00:00:00 Safety stocks aren’t safe: episode kickoff
00:05:37 Classic safety stock: normals plus service levels
00:11:14 Service level targets mislead; fashion-season example
00:16:51 MOQs, pallets, trucks demand smarter allocation
00:22:28 ERP ubiquity, Excel overrides, orthodoxy’s mismatch
00:28:05 High-quantile thinking breeds dead inventory
00:33:42 Lead times: bimodal reality, fat tails
00:39:19 Rate-of-return lens replaces KPI chasing
00:44:56 98% service level can be disastrous
00:50:33 Zero inventory isn’t a universal win
00:56:10 Manual tweaks reveal automation impossible
01:01:47 FMCG: full-truck constraint, promotions missing
01:07:24 Finance alignment matters; safety stock still fails
01:13:01 Retail penalties require client-by-client arbitration
01:18:38 Service level vs fill rate, demand confusion
01:24:15 Robotize ordering; buyers stop spreadsheet babysitting
Summary
Safety stock looks scientific, but it substitutes a target percentage for economic thinking. It optimizes “service level” instead of return on scarce capital, and it ignores real constraints like MOQs, truck capacity, price breaks, perishability, and wildly uneven stockout costs across SKUs. The math assumptions often don’t match reality, so planners override outputs in Excel—proof the model doesn’t work. The fix is to start from economics: allocate resources by expected payoff, then automate decisions that are sane by default.
Extended Summary
Safety stock is sold as a “scientific” way to be prudent: assume demand and lead times behave nicely, pick a service level, and compute the minimum inventory needed to avoid stockouts at that probability. The problem, Joannes argues, is that this is not economics—it’s arithmetic dressed up as wisdom. It optimizes a made-up target (a service level percentage) rather than the purpose of a business: allocating scarce resources to get the best return.
Once you look at inventory as capital, the holes become obvious. Safety stock offers no guidance on how to allocate money across thousands of SKUs, how to handle minimum order quantities, price breaks, truck capacity, or the everyday reality that replenishment decisions must fit hard constraints. It tells you a “target level,” and then the real world immediately forces rounding, bundling, and compromises—precisely where prioritization matters, and precisely what the formula cannot supply.
Service level itself is exposed as a poor proxy for “good service,” much less profitability. In fashion, high service levels near season-end are a recipe for dead stock; stockouts can be desirable if they clear space for the next collection. In aviation, a blanket 98% is absurdly low for cheap parts (where missing one can ground an aircraft at enormous cost) and absurdly high for multi-million-euro components (where stocking them ties up capital better spent elsewhere). The right answer varies wildly by item, and the asymmetry between “too much” and “too little” inventory is not constant.
The mathematics also fails to describe reality. Normal distributions imply negative demand and negative lead times—nonsense. Lead times are often bimodal: either things arrive as promised, or they go very wrong, sometimes never arriving at all. On top of that, safety stock typically ignores other uncertainties that matter—returns, scrap, regulatory shocks like tariffs, and nonlinear penalties in retail agreements.
The practical evidence is the “army of clerks” overriding outputs in spreadsheets. If a system produces so many exceptions that humans must review everything, it’s not automation; it’s busywork. The proposed alternative is to start from an economic view—rate of return—then “robotize” decisions so they are sane out of the box, with manual intervention reduced to the exceptional, not the routine. In short: stop worshipping a percentage and start measuring what it costs, what it earns, and what it prevents.
Full Transcript
Conor Doherty: This is Supply Chain Breakdown, and today we will be breaking down why safety stocks are not, in fact, safe. You know who I am. I’m Conor, Communication Director at Lokad, and to my left, as always, the indomitable Joannes Vermorel.
Now before we get started, comment down below: what is your position on safety stocks? Do you think they are a machine for generating dead stocks? We’ll get to that later. Let us know your comments, your questions, and I will pose them to Joannes a little bit later.
And with that, Joannes, let’s waste no more time. Today’s topic: safety stocks are not safe. I know, having worked here for many years and having had both public and private conversations with you, having read your new book, having read articles before, you’re not a fan of safety stocks, I think is fair to say.
So before we get critical, let’s be descriptive. You’ve written about the classic safety stock position or perspective. What is that, and what are the promises it makes that you think it doesn’t cash out?
Joannes Vermorel: The classic safety stock is a model that gives you an inventory position. That’s it.
How is this model built? It assumes that you have a normal distribution over the uncertainty of future demand, a normal distribution over the uncertainty of future lead time, and it assumes that you have a service level target defined as the probability of being out of stock in your next replenishment cycle.
And it will give you the target inventory level that, according to this model, you should look for if you want to achieve this service level with the minimum amount of inventory.
So that’s, in essence, it tells you: we have an optimality here. It’s the minimum amount of inventory for a given service level, but with a whole series of gotchas.
Conor Doherty: Actually, it’s quite extensive. Give some of the gotchas.
Joannes Vermorel: I think the core problem with safety stock is that it’s a non-economic perspective. Which simply means that it is not optimizing the profitability of your company.
In fact, I would even go as far as: it has no correlation whatsoever with whether your company makes profits or not. And that’s a grand illusion, that people operate under the impression that safety stock gives them something optimal, or at least something safe, something good. But my argument is: absolutely not.
The economic aspect is entirely absent from this model, and thus, just because it is entirely absent, why would you ever expect that safety stock will give you anything profitable, or even anything good?
My experience—again, we go from the argument of “we have no economics”—and then, in practice, Lokad tried for a few years doing that and it was complete nonsense.
Yes, occasionally, the number that you get out of the safety stock formula is going to be correct, just like a broken clock displays the correct time twice a day. But otherwise it is extremely bogus.
Conor Doherty: Okay. Again, I do want to represent the other side a little bit here. There are certainly good arguments for using safety stock formulas. You’ve been in the business a very, very long time.
What is the steelman that you’ve heard to defend the perspective that you are challenging?
Joannes Vermorel: The most convincing argument to me is: “We don’t know any better.” Okay, fine.
But then this argument is a little bit like using astrology. If you don’t know anything better than astrology, you can, I guess, use that to predict the future. It’s not going to be very good. But if you have nothing else, maybe that’s a reasonable fallback.
