00:00:00 Decision definition, irreversibility, and forecasts contrasted
00:05:50 Opportunity cost missing from mainstream planning
00:11:40 Robust, fragile, and extremized decisions explained
00:17:30 Optimization iterates by finding insane decisions
00:23:20 Holimization, educated guesses, and retrofit lessons
00:29:10 Data semantics pitfalls and absurd recommendations
00:35:00 Rate of return at decision granularity
00:40:50 Window of responsibility frames decision horizons
00:46:40 Heuristic windows replace cumbersome policy models
00:52:30 Load measures capital committed over time
00:58:20 Half-life captures company inertia beyond lead times
01:04:10 Half-life reports decision duration without full simulation
01:10:00 Liquidity lens: consumables versus rotables
01:15:50 From KPI rituals to decision mechanization
01:21:40 Automation compresses latency and boosts productivity
01:27:30 Robotized decisions enable measurable continuous improvement
Summary
Joannes Vermorel and Conor Doherty continue their chapter-by-chapter discussion of Introduction to Supply Chain. Chapter eight turns to decisions: the operational choices that allocate resources, expose companies to risk, and ultimately determine whether supply chain theory creates economic value.
Extended Summary
This discussion is, at bottom, about one idea: in supply chain, only decisions are real. Forecasts, plans, KPIs, and dashboards may be useful, but they are not the thing itself. A forecast does not move inventory, commit cash, or disappoint a customer. A decision does. That is why Joannes insists that supply chain should be understood as a series of resource allocations under uncertainty, judged by expected risk-adjusted return.
This immediately separates his view from the mainstream. The conventional approach treats the future as if it were sufficiently knowable to justify a plan and then an orchestration of that plan. But once you assume one knowable future, you stop thinking seriously about alternatives, and once you stop thinking about alternatives, opportunity cost disappears from view. That is not merely a philosophical error. It becomes a software error. Companies end up operating with systems that can process vast quantities of data, yet remain blind to the very trade-offs that matter most.
Another central point is that “optimization” is often a misleading word. What most systems do is not discover some final economic truth. They merely push toward an extreme within a scoring system that may itself be flawed. Hence the real work is iterative: generate decisions, inspect the absurd ones, discover what the model failed to understand, and revise the economic logic. The issue is often not the solver, but the objective.
The conversation also makes clear that time matters economically, not just operationally. A decision is not merely a quantity; it is a commitment of capital over time. Hence the ideas of load and half-life. Load measures the monetary weight of a decision across time. Half-life measures how long it takes to recover half the capital tied up in that decision. These are attempts to describe something many supply chain theories obscure: inertia. A firm may think in terms of lead times, but lead times alone do not tell you how quickly the business can change direction.
Finally, the practical implication is organizational. Much of legacy supply chain management exists because human beings once had to divide the labor manually. But software changes that. If decisions can be generated daily, economically, and at scale, then many entrenched routines persist not because they are rational, but because they are inherited. The real contest is no longer between one manual process and another. It is between mechanized decision-making and organizations still mistaking paperwork for action.
Full Transcript
Conor Doherty: Welcome back. This is episode eight of a special series where Joannes and I take his new book, Introduction to Supply Chain, and we discuss the ideas chapter by chapter. Now, as you probably remember, for this series, I pretend to be one of the 10 million or so practitioners in the world who might see this book, pick it up, start reading, and naturally have some questions.
Now, this is chapter 8, which means we’ve already recorded the previous seven chapters. I highly recommend that you watch those first because some of those ideas will be referenced today. And with that, I give you today’s discussion on chapter 8, “Decisions.”
We are talking about chapter 8, which is unpacking, I think, the core concept of the book. I know it’s chapter 8 of 11, but based on your philosophy, Lokad’s philosophy, everything we do, decisions are actually just the most fundamental concept.
But in chapter 8, you give your fully formed definition, which is, and I quote you, “a flow commitment that allocates scarce resources among admissible options for closing alternatives in pursuit of the highest expected risk-adjusted rate of return.” Very precise. You started in chapter 1 with a slightly less verbose, less precise definition, but I’m curious: you’ve given a lot of examples already, and people can rewind and review them. What does your definition now of decision actually bring to the table in terms of supply chain operations?
Joannes Vermorel: First, it clarifies the applied economics perspective. It literally says we are narrowing our focus onto things where we are explicitly shooting for improving the rate of returns. So that’s a statement. There might be classes of decisions that don’t even belong to this paradigm. I’m saying it’s fine, they are just not supply chain decisions.
So here, we are really embracing fully this paradigm, and then I make it very clear that we are talking of something that is tangible on the flow. You see, what I want to eliminate with this “decision” is the confusion with potential artifacts. A forecast is not a decision. A forecast, you can project whatever you want. As long as it’s a projection, it has no bearings on the profitability of your company.
You can forecast growth 200% or degrowth 10%, whatever. It’s a projection. Ultimately, I would say those projections are inconsequential. They only become consequential when you actually do something tangible on your flow. And that’s why I say this is the decision. The decision is going to be an allocation of resource, and that will have an impact that is, to some extent, irreversible.
I’m saying “to some extent” because most supply chain decisions can be reversed, just for a price, at a price. So if you decide to move something from place A to place B, you can reverse this movement. It just costs money. So I would say I’m saying that the decision is irreversible in the sense that it is always reversible, but you have to pay for it.
Conor Doherty: Just to add a little bit of color to that, because again, obviously I do work here, but also having read the book almost a second time now, it’s irreversible in the sense that you can’t go back to the exact state of affairs that existed at the moment that you took the decision.
So, you can go back to the sort of superposition of, “Well, I can choose again,” but what you cannot do is rewind time like Superman reversing the rotation of the Earth and say, “I’m exactly back at 9:01 a.m. on Tuesday morning when I first had the market as it was, the Suez Canal as it was, the Strait of Hormuz as it was.” No, that doesn’t exist anymore. That’s gone.
So there’s irreversibility in the absolute sense and the relative. You’re talking about the relative sense.
Joannes Vermorel: Yeah, exactly. And as opposed to, again, a forecast, I can program on my computers an algorithm that would, let’s say, revise the forecast a million times per second. It will have no consequence on the supply chain, this thing. So I’m tweaking this number a million times a second, nothing is happening physically in the supply chain.
But if I decide to move stuff, I mean, obviously, if I try to reverse my inventory movement decision and I do it a million times per second, the thing has not even started moving yet. But let’s say, as soon as the stuff is there, there is an item on a truck and the truck is driving along the highway, then reversibility is lost.
It will take time to bring back the thing, and the supply chain will never go back to the exact position of the market, the exact position it was before this decision.
Conor Doherty: Well, again, this gets to the point of opportunity cost, which, again, if you break down the definition, “allocating scarce resources among admissible options,” essentially you can do A, you can do B, you can do C, you can add almost an infinite item, you have almost an endless number of options, and each one of those will have a different projected rate of return.
How common is the understanding of, or how present is the idea of opportunity cost in, let’s say, what you call the mainstream perspective on supply chain decision-making?
Joannes Vermorel: It’s completely absent. It’s completely absent, and it’s by design. You see, the mainstream perspective sees the future as something that can be captured within tolerances, just like this is a natural science perspective. I can project the position of Mars one year from now super, super precise. The only uncertainty would be the tolerance due to the fact that my telescope has some very, very minimal imperfection.
