on it.

It was through the decision generation mechanism that they could identify errors in their data interpretation. Supply chain practitioners could review the generated decisions and point out inconsistencies, allowing Lokad to correct their understanding of the data.

Lokad’s “decision-first” approach to supply chain optimization emphasizes the importance of making tangible supply chain decisions before diving into data analysis and forecasting. This counterintuitive method allows for a better understanding of the data and ultimately leads to more effective supply chain optimization.

They explore the challenges and process of generating optimal decisions in supply chain management.

Vermorel explains that when attempting to generate decisions, they often run into issues that conflict with the reality of the supply chain. These problems are usually mundane and repetitive, but identifying and addressing them is essential for creating effective solutions.

To better understand data in the supply chain, Vermorel suggests generating decisions based on that data. If the decisions are approximately correct and reasonable, then they validate the semantic understanding of the input data. This back-and-forth process between the mental model of data and decision generation helps improve accuracy and better align the decisions with the reality of the supply chain.

The interviewees also discuss the challenge of trial and error in generating optimal decisions. Vermorel points out that the problems are not limited to data but also include the very metrics being optimized. Adopting a Cartesian perspective, one should optimize in terms of dollars of error instead of percentage of error. This involves applying economic drivers such as carrying cost, gross margin, and stock-out penalties to express the performance of decisions in dollars.

However, even when applying seemingly sensible economic drivers, the initial decisions often end up being nonsensical. Vermorel explains that subtle problems lie within the metrics themselves, requiring a back-and-forth process between understanding the data, generating decisions, and refining the economic drivers.

The end goal of this decision-first approach is to create a numerical recipe capable of generating mundane decisions automatically. This is essential for managing the massive volume of decisions required daily in large supply chains, as it allows companies to avoid employing an army of clerks and focus on the ongoing improvement of the numerical recipe itself.

They discuss the importance of a “decision first” approach and the limitations of the industry’s focus on forecasting accuracy.

Vermorel explains that the traditional top-down, waterfall approach to supply chain optimization doesn’t work. This method involves upgrading systems, documenting processes, and conducting extensive studies to create a comprehensive plan. However, Vermorel argues that until companies can generate actual decisions, they cannot know if any of their previous steps were effective.

The industry’s focus on forecasting accuracy is intellectually seductive, but Vermorel suggests that it is flawed. While forecasts are important in anticipating future market states, they are merely numerical artifacts without any direct influence on the supply chain. Improving forecasts alone does not lead to real-world learning or optimization. Instead, Vermorel emphasizes that companies should prioritize making decisions that align with reality.

To illustrate the limitations of focusing on numerical artifacts, Vermorel describes how companies may create short-term, mid-term, and long-term forecasts. Instead of solving the initial problem, this approach creates multiple forecasting problems and requires additional effort to reconcile the different forecasts. This only serves to make the situation worse and still does not provide a clear connection to reality.

Vermorel champions a decision-first approach, stating that decisions are atomic and well-defined, which can lead to real-world learning and effective supply chain optimization. He emphasizes the need for companies to confront reality through making decisions, which then allows them to assess the effectiveness of their optimization efforts

The Founder shares his experience of initially trying a more classical, Cartesian approach, which failed repeatedly, and emphasizes the need for a decision-first perspective to avoid making mistakes.

Vermorel believes that organizations can only learn by doing and stresses the importance of generating decisions and iterating on them, rather than attempting to develop perfect solutions in a top-down manner. He acknowledges the difficulty of convincing companies to change their approach, as many have tried various supply chain optimization systems without success. In terms of readiness, he thinks most companies are open to adopting a new approach but must first acknowledge the failures of previous attempts and the limitations of traditional Cartesian methods.

Full Transcript

Kieran Chandler: Today on Lokad TV, we’re going to understand the alternative approach of putting decisions first and understand how this can improve the way an organization operates. So, Joannes, perhaps you can start off by telling us a little bit more about what you mean by a decision-first approach.

Joannes Vermorel: The decision-first approach is a very specific angle that we uncovered after a couple of years doing the job that Lokad does, which is basically predictive supply chain optimization. When I started Lokad, I had a Cartesian perspective where you want to optimize something, so you say, “I’m going to have data, clarify the semantics, then I’m going to apply a series of well-defined numerical recipes: forecasting, optimization, and then I’m going to specifically go for certain metrics, and we are going to apply all that, and then we are going to get good decisions.” However, it turned out that this approach does not work. This is absolutely not how we carry out projects nowadays, and the way we do it is very deeply counterintuitive.