If that is literally the strongest argument, the rest—when you delve into the math, the technicalities, everything—those arguments are extremely weak.
We can go on and on, but for example: what does that mean, an economic perspective? There are layers of criticism I can make, but economics is the study of allocation of scarce resources that have multiple uses.
So what are we talking about? First, we’re talking about money and inventory. Safety stock is about inventory replenishment, so it’s fundamentally about allocating your money into inventory.
We have alternative uses. What are those? We can order more for many, many SKUs.
First thing: if I look at safety stock at just one SKU, does it tell me how much I should order? Not really. Why? Because first, it just gives you a target stock level. So you could say, “I’m just going to reorder up to stock.”
But the reality is: unless your suppliers are retailers, they most likely do not sell stuff unit by unit. Ninety percent of B2B business is not unit by unit. Otherwise, that would be retail.
So when you’re buying, there will most likely be MOQs. They may be quantities that are more interesting: having a full box, having a full pallet, having a full truck.
So this idea that you can just reorder up to quantity and that’s it—no. There will be constraints. You can also have price breaks from your suppliers.
So first, you see that the quantity that you’re talking about is going to be rounded upward, potentially by a lot. You have your model, and then you do a massive up-rounding, and suddenly you’re nowhere near the optimal because very frequently you end up with: “My safety stock is 15 units, my MOQ is 100, what do I do?” It’s absolutely not clear.
That would be just one example where it doesn’t even tell you what to invest.
But then the question: it’s incorrect to think of “how should I allocate my money” just for this one product, because I have many SKUs. Unless you’re a tiny, tiny company, you have many, many SKUs.
Thus, the question is: how do I allocate my dollars or euros across the board, not just how much I allocate for this one SKU?
There is a question: should I put one extra euro of stock on this SKU or this other SKU? Safety stock does not tell you that at all. It just tells you: “You should have this much on all the SKUs.” But the reality is: what if you set your service levels and then you end up with a budget that exceeds how much you’re willing to spend? How do you prioritize? Again, the safety stock formula does not tell you how to prioritize.
Sometimes you have even more mundane situations. Let’s say you’re passing an order for a supplier and the capacity of the truck is, let’s say, nine tons, and you realize your order is nine tons and a half. It doesn’t fit in the truck. It’s over capacity.
They don’t want to send two trucks because the second truck would be driving nearly empty. So you need to reduce your purchase order by this half a ton of excess, but it has dozens of different products in the truck. Which one do you pick? How do you reduce intelligently your quantities? Safety stock doesn’t tell you.
So you see the lack of economic prioritization, due to the fact that it’s a non-economic perspective, it doesn’t tell you many, many things.
Same if you have to discuss with the finance director: we could invest, let’s say, $200,000 in having permanent more working capital frozen in our inventory, or we can invest in a new conveyor belt that costs $200,000. How do you arbitrage between that? The answer is rate of return. You want to invest every dollar where it delivers the highest rate of return.
Does safety stock tell you anything about the rate of return? Absolutely not.
Conor Doherty: On that, you’ve mentioned points there that I could summarize as assumptions. I want to come back to assumptions in a moment.
But one of them, and I do want to again fairly represent criticism that has been brought forward to your perspective: you and I have had conversations with other practitioners in the past and they’ve made an argument something like this.
Safety stock is essentially a statistical parameter. It lets people hit desired service levels at minimal inventory cost. Now granted, most practitioners would not say it’s a perfect policy. It’s a somewhat imperfect, a bit of a rough heuristic.
But to say that it has absolutely no economic dimension to it is a bit of an overstatement. How do you respond?
Joannes Vermorel: No. Again, I think that is a very deeply flawed understanding of what economics is about.
Let’s go. First, if we’re not talking for a company about the maximization of rate of return, we haven’t even started to discuss economics. It’s not because there is a dollar sign on your dashboard that suddenly it becomes an economic dashboard.
We have twofold. First, we have the service level target. People assume, directly, “Oh, this is a correct target.” Why is that? It’s not.
The fact that you cherry-pick a percentage out of thin air doesn’t make it anything that is economically relevant or good.
For example, why would going for high service level, so low probability of out-of-stock, be something that is even reasonable?
Let’s have a look at a simple case: a fashion store. This is the end of the season. This is the end of the winter season. We are in spring now.
Do you want to maintain high service levels for your winter garments? We are in May. No. On the contrary, you want your service level—so probability of stockout—to go quite high, because you want to liquidate the winter collection so that there is free space in the store for the summer collection that goes in.
So you see: the problem is that service level is an extremely bad proxy for quality of service. The implicit assumption is: if we have high service level, clients will be served well. This is absolutely not the case. There is no correlation whatsoever.
Conor Doherty: You have to unpack that because a lot of people would challenge you if they were in the room when you say that.
Joannes Vermorel: As we’ve seen, for a fashion store, the right move is to let those service levels drop so that we evacuate the winter collection to make room for the summer collection.
But if we take another case: imagine a B2B distributor of electrical equipment. A company passes an order for a construction site five months from now, and they say: here are 300 product references, and for each reference we need multiple numbers of units from 10 to 5,000, because it’s going to be light switches, cable, lighting fixtures, whatnot.
The client is passing the order five months in advance because they know it’s a complex order. There is tons of things. They want to give enough leeway to the distributor to organize all of that.
But then comes the due dates, and on these dates the client company needs to have everything, because otherwise they are going to be blocked for the construction site.
If they are short on cables, then the rest of the construction site cannot proceed. They need to have all of it.
If you say, “But you know what, you have 98%,” 98% is no good. The construction site is going to be blocked. You will not be able to plaster things. You will be stuck.
So that is a situation where you need 100%, not a probabilistic thing that is only a few percent off from the target.
But again, you were granted many months to do that.
That’s why I say: this idea that service level is a correct proxy is completely bogus.
Then you also have the fact that you do not factor, as part of the safety stock formula, properly, the inventory cost. When you say you minimize cost, you don’t minimize cost. What you minimize is a very loose proxy of cost.
