00:00:07 Discussing SKU management for supply chain planners.
00:01:09 Dependence of SKU management on industry verticals.
00:02:14 SKU management in retail and impact on demand planners.
00:04:00 Factors driving SKU management and classic rules followed.
00:07:01 Comparing classic approach with the way Loca handles SKU management.
00:08:01 Discussing the difference between traditional demand planning and supply chain science.
00:09:37 The importance of making smart inventory replenishment decisions.
00:11:00 The ability for one person to manage millions of SKUs using numerical recipes.
00:12:19 How numerical recipes help build an asset for companies.
00:14:00 The necessity of automating repetitive white-collar tasks and the limits of automation.
00:16:02 Limitations of supply chain scientists and managing complexity.
00:17:42 Diminishing returns in productivity and coordination among scientists.
00:19:33 Comparing effectiveness and speed of a single scientist vs. a team.
00:21:09 Continuous improvement and capitalistic approach of supply chain scientists.
00:22:45 Barriers to implementing capitalistic approaches in supply chain management.
In a recent interview, Joannes Vermorel, founder of Lokad, discussed the challenges faced by modern supply chain planners, particularly in managing stock-keeping units (SKUs). Vermorel explained that the number of SKUs managed usually ranges from a few hundred to a few thousand, depending on the industry vertical. Traditionally, demand planners use spreadsheets with dozens of columns to make inventory decisions, but Lokad employs supply chain scientists who develop numerical recipes to make inventory decisions. The goal is to create a solution with “zero percent insanity,” ensuring the decisions are sensible. Vermorel argues that treating every problem as a bug to be fixed encourages a mindset of constant growth and improvement, leading to better overall performance.
In the interview, Kieran Chandler, the host, discusses with Joannes Vermorel, the founder of Lokad, the challenges faced by modern supply chain planners, particularly in managing stock-keeping units (SKUs). Vermorel explains that the number of SKUs a planner manages usually ranges from a few hundred to a few thousand, depending on the industry vertical. Retail is an outlier, as planners may deal with numerous SKUs at the warehouse level, but at the store level, they typically manage min-max templates instead of individual SKUs.
The number of SKUs a planner manages is often determined by the time it takes to complete a cycle through the list of references. Planners typically work with spreadsheets, adjusting quantities and min-max values, and classifying SKUs into categories such as top sellers and slow movers.
Vermorel emphasizes that the relationship between SKU volume and erraticity is inversely correlated. Large FMCG (fast-moving consumer goods) companies with high volumes have lower erraticity, while industries with low volume and high erraticity, such as automotive parts, may have a more challenging forecast, but the economic value is not as significant. The number of SKUs a supply chain planner manages depends on the industry vertical and the nature of the products. The process typically involves managing a few hundred to a few thousand SKUs and using spreadsheets to monitor and adjust stock levels, considering factors such as volume, erraticity, and economic value.
They contrast Lokad’s approach with the classical method of supply chain management.
Traditionally, demand planners use spreadsheets with dozens of columns to make inventory decisions, focusing more on higher priority items (A) and less on lower priority items (B and C). This approach involves significant operational expenditure (OPEX) with little capitalization. The only capitalization comes from designing the spreadsheet, which becomes useful for subsequent months.
Lokad, on the other hand, employs supply chain scientists who develop numerical recipes to make inventory decisions. Their first goal is to create a solution with “zero percent insanity,” ensuring the decisions are sensible. For example, a poor decision might be only stocking a fashion store with brown and black handbags because they sell the most, while neglecting other colors needed for merchandising purposes.
By establishing numerical recipes that capture expertise, Lokad’s approach allows a single supply chain scientist to manage vast numbers of SKUs and huge amounts of stock. This approach represents a significant departure from the classical method, which would require dozens or even hundreds of demand planners to manage the same workload. Lokad’s focus is on building an asset (CAPEX) rather than just consuming resources (OPEX).
The Lokad approach questions the need for revisiting spreadsheets every day, as decisions are based on available data. Instead, it implements the thought process of demand planners through numerical recipes, possibly involving specific machine learning techniques.
Vermorel explains the importance of utilizing machine learning techniques for companies still employing white-collar workers to perform repetitive tasks, as this could lead to improved efficiency. However, he acknowledges that some jobs, like warehouse cleanup, remain difficult to automate due to the limitations of current technology.
Vermorel emphasizes that many supply chain decisions, like order quantities and price points, can be entirely automated through numerical recipes. He clarifies that this does not mean automation without human supervision, but rather deploying human insights at scale and allowing computers to handle mundane numerical work. The limitations, he says, lie in the complexity of the supply chain itself and the need for approximations to ensure that the numerical recipe remains manageable from a software perspective.