So I don’t know exactly the position of Mars, but within a tiny, tiny, tiny uncertainty relative to the size of Mars, I know with almost perfection the placement of Mars in the sky and in the solar system one year from now. So, you see, that’s a natural science perspective. And for the mainstream theory, that’s the exact same thing.
They embrace their forecast, they build a plan, and the plan is an orchestration of that. So the opportunities do not exist. The opportunity costs do not exist because what you have is one future and then an orchestration so that everything happens to make that future happen. So there is no question of alternatives.
Conor Doherty: Again, sorry, just to be clear, I might completely agree that in execution people don’t necessarily weigh up every conceivable permutation of options, but are you suggesting that it’s absent, like no one considers it at all? They’re not even aware that opportunity cost exists?
Joannes Vermorel: The problem is that, no, I mean, obviously in the mind of the manager, the practitioner, obviously people have this, what I described in the book as the “rugged vision” of the future. Obviously it is, because the rugged vision, and we discussed that in the chapter about the future, is this idea that you are keeping in your mind a rough understanding of the fact that there is this irreducible uncertainty of the future, that things are apt to change.
In fact, the teleological vision is super strange when you think about it. There is nothing that really works like that in human society, certainly not in a world which is not the world of natural sciences. It’s a world that is praxeological, which is a complicated word to say it’s human action. So the future is made of human actions, future actions that have not been decided yet.
So everything is apt to change because people can change their minds, your clients can change their mind, your suppliers can change their mind, your government might change its mind, et cetera. So fundamentally, when I said that the mainstream perspective is blind to opportunity cost, yes, the people in the company themselves are typically not, but the reality is that in modern supply chains, what you can do and cannot do is very much defined by the software that you operate.
Because, you see, your supply chain is entirely software-driven for any company past, let’s say, half a billion euro or dollar of revenue per year. I know no such companies that are not software-driven. Those companies are too large. Nobody knows every inventory position in their head.
We are talking of thousands of products, potentially millions of inventory movements per year. And that’s why supply chain is this strange thing where you do not observe directly a supply chain nowadays. You only observe it indirectly with the mediation of the software. And now, if you have a piece of software for that which very much embraces mainstream theory, even if in your head you’re thinking opportunities and opportunity cost, if this thing has no counterparts whatsoever anywhere in your applicative landscape, then it becomes such an uphill battle to do that.
Yes, you can hack a few spreadsheets right and left to still have that, but it is difficult and you are fighting an uphill battle because, again, if we look at the paradigm, what does it say? That’s going to be software implementation that says you forecast, you have time series, and then you have some allocation logic that just adopts this deterministic plan and just does an orchestration.
So, okay, if you say we have an opportunity or a risk, how do you inject that into those time series and orchestration? People would say, “Oh, you can tweak buffers.” Yes, that would work for a few opportunities, a few risks, but that’s really thin. Many opportunity costs won’t be reflected by your safety stock parameterization.
An example of that would be: what if there is, we need to extend the assortment? There is an opportunity cost in not widening the assortment because that’s what competitors are doing, and those clients are leaving our company because they do not find an assortment wide enough. So even if we are competitive on price, they find more value in rivals that can have the better assortment.
Those sorts of things, if you take the mainstream perspective, are completely absent from the theory and from the software, which makes it very difficult to hack any kind of spreadsheets because even if you can do that, you will have no point to inject those modelizations of your spreadsheet into your software. It will have to live side by side, and the communication between the two is very, very unclear, to say the least.
Conor Doherty: What I want to talk about more is, you don’t actually present it as a formal classification in the chapter, but it is something that you do sprinkle throughout all the chapters, at least up to this point: certain terminology like, you talk about optimality. You do talk in chapter 8 about optimality.
Right at the start, you talked about robust decisions, and you also talk about fragile decisions here. So these are not presented in, like, your classification of systems of record, reports, and intelligence. That’s very tight. That’s very clean. What is less clean, I would say, but again still very useful, is a very clear side-by-side explanation of, okay, what exactly delineates a robust decision that is generated by a system of intelligence from a fragile decision generated by a system of intelligence also, and an optimal decision, or “optimal,” we can debate the operations research nature of that.
So please expand on that, because we’ve used all three of those terms so far.
Joannes Vermorel: In fact, I realize that, and I posted something like a couple of months after publishing the book about a concept called “extremization,” and then it clarifies everything. It should have been in the book. One day, there’ll be another alert. For the second edition, I will rewrite chapter 8.
So fundamentally, what we are calling optimization right now should be called extremization. The term optimization is wrong. We should call it extremization. What does extremization mean? You take a mathematical function that is your criteria and you extremize. So you’re looking for the maximum or the minimum. And the variables that you can move to seek this maximum or minimum values, they are your decision variables.
That’s extremization. And when people say something is optimal, in fact what they’re saying is that something is extreme according to this function. Now, the point that I’m making with this optimization concept is that the challenge in supply chain is that the way you’re keeping the score, the way you are counting your dollars, is itself a moving target.
We see that in the chapter, for example, about economics. I say we have plenty of shadow prices, of valuation concerns. There are plenty of things that are very elusive, very important. And thus, when you’re counting the dollars, it may sound strange, but you have a lot of things that are very subjective.
Obviously, given enough time, like one decade or two, those things will come to pass and you will have the true accounting vision that will come into play. But for many, many important things, you need to look so far into the future that you cannot have this sort of accounting perspective.
An example: a growing forest, essentially. Or, for example, let’s say you are Boeing or Airbus. Your reputation is built on decades of keeping your aircraft safe, and that could be undone in a moment. When they bring a new aircraft to the market, it’s like a one- or two-decade effort of research, of safety engineering. So you need to look literally 20 years into the future.
Everything, any kind of business projections, when you’re looking so far into the future, just dissolve. The power of statistics becomes very weak. And that’s what I described in the chapter about economics: that’s why you need to make opinionated statements in your economic modelization to reflect the things that are extremely important to you because they are strategic. Although, if you don’t do that, you will destroy your company guaranteed within a decade or two.
So that means that extremization is weak because fundamentally you’re assuming that your objective function, what you’re trying to find an extreme of, is a known quantity. It is not a known quantity. And now you end up with a problem that if you change the function, if you change the way you count your dollars, you can have a decision that looks good under one score and not so good under a different score.
So suddenly, whether a decision looks good or bad is really depending on what you use to assess, to count those dollars. And I refer to this process as optimization. That is essentially a process where you repeatedly extremize, look for the flaws in your modelization, and then redesign your scoring function, your objective functions, and rinse and repeat.
So it’s an iterative process where you say, “Optimization,” you know what people call optimization, is just going to be a step. You do that, you look at the decisions that come out, you see the problems. But that requires a high level of intelligence in the sense that it’s not going to be mathematical problems. It’s going to be problems of perspective, of framing, of anticipating things where your quantitative model is blind, and then revising your objective functions based on that.
That’s really why I introduced this term, extremization, because that’s really what it is about. It is not about, “We have an objective function,” just like what academia does, and then we iterate over a solver to improve the solver endlessly so that we have near-perfect solution. That’s absolutely not the problem.