Just like service level is a bad proxy of quality of service, and certainly not a proxy of quality of service expressed in dollars, the safety stock perspective on inventory gives you an incredibly shallow perspective on stock.
It is minimizing the number of units in stock. That’s it. Then you can multiply by unit acquisition price and you would get something expressed in dollars, but it’s still not the cost.
What about expiration dates? Let’s say you are a manufacturing company, you do cosmetics, you buy a lot of products, chemicals, organic compounds, and they come with an expiration date.
If I have 100 units in stock today but they all expire tomorrow, it’s not the same cost as if I have 100 units in stock and they expire one year from now. Very different situations. Yet, from the safety stock perspective, it would say those are the same.
That’s why I say it is absolutely not an economic perspective.
Even when you look at cost: safety stock just gives you the lowest inventory position that will satisfy your service level target, according to a very simplistic take on what the future looks like.
That’s my problem: there is no economics anywhere in this model.
Conor Doherty: You teed me up for what I wanted to ask earlier, which was my second point on assumptions.
You’ve repeatedly pointed out service levels as essentially a KPI, a target, and then safety stocks exist, in your words—I’m paraphrasing—to satisfy that KPI, because people in companies have to justify the KPIs.
You’ve mentioned rate of return. You’ve identified two separate assumptions underlying decisions. One is: deploy safety stock so that my high service levels—maybe they’re arbitrarily taken, you would say—okay, fine, we can grant that.
Another assumption is: I make decisions that drive maximal profit per dollar or euro or yen invested—what you call rate of return.
Why is that not the norm, in your opinion? Why is one assumption so much more common, and the other, which sounds to many people very straightforward and intuitive, why is that not the norm?
Joannes Vermorel: First, that’s the folly of supply chain simplistic models that were developed in the early ’70s by software vendors who were overly enthusiastic on what would actually work.
It just became the orthodoxy. It became like the Bible, and it’s complete nonsense.
Why is safety stock so widespread? Because it has been implemented in every single ERP.
Why has it been implemented in every single ERP? Because it can be implemented in two hours by a semi-incompetent software engineer. That’s it.
So every single enterprise software vendor was able to say, “I’m going to tick the box for safety stock. Give me two hours. I give you an implementation.”
It became ubiquitous, but the reality is that it didn’t work. That’s why people are still using so much Excel in companies.
If safety stock was working, there would be no spreadsheets. You would just let the safety stock pilot your replenishment.
Yet the vast, vast majority of companies—where I’ve seen that they have safety stocks—it’s used with a lot of override. People are overriding the purchase orders enormously on top of the safety stock.
Some clients where we implemented Lokad started with their safety stock: they had over 90% manual overrides.
When you have something that generates orders where you end up doing 90% plus manual overrides, we are back to: the clock is correct twice a day. Occasionally the number that comes out of the formula would be correct, but most of the time it is not, and then someone needs to do an override.
For me, that’s where we have this massive discrepancy: we have the theory—the mainstream supply chain theory and the orthodoxy—implemented through enterprise software products, that say safety stock is a gold standard.
And we have the actual practice, where people do all sorts of things on Excel spreadsheets out of necessity, because the numbers—the inventory replenishment numbers—that come out of the safety stock formula are just plain nonsense.
Conor Doherty: Well, again, to build on what you’ve said, someone could point out that in the scenario you just described—people with what you have often deemed an army of clerks working with Excel—those people look at their safety stock formulas, they decide whatever quantity has been recommended, “I don’t like that,” and they override it, they move it up or they move it down.
Those people, in that moment, are making an economically influenced decision. Isn’t that what you’re advocating?
Joannes Vermorel: Yes, in their head that’s what is happening, because in their head they go back to: is it wise for the company? Will this bring money or cost money?
The economic perspective is very close to the super intuitive one: “Am I going to make a profit or not on this?” It’s just this intuition.
If there is perishability, they would think: not going to fly. If we are at the end of the winter collection for a fashion store, they would say: not going to work.
If we have this client B2B setting where an important VIP client is passing a massive order to a B2B distributor, the person managing the inventory will see: this client is VIP, we really need to get that done. I will even reserve the stock to make sure it’s done, and not rely on safety stock.
So yes.
But the problem is where the community needs to acknowledge that safety stock is broken. It is at a paradigm level. The perspective that goes into safety stock is wrong, and thus no matter the sophistication you bring to the table, it’s still going to be wrong.
For example, a big problem with safety stock is that it uses normal distributions for demand and lead time. This is complete nonsense.
This assumption gives positive probabilities for negative lead times. Pure nonsense. It also gives positive probabilities for negative demand. Again, pure nonsense.
Okay, let’s assume we fix that. We use a fat-tail distribution for demand. We use a distribution semi-realistic for lead times. We say either everything goes right on time, or the supplier has a problem and it can be much, much, much longer.
You’re still operating in the wrong paradigm. You would be fixing the technicalities, but you’re still moving in the wrong direction.
It’s like a software engineer saying: “Your safety stock formula takes half a millisecond to compute, I can do it in 10 nanoseconds.” Fine. It makes no difference because the formula is garbage.
Conor Doherty: Well, again, someone might say—building on what you were saying—to provide a bit of pushback.
A company could say: yes, Joannes, you’re right, there is a lot of manual override, but we’re still making money with that. Our skilled practitioners are making manual overrides that reflect the underlying financial interests of the company, and we’re making money.
So what is it you want us to do exactly? What’s the problem with what we’re doing, and what do you want us to do?
Joannes Vermorel: Companies can be profitable for a whole variety of reasons while having very poor supply chain practices.
If you look at the life of a fantastic entrepreneur—Steve Jobs—he died unfortunately quite young from an untreated cancer because he was believing in very strange theories about how you can approach cancer. He followed very strange alternative treatments, and then finally got to classic treatments quite late.
It’s a tragedy, but it illustrates: you can have an incredibly brilliant individual creating Apple, fantastically profitable, making all the right calls on many things, and yet doing very strange calls on some others.
A company can be fantastically successful because they have the perfect product, perfect technology, perfect this and perfect that, despite their supply chain practice being substandard. It’s not incompatible.
