Summary
SMCP and Lokad described supply-chain management as economics, not ritual: scarce inventory must be allocated where it earns the highest return. Instead of backward-looking “weeks of coverage” and cluster rules, they use probabilistic, store-SKU decisions, explicit about uncertainty, substitution, and capacity. Automation frees people from exception-chasing to merchant work—setting priorities, measuring results, and iterating. Transfers and buffers are planned as opportunity-cost choices, not habits. Success is defined less by dashboards and more by behavior: teams drop spreadsheets, question assumptions, and seek continuous improvement. The claimed payoff 10x payback and faster cycles-reflects incentives aligned with long-term brand value rather than short-term numerology.
Full Transcript
The original speech in French was translated to English.
Maxime Rabillet: Hello everyone and welcome to this new case study and testimonial—an intervention that promises to be stimulating. This time we’re in the textile sector, if I’m not mistaken. It’s the only conference for this sector on today’s program. I’m Maxime Rabillet, a journalist at Supply Chain.
We tried to keep a firm hand—or at least a clear perspective—to ensure diversity in this program and to stay focused on real-world experience. As you can see, there are four speakers on stage, so I have no intention of monopolizing the conversation. I’ll step back quickly after giving you the information I have about the content.
The title is: “Precision at Scale: the transformation of SMCP’s supply chain with Lokad.” We’re set for 45 minutes and, ideally, we’ll reserve the last few minutes of this session for questions from you—or from me. I’ll hand the floor over now.
Swann Bareilhe: Perfect. Hello everyone, and thank you for the introduction. My name is Swann Bareilhe, I’m a Partner Supply Chain Scientist at Lokad, and I have the pleasure of moderating this round table with Carole Thomazeau, Yuting Chang, and Joannes Vermorel.
Just to briefly introduce our speakers: Carole Thomazeau is Director of Business Planning & Supply at Sandro. Yuting Chang is Global Group Transformation Leader at SMCP—SMCP includes Sandro, Maje, Claudie Pierlot, and Fursac. And Joannes Vermorel is the founder and CEO of Lokad.
Today’s discussion is about the supply chain transformation that we are jointly implementing at SMCP. We will notably cover how we evolved automated replenishment for the various points of sale, and how stock is rebalanced, more generally, across different omnichannel platforms.
We’ll focus a bit on the technical side, but not too much. We’ll spend a lot of time on the human dimension of the project and the lessons other companies might draw from the project we’re conducting together. In general, I’ll ask several questions to our guests, and we’ll take a few questions at the end.
Before diving into the heart of it, let’s set the stage and the context in which we operate. This one is for you, Carole and Yuting: could you give us an overview of your supply chain? The different sales channels, the scale of replenishment operations, orders of magnitude for references, the number of SKUs?
Yuting Chang: Hello everyone. Can you hear me well? Great. I’m at SMCP as Group Transformation Manager, as mentioned. For those who don’t know SMCP yet, we have five brands: Sandro Women, Sandro Men, Maje, Claudie Pierlot, and Fursac.
We have roughly 1,200 points of sale worldwide and four business units across the globe. To answer Swann’s question about the context that led us to seek an optimization solution: we have a current tool we use across the group that allows us to manage allocation all the way to the boutique.
However, it lacks the forward-looking piece—the anticipatory element via sales forecasting. As you may know, in fashion we have a lot of seasonality and many seasonal products. The challenge is therefore to bring in this sales-forecasting piece and to seek optimization so that our end-of-season stock—the residual—which is a very important KPI for the group, is controlled, and consequently our margin is controlled as well.
Carole Thomazeau: I’d add that Sandro was, and still is, the pilot for the group on this Lokad rollout. We chose Sandro for two reasons. First, because it has both Men and Women, which are two somewhat different businesses with sales that don’t behave exactly the same way. Second, because the teams had a certain level of maturity and were ready to be challenged.