The problem is that usually, as soon as we have an objective function, a goal that we’re working towards, yes, we do the optimization, so the extremization, and then we realize, okay, we still have problems. We need to change the very definition of the problem because thanks to those extreme solutions, we realize that we have defects.
And very frequently, the defect is not in the extremizer. The defect is in the objective function that is being sought by the extremizer, or a solver in the literature. So if we go back to fragile and robust decisions, first, the thing is that if you have decisions that are generated as a result of this orchestration process of the mainstream theory, you end up with things that are, in practice, incredibly fragile. It’s by design.
If you assume that the future is known and you just do this orchestration and that will give you those decisions, then those decisions are completely blind to any kind of variation.
Conor Doherty: I think you give a good, just to jump in with clear context there, because you do talk about that very clearly in the chapter, and again it’s an example we’ve given before, but if you’re working in MRO, for example, and you think you know all the steps and all the things you’re going to have to repair in an engine, and then you write up, you forecast it, “Okay, I will need to fix parts A, B, C, G,” and then you open up the engine, it’s like, “Oh, actually, I need to do absolutely everything. There’s rust here, there’s corrosion there, everything.”
You treated it like it’s known, but it’s really variable, and you actually open up the engine and have a look. So again, you had decisions, but minor deviations suddenly invalidate the whole schedule.
Joannes Vermorel: Exactly. Because if you can repair almost entirely an aircraft engine, you don’t have zero engine. You have zero engine.
Conor Doherty: Zero engine.
Joannes Vermorel: Exactly. You cannot have an engine minus one joint or one small pump and it’s still an engine. It won’t fly.
Conor Doherty: Quite figuratively, it will not fly.
Joannes Vermorel: Yes, yes. So that’s the sort of thing where you have hard dependencies. And sometimes it can be small things. Just like after the lockdowns of 2020 and 2021, the automotive industry realized that they were blocked because they were lacking a few semiconductor components.
And literally, they had everything for the car. They had like 1.5 metric tons of car that were ready, and what was lacking was 20 grams of semiconductor. That is all that is lacking, but yes, the car cannot be sold to the client if those things are missing.
Conor Doherty: It’s not only that. It’s not only the fact that, in that specific example, the schedule will be disrupted. It’s also the fact that the costs of that could be asymmetrical. So, yes, I lack, I don’t know, I think a $2 screw. I could have stocked it for 10 years at a very agreeable price. I don’t have it on hand right now. I’m lacking that $2 screw, but the cost of this potential AOG event is $100,000, $200,000, $300,000.
So again, it’s not just the fragility of the decisions, but the cost of being wrong on that decision, that minor deviation, is asymmetric.
Joannes Vermorel: Exactly. It takes many shapes or forms. If we go into, for example, B2B retail, let’s say you have one of your VIP clients. They pass an order six months in advance saying, “Here is a list of 200 distinct items. We want to have them at this date because it’s very critical for us.”
And because they know that you won’t have that in stock, they pass the order months in advance. That’s great. They give you a lot of leeway to have everything completely ready six months from now. And now comes the date and you’re short on stuff. So you break the promise, and your VIP client is pissed off because they gave you a lot of leeway and you still did not manage to fulfill your promise.
That’s the sort of thing that would be a breach of trust. And that can be extremely damaging. So that’s the sort of thing where, again, if you steer your supply chain according to the mainstream theory, your instruments are completely blind to that. The people are not, but your software layer is, and that’s the problem.
Because, again, any sizable company is well beyond the sort of flows that can be managed just in the head of people. Most of the stuff only lives in the software layers that you have.
Conor Doherty: Well, this is the thing. It’s perfect that you said that, because it was the thought I had when rereading this. I said this right when I set up the fact that a robust decision, an optimal decision, and a fragile decision could all still come from intelligence, because you might have a system of intelligence that initially generated the schedule, again in the case of an MRO, that generated the “Oh, you need this part, this part, and this part to complete steps one through 25,” or whatever.
You open up the engine: “Oh, wait, no, that’s completely invalidated. Fat-tail event. I need parts I don’t have.” Okay. You just said that basically the better decisions live in the instrumentation layer. You might have a better instrumentation layer, the system of intelligence that you yourself advocate, and still make fragile decisions.
So yes, how does one go from fragile decisions without a system of intelligence, fragile decisions with systems of intelligence, to robust decisions with systems of intelligence? What’s the trajectory there?
Joannes Vermorel: That is the whole “optimization.” Fundamentally, the process is iterative. A lot of it is guesswork, educated guess. You do an educated guess on your objective functions, which in practice means you are putting your dollars and saying, “Okay, this must be worth this.” You’re describing expiration, and again, you’re sprinkling your dollars in your assessment.
Again, you try your best, and then what I say is that at some point you just stop and say, “Now I’m going to extremize.” You see, that’s where I say don’t think of a decision as optimal, think of it as extremized. So that means that you will have a machine, a software, a piece of software, that will just steer the decision towards extremizing this criterion.
Now you have, again, not an optimal, but an extremized decision. Is it good or bad? According to the math, it is extreme, but that’s not the true criterion. The question that you want is: is it really good and profitable for your business?
The thing is that, in practice, when you have your objective functions, you’re not going to extremize one decision. You’re going to extremize probably thousands of decisions. Again, because we are not making one grand decision. It’s the flow for the company. So there are thousands of products that need to be replenished, thousands of articles that can be produced, thousands of prices that can be revised up or down, et cetera.
Now, in practice, for this iteration, you look for insane decisions. That’s an anecdotal search. You look for the case where common sense dictates it is a very poor decision. And here, what guides you is not the math, because if the math was helping, it would have already been taken into account in the extremization process. Your solver would have already avoided this situation.
So most likely, what you’re looking for is not weaknesses in your solver, the thing that extremizes your decision, but the gaps in your understanding of the business. And that’s why, typically, if you want to identify those very poor decisions, you need to have people who have experience, who have seen a lot of situations, and who say, “Ah, this decision just doesn’t look right. Something is fishy here. Are you sure?”
For example, in the Lokad experience, the first time, that was more than a decade ago, 12 or 13 years ago now, we started to work for aviation. We started with suggesting to replenish some parts, and then the feedback that we got from our client at the time was, “But no, those are retrofits.” And our answer was, “What is a retrofit?”
We didn’t even know at the time what a freaking retrofit was. That’s okay. For the audience who doesn’t know, it’s when the aviation OEMs have even the slightest doubt that one of their parts might represent a danger for aircraft. You see, in aviation, security is paramount. So they don’t wait for accidents to happen.
When the engineering team just realizes that maybe, even just maybe, there is a weakness, what they will do is redesign the parts and then, for every single part that has ever been sold, they will push a replacement free of charge. They would say, “If you don’t send back all the old parts as proof, and the plane ever crashes and it’s because of this part, it’s on you, not on us.”
So obviously, the effect supply-chain-wise of a retrofit is you have a massive push of parts, and then all your aircraft in your fleet are now synchronized with fresh parts. So immediately you have a big drop in repairs being needed for those parts because they are all completely synchronized, brand new.
Conor Doherty: You’ve affected the supply.
Joannes Vermorel: Yes, you’ve completely mechanically impacted the demand because all your fleet now has parts that have exactly the same point in their life, same flight hours, same life cycle, same number of flight hours, flight cycles, et cetera, almost.