If you tell me: “We crush the competition thanks to our supply chain,” then yes, I would say you’re doing something right. If you can deliver faster than the others, your supply chain costs are way lower, you’ve been extensively robotized for the last 15 years, I would say: okay, you probably are doing something right.
That would be Amazon.
Amazon is very much on this rate-of-return optimization that I’m talking about.
But if you give me a company where there are as many planners today per dollar of turnover as there were 20 years ago, where nothing has fundamentally progressed in the last 20 years conceptually, I really challenge that what you’re doing is state-of-the-art.
If you’ve been stagnant for the last two decades, considering how much software, statistics, optimization has progressed over the last two decades—which is enormous—if you’ve been stagnant, you cannot reasonably claim you are state-of-the-art.
You should assume your practices are vastly obsolete, and that’s a very reasonable assumption.
Conor Doherty: Well, pushing forward in terms of the actual economics implications of safety stocks, because again the topic is “Safety stocks aren’t safe.”
It seems what you mean there, obviously not in a physical sense, but from the perspective of maximizing return—rate of return—financial return on your investment, which you talk about in your book and in your lectures.
What are the common unseen financial symptoms of loss due to safety stocks? Not just the carrying cost of having extra stock, but what are the other economic perils of this?
Joannes Vermorel: Safety stock is a machine to generate—this model, in practice—overstocks and dead inventory, inventory write-offs.
Why? Because fundamentally it says: “I want to go up to a very high quantile of future demand.”
That’s essentially what safety stock is: take, as an inventory position, a very high quantile. A quantile is a point in a probability distribution.
Let’s have a look at two distinct probability distributions of demand that have the same high quantile.
I say my optimistic scenario—90% up—is I sell 100 units over my next order cycle. That’s my high quantile, and that’s going to be my inventory position.
Now, I can describe two variants.
Variant number one: if it’s not 100, then on average, in the other situations, it’s going to be 80. I target an inventory position at 100, and if it doesn’t happen, it’s going to be on average 80.
Variant number two: if it doesn’t happen, in the other 90% probabilities, the average demand is zero. Zero, hit or miss.
So you have two situations: one where, if you take a beefy inventory at 100 units, most likely you will sell 80 units, and you will liquidate the bulk of your inventory. The other one is either it’s a hit—you sell 100—or it’s a miss—you sell zero—and you’re left with 100 units of dead inventory.
Should you approach those two situations the same way in terms of inventory optimization? The safety stock theory tells you yes. I say no.
Those two situations are nothing alike. They should be treated very differently.
Fundamentally, safety stock is only looking at the high quantile—the very optimistic event where you have a surge of demand.
But what if you have a possibility of a drop of demand? Safety stock doesn’t tell you anything about the risk that demand collapses. It does not.
That’s why I say it’s a machine, and that’s why I say it’s very unsafe when it comes to inventory write-offs, because it is, by design, entirely blind to potential collapse of demand or drop.
We have another problem. Safety stock assumes that the only two sources of uncertainty are demand and lead time, but there are so many others.
Returns, scrap rates—exactly. If you’re in e-commerce, returns.
Look at what happens with tariffs with the US administration over the last year: nobody can predict what the US administration will do in terms of tariffs in the next two months, 12 months, but what we know is most likely it’s going to be a bumpy ride.
Now safety stock says: “I don’t care about those other uncertainties.” But they’re consequential. You need to factor the relevant and consequential sources of uncertainty, not just demand and lead time.
My problem with safety stock is that it just, by design, ignores those. That’s why I say they are unsafe: they will generate massive costs that would have been entirely preventable if you were simply not using safety stock.
Conor Doherty: Getting a little bit of pushback, both in private messages and I can see in the chat, there’ll be questions you’ll be pushed on.
You mentioned lead times, and classic safety stock formulas treating lead times as a constant. Can you unpack more the problem with treating lead times as fixed, instead of something that varies?
Joannes Vermorel: The classic safety stock models assume that lead times are normally distributed—a bell curve.
Many companies don’t even do that. I understand why they don’t, because you end up with negative lead times when you do that, which is super strange.
If you go for a fixed lead time or a normally distributed lead time, the problem is that it does not reflect what is happening in a real-world supply chain. Not at all.
It means you are making a projection about the future that is completely bogus. It is just not the way things will unfold.
If the way you look at the future is completely bogus, why do you think the decision that will come out of this analysis will be correct? That is very strange.
Let’s go back to how lead times behave in practice. Lead times are very frequently bimodal.
You have one mode: everything, stars are aligned, everything goes smoothly, and the supplier says 11 days, and you get the stuff in 11 days. That’s the first modality.
Everything goes right. The supplier has everything in stock. They can ship right away. Then it’s just the time it takes to transport stuff.
Then we have the second modality: something goes wrong. The shipment is lost, the supplier doesn’t have the thing, the supplier has a strike, your container gets lost in a storm at sea—anything is possible.
Then the time to get what you’ve ordered suddenly becomes extremely long. In fact, this distribution doesn’t even have an average, because a certain percentage of orders will just never arrive. Lead time: infinite.
That’s why you end up with fat-tailed distributions where you can’t even compute an average, because you would have to take into account the fact that sometimes things never arrive, and you can’t average that out with the rest.
Conor Doherty: Okay. Well, I hope that helped. I’m not going to say who sent that, but I hope that helped.
I’m going to push forward. We’ve been talking for about 35 minutes, so we will get to the audience comments in a moment. If you have any other comments or questions, get them in now.
But before we do, to be somewhat more constructive—and bear in mind that next week we will have a discussion on KPIs—as an amuse-bouche, as a prelude to that discussion, underlying all of this is choosing poorly, I guess, KPIs.
So what are the KPIs that you think people should be focusing on more now, and what are some concrete steps for pushing forward?
Joannes Vermorel: It’s not even about KPIs. It’s more like the paradigm. You are not even looking at the problem as an economic problem.
That’s the bulk of my criticism for safety stock: profitability has no place. It does not even exist.
People can say: “But you can do shenanigans to make the safety stock behave in a way that would be slightly more aligned with profitability.”