Putting a framework in place also means questioning yourself and moving away from “coverage” notions toward probabilities of selling the best product as quickly as possible. So there are a lot of ingrained habits to revisit, and that’s why we chose Lokad.
Swann Bareilhe: Very good. We’ve already started touching on Lokad’s vision—the Quantitative Supply Chain. Turning to Joannes: in simple terms, what is the Quantitative Supply Chain and why is it suited to SMCP’s challenges?
Joannes Vermorel: Lokad’s approach is to view the supply chain as a resource allocation problem. At SMCP, every euro is in competition—across all the products you could invest in. Every item in a warehouse that could be sent to a point of sale is competing across all points of sale. Once you place an item somewhere, you can’t place it somewhere else.
Then, each point of sale has a limited capacity to showcase products. That’s Lokad’s perspective: we take an economic view of the supply chain. We think in terms of resource allocation, and we want to serve the brand’s economic interest over the long term.
I emphasize “long term” because this isn’t short-termist financial optimization with the bad reputation it can carry when you’re very myopic and end up doing unwise things. Here, we’re clearly talking about great brands with customers who have sometimes been loyal for decades. These houses are built over half a century, so you need to look very far ahead.
That’s the economic perspective. The other aspect: SMCP sells beautiful pieces, a bit more expensive than bottles of shampoo in a hypermarket, so the volumes aren’t the same as a hypermarket selling hundreds of units per day. We’re dealing with mechanically smaller volumes, very deep catalogs that renew frequently, and therefore a great deal of uncertainty.
To round this out: we have an economic view of resources, and we must handle very high uncertainty that’s structural to the business. If you have a very extensive catalog of luxury items, by construction you accept that you won’t sell ten units per article per point of sale per day. You’ll have much more intermittent sales.
Lokad’s aim is to bring these two aspects together into a numerical recipe that, even if we use somewhat complex things like probabilities, still reflects the common sense of an experienced store manager who would make judicious choices about assortment and store stock levels.
Yuting Chang: Building on what Joannes just said: since the Covid year, we’ve had a fundamental challenge—optimizing our buying strategy via the OTB (Open-to-Buy) budget, which is increasingly constrained. We need to buy better and, of course, allocate better.
Each piece, even if its probability of sale is fairly uncertain and fairly low for each point of sale, still requires a decision: to which boutique should we send this precious stock? Given fairly low sales during the full-price period, the challenge of optimizing sales at full price becomes even more important.
Swann Bareilhe: We’ve talked strategy and product/supply chain vision. Before getting into the project itself, were there other elements that led you to choose Lokad as a partner? Vision is one point; are there other job-centric aspects you want to highlight?
Carole Thomazeau: Yes. As Yuting said, shortage management lies at the heart of our replenishment problem: we buy short, so we have to allocate to the best place. In managing shortages, we realized there are lots of criteria to consider.
For example: is the store touristy, and will a stockout cost me more there than in a store with a clientele that will come back and wait a couple of days for the piece to return? Is my store “Ship-from-Store,” meaning its stock can also serve a second channel? Is the breadth of offer greater or smaller; if I don’t have the piece, will there be a switch to another piece?
These are many constraints that, manually, a replenisher can’t all take into account—the human brain isn’t wired to be multi-dimensional. So we end up reading and managing shortages based on a single criterion. Lokad allowed us to rank all these points according to priorities that we discussed together, and that really optimizes allocation at that level.
Another major point: transfers. Even if allocation is optimized, at some point transfers become obvious. This used to be very time-consuming; teams did it reluctantly. Now we get proposals every morning, and we decide whether to execute them or not depending on what we want to do.
A final point, under construction: we have stock states, KPIs that let us very quickly measure the entire situation, without spending days crunching data. The data is there, and we can analyze the problem. Often, it took two days to get the data, and by the time it arrived, the problem was already obsolete. Now we can build dashboards on the fly, see real ROI, and see what each decision brought in. We can simply measure each action.