That’s the sort of thing where you start to realize that you need to extremize a decision to realize what you’re missing. And that can be a lot of problems. It can be, sometimes, misunderstanding the semantics of the data. An ERP can be like 5,000 tables, 100 fields per table. It’s very confused. The data is correct, but is your understanding of the data correct? You don’t know. Nobody knows. The ERP doesn’t know. People don’t know.
The people who wrote the ERP are dead or absent or far away. Again, it’s very messy. The data is correct because if it was incorrect, the company would not even operate. If you don’t know what to pay to your supplier or what to charge to your clients, you’re in deep trouble. So most companies who are still in existence actually have a fairly accurate modelization of their flow of physical goods that starts from “I purchase stuff, I transform and transport stuff, and then I sell it.”
So usually the systems of record are basically correct, not necessarily perfect, but basically correct.
Conor Doherty: Yeah, but your understanding, when you’re doing this system-of-intelligence layer, can be completely off.
Joannes Vermorel: Exactly. And this extremization is the only way to actually realize that your understanding is off, because it will become a lot more evident. You will have decisions that just look absurd to practitioners, and that’s what you need to fix, and that will change your optimization function, and you will just reiterate. And you stop when you don’t have any decision that looks insane anymore.
Conor Doherty: Well, we’ve covered how to think about decisions quite a bit, and we’ve revised some of what is needed to make the decisions, but what we need to cover now, I think, is evaluating decisions. And one of the things, again, we covered it a bit in, I think it was economics, chapter 4, but again, obviously it’s relevant because in chapter 8 you reintroduce it as an element of your definition: rate of return.
You apply that as the sort of north star for evaluating the efficacy of your decisions, fragile, robust, optimal, whatever you want to call them. Are those decisions working? Yes or no? How? Rate of return.
Again, people can go back to the podcast on chapter 4 if they want, but relevant to have a primer here. Again, rate of return: what is it, and why is it the best guiding star?
Joannes Vermorel: The rate of return is your guiding star because, ultimately, the game of supply chain is to convert your precious dollars into stuff, transport, transform, do things, and then turn those atoms back into dollars. And you just repeat the cycle, and that’s how you make money.
If you were just a pure trader in finance, you would just buy low and sell high and make money. But the game of supply chain is: we introduce intermediate steps, which is you want to essentially buy low, transform, transport, and sell higher. And that’s your margin. Repeat.
So rate of return is really about how can you have this magic loop happening with more leverage. You just generate more profits at the end of the day. And because, in fact, your flow is not one thing but thousands, potentially tens of thousands, millions of things that compete for your money, what you want is to inject the dollars, your precious dollars, into the areas where the loop gives you the maximum rate of return, because at the end of the year, if you count your dollars, those are the dollars you’ve generated by having something profitable.
You want to allocate the dollars where the returns will be maximized.
Conor Doherty: Isn’t that what people already do? Track profitability? They have entire divisions. Most companies over half a billion have a finance team. They tell you you’re making more money or making less money.
Joannes Vermorel: No, they don’t do it at the decision layer. That’s the problem of supply chain. Yes, the finance division will, at a super macro level, assess. They would say this division, look how much working capital they need, what is the profit that they generate, et cetera.
So the finance department will do that at the sort of business-unit level, very aggregated. The problem is that it gives you the capacity to steer investment, but it’s going to be extremely aggregated and extremely infrequent. That’s typically something that you’re going to revise once per quarter at the business-unit level.
But here that’s not the sort of granularity I’m talking about. I’m talking about every single unit, every single day, every single location, every single option that you have. The super minute aspect of the flow.
And what I’m saying is that those economic principles should also preside at that level. Again, if we go back to the mainstream supply-chain vision, it’s completely absent. The mainstream supply-chain vision just says we have one future and then we execute, and that’s it.
And the only thing it will tell you is that you have your plan, then you have the orchestration, you can project the cost, you can project the earnings, and you will know your profits. What I’m saying is that it’s not good. It’s not good because, as I was telling you, this paradigm is blind to opportunity cost. It is also completely blind to the steering of your investment toward the future that maximizes your future profits.
It’s just one static future, and that assumes that this future is already good, profit-wise.
Conor Doherty: Essentially, again, just to sketch that out just a little bit, to take an example, if you’re talking about allocating units, so it’s like, “I have 100 units, I have five stores, I can give 20 to each and that will have a projected or estimated rate of return. I can give 10 here, 10 here, 10 here, 10 here, and 60 there, that will have a…” and you can play with the numbers. I can send none. I can do nothing.
And you talk about that, like doing nothing, sitting on your hands, is a financial decision. You’re not allocating, but you’re choosing not to pursue. You’re choosing to retain capital. All of these things represent a different potential rate of return. They’re all at your disposal simultaneously.
Joannes Vermorel: Yes.
Conor Doherty: And this is why, again, coming back to the example I gave earlier about reversibility, and I said you can unwind a decision to a degree for a cost, obviously, but what you can’t do is go back to the exact position you were in last week, 9:01 a.m. Tuesday, and I had that level in this store, that level in this store, that level in this store, this potential price, these lead times, all those things.
All the combinatorial complexity of all of that only existed at that moment, at that time, and the rates of return for all of those possible combinations only existed at that moment in time. So I think that’s the granularity that you’re talking about, and you need sophisticated instrumentation for that.
Joannes Vermorel: Yes. Although, again, it’s an illusion to think that your software stack is not sophisticated already. Your operating system, your Windows or Linux, they are fantastically complicated pieces of software. What I mean is that there is this illusion that supply chains are operated with simple stuff. It’s not the case. It hasn’t been the case since the ’80s. We have a huge amount of software complexity.
What I’m saying is that, decision-wise, we have just a few basic percentages that are incredibly simplistic.
Conor Doherty: KPIs and artifacts and whatnot, you mean?
Joannes Vermorel: Yes. And when you look at the capacity of the software layer that exists to steer actual supply chains nowadays, if we put aside the clients of Lokad, usually it’s just incredibly, incredibly simplistic. Thus it is completely blind to all the nice opportunities where they could have much better decisions if the software that they were using to operate the supply chain was not built upon incredibly simplistic assumptions such as deterministic big time-series forecasts, these sorts of things, point time-series forecasts.
Many practitioners in supply chain don’t realize how incredibly rigid their planning layer actually is. And that’s why they end up revamping all the numbers in their spreadsheets all the time. It’s because, in fact, the analytical layer that they’re using is just incredibly simplistic. It just ignores everything. There is not a probabilistic view of the lead times. The lead-time uncertainty is ignored.
The probable variation of the prices of your suppliers is also ignored. You just assume that you will have constant price forever for what you’re buying. It’s crazy. You have zero modeling of any kind of geopolitical effect, such as, will the tariff jump again? This thing is just absent from the system.
Not very smart, because you don’t have to be a genius to see that there have been administrations in the world who have been very trigger-happy in touching the tariffs. Those are not fantastically sophisticated ideas. That’s where I push back on the idea that what Lokad is doing is very sophisticated. I believe that what we are doing is the thing that is least sophisticated that still passes the bar of not being idiotic with regard to the situation.
Conceptually, yes, you want a non-insane decision. You want something that is not making crazy assumptions about the future. And my main reproach about the mainstream supply-chain perspective is that it puts front and center this plan, which is very rigid and which represents a degree of confidence about the future that is completely unwarranted. It’s a house of cards.