When I say decision-first, I mean it’s literally about producing a decision, something that conforms to reality. What do I mean by decisions? I mean mundane supply chain decisions like deciding to purchase one more unit from a supplier, deciding to manufacture one more unit in your production line, deciding to move one unit of stock from location A to location B, or deciding to move the price of some product up or down. These are tangible, physical decisions that have a real economic impact on your supply chain. When I say decision-first, I mean that the first step is literally to start by taking one of those decisions before doing all the rest, which sounds very bizarre because you would think that all the rest comes first, but no, it’s the decision that comes first.

Kieran Chandler: Let’s talk a bit about how you came about that idea. What was it you experienced at Lokad to come up with that approach?

Joannes Vermorel: We realized that when we started, we needed to do data preparation. For example, if we wanted to optimize the supply chain, we needed data, just very basic historical data about historical sales, stock movements, and these sorts of things. You would think that you can document the data to make sure that you understand it, and that’s what we did. However, the problem was that every single time we were documenting the data, we were kind of getting it wrong. But we would not realize that we were getting it wrong before we got to the point where we generated the decision. It’s only thanks to the decision generation mechanism, where we produce a decision such as, “Let’s move X units of stock from this location to this other location,” and then we have a supply chain practitioner look at the decision and say, “Well, that’s just plain wrong. You clearly misunderstood the data.”

Kieran Chandler: In situations where we shouldn’t be doing that, there is a very valid reason for not doing it. It’s not necessarily advanced; it’s very mundane. For example, you don’t have the capacity, or you think you have 1,000 units left in the first location, but in reality, there are only five left. So, you can’t even move the 50 units that you want to move. There are problems like this.

Joannes Vermorel: It’s interesting that you start by generating the decision, and then you have supply chain practitioners with some experience who can, at a glance, tell you that it’s a flawed decision. Then you realize that you have many problems that you hadn’t identified.

Kieran Chandler: It’s certainly counterintuitive, isn’t it? Because you’d think that by doing the data cleansing and understanding how the data is structured, you’d already eradicate some of those flawed decisions.

Joannes Vermorel: When I say we generate flawed decisions, I mean that we generate decisions that conflict with the reality of the supply chain in a relatively mundane way. There’s a lot of intelligence needed to generate the decision itself. The key takeaway here is to understand that the only way to make sure you understand the data properly is to be able to generate a decision out of it. If this decision is approximately correct and reasonable, it validates the semantics that you believe are applicable to the input data and that you’ve used to create the model in the first place.

However, it’s a mechanism where you have to go back and forth between the mental model, which is just the semantics of what you think the data means, and the decision generation mechanism that generates the decision, which is where reality gives you some feedback on what you’re doing. Then you’ll realize that there were plenty of things that you were getting wrong about the data, so you go back and forth.

Kieran Chandler: How do we get to the final decision then? It seems like a lot of trial and error. How much trial and error does this take?

Joannes Vermorel: It’s even worse than that because I’ve only discussed the problems that you’ve identified about the data. It’s not just the problems with the data; it’s also the very metrics that you optimize. If you go for a Cartesian perspective, you’d say that you can’t optimize what you don’t measure. At Lokad, we advocate for an optimization done in dollars of error, not percentage of error. So, you need a metric expressed in dollars that represents all the economic drivers that are applicable. For example, in the case of stock, that would mean taking into account economic drivers such as carrying cost, gross margin, and stockout penalties.

Kieran Chandler: Can you explain how you measure the performance of decision-making at Lokad?

Joannes Vermorel: Yes, this is how you measure the performance of your decision expressed in dollars. Logically, you would say, “I take the DR that I understand, I take a statistical high dimensional learning capabilities that’s a forecasting part, and then I take high dimensional optimization numerical optimization capabilities, and I apply a metric that reflects my business strategy and I understand my drivers.”

Kieran Chandler: And what happens when you apply this metric for the first time?

Joannes Vermorel: You will end up with tons of nonsensical decisions, and that’s very puzzling because all your economic drivers, they look kind of not self-evident, but almost, you know, carrying cars. Gross margin, I mean, we are not talking about super advanced stuff. Yet when we apply that to decision, we invariably end up with very dumb poor decisions. And what do those decisions reflect? They reflect that mistakes, thoughts, are subtle problems, lies in the metrics themselves. And thus, just like we had with, they are, you know, a lot of back and forth between your understanding of the data and the decision that you generate on top of that. You have a lot of back and forth on the economic drivers, how you understand them, and what sort of decision that you generate. We always talk about this idea of you can’t optimize what you don’t measure. So what’s the end goal? Is the end goal that the data validates the decisions that you’re making?

Kieran Chandler: And what is the end goal according to supply chain practitioners?