Yuting Chang: To add an idea of the scope of SKUs handled by Carole’s allocation team: we have Sandro Men and Women. For these two lines, that’s about, per collection—we have two collections per year, Spring/Summer and Fall/Winter—roughly 3,500 SKUs to manage across 250 points of sale.
This volume of points of sale, and the weekly cycle of managing and evaluating: which products to prioritize, to allocate, etc. The team does very well with the current tool, but we never go deep enough, SKU by SKU, to make the decisions needed to optimize revenue. This volume is the key to success for better managing our residual at the end of the year.
Swann Bareilhe: Which brings us to the question: how did you handle replenishment before Lokad? Remind us what you did before—what was manual, the kinds of rules in place—so we can visualize the evolution the project brought.
Carole Thomazeau: Everything was manual. The accuracy of allocation depended on the seniority of the people doing it. And, as we know, there’s a lot of turnover in these roles, so there’s a learning curve each time, which is a bit long.
We used stock coverage—weeks of coverage. As Joannes said, when sales are erratic—maybe one per month, 0.8 or 1.2—that still rounds to 1. And 1 is 1, but we couldn’t see whether “the 1” at La Varenne Saint-Hilaire or “the 1” on the Champs-Élysées would bring in more revenue. We didn’t have that fine-grained decision.
Transfers were also manual and very time-consuming; we did them once in a while. Shortage management was based only on revenue, which is overly simplistic compared with all the criteria we could consider.
Yuting Chang: A key element of the tool we used before Lokad: we focused mainly on past sales. We’d say: “Based on the last few weeks’ sales, that’s the average number of pieces,” and then: “We want to cover two weeks of coverage,” because that was the timing. This mechanics, based on past sales, drove our decision, whereas in fashion seasonality is strong.
We have to anticipate—pre-send stock to meet upcoming seasonal needs. Sales forecasting is an important element: we’re looking for a deviation from what we’re doing today.
Swann Bareilhe: Joannes, you’re very familiar with all this. Could you summarize the kinds of pitfalls teams face when they rely heavily on manual rules and static forecasts, especially with fast assortments like at SMCP?
Joannes Vermorel: For me, there are a few very different classes of pitfalls. The most important: treating your best collaborators like exception co-processors. You have a system—an ERP—that can implement simplistic rules, but that’s not enough. You end up with “exception” alerts.
What then? You take your best collaborators, those with the most domain expertise, and you treat their expertise and time as disposable. They repeat the same gestures every day. Yet high domain expertise is a scarce resource in the company. That’s the first pitfall: this rare expertise, which should be cultivated, is used like a throwaway component. Each day you need X person-days, and that gets consumed. It just lets the company operate one more day; nothing is capitalized.
Our view, with supply chain scientists, is to robotize decision-making not to remove human expertise, but rather to give these people time to improve the numerical recipe, to refine the business nuance, and to get them out of exception management that is extraordinarily time-consuming and on which we don’t capitalize anything. That’s the most fundamental point.
Second class of pitfalls: time series models—very popular, but totally unsuited to supply chain. Why, especially in fashion and luxury? First, collections: a time series is supposed to have no beginning or end; collections have a beginning and an end, so that doesn’t work.
Then, you measure very precisely and end up rounding to the unit. As you said: a forecast of 0.8 or 1.2 becomes 1. But rounding to the unit, in your business, weighs heavily—many products in store have only a single unit. That rounding is considerable.
Another problem: from the customer’s point of view, you don’t walk into a store with a barcode in mind. There’s a halo of items that interest you; what matters is serving the customer well. A suit that’s a bit darker or lighter might be a good substitute. But if you don’t have the right size, you don’t have the right size: that’s much less substitutable.
So we have two big failure modes: squandering your most important expertise, and using simplistic time-series models—popular but unsuited—maybe fine for shampoo, not for luxury pieces. Lokad goes in very different directions relative to those problems.
Yuting Chang: Building on that second point for textile: sizing is a big issue. We sell a lot in sizes 36/38 for women, for example. But we also have clients who need 34 or 42.