The balance between the number of lines of code and the workload of a single supply chain scientist is also discussed. Vermorel suggests that splitting the supply chain into smaller parts managed by multiple scientists can help improve the refinement of numerical recipes. However, this can lead to diminishing returns in terms of productivity, as the additional scientists contribute less to overall output.
The interview touches upon the paradox of productivity, with one person responsible for managing a vast number of SKUs, and the need for more people to handle large supply chains. Vermorel concludes by highlighting the importance of mitigating “truck factors” by having backup personnel who can take over in case someone leaves the company.
The discussion focuses on how supply chain decision-making can be made more effective, efficient, and capitalistic by leveraging technology and continuous improvement.
Vermorel argues that traditional demand planning approaches, which rely on human decision-making and spreadsheets, are limited in their ability to drive continuous improvement. After the initial setup of a demand planning system, improvements usually stall, and the team becomes stuck in a cycle of merely maintaining the system. This prevents them from having the time and resources to focus on continuous improvement.
On the other hand, Lokad’s approach aims to automate 100% of supply chain decisions, allowing supply chain scientists to dedicate their efforts entirely to continuous improvement. Although setting up this kind of system may take longer than a traditional demand planning system, it ultimately results in a more efficient and effective supply chain.
Vermorel emphasizes that treating every problem as a bug to be fixed encourages a mindset of constant growth and improvement. This capitalistic approach to supply chain management ensures that supply chain scientists are continuously building upon their improvements, leading to better overall performance.
However, there are challenges to implementing these capitalistic approaches. For decades, the technology and software necessary for such systems did not exist. Additionally, many companies did not view supply chain management as an essential function, treating it as a mere support function or cost center. As a result, there was little incentive to invest in new technology or practices to make supply chain management more capitalistic.
To overcome these barriers, Vermorel suggests that companies need to change their mindset, recognizing the value of supply chain management as an asset, rather than just a cost center. This, combined with the availability of new technologies and software, can enable a more capitalistic and effective approach to supply chain decision-making.
Kieran Chandler: Hey, with modern companies providing increasingly large catalogs and technology facilitating easier stock management, a modern supply chain planner must spin many plates. We’re going to ask just how many SKUs a supply chain planner should manage and just how many is too many. So Joannes, it seems that supply chain planners have an awful lot on their plates these days. How many SKUs does a supply chain planner tend to manage?
Joannes Vermorel: From what I’ve observed, it depends on the industry, but most companies typically manage a few hundred to a few thousand SKUs. Though there are some situations where companies manage tens of thousands of SKUs, that’s more of an exception. The typical range I’ve seen is around 500 to 1,000 SKUs across many industries.
Kieran Chandler: How much does it depend on the industry? I’m imagining that in luxury retail, you’re not managing many SKUs, but in a hypermarket, you’ll manage many more.
Joannes Vermorel: Yes, retail is probably the outlier where demand planners deal with a larger number of SKUs. However, even in retail networks, demand planners typically manage only a few hundred SKUs at the warehouse level. At the store level, they usually use min-max templates that they replicate across a large number of stores with similar characteristics. This way, they are not directly managing SKUs at the store level, but rather managing a meta-SKU or template. If you multiply the number of stores managed by the number of products, you end up with a large number of SKUs, but that’s not how the work is typically done. So, managers usually handle a few thousand SKUs at most per person.
Kieran Chandler: What factors drive the number of SKUs a person manages? Are there any classic rules that people follow?
Joannes Vermorel: The classic approach most demand planners and supply planners use is to go through a long spreadsheet, with one SKU per line and various instrumented columns that provide indicators. These indicators can include how much was sold during the last couple of weeks, the last year, or in the same period but last year to account for seasonality. Planners go line by line, adjusting quantities and min-max levels based on the data in these columns.
Kieran Chandler: Spreadsheet, well you go back from the stop, and you iterate. Potentially, you slice and dice your SKUs into classes like ABC, whatever, so that you spend more time on the top sellers and less time on the slow movers. That’s kind of it, and you see that the number of SKUs becomes very much defined by the time it takes for the demand planner to do a cycle through the list of references that this person is managing. So in that example, I’m imagining that it depends on a lot of product kind of variability. So if you’re in a company like Coca-Cola where there’s just one product, does that mean we’ve only got one demand planner?