You’re not even close to knowing the future as well as your systems pretend they do. And that creates all those fragile decisions because those decisions are anchored into something that is just an incredibly, incredibly naive take on the future. The consequence is that it’s just bad business because it’s fairly dumb.
Conor Doherty: Well, we’ve covered the what of decisions. We’ve covered the how to evaluate decisions. What we haven’t done is discuss the time dimension to this, because obviously, whether robust or fragile or judged by rate of return, whatever, if I order 100 units today, I need to have a time horizon to evaluate whether or not that was a good decision.
In the book, you talk about “window of responsibility.” It’s a term that I had not come across, certainly in a supply-chain perspective. So before we scrutinize it, what is the window of responsibility?
Joannes Vermorel: It’s a simplification that tries to essentially put a horizon to the analysis of your decision. If you go really first principles, you would say a decision that I take now will have consequences until the end of time, in absolute terms.
That’s not a super practical perspective. That’s very much the perspective of a mathematician.
Conor Doherty: And a philosopher.
Joannes Vermorel: Yes, a philosopher. So, okay, we need something that is a little bit more pragmatic.
Conor Doherty: Operational.
Joannes Vermorel: And what we say is that, very frequently, you can have a definition that says I take a decision now and I know that I can revisit this decision. Once I take this decision, I need to look at a certain period into the future that is of really, really prime relevance. That’s the essence of this window of responsibility.
For example, if I order now, anything that happens between now and the arrival of the goods, it’s just not really my responsibility. It will be lost. If there is a lot of demand that happens between now and the time when the goods are landing, I will not be able to serve those clients beyond what I currently have, which is the result of a previous window of responsibility.
So essentially, you would say, if I’m ordering now, it is not to serve my clients tomorrow. It is to serve the clients when the thing lands. So that means that my window of responsibility for this ordering does not start right now. It starts a little bit into the future.
And then, should it go indefinitely into the future? Well, not really, because you will have reorder points, you will have another opportunity to reorder. Ultimately, do you want to cover the demand indefinitely? The answer is no. You want to cover it up until there will be another opportunity in time to revisit the decision and just do that.
Effectively, what we’re doing with a window of responsibility is essentially a heuristic to bypass the need to have a more sophisticated mathematical object, which is a policy in math. So, with a window of responsibility, we can remove the need to introduce policies for entire classes of decisions in supply chain.
This is not perfect. This is a heuristic, but in practice it is, first, computationally much less expensive. Operationally, it is something that is much easier to white-box. There is a whole variety of reasons why going through this sort of heuristic is a good thing. That’s typically what Lokad does.
As a rule, we try to avoid operating a supply chain through full-blown policies because those things can become extremely difficult to debug, to white-box, to make the situation understandable by practitioners. And that’s very important because sometimes, when something disruptive is happening in the market, you need to tweak your numerical recipe in a very short time frame. And if you don’t have a numerical modelization that is really, by design, by construction, extremely white-boxed, that becomes a very, very risky undertaking, because suddenly the risk becomes your logic, your extremization. You get it wrong, you have bugs, and you do crap decisions just because you got the code wrong.
Conor Doherty: Well, so again, we were talking about policies there. You were talking about, in the mathematical Warren Powell sense.
Joannes Vermorel: Exactly. Friend of the channel.
Conor Doherty: We’ll have him back to debate that, but right, it’s essentially a mathematical object that will simulate your decision-making process over time, and you can simulate that indefinitely. Conceptually, that’s the cleanest way to do it. In practice, if you can, with heuristics, bypass having to reify an actual policy, it is operationally much more tractable.
Sorry, what was the point I wanted to make there? Oh yes. Sorry, I just wanted to plant a flag there just to clarify. But yes, so the issue there is with the concept of window of responsibility, which I do like as a terminology. Are you naming a thing that people already do intuitively, or are you naming a concept that nobody does? And if you’re saying that no one does it, are you maintaining that when people place orders and reorders, they don’t realize that it is to cover future demand?
Joannes Vermorel: No, the point is that they approach it, again, with very weak models. Like time series and safety stock. Those models are blind to so many things. So they’re doing it, let’s take a simple situation: I have a supplier overseas. I order 100 distinct items from this supplier. But there is a catch: to make it economically viable, it has to be full containers.
So I cannot, just because I have one item that is out of stock, pass an order. There is not enough volume on any of those 100 items. It’s only the joint volume that gives me a full container. Now the thing is, if I forget a few units, if, for example, one of the items I forget to put in the container, then I won’t have any opportunity to reorder a container before a certain amount of time.
Because I will be stuck. I have a stockout, and I cannot reorder this item in isolation. I am stuck. I will have to wait until the other items are consumed, and then I can make another container order in bulk, essentially. Yes, but it’s in bulk across items.
So if we start thinking about the situation, and this is really not super fancy, having an MOQ at the supplier level, so it’s not like you need to order items by a minimum of 100 units. It’s the supplier saying, “To pass an order, it should be at least $20,000,” or whatever. But it has to be a certain size. Those sorts of constraints are ubiquitous. Every company past a certain size has tons of situations like that.
If you look at those situations, what does the window of responsibility tell you? It tells you: beware. It is not the lead time for this item in isolation. You need to take into account the life cycle of the ordering process itself, which comes in containers, so that you will have to cover the demand until the next container lands.
So suddenly, you have to understand that what I am trying to cover is, I am responsible for when the container lands, and I will be responsible for the quality of service through this purchase order until the next container lands. And again, what happens in between, for example between now and my first container landing, let’s imagine a situation where there is no backlog, clients if they see that you don’t have the product, they just walk away.
It means that you should be very careful because the demand that you will observe for an item will bring your stock to zero, but then it will be lost. So you should not assume that the demand between now and when the container lands is something that is just the sum of all your buckets over time. It’s not true. It’s going to consume your stock and then drop to zero.
Then you will have a probability distribution for what will be the stock level that remains when the container actually lands. So again, the window of responsibility clarifies this sort of situation, while the classic safety-stock perspective is completely blind. You cannot, through safety stocks, have a proper modelization of the life cycle of your purchase orders that come through containers, for example. The framework cannot let you express those sorts of things. You’re stuck. You can turn the problem the way you want, tweak your service level up and down, whatever, this thing doesn’t fit the time-series perspective.
Conor Doherty: I very much enjoyed that example, actually. I was nodding along in quite a bit of approval there. That’s very, very clear. My follow-up to that is, while I understand, I think most people will follow that, it’s a very intuitive example. How does the half-life of a decision fit into that equation?
Again, responsibility, very clear. That was very, very clear. We are very much into Lokad territory. Sort of, it’s in chapter 8.
Joannes Vermorel: Yes, exactly. So we have seen window of responsibility. It’s a heuristic. And again, the point of this heuristic is to bypass the need to introduce policies so we can go back to assessing: do I order zero, one, two, three, four, five units, and I have an economic assessment, instead of having I assess, as a decision-making algorithm, my policy, and I see what is the rate of return.
Now, to understand the half-life, we have to start with the load, which is a simpler version. The load of a decision is working capital over time. One of the mistakes, and again the classic theory got it completely wrong because it’s a non-economic theory, is that the classic theory of supply chain does not understand the difference between “I need to allocate $1 million for a day” or “for a year.” It’s not the same.