It essentially amounts to: I’m going to use another technique to decide how much I should order, and then, once I have my answer, I’m going to reverse engineer this answer into a safety stock parameter that kind of makes sense.
By the way, at Lokad, we sometimes do that just because we have ERP constraints. DRP doesn’t support anything but safety stock.
In this case, we do some dynamic reverse engineering of the safety stock so that we adjust dynamically the parameters of the safety stock so that it generates exactly the purchase order that we intended in the first place.
But that is just overcomplicating things for no upside, except sometimes you have to do it because you’re suffering from a lock-in effect at the ERP level. But I digress.
Back to safety stock. What I would advise the audience: start looking at your supply chain from an economic perspective.
What does that mean? You are allocating resources: dollars, shelf space, trucks, inventory that is going to be consumed to drive a production, etc. You have resources that have multiple possible uses.
Whenever you make a choice, you need to think: I am doing an allocation. What is the rate of return? What is my resource worth, and how much value do I get out of this allocation?
If my resource is worth $1,000 and my return is $500, why do you do this allocation in the first place? It’s not reasonable.
You need to think in economic terms. Once you start embracing this economic vision, you will see that safety stock does not make sense.
That would be the correct starting point: understand that many things that are taken for granted don’t make any sense.
That’s why there is so much friction between the software system that implements safety stock and the poor practitioner who is struggling with Excel spreadsheets, where they have to move all the numbers up and down all the time because otherwise it just does not make sense.
We have this schizophrenia, as if the person manually tweaking the numbers was wrong. No. The person manually tweaking the numbers is doing the correct thing, because in their head they are doing this mini economic calculation.
It’s rough, it’s dirty, it’s not precise. That’s why we can do it better. But at least it makes sense.
In contrast, safety stock is essentially mathematical nonsense. It is scientism: it looks scientific, it gives an aura of credibility to a software system, but that’s it.
Conor Doherty: All right. Well, Joannes, thank you.
I’m going to push on to DMs and comments. There are comments to get to. I’ll go with DMs.
As always, as a journalist, I preserve anonymity, but it is someone we know, so play nice.
Thanks. We already hit 98% service level with classic safety stocks. Why swap a clear KPI for your probability maths?
Joannes Vermorel: First, you say you have 98%. Is it good or is it bad?
Is it profitable or non-profitable? I can give you situations where 98% is extremely unprofitable because way too high, and situations where it’s extremely unprofitable because way too low.
Let’s have a look at both.
98% in aviation: an AOG—aircraft on ground—let’s say it’s an A320, it costs €250,000 per day. There are 300,000 distinct parts in an aircraft.
If you have 98% service level, you’re going to have crazy cost in AOG. It’s way, way, way too low.
Now another case: fast fashion. You are a Zara-like company and you’re pushing new stuff every month or every two months. 98% is way too high.
Your clients, when they walk into the store, they don’t know what you’re going to present. It’s pointless to go for super high service level.
What matters is to have a very appealing assortment where people who come into the store will find something they like and purchase.
Your assortment is a construct of your mind. There is no strict requirement. You can dynamically extend the assortment or shrink it. It varies.
If you want your collections rotating fast, you need to make room in the store, which means you cannot afford 98%. It has to be lower, otherwise you clutter your store with old stuff that are not trendy anymore.
So again: service level is an extremely bad proxy of quality of service and also an extremely bad proxy of profitability.
When people say, “We are already at 98,” what I hear is: “We could be at 99 next year,” and very frequently that’s what we did for our aviation clients: we absolutely collapsed the service levels.
How do you actually get a very high quality of service? The answer is: for anything that is cheap—like a screw, tape, whatever—you want to have 99.999% service level, extremely high.
And then what about an APU, auxiliary power unit? That’s like an engine you put in the tail of the aircraft. Those are worth like six, seven million euros a piece. Do you want to have that in stock? Probably not. Maybe for this part you accept 70% service level.
Why? Because by not having an APU in inventory, you free €6–7 million that you can use to buy tons of cheaper parts.
That’s why this idea of saying “I have my service level target and my goal is to bring it to 99” is complete nonsense.
This assumes everything is uniform, that the forces between too little inventory and too much inventory are symmetrical. They are not. They are widely asymmetrical, and the asymmetry varies widely from one product to the next.
That’s why it’s nonsense.
Typically, clients say, “We have 98%,” and that’s where Lokad generates huge ROI, usually not because we have fancier technology, but because we are the first to say: we are going to look at this thing from an economic angle.
Then we realize there is huge money left on the table. Usually, with a Lokad initiative, half of the ROI is unlocked simply by adopting an economic perspective.
Because people have not done the economic calculation—rate of return—factoring stockout penalty, factoring a true proxy of quality of service in euros or dollars, not a made-up metric like service level, you realize you were nowhere near economically optimal. You had a paper optimality.
Conor Doherty: Okay. Well, thank you. I hope that was helpful.
I’ll push on to the named comments. This is from Lucio. I’m going to read this.
This was right at the start, when you talked about safety stocks assuming a normal distribution. Context: assuming a normal distribution isn’t mandatory when calculating safety stocks, and CSL—I presume it means cycle service levels—is just one of many possible approaches.
It’s true there’s always a trade-off between the risk of stockouts and the risk of capped revenues. What are your thoughts to that?
Joannes Vermorel: That’s what I said. In theory, you could substitute fat-tailed distributions to replace those normal distributions. Conceptually, you can.
Do software vendors do that? Mostly no.
When they do, people are kind of lost, because those fat-tailed distributions are brutal and confusing.
The only way to remove the confusion—Lokad experience—is to bring back the economic perspective, because then people see, in dollars or euros, what the hell is going on.
Fat-tailed distributions are weird. It can be a probability distribution that has no average. That is what I’m talking about with lead times. They have no average.
By the way, that means the average lead time—if you take into account things never delivered and you’re still waiting—when you have a data set and you want to compute the average lead time ever observed, it only goes up and up and up as time goes, because you have things that were not delivered 10 years ago that keep being not delivered, etc.
So you can have rules that say: we cap it at one year, and this and that. But then you’re not looking at the mathematical average anymore; you’re looking at something strange.