Joannes Vermorel: The end goal is to have something that is capitalistic. You see when we say decision first, the idea is that we want to deliver a numerical recipe that generates all those super mundane decisions automatically. Why? Because you have so many of them. Our largest clients have millions of them to generate every single day. So either you have an army of clerks, which many companies still have, or you decide, “Okay, I’m going to have like a miracle recipe that does this work,” and then all the efforts that I’m still pouring into this area are for the ongoing improvement of the numerical recipe itself.

Kieran Chandler: And what’s interesting about this approach?

Joannes Vermorel: Interestingly, until you actually generate those decisions, all the other things that you can do, you’re not even sure that they work. That’s the thing because if you take this kind of cartesian top-down perspective, you would say, “Okay, my plan is first we upgrade the ERP to new capabilities XYZ. Okay, that takes six months, and then we are going to document and clarify order they are, and then that’s XYZ stuff again and even push everything into a data lake, and then we are going to do another six-month study with maybe external consultant to completely clarify the strategy and have a complete modalization quantitative.”

Kieran Chandler: Of the business drivers, and then we are going to finally put all those things together to generate this automated execution of the optimized decision. That looks like, you know, a plan with a nice waterfall outlook where you go from phase one to phase two to phase three. But literally, this thing does not work at all, and that’s probably the most frustrating lesson. Until you get to the point where you actually generate decisions, you don’t have any idea whether any of the steps that you took before is actually working at all. And that’s something a bit shocking. You would think that you could have a plan that you can execute reliably, but no, the lesson is if you don’t have this contact with this reality, this feedback loop, you don’t know, you really don’t know.

Joannes Vermorel: Yeah, I mean the industry is very focused on this idea of forecasting and forecasting accuracy, isn’t it? Why is this kind of view that enterprises, companies, and consultants push so hard?

Kieran Chandler: Because it’s very seductive intellectually. It looks like something very reasonable.

Joannes Vermorel: I mean, yes, forecasts are important, obviously. Because every time you take a supply chain decision, it’s basically a statement that you’re making about a future state of the market. If you pass a purchase order for raw materials to a supplier, basically you are implicitly making a statement about the state of demand in the future. So supply chain, because we can’t teleport and because we can’t just instantaneously 3D print everything, it’s all about anticipating the future state of the market. So companies try to approach that rationally and say, “Oh, let’s do a round of forecasting improvement,” that’s basically what we discussed in our Naked Forecast episode. And then you end up with all sorts of problems, and actually, when we revisit from a slightly different perspective, which is this possibility of this episode of decision first, I would say, well, the problem is that you’re not decision first when you’re doing that. If you say, “Oh, let’s push for better forecasts and then we will see what we can make with those better forecasts,” you’re not decision first. You’re starting with a numerical artifact. All sorts of forecasts are just numerical artifacts. They don’t have any direct influence on your supply chain. And then you say maybe it will improve, and my feedback is the lesson learned of a decade now running supply chain projects at Lokad is that no, you’re not going to learn anything. Learning only comes from reality, which is the ultimate arbiter of who is right or wrong. And when I say reality, I mean the way to make sure that this supply chain optimization project is actually doing something that is in the right direction are those decisions, because those decisions are really the crux of putting yourself at risk with regards to what the reality can tell you whether it’s actually working or not.

Kieran Chandler: Okay, and putting decisions first is certainly the way that we would look at it, but what about these alternative approaches that are out there in the industry? And how about sort of like splitting up those decisions?

Joannes Vermorel: Yes, I mean, for example, what typically happens is that when you start focusing on artifacts, especially numerical artifacts, you have very little constraints attached to them, and thus you can basically split them or divide them. For instance, you can have a short-term forecast, a midterm forecast, and a long-term forecast. When you do that, you create more problems. You had one problem of doing forecasts, now you have three forecasting problems, and you also have the problem of different teams needing to reconcile their work. So, you will need to reconcile the long-term forecast with the midterm forecast, and then you will need to reconcile the short-term forecast with the midterm forecast, and maybe even the short-term forecast with the long-term forecast. You had one problem, now you have six. You just made the situation worse, and it’s still very unclear whether what you’re doing has any connection whatsoever with reality.

Reality will not tell you whether what you’re doing is wrong in a very indirect way. And that’s the beauty of decisions. If we go to this decision-first approach, decisions tend to be highly atomic and well-defined. The way we see it at Lokad, they are atomic; you cannot subdivide them. If I say, “Purchase one unit from this supplier today,” you know it’s as atomic as it can get. Sometimes you can refine it, like “Purchase one unit from this supplier today and have this unit shipped by truck,” because maybe there’s an option to ship it by train or something else. So, at Lokad, we have the decisions that are, by definition, completely atomic. You can’t really subdivide them, which is very nice because it also puts limits on the stuff that you can make up.