So the question is: I have my pool of dollars. I told you the game being played is I convert those dollars into atoms and then atoms back into dollars. So how long will it take for the game to play? Obviously, a decision might block working capital, let’s say $1 million. If it’s just a day, that means that I allocate this million dollars and maybe tomorrow I have $1,010,000 and I’m back in the game, where I can reallocate those dollars to something else.
Or I can have another decision, which is I allocate $1 million and then it’s blocked, and in three months I have $1,050,000. Okay, but then it’s three months of my capital being blocked. So the load is dollar-seconds. It is money times time. In terms of units, it will be euro-seconds, dollar-seconds. It is the amount of money multiplied by a duration.
What I’m saying is that if you want to compare how much pressure a decision is putting on your cash, you need to look at the load, not the working-capital requirement. Because if you think, “Oh, I have a decision that is again $1 million, but it’s just one day,” that would be giving me $1 million-days.
Now I do another one where it’s just, let’s say, $100,000. So it looks much, much smaller in working capital, except that it’s for 100 days. That would be 10 million dollar-days in terms of load. So you see that the second decision, although the working capital looks smaller, $100,000 versus $1 million, if you think in terms of load, which is dollar-days, it is actually 10 times bigger, because you need to retain those dollars for a much longer period of time.
Conor Doherty: It is an expression of how long your capital is invested in a decision that has yet to pay out.
Joannes Vermorel: Exactly, yes. So what I’m saying is that the load is the critical thing that you need to look at if you start thinking at a very granular level for cash-flow optimization. You need to think of how long, how big a commitment is it really?
And here, the assumption is that we are dealing intellectually with many thousands, tens of thousands, of smaller decisions, smaller allocations, and they have various loads. What I’m saying is that if you want to understand how much pressure they add to your cash situation, you need to think in terms of loads.
And that’s where the load, in terms of time, will need you to think of all the possible futures to see how the thing that you’re deciding now, such as purchase orders, will be liquidated over time through sales or service. And the half-life is how long it takes for half of that investment to free itself, to come back to you.
Conor Doherty: Yes, exactly.
Joannes Vermorel: Now that we have seen those dollar-days as the load, we have exactly the half-life. Why is it very critical? Again, what sparked this analysis for me was I realized that there are many authors who completely misunderstand what is going on with lead time.
Lead time is something where, if you take a simplistic case like purchase orders, the intuitive understanding that people have is kind of correct. However, it is very not correct when you start thinking about the company as a whole. You want to think of something that would be close to company inertia versus variation of the flow. That’s what you want to think of.
So you want to think, if the market shifts dramatically, what will be the characteristic amount of time for me to just purge what I’m doing now? Think of it like you have a ship, you want to turn, and because it’s a very big ship, turning is itself a slow maneuver. It takes dozens of kilometers to do a turn. A U-turn would take you maybe 100 kilometers just because your ship has so much inertia.
Now the question is that you want to think of, if you want to do a very sizable turn, what is the characteristic time? Obviously lead times are an aspect of that. Obviously you would think, okay, if I have stuff that is inbound and my orders are taking three months, obviously my characteristic time will probably be something that is bigger than that, because those things are still inbound. So even if I decide to do something completely else, completely different, for the next three months I will keep receiving the stuff that I ordered.
But now, if you have an actual supply chain, you can have many steps, and this characteristic time becomes something very fuzzy, because you have a lot of additions, and also how do you mix the fact that you have stuff that requires short lead times, long lead times? Again, can you just add all the lead times together and say that’s my longest time frame and that’s the characteristic time of the company? Not really.
Because maybe you have, for example, something that you order and you need to order it one year in advance for some reason, but it’s super cheap. Actually, that’s very much the case in car manufacturers. You have a few things that you need to order like a year in advance, but you do it because, if you do that, you can have economies of scale and it’s super cheap. But it’s not really a commitment because those things are super cheap.
So, yes, you have a very long lead time for those, but they don’t really prevent you from picking a completely different direction because even if you were to just decide, “Screw those things, I just discard them completely,” yes, they will still arrive, I will still pay for them, but the amount of money is inconsequential. So it doesn’t matter. I can still make my metaphorical U-turn on my supply chain, and it’s not a big deal.
So you see, the lead times don’t tell you this characteristic time of your supply chain. It doesn’t tell you something that would be the true inertia of your supply chain. It gives you a very specific local time measurement. But it’s wrong. And then you end up with also a lot of non-mainstream theories that completely misunderstand lead times in ways that say, “Oh, we are going to divide your lead times by five.” That would be, for example, DDMRP, where they say, “If you introduce decoupling points, look, I have just divided the lead times by five.”
I say, wait a minute. If you introduce a decoupling point, that means that if you think in terms of inertia, you’re creating a big pile of stock. And because the assumption is a decoupling point never runs out of stock, you’re creating a big pile of inventory. There is no alternative.
It might make sense from an economic perspective. I’m not saying the opposite. I’m just saying that if you introduce a big buffer of inventory, you are creating inertia in your company. The characteristic time that it will take you to do a U-turn on the market and do something completely different has to be longer, not shorter.
Or it could be shorter, but you would take a massive financial penalty for that, like if you just liquidate everything away. Usually you won’t do that.
Conor Doherty: Yes, exactly.
Joannes Vermorel: If it’s really inconsequential, yes, you can do that, just cost of doing business, move on and do something else. But if it’s really substantial, no, you would need to purge that. Companies who are doing significant reorientation would typically be very careful in making sure that they liquidate all the stuff so that they can effectively do their strategic reorientation without incurring massive inventory write-offs for the stuff that they stop doing.
So that’s what the half-life tells you: how much time it will take to essentially recover half of the load, again the load, the capital-days. For example, that can be a little bit tricky because you can have situations where your inventory will keep running almost indefinitely.
Let’s say you order 100 units, and in 19 days you’ve sold all the units but one. And then it’s out of fashion and you have this one unit that is left. In fact, you never manage to sell this one unit left. The other ones were sold profitably, but you have an inventory write-off.
That’s why it’s interesting to think in terms of half-life. You don’t want to say, “Oh, my characteristic duration is time to really liquidate that,” because it might be like three years, because then I will need two years to finally decide that it’s a write-off and just purge it. But maybe after 30 days I had already liquidated two-thirds of my inventory. So it would be unreasonable to think that I have like a three-year inertia, while in fact most of my commitment was gone within just 30 days.
Conor Doherty: It’s a reporting mechanism, essentially. Gives you insight into the time duration of your decisions.
Joannes Vermorel: Exactly. Because, again, the load gives you an idea on how do I factor the cash, without resorting to some sort of policy mechanism where I’m going to simulate everything and know what would be the point in time where I have the lowest cash possible. It gives you an instrument that lets you prioritize the decisions that you’re taking now without doing a complete simulation of the future.
Again, that’s an instrument to essentially bypass the need for full Monte Carlo simulation of the future with policies. So the load gives you a very interesting instrument, and then the half-life answers another question, which is if suddenly you think that there is trouble brewing ahead, you don’t know exactly where, but you think that the market is very volatile, you’re thinking, okay, things are not looking…
Remember the movie The Big Short?