Joannes Vermorel: No, I mean Coca-Cola has hundreds of products. And if you start looking at the sort of things that need to be planned from a Coca-Cola perspective, first they will need to pretty much plan every single channel because their channels are gigantic. Typically, the planners are going to be organized by geographies or channels, and so you end up with a planner for one channel. They have a few hundred SKUs, and so you end up being back to this sort of a few hundred SKUs per planner. They also frequently, I mean companies, very large FMCG companies, would also have to do a bit of VMI, vendor-managed inventory. So again, you would fall back to a couple of hundred SKUs per planner.
Kieran Chandler: And then how about forecasting difficulty if you’re in an industry where there’s lots of new products, something like fashion? Does that mean you’ll be able to manage fewer SKUs per person?
Joannes Vermorel: That’s an interesting thing, and again, I’m describing here what I consider to be mainstream nowadays, not the way Lokad operates. But the thing is, when you go for very high erraticity, it’s typically products that are very low volume. You see that there is an inverse correlation between erraticity and volume. If you are a large FMCG, you have high volumes, lower erraticity. But also, you would think, “Oh, if I have less erraticity, maybe the forecast is easier.” Yes, but also what you’re forecasting is, from an economical perspective, very important because we are talking about a large mass. On the other hand of the spectrum, if we are talking about a super erratic forecast, let’s say automotive parts in the long tail, then yes, it’s incredibly erratic, but also the volume is very low and the value is not that great. So even if, yes, technically the forecast is more difficult and the erraticity is much higher, the economic reality is that the economic weight of this article in your supply chain is low, and thus it doesn’t really matter whether it’s more difficult or not. It is not very reasonable to spend more time on those articles.
Kieran Chandler: Okay, and so then let’s maybe contrast the way that Lokad does it compared to the more classic approach. How does it vary compared to what a supply chain scientist would manage compared to what you can manage in a classical way?
Joannes Vermorel: So the classical perspective, and that’s why we end up with this number of SKUs per demand planner, the classical way is literally people consolidating information.
Kieran Chandler: So, people use spreadsheets with dozens of columns that explain what they should see, and then they make decisions by going through the spreadsheet. They start with the most important items and spend less time on the less important ones. How often do they revisit these items?
Joannes Vermorel: Well, they might revisit all the important items on a daily basis, while revisiting the less important ones only on a monthly basis. The time spent by the planner is operational expenditure. The work you consume, the time of your demand planner, just to do the demand planning work, there’s nothing that is capitalized. The only capitalization comes from having a well-designed spreadsheet with all the relevant columns. This part of the work, having a nicely instrumented spreadsheet, is capitalistic in the sense that you do it once and then your work is faster for all the months that follow. However, this part is just a few weeks at the beginning, and then it’s done. You don’t capitalize beyond this point.
Kieran Chandler: Can you tell us about the approach of Lokad and how it differs from traditional methods?
Joannes Vermorel: Lokad’s approach is very different. A supply chain scientist is fundamentally going to craft a recipe where you want all your decisions, right out of the box, to be non-stupid. You want to have zero percent insanity. That’s the first milestone we target when we want to go for production.
Kieran Chandler: Can you give an example of what a stupid decision might be?
Joannes Vermorel: A stupid decision would be, let’s say, you have a fashion store that sells handbags. You only put brown and black handbags in the store because those colors sell the most. As a result, the shopping window looks sad and lacks color variety. You would like to have touches of other colors, like white or yellow, for merchandising purposes. A smart inventory replenishment decision needs to take into account factors beyond sales and service-driven aspects. You want to consider the store’s appearance as well.
Kieran Chandler: So, you’re saying that the numerical recipes should capture these insights and reflect the sort of expertise that someone doing the job manually would have, rather than sticking to naive safety stock formulas?
Joannes Vermorel: Exactly. First, you want to establish a numerical recipe that captures these insights. Once you have that, you realize that you can operate pretty much at any scale. At Lokad, we have supply chain scientists who individually manage over a billion euros worth of inventory.
Kieran Chandler: of stock just one person and one person is managing something like four million skews individually. So you see there is suddenly a complete disconnect in how many skews and the amount of people one person can scale to levels that would represent, if done classically, dozens if not hundreds of planners the classical way. And by the way, we had massive conduct of change that did happen on our clients when we did roll out this sort of techniques because suddenly…
Joannes Vermorel: It doesn’t mean that all those planners were, by the way, fired. There are tons of things where you can have more added value. But the question is, if you are part of a company and what you’re doing is just cycling through a spreadsheet every single day, how is it creating accurate value for the company? Does it really, you know, do is the company really making an investment in your work, where the work you produce is generating capital for the company, you know, something that is an asset or is it just something that gets consumed? That’s capex versus opex. And the Lokad approach fundamentally is really to focus on things: capex, capex, capex. We want to have an asset.
Kieran Chandler: So how is that numerical recipe kind of building that asset then? How does that work?
Joannes Vermorel: The idea is that why should you actually revisit your spreadsheet every single day? You know, if you take a decision, you take this decision based on the data that you have. So you see, as a demand planner, when you have hundreds of products, you don’t know every single product by heart, you know every single thing that there is. No, I mean, it may happen in some very specific fields, but it’s very rare. Usually, you just correctly engineer your dozens of columns that explain what you should be looking at, and then you take a numerical decision based on that. Well, the approach of Lokad is to say, let’s implement what you’re doing in your head. And yeah, it requires some maybe some tidbits of very specific machine learning to do that. Yes, maybe there are relationships that are not easy to express in terms of just plain numerical formulas, classic numerical formulas because maybe you’re doing, in your head, a risk assessment.
So the way Lokad, for example, numerically tackles risk assessment is through probabilistic forecasts with economic drivers. But you see there are a whole series of problems, and the idea is that whenever you end up with a number generated by your numerical recipe that looks just wrong, you need to treat that as a bug and fix it. And there should be no exception, no alert.
Kieran Chandler: So you mentioned those tidbits of machine learning techniques. Should all companies be looking to leverage those sorts of technologies in their approaches?
Joannes Vermorel: I would say any company that is still employing nowadays white collars to do exceedingly repetitive tasks is just making a mistake, period. You see, no exception. There are areas where, in terms of physical tasks, some operations are still very, very difficult to automate. For example, robots tend to be rigid sometimes, and having somebody just to do something, for example…
Kieran Chandler: Operations as simple as cleanups, where there is, let’s say, a spillage of oil in your warehouse, and you just need to clean it, are actually exceedingly difficult to automate. To have a robot that can do a bit of cleaning, take a sponge, and do that is very, very difficult. So, there are some jobs that look simple, like just take a bucket of water, a sponge, detergent, and clean it. These are the things that are exceedingly difficult to automate and are kind of beyond the capabilities of our most sophisticated robots at present time.
Joannes Vermorel: In this case, we have people to do this work, but automation is just beyond our technical capabilities. When it comes to supply chain decisions, such as numerical decisions like what should I order, how many units should I order for every single SKU that I’m managing, should I move my price point up or down, or should I do inventory transfers from location A to location B, all those questions can be entirely automated. I’m not saying it’s an automation without human supervision; that’s not what I’m describing. I’m describing a numerical recipe that has been engineered by a human, where people understand what is going on. It’s just that you deploy the very human insights at scale, letting the computer do the mundane numerical work for you.
Kieran Chandler: So, where do the limitations lie then? You mentioned that supply chain scientists will be managing billions of dollars’ worth of stock. What’s the limiting factor then?
Joannes Vermorel: The limiting factor becomes the complexity of the supply chain itself, where at some point, your numerical recipe is going to be an approximation of your supply chain. You want to be approximately correct and not exactly wrong. The supply chain scientists cannot model the reality exactly; you always have to make choices so that your numerical recipe remains manageable from a software perspective. You have lines of code; if you are one person, you have to maintain 20,000 lines of code, which is manageable. If you are one person and you need to maintain half a million lines of code, it becomes unmanageable. So, there is a balance in the number of lines of code that are involved.
Thus, if you have one person at some point, it becomes interesting to introduce, especially if you go for very large supply chains, ways to split your supply chains between various supply chain scientists so that individually they can spend more time on certain problems. For example, if you have one supply chain and you have pricing decisions and purchasing decisions, at some point, those two things are going to be heavily entangled, but at some point, it
Kieran Chandler: So, Joannes, you mentioned that having two people responsible for pricing and purchasing could be beneficial, but there are diminishing returns in productivity. Could you explain that a bit more?
Joannes Vermorel: Yes, it makes sense to have two persons just to have a higher degree of refinement in your numerical recipes for pricing and for purchasing. Nonetheless, those two persons are going to be discussing a lot and coordinating their action, which means that also you have diminishing returns in terms of productivity. So, you see, it’s just that at some point, to get this extra one percent of performance, it is very reasonable to add more people to the case, even if that means that, in terms of productivity, adding the second person only improves the productivity by a plus. If you were having like completely linear, you would say plus 100 percent production output if you were adding a second supply chain scientist. In reality, you’re only going to have like fifty percent, and then you are the third person, and this third person is only going to add thirty percent. So, it’s going to be, you know, very rapidly decreasing in terms of throughput. You have very strong these economies of scale. Nonetheless, if you’re operating with a very large supply chain, it is worth doing that if only to mitigate truck factors when there is somebody that leaves, you have somebody that is ready to take over.
Kieran Chandler: It’s interesting you mention the word productivity there because it all sounds like a bit of a paradox. You’ve got this one person that’s responsible for so many more skews. How can they be more effective and more quick in their decisions than a team of people that are responsible for a smaller scope?
Joannes Vermorel: Because the team of people who are responsible for a small scope, there is nothing that is very capitalistic in what they do. You know the only capitalistic part is the setup of a clean nice spreadsheet, your work environment, which happened, you know, during the first couple of weeks. And then you don’t pro you don’t capitalize anymore. So, you’re stuck, you’re stuck in the dead-end where suddenly all your efforts are completely consumed, and you don’t inject more intelligence in the system. You just cycle through a process that consumes all your time, and so you don’t have any time left for continuous improvements. And you see the approach of Lokad is to say that 100 percent of the effort of the supply chain scientist must be dedicated to continuous improvement. So, the setup is taking a bit longer than those again from the classical demand planning perspective. You could probably have a setup in two weeks where you just instrument your spreadsheet, and then you’re done. A setup from the Lokad perspective is probably going to take, you know, more a few more weeks than that. But, in exchange, you end up with something where you have something where you get 100 of your decision that needs to be taken on any given day done automatically, which gives you almost all your time to focus on continuous improvement.
Kieran Chandler: I see. And treating every single problem as a bug that needs to be fixed must be time-consuming.
Joannes Vermorel: Yes, and you see the thing is that if you treat every single problem as a bug that needs to be fixed, that means that you put yourself in a position where when you walk into the office every single day, essentially, you just spend a few minutes to make sure that there is nothing like you don’t have a fire to extinguish.
Kieran Chandler: Just because something completely unexpected happens, like a warehouse being flooded, there’s nothing you can do. This sort of thing happens in the supply chain and then you can spend your entire day improving your numerical recipe. And that gives you an incredibly capitalistic approach. If you think that, because then the supply chain scientists every single week is going to add a layer of improvement, that is exactly the sort of thing that was happening from a classical perspective with the demand planner during the first two weeks, but then it stops. And with the rocket perspective, this super capitalistic work never stops. And that’s why, fast forward a couple of months down the road, you end up with something where it’s just one person but it’s much more productive than the classical approach. And it is also in terms of supply chain performance, it’s also much better. Just because again, you’ve built upon your improvements in a way that is highly, highly capitalistic. Okay, we’ll start sort of wrapping things up then. So, where do the biggest barriers lie to introducing these kind of capitalistic approaches? What are the big challenges you kind of need to open?
Joannes Vermorel: I mean, the biggest challenge is that for decades we didn’t have the sort of software recipes, software technologies to make these sort of things possible. So, we had those spreadsheets, and then, for example, at Lokad until we figured out the probabilistic approach, we had a really hard time to express numerically what was happening in the head of a demand planner who was doing some kind of risk assessment. You understood intuitively what was happening. People could describe what they were doing, but then how do you translate that into formulas? That was an open question. And there are formulas, there are approaches like safety suck that just try to do that, but it’s just not working. We needed a better class of numerical recipes. So, that was one class of barriers. Another class of barriers was that many companies were not thinking of supply chain as a function of any importance. So, supply chain was just like a support function. The fact that it’s not a core function wasn’t a problem. It was a support function. It costs money, just like most support functions. You don’t expect your support function to create value for the company. So, it’s just a cost center. And as long as this cost center keeps its own cost under control, then so be it. So, you see, it was two-fold. First, the fact that there were no real technological ways to make this thing capitalistic. And then, the second thing was if people don’t realize that it’s an asset, then they don’t enter the sort of mindset it takes to really change the practice so that it becomes capitalistic. Because you see, for the supply chain practice to become capitalistic, it needs to start with an act of belief, an act of faith, if you want that it can become an asset. As long as you think that it’s just a support function that is just a cost center, then guess what? It will never grow beyond being a cost center.
Kieran Chandler: Yeah, it’s an interesting concept, this idea of changing things to be more capitalistic. So, we’ll have to wrap it up there, but thanks very much for tuning in and we’ll see you again in the next episode. Thanks for watching.