There are plenty of ways, in the paradigm of safety stock, to refine the model, just like academia has been doing.
You replace normal distribution by fat-tailed, pick one. You include a third uncertainty. You could do that: analytical model and so on.
But the problem I’m saying is: you’re operating from the wrong paradigm. It’s the wrong pursuit. You’re going faster, but you’re moving in the wrong direction.
Will you get to your destination if you’re moving in the wrong direction? No. If you move faster, still no. That’s the problem.
Conor Doherty: Okay. Thank you.
I’ll push on to Miguel’s comment: the goal is to reach zero inventory at all times. Lead time depends on how firm the demand is. Understanding that firmness helps find the best way to manage inventory levels. Thoughts?
Joannes Vermorel: Why do you want to achieve zero inventory?
That’s the problem. It’s not an economic perspective. People say we should be going for 98% service level; I say nonsense. People tell me we should have zero inventory; I say why?
Do you have concrete proof that it would be economically more profitable? Why do you do that?
We have seen a very simple business model used in many industries: you have producers, and you have intermediaries that pass big orders to producers.
Let’s say you are a manufacturing company and you can only manufacture things by 10,000 units, but your clients need one unit at a time.
You don’t want to deal with those small orders. You sell your batch to a wholesaler that will manage that for you.
The whole value of the wholesaler is to keep the stock and serve them because stock can only be acquired by 10,000 units and sold unit by unit.
There is one company that deals with production and they don’t want to deal with individual clients ordering one unit or small numbers. Then you have the wholesaler that deals with the split and the large number of clients.
So the idea that you want to have zero inventory—again, if everything else is equal, so you can serve the exact same clients in the exact same way, acquiring goods at the exact same cost, and you can do everything else the same way except your inventory level is lower, yes, most likely it’s the correct answer.
But in practice it is never everything else being equal.
If you lower your inventory levels, it means you are ordering smaller quantities that are very frequently more expensive. You order more frequently, transportation cost gets more expensive.
Batching is very frequently not an option. It is an element that is essential in your supply chain strategy. Whenever you have batching, you will have stocks.
That’s why you should not have a non-economic take like “98% service level,” or “less inventory is better.” That’s the opposite of what I’m saying.
You should consider the rate of return of your various options. If more inventory is more profitable, good for you: have more inventory.
There are even situations where that happens. Some products have a very high probability to be much more expensive next year, for specific reasons. Everybody agrees the odds are high.
But you have cash. What do you do? You buy a lot more and you stock that. The product is not perishable. You already have the storage facility. So beyond the stock, it doesn’t require much more investment.
In this case it’s reasonable to time the market: you order more, and next year you will be making a nice profit because everybody will raise their price, but you purchased earlier at a lower price.
That’s why reducing inventory is not necessarily a goal. It is only a goal if it can be done profitably.
Conor Doherty: All right.
We’ve been going for almost an hour, but there’s still a lot. I’m going to presume you want to keep going. I know your game. Important topic. That’s how Lokad rolls. Let the record show.
Let’s end those safety stocks.
One for today: we will close the book. I want to be very clear: this is not a plant, but I’m going to have to read something, and it ends in a compliment for you, so bear with me.
This is from Murthy. I know he’s been a regular attendee. Comment:
While most companies advocate the use of safety stock, very few actually measure how often it’s used. In several studies we conducted—we, not Lokad—across both finished goods and raw materials, we found that over 50% of SKUs, a conservative estimate, didn’t draw on any safety stock in the past 90 days. In my opinion, this highlights a major gap.
While safety stock is necessary—and I’m sure you’ll challenge that—it must be regularly reviewed and adjusted to reflect real demand and usage patterns. Joannes is absolutely right.
Did you like that, sir?
Joannes Vermorel: Thank you.
Let’s pause for a second. If you have a numerical recipe that needs to be manually updated even half of the time, this is absolutely not good. This is dismal.
Let’s go back to what is a realistic expectation. At Lokad, when we create a numerical recipe that generates inventory decisions, or production decisions, or pricing decisions, or store allocation decisions, what we want is 0% insanity.
Out of five million decisions, we don’t want a single line that is nonsense.
Yes, some of those lines will not turn out to be great because the future didn’t unfold as expected. This is fine. But as it stands today, based on what we know, even if we generate millions of decisions, we want essentially zero decisions that are nonsense.
If we have a numerical recipe that generates even 1% of garbage decisions, it is unusable. People will go nuts. It’s not even close to usable.
So if you have a numerical recipe—safety stock—that generates, they say, at least 50% insane decisions that people have to manually correct, that’s because the numbers initially don’t make sense.
Very frequently, people tweak half the numbers, but then they tweak enormously the other half so that the safety stock “kind of makes sense.” They toy with settings just to get safety stock where it should be.
An outtake is that very frequently, when you take all the manual tweaks into account, you’re closer to 90%. But let’s take this 50%.
We are again 50 times above what would be a reasonable quality for your numerical recipe. It is fundamental that your model does not give you insane decisions. If it does, you need to change the model.
Do not accept something that gives you more than a vanishingly small percentage of insane decisions. Your model should have 0% insanity. This is fundamental.
Otherwise you can’t automate anything, and you can’t even improve anything.
Why? Because if you have a process that is manual for a big part, if you have 1% insanity, in practice supply chain practitioners need to review all the lines because you don’t know where the crazy lines are.
If you have 1% insanity, people review close to 100% of the lines. It’s as if you didn’t have any automation.
And if they review all the lines and tweak many lines, how will you ever be convinced that model B is better than model A? You have so many manual tweaks in between that comparison becomes impractical.
If you compare two systems, you won’t even have the same practitioners doing A and B. If you have Bob who is super good against Roger who is slacking, Bob will have much better results. Is it because the model is different or because Bob is better than Roger?
That’s why you need to robotize. But if you want to robotize, you need 0% insanity. Here it’s 50% manual override. It means the thing is not working in practice.
Conor Doherty: All right, thank you.
I’ll push on. This question is from Manuel. Hey Manuel.
You’ve touched on this earlier, so I’m going to add context to broaden the question. The question was: can you please comment on the use of safety stocks in FMCG?