Conor Doherty: Yes.
Joannes Vermorel: People were knowing that there was something really wrong. The bubble was going to burst any moment. There is something that is going to go wrong. We are heading for trouble. We just don’t know when. We don’t know exactly. We can’t really make tons of plans because we don’t know exactly how things are going to unfold.
But what you can do is be a little bit defensive in the sense that I’m going to stop positioning myself on stuff that needs to go very far into the future. I want to shorten a little bit my position. And that’s what the half-life gives you. It’s a way to prioritize stuff that is not committing you too far into the future.
So, in a sense, to protect you against all possible variations in the future, even those that are truly, truly completely unexpected, like lockdowns, the half-life gives you a way to shorten your horizons. Or conversely, if you’re very confident and the business is good and you have a multi-year trajectory ahead of you that is very clear, then you can say, “Okay, I want to take advantage of the fact that it’s not so much a problem of me committing myself to stuff that will take longer to cycle out, if the rate of returns warrants those longer commitments.”
Conor Doherty: So, in summation on that point, the rate of return will be the way of evaluating, when you’re taking a decision, what the estimated profitability or return on that decision will be, whereas the half-life will tell you how long it will take for you as a company to liberate half of the capital that’s been tied up in that decision.
Joannes Vermorel: Yes, exactly. And all of that is dependent on how you model the future. So the half-life essentially tells you that, even if everything goes to plan with your fuzzy view of the future, it will still take you that much amount of time. So the half-life would be your minimum inertia, your baseline inertia.
Assume that if the market, if there is turbulence or something, this inertia will increase. So that would be a way to say, if you think that trouble is brewing ahead, then you really want to reduce those half-lives, because fundamentally they represent a baseline for your inertia, which is only unfortunately going to increase if there are problems, because it will take you more time to liquidate those positions.
Conor Doherty: Two different insights in terms of the risk of a given decision: one being how much you expect to earn, or how much profitability you can expect, for the risk of having made an investment according to your model, which is kind of fuzzy knowledge about the future, and the other is the risk of how much time, how much of the company’s time, it will take for the company to liberate half of the capital that was actually invested in that. There are two different forms of risk.
Joannes Vermorel: Exactly, exactly. And here, the second one, when you’re thinking of half-life, it’s really an opportunistic mindset. I want maybe to preserve my capacity to move and do something completely different. You don’t know when, but you just want to have something that is more liquid, in a sense.
Conor Doherty: Exactly like assets, essentially. You don’t want low liquidity. You want to have shares on a market that you can get out of at any given moment. You don’t want to have loads of money tied up.
Joannes Vermorel: Exactly. And the thing is that in supply chain, because it’s atoms, it’s a spectrum. It’s, “I want to get out.” Well, you can’t. You have those atoms that you need to liquidate. It will take time. But it’s not because it takes time that…
So you don’t have this binary thing, liquid/illiquid. You have a whole spectrum. And what I’m saying is that the half-life lets you characterize the spectrum on how illiquid your decisions are when it comes to getting back the cash so that I can reinvest into something potentially completely different, undo this decision.
And just to give you a very concrete example, it would be, for example, if you’re considering in aviation a consumable and a rotable, that’s going to be so completely different. A consumable is that whenever you have 10 units in stock, one is consumed. If I replenish plus one, next time I consume one, it undoes my decision. The stock is back to the previous position, so the characteristic amount of time, the half-life, will be up to the consumption to just undo the purchase.
Now, if I go for a rotable, that’s going to be much longer because if I purchase one, this thing is going to be serviced, and then fly, and then at some point be dismounted from the aircraft, repaired, and become serviceable again and again. Again, for the audience who is not familiar with aviation, the idea is that aircraft are repaired all the time. You have preservation of mass of the aircraft. So whenever you take a component out of the aircraft, you put another back in.
A component that you can put back in is called serviceable. The other one is unserviceable. It needs to be repaired. Usually it’s just an inspection, sometimes more, and then it goes back into the stock and then it will be serviced again. But because some parts are incredibly long-lived, like multiple decades, if you buy a rotable, a part that is repaired, this part will live in your fleet and your inventories for decades cycling.
Which means that if you get a part that is rotable that you don’t need, well, those parts will only exit your pool when they are scrapped. So scrapping is: you attempt to repair and it’s not possible, so you scrap it. But it might be three decades from now. It might be a long, long, long time.
As opposed to a consumable, where next time it’s consumed, boom, it exits, and we are back to the previous situation. The half-life of those two decisions is going to be dramatically different.
Conor Doherty: All right. Well, we have been going for quite a while now, but what we’ve said has actually set up the last question, which is not even my own. It’s actually one that I can read verbatim. It’s from a friend of the channel. I won’t say his name because parts of the question do reveal a little bit about the makeup of not only his company but, I think, a lot of companies that we’ve discussed today.
And I also think, even like the last couple of examples, we started talking about opportunity cost and now we’re talking about the half-life of a decision. I think if people sit and they listen for, I don’t know, I think it’s been 80 minutes, genuinely I think the pieces are nicely moved, things are built upon. Obviously, chapter 8 is my favorite chapter of the book. I’ve now read it at least three times.
So I think it’s quite intuitive when you’re listening here, and the examples I think were very helpful. However, if you’re coming in green and you’ve never heard of load and you just hear, “You need to start thinking about windows of responsibility and load and half-life of decisions,” that can be brand new, and it would also be a little bit troubling for people.
So I’m going to read the question verbatim, and you give me your response. So, quote: “Joannes, how do you shift an organization from KPI supply chain thinking to economically driven decision-making like you described, especially when service level, safety stock, APICS, and S&OP habits are deeply entrenched? Is the best path a POC that compares outcomes and lets the quality of decisions create the penny-drop moment, particularly in a reactive business where even basic MRP discipline is absent and executives still respond by simply changing policies rather than improving decisions?”
It’s a very detailed question, but I think a lot of that will resonate with you.
Joannes Vermorel: Yes. So the first thing is that the higher-ups need to understand a simple idea: the only thing that is real are those decisions. What I decide to buy, produce, transport, service, that is the reality. The plans, the artifacts, all of that is ideas. Again, that’s a simple idea, but once you understand that, we say, okay, what is the connection between what we’re doing and those things?
So here, the next idea is that until, I would say, we had the right software technologies, we needed to have a many-stage process because it was a way to manage the division of labor. So we needed to slice and dice a problem into a lot of people just to be able to deal with the flow.
Essentially, all the S&OP theory, the way it slices and dices the responsibilities, is just a perspective for the division of labor, and it will end up with inventory managers having a slice of 1,000 SKUs per person or something. No, it will vary. In some companies it’s going to be 100 SKUs, in some companies that’s going to be 5,000 SKUs. But the bottom line is that it’s fundamentally a mental model for a division of labor.
Software makes that completely obsolete. One piece of software can do all of that at scale. So I would say, if you have to convince the higher-ups, the first thing is to understand that what they’re looking at is something that is just a legacy of a division of labor that makes very little sense. When you start thinking about the fact that you have software and that you don’t need all this division of layers where you have a matrix with so many people involved, it doesn’t make sense. So software should be steering the decisions because that’s what is tangible. That’s the first thing.
And then the thing is that when you say that, yes, we can go for a proof of concept, but we need to understand right from the start that this proof of concept will not really be comparable with what people are doing. Because 99% of what S&OP organizations are doing are those numerical artifacts.
So when you have a proof of concept that goes for those decisions, you cannot compare that to what your S&OP is doing, because the software just says, “I don’t care about all your intermediate numbers, your projections, the way you slice and dice your budgets, et cetera. I’m directly going to tell you what you should produce, what you should buy, where you should put your inventory.” Straight. No intermediate steps.
Obviously the intermediate steps, they exist inside the software, but they are completely second-class citizens. It’s not the same focus. So my suggestion is that, and again, one hour is not enough to understand, go back to: don’t be distracted by this, I would say, obsolete division of labor. The only thing that is real is the allocation of resources that governs the flow. Those minute decisions, again: what do I buy? What do I produce? Where do I put the stock? Do I move my prices up or down? It’s literally half a dozen very simple basic decisions.
And then understand that the entire organization, all the slice and dice, all of that is obsolete. So if we, there is an opportunity to just do it simply with a pilot that generates the decisions. Now, once we accept that, the good news is that the pilot is going to be cheap compared to this enormous organization. It’s going to be very fast because, again, instead of having people meeting quarterly, struggling like hell to do it weekly or monthly, now you have to say your criteria to assess is that you should convince your top management that the opportunity cost is too great to not do it.
Because if it works, it’s a massive productivity improvement. And then, why should it work supply-chain-wise? The answer is just because it will be so much more reactive. Just think about it: the further away in the future, the more difficult it is to predict it accurately. If you have something that just runs and refreshes your decisions daily, this thing will have like a three-month advantage over any kind of S&OP process just because instead of trying to operate on something that is three months old by design because you have your quarterly cycles, this thing will be fresh from yesterday.
So even if your predictive technology is kind of the same, you would still have a massive advantage, structural, because you’re refreshing your decision every day. And refreshing decisions every day doesn’t mean that your stuff is completely unstable. You can take into account the fact, again that’s the economic modeling, that changing your mind has a cost, et cetera.
But again, the thing is that you have a massive structural gain that is just compressing the delays, the time it normally takes to reach a decision, to something that is super short, like one day, and then removing a massive piece in the organization that keeps busy typically dozens of people, and having it collapse into something that is a massive, massive productivity boost.
Yeah, that’s… again, if the higher-ups don’t want to hear about massive productivity gains and, I would say, low-latency decision-making, it’s going to be a very difficult uphill battle. But again, I think in these days of generative AI, it becomes very obvious that mundane repetitive processes must be mechanized.
And let’s be honest: mechanizing those replenishment decisions, master scheduling for production, dispatch inventory allocations for retail, and price, they are very mundane and repetitive. They are insanely repetitive. So they are super prime candidates for complete automation.
And again, to the boss, I would say, you should not pass, as a rule of thumb, an opportunity to massively mechanize an entire segment of your company because your competitors will be doing it, and you will not be able to sustain an advantage if you have a delta of productivity, 10 to 1, to your competitors.
Maybe a tiny few companies on Earth can say, “I don’t care about productivity.” That would be Rolex, Ferrari, LVMH. The luxury markets are not really representative of the whole. Hermès could say, “This bag, it took 100 days for one person to do,” and that’s a mass of craftsmanship to do it, and it’s fine. Okay, Hermès can do that.
But any other company? No. If your competitors can actually operate their white-collar workforce with a productivity that is 10x your productivity, they will outcompete you so thoroughly that there will be no question about the outcome.
Conor Doherty: Well, that’s the value of a POC because, again, there are multiple dimensions by which you can evaluate, and I like you use the term delta. The exact term that was in my head while you were talking. The idea of the delta between good decisions and bad decisions can be dismissed as long as it’s theoretical. If I tell you, “Oh,” again, if you’re just listening to this conversation, you go, “It sounds really, really nice, but maybe no, because I haven’t seen it.”
When you see, in just black and white, be it financials, be it in productivity reports, when you see in real time, “This is the delta that I was talking about,” it’s a lot harder to dismiss.
Joannes Vermorel: Yes, but again, you will not see very clearly this delta. Why? Because if you go for a pilot that generates unattended decisions, that’s typically what Lokad does, the problem is that the counterpart is not unattended decisions. It’s people who are very painfully manually doing it. So you don’t have any clear baseline. Because you’re comparing two processes.
The good thing about Lokad is that it is something completely unified. You can AB test a version of Lokad versus another version of Lokad. And you will have an assessment of this optimization process, and we can only do that because it’s robotized. So if I have one numerical recipe, I modify it, I have a second numerical recipe, I can AB test through backtesting, almost, I would say, in just one hour, or the time it takes for the compute to happen, and I will have an assessment.
Now, what about the baseline, the super manual process? You cannot benchmark anything. And people will come up with all sorts of objections. They would say, “Look, Lokad massively outperformed on this segment.” They would say, “Ah, it’s unfair. It was just an intern. We had a guy that was sick. It was not the normal person. It was an intern doing it, and it was done crappily. Yes, we admit, but it’s unfair to do this comparison because you cherry-picked our worst-performing team member because he was so green.”
Or there will be infinite objections such as, “Oh, but here, yes, you outperformed, but it was just because the forecasts were wrong. Next year we will have the forecast right,” et cetera. The problem is that if you compare something that is atomic, a piece of what you have when you have unattended decision-making with Lokad, this thing takes data fresh from yesterday, let’s say, and produces a decision for today, bam, and you can simulate that cleanly through backtesting through the past, versus something where it’s a workload spread over dozens of people in a many-stage process, it becomes extremely difficult to compare.
At the end of the day, you will see allocations and you would say, “Oh, but look, Lokad has much better allocations.” But then people would say, “Yes, but it’s only because of that, because of that, and thank you for pointing out, we’ll just fix it.”
My take is that, if you do not have something that is robotized, you can’t even start improving. You should assume that your manual process that is steering those allocations of resources, all your supply chain decisions, is just as good as it will ever be manually. We are talking about companies like half a billion and above. They have been operating for potentially decades. They had decades to improve their manual process, decades to improve their S&OP.
There is nothing really fundamentally that has changed software-wise over the last three decades for those systems of record. Even companies in the ’90s knew electronically what they were buying, what they were producing, what they were selling. So do not expect that those things that have been stagnant for decades will next year give you anything really superior. All the low-hanging fruit such as “we need to train people,” it’s already done. “We need to have standards of quality,” already done. “We need excellency this, excellency that,” already done. “Hire the best people we can,” already done.
Those boxes have been ticked, and usually decades ago. If you go for unattended decisions, then you have a numerical recipe and you can ruthlessly iterate. Usually companies have not ever even started doing that, and that’s why you have such a huge potential to improve against the manual process. It’s because having something that is unattended, that is robotized, gives you levers for improvement that just did not exist with the manual process.
Conor Doherty: All right. Well, I’m convinced. I don’t have any other questions. We’ve been going for, I think, about an hour and a half. I really enjoyed that one. Let’s not wait so long for the next one.
And thank you all for watching. As always, if you want to get in touch with Joannes and me, you can connect with us on LinkedIn or send us an email directly at contact@lokad.com. And with that, we’ll see you next time for chapter 9. And yeah, get back to work.