If the average sales forecast is 0.3 and, as a result, we send one size-42 piece to every store, that means we need to buy 250 size-42 pieces to feed all stores. Then our DC has no remaining stock to replenish as soon as a sale occurs at one point of sale.
What matters is knowing when we need to send that 42 to the boutique, and holding stock centrally to feed the next point of sale with the highest probability of selling that size 42. And, based on the management rules and principles we set up with Lokad, asking: in that case, how many should I buy upfront—buying more intelligently relative to how we allocate that 42.
Swann Bareilhe: We’ve talked a lot about automation and capitalizing on human expertise. That brings us to the successes from the project we’ve been running jointly for a year. Do you have examples you’d like to highlight—either from the project phase (how we worked together) or now in operations (which is slightly different from the pure implementation phase)?
Carole Thomazeau: The project: we started in January, and by June we already had solutions integrated into our systems. It was quite fast and very collaborative, because at Lokad they aren’t “just scientists,” they’re supply chain scientists—and that word is super important.
They challenge our habits—“why do we do this”—with their best practices, etc. We have real discussions that keep the project alive; it’s not just parameter setting and “plugging in a machine.” We co-build it.
I remember a challenge I gave Swann a month before the sales: I told him, “I want transfers before the sales so I can deliver them.” He said, “Well, still…” I thought, “You can do it.” And I got my transfers for the sales. There’s strong commitment and responsiveness in the team.
What I want to say is that a key to success is team buy-in. I told them: “Stop doing. Watch, analyze, and think about how we can do better.” That could unsettle a team, but not at all. They’re super engaged. Every morning they have ideas and ask Swann’s team: “Could we improve this, that?” The fact they have ideas shows the tool works and they’ve made it their own.
They’ve upskilled: now they think about what they do. They don’t just do; they think about what they do and why they were doing it. We’ve had ROI on transfers. We block less stock for e-commerce—though I try to block a bit more—but we used to block too much and didn’t send to retail even though there was revenue in store. Old habits have been broken.
Yuting Chang: Operationally, I know the previous tool very well. Picture this: Monday morning, each allocation manager sits down and spends two days checking products. In the end, because we have so many SKUs and stores to check, after two days you say: “I got through the top-20 of our ranking and managed to cut the prod of our flop-20.”
Everything in the middle—we sort of know, but not with certainty. With the previous tool, it’s very time-consuming. Today, what we ask of the team is a mindset change.
On Monday morning, you open your computer, look at the KPIs we’ve put in place: check the out-of-stock rate, stock coverage, and tell us which actions are missing for the week. Instead of immediately diving into SKU-level allocation, typing coverage, handling exceptions store by store—“this product, that store”—and saying: “In a more routine localization, I need to ensure there’s size 42, because once the customer has come and gone, she won’t come back.” It’s a control-tower mindset for operations rather than always having your head down.
Carole Thomazeau: To sum up, they’ve become merchants. Before, they were technicians. They ask the right questions: how to animate the business every Monday and keep like-for-like growing. That’s huge. They’ve gained productivity, and now they have time to do it.
Swann Bareilhe: If you had to talk about Lokad to colleagues at SMCP who don’t work daily in operations?
Carole Thomazeau: First, all the brands want to adopt it. We’ve had to make a waiting list in our roadmap. How do I present it? For me, the probabilistic approach is really important: “the 1”—what is a “1” worth at La Varenne Saint-Hilaire versus a “1” on the Champs-Élysées? I hadn’t seen a tool that offered that otherwise.
Then it’s the quality of support, the quality of the people we talk to, and the fact it’s continuous improvement. We have a new idea—we’re not going to break everything—but a new parameter, a small tweak, because the context means what we did a month ago no longer works. It’s easy to put in place.
Once a week, we have meetings with the Lokad team—with Tristan and Cyril—and we say: “We’d like to push stock here, can we increase some settings?” There are lots of parameters given to the user—we don’t depend on them—but we do have continuous improvements on the tool.
Yuting