Focusing on decisions prevents you from making entire classes of mistakes, such as splitting things or creating made-up work. It’s a real change in perspective.

Kieran Chandler: And so, what do you do to convince those organizations that are very used to taking a more classical approach?

Joannes Vermorel: That’s the crux of the…

Kieran Chandler: The problem is that it’s very difficult to convince because first, I wasn’t convinced myself. You know, I didn’t start with this approach back in 2008. I tried the more classical way, which I would describe as the Cartesian way. You just try this waterfall principle or engineering principle: clarify the input, clarify the metrics, clarify the model, and then put all that together to do an optimization in this order and it will work. But no, it’s absolutely not the sort of way it works.

Joannes Vermorel: I can communicate my experience that it had failed over and over, and we had a vast series of very painful initiatives. I mean, ultimately, when you do it the Cartesian way, what happens in practice? Well, you’re just going to do your nice project and then, once at the day you were supposed to be done, you just realize that nothing is working and you have to do it all over again.

So ultimately, it will work because when you actually try to go to production, you will see all those things and then you will have to revisit all that you’ve done before. And that’s the trick. If you don’t embrace this perspective, what will happen is that your projects will just take years because you will just do a one-year project to get it done in this waterfall perspective, nice plan and everything. And the day you try to turn on the system, you just realize that everything falls apart. So you shut it off again and you repeat. And then it takes years.

My philosophy is at first, I try just to communicate on this experiment. And then, on the more philosophical perspective, I would say everything that you learn, you don’t learn from textbooks. For the vast majority of things in life, you learn by doing. It’s actually exceedingly hard to learn anything by not doing it.

Yes, you can, in theory, learn a foreign language by just picking a book, reading it for six months, just memorizing it, and then you’re fluent. I don’t know if I’ve ever seen anybody capable of doing that. It might, in theory, be possible, but in reality, no. You would try, you would stumble, and sometimes people would just not understand what you’re saying, and then you would gradually get better. But you see, you learn by doing. That’s the kind of the old Greek concept of praxis. And I think for something that is a very complex system like supply chain, that there is any other way to learn, I think this is folly.

Kieran Chandler: Okay, well let’s start sort of wrapping things up a little bit. But all sounds great theoretically, but it’s far from established. So do you think the industry is really ready to embrace a whole new approach, or do you think it’s very set in its ways and the old approaches are too far ingrained?

Joannes Vermorel: The funny thing is that, in terms of readiness, I see the vast majority of companies who have been deploying supply chain optimization systems for decades, especially large ones, and I mean, it’s insane. They have, I mean…

Kieran Chandler: Even the term, you know, ERP (Enterprise Resource Planning) stands normally for some kind of supply chain optimization thing. The vast majority of the ERPs that have been deployed over the last three decades do not deliver any value in this area. Yes, they deliver a lot of value on the management side. You know, you can keep track of your stocks, you can have real-time visibility, which is very nice. You can have a lot of workflow automations for invoices, payments, keeping track of delays, all those sorts of things. Very good. But when it comes to predictive optimization, the state of the industry is almost nonexistent. There are very few things in place that work, and it’s not for a lack of trying. Frequently, the large companies that we serve, we are typically their attempt number five, six, or seven at this thing. So clearly, in terms of readiness, I think a lot of companies are ready to do it because they have been trying for the last three decades. It’s not for a lack of trying.

Joannes Vermorel: I think the crux of it is that most companies have not even tried to acknowledge the fine print of their failures. That may be related to the power of negative knowledge, which is something we discussed in another episode. Maybe some companies need to start focusing on those complex problems, those wicked problems, or problems that resist the nice Cartesian approach where you can just take a pen and paper, sit at the desk, think hard, come up with a solution, deploy it, and it works. That’s not how it works in supply chain management. The supply chain is way too messy. You need to generate those decisions, not just act on them. You can generate the decision, have some people look at them, and then they will tell you, “No, it doesn’t work.” That’s already enough. But in terms of readiness, back to your question, I think a lot of companies are ready, and that’s going to be an uphill battle. For Lokad, we still need to convince them that there is no other way but to put their initiative at risk with regard to reality, instead of trying to do it right the first time from a Cartesian, top-down perspective, which unfortunately does not work.

Kieran Chandler: Okay, brilliant. We’re going to have to wrap it up there, but thanks for your time. So that’s everything for this week. Thanks very much for tuning in, and we’ll see you again in the next episode. Bye for now.