Let’s add in the downsides of the approach when it comes to perishability, write-offs, etc.
Joannes Vermorel: FMCG: you have suppliers and you want typically full trucks.
If we are talking of something massive—shampoos, cereals, whatnot—you are talking of mass-produced goods, and your suppliers will typically deliver in full trucks.
Each supplier, you will typically order more than one product. Maybe a dozen of things at least, any given day you pass an order. It might be one shipment per week, depends.
But fundamentally, assume it’s common that you want to fill a full truck.
Safety stock will give you quantities for those 10 products that you’re ordering from the supplier, but it doesn’t total into a full truck.
What do you do? You end up with a quantity that is half a truck, or a truck plus 5%. How do you deal with this situation? Safety stock does not tell you.
People say: “Oh, but that’s a starting point.” That means you will need an employee that tweaks every single number, one number at a time, until it matches constraints.
For me, this is nonsense. If you have a model that creates complications that make it mandatory to have someone review all numbers one by one, you need to change the model.
Same thing: safety stock cannot comprehend that you have the possibility to organize a promotion. You have the option, not mandatory, to organize a promotion to liquidate excess inventory.
Where is this option in your model? It’s absent.
If you have a supply chain model where prices—either what you buy or what you sell—are absent, this is bogus. This is again a non-economic perspective.
For FMCG, margins are typically thin. You cannot afford to not look into that. Your margins are too thin to leave so much money on the table.
Maybe if you are an incredibly profitable company—Louis Vuitton—you sell things with comfortable margins. If your supply chain is not super optimized, you still have fat luxury margins, so it’s not as much a problem.
But if you sell shampoo in the supermarket, this is a game where price matters. Every single cent you can get matters.
Conor Doherty: All right. I will push on.
This is from David Rollington. Hey David, friend of the channel. Everyone’s a friend of the channel—what am I saying?
But hey David. This is a long comment. You can take another sip if you want.
On this topic, safety stocks: there needs to be full discussion, communication and agreement on KPIs before anything else. When KPIs are handed down and all purchases come from outside the country, safety stock becomes necessary.
It’s similar to how S&OP process agreements work. Finance must be involved to ensure that all trade-offs are properly considered. Your thoughts?
Joannes Vermorel: Should finance be involved? Yes.
Should consequential economic trade-offs be factored in? Absolutely.
Is safety stock doing any of that? Absolutely not.
For me, that’s again the problem. It’s completely non-sequitur.
People tell me: “Finance should be involved, we should consider all economic factors,” and then: “thus we should adopt safety stock.” How do you transition from that?
There is nothing economic about safety stock. It assumes you go for an arbitrary percentage point service level. You take distributions, lead time, demand, and you don’t consider price breaks, perishability, probability of write-off, cost of storage, opportunity cost of storing something instead of something else, etc.
People give me tons of very good reasons, and then there is this jump: all economic factors must be factored, thus safety stock. How do you end up there? That baffles me.
That’s why I say to the audience: you need to look seriously at what economics means. Think in terms of allocation of resources, and what is the rate of return.
When you start thinking like that, you will realize: why do we have safety stock in the middle of our system? It adds nothing, just complicates for no good reason.
Conor Doherty: Okay. This is just a comment for me, but it adds context.
You’ve talked about leaving money on the table. I personally love that phrase because it visualizes what you’re talking about.
You can compare two things and say process A works. We can discuss to what degree. Process B maybe works better.
What does better look like? Less money left on the table, or more positively, more money in your pockets at the end of the month.
Joannes Vermorel: Exactly.
Two things can be true simultaneously. If you have very bad inventory practices, very bad production practices, supply chain wise, but your products are absolutely fantastic, you will still make money. But it’s not thanks to your smart inventory management system or production schedule.
There are many historical cases: companies having atrocious practices in this area, but their products were so good that they survive.
For example, Nike faced a massive supply chain disaster despite having one of the most loved brands worldwide. You can look it up online: it’s the Nike 2004 disaster with one of our ex-competitors, i2.
Nike survived. They’re still profitable. But if you look at their track record for supply chain, at least at the time it wasn’t great.
So you can leave tons of money on the table, and if for other reasons your company is extremely profitable, you survive. But it’s not thanks to your planning team or supply chain team.
Conor Doherty: All right. Thank you.
Joannes, this is the last question, and I’ve looked through it. A bit of disagreement—something you like. Let the record show: we welcome disagreement, or at least friendly banter. You can always reach out to us and challenge our positions.
This is from Con. Hello, Con.
When supplying retail, customers typically demand very high service levels combined with very short lead times, nothing like the five-month horizon in your building industry example. They also provide little to no visibility into future demand, and failure to meet the agreed service levels usually results in significant penalties.
Question: how do you maintain such high service levels without relying on traditional safety stock concepts?
Joannes Vermorel: Here you have a problem: your customer, typically in mass retail with FMCG, goes with an agreement where they literally set quality of service expressed as service level. That’s the setup.
Now the problem is that it still does not answer what you should do.
Let’s face it: you are an FMCG and you have those service-level-driven agreements from your clients. You say: because my client is going for this service level, I need to reflect that in my organization. That’s a strong argument.
But unfortunately, if you look again at the economic alternative, the economic alternative is better.
Why? Imagine you are in a situation where you don’t have enough stock to serve all your clients.
How do you decide which one do you serve, and in which order? Do you do first-in-first-out? Do you do VIP clients first? Do you want to spread the shortage across all your clients?
Is your client truly equally sensitive for all your products? There are plenty of shenanigans here.
For example, if you have a client where you’re not going to hit 98%, is it important to stay very close to 98% or not?
If a client company says: “I order 1,000 units, and if it’s not everything in full, it counts as a miss. If you give me 999, it counts as a miss.” In this case, no matter how close you are, it counts as a miss.
Then it’s better to not send anything. We have that with our clients: the retailer says if you don’t have it 100% in full, it counts as a miss. If you send zero, it’s the same. Then you don’t send anything and you use remaining inventory to serve other clients and meet their bar.
So we have arbitrage, and safety stock doesn’t tell you how good you are at arbitraging those clients.
Another example: your KPI, 98% service level, is computed monthly. You had a rough start of the month, and you are right now at 97%.
If you want to be at 98 by the end of the month when clients compute their KPI, you need for the remainder of the month to be at 99 to compensate. I’m simplifying, but that’s the idea.
Is it worth it? Is it profitable?
Should you bump your stock levels and go at 99% till the end of the month so that at the end you have your target? That’s a question.
You had your 98 but you didn’t deliver because things were beyond your model. The only way to answer “Should we try to go at 99 to compensate?” is to make an economic analysis on the rate of return.
If, in order to go from 98% service level to 99%, you need to multiply your inventory cost by three, you’re not going to do it. If you can do that by increasing inventory cost by 10%, maybe you should.
That’s the point: because it’s not an economic perspective, you lack all the nuance to know whether you should take the opportunity or give up.
Safety stock, by design, tells you nothing. You’re blind on that.
That’s why we go back to leaving money on the table.
Imagine: first week you are at 97.9, so you’re 0.1% below target, and then the other three weeks your system goes back to exactly 98, and you end up missing your target by a tiny percentage. It’s dumb.
You know that if you had pushed a tiny bit, you would have been above. It lacks nuance and any capacity to apprehend nonlinearity.
A nonlinearity would be: if you don’t hit exactly their target 98% for the whole month, they classify you as a bad supplier.
Safety stock does not factor any nonlinearity of any kind.
Joannes Vermorel: A follow-up comment to this, from Manuel: service level does not measure on-time delivery rate.
In addition, that’s another problem: how do you distinguish between if I serve my client on time or one day late? How much of a problem is it? The short answer is: it depends. It depends quite a lot.
It also does not differentiate between if you do not deliver in full.
Service level only tells you whether it’s a hit or a miss. It does not tell you by how much you miss.
Let’s say I’m looking at a bookshop and clients mostly come and buy one book at a time. On average you get five people a day that buy one unit of the book.
Then there is a professor that comes in and says, “I want 30 copies.” That will count in the safety stock analysis as one request not being served, but this one request is worth 30 units.
That’s the distinction between service level and fill rate.
Now, in retail and FMCG, you fail to serve 1,000 units from a client today, and the next day the same client comes back with another order, asking for 1,100.
Yesterday they asked for 1,000, you did not deliver. Today they come back with 1,100.
Is the total demand 2,100—yesterday plus today—or is today’s demand just the same that was unserved yesterday, being asked again?
Long story short: if you do not enter the realm of economic analysis, you cannot make sense of that.
You have an ambiguity on what demand actually is, and the only way to make sense of it, in practice, is to do guesses in euros or dollars about what is profitable.
The problems with safety stocks are literally endless. It is terrible.
Conor Doherty: All right.
One last comment that has come through. I don’t reveal private comments.
How do we retrain buyers who trust a single safety stock number? How do you get people off the mentality of: “I want one simple number. Don’t give me distributions. Give me one simple number.”
Joannes Vermorel: You robotize the decision-making process.
Why do you have buyers who go through a spreadsheet and do plus one, minus one on thousands of lines in order to produce a purchase order? This is insane.
That’s the problem.
People see everything as the increment compared to status quo. No. That’s not the correct way.
The correct attitude is: I need a model for this allocation process that gives me sane decisions out of the box.
Not through five layers of people tweaking numbers manually. I want a numerical recipe that generates purchase order decisions that are good, satisfying, with 0% insanity.
Once you have that—and that’s what Lokad does—how much retraining do you need? Very little.
Buyers are very happy. They look at the number and say: “All good. Perfect. Done.” They can spend their time talking to suppliers on how to make things more collaborative, instead of spending half their time revisiting spreadsheets.
If you approach the problem as: the buyer is going to revisit numbers line by line, and I want to give numbers that fit their traditional paradigms—the traditional paradigms are broken. Don’t play that game. It’s not winnable.
At Lokad, what we do is go straight to complete robotization of decisions that do not require manual input to be generated, and do not require manual input to be post-generation corrected.
They are 100% good, satisfying, with 0% insanity. Maybe the purchaser would say: “I would have done it a little bit different,” but it’s not worth my time, just ship it.
That is the right situation.
Then retraining becomes: train the team to reallocate their time to higher value-add instead of spending time endlessly tweaking numbers.
But for that, you need a model that is not broken by design.
Safety stock is broken by design, and thus safety stock will never let you escape this hellish situation where people are the human co-processors of your system.
Your system generates nonsense, and humans go line by line to fix the nonsense.
If you want to escape, you need an alternative. It will not be a variant of safety stock. It will be something much simpler, but radically different.
Conor Doherty: Agreed.
My thought when I heard the question—how do we retrain buyers who trust a single safety stock number and don’t want distributions—there’s an assumption there.
The world you’re advocating: instead of looking at an Excel spreadsheet and rounding a number up or down, you still, in the automated world, have a number in a dashboard.
The idea of distributions is how that number is calculated. That’s under the hood.
We’re not advocating that supply chain planners have to become advanced engineers and statisticians.
It’s one number in the current model, and it’s one number in the alternative model we’re proposing.
Joannes Vermorel: Exactly.
When you pass a purchase order for a product, you have to state the quantity.
So yes, the probability distributions are fundamental, but they are under the hood. This is the way the calculation is done.
Nowadays, with our clients at Lokad, they don’t need to be very competent with probabilities, just like if you drive a car you don’t have to be super competent in thermodynamics. It just works.
Conor Doherty: Well, on that note, we’ve been going for 90 minutes. We’re out of questions, we’re out of time.
As always, Joannes, thank you very much for joining me.
And to everyone else, thank you for attending. Thank you for your comments, your questions, your DMs. They’re lovely, as always.
If you want to continue the conversation, I say it every single week: reach out to Joannes and me. Connect with us on LinkedIn. We’re always happy to discuss.
If you want to challenge Joannes, we can set that up as well. We are open to feedback and to debate, or just friendly conversation.
You can also, if you want, send us an email at contact@lokad.com.
And on that note, we’ll see you next week when we talk about KPIs. But for now, get back to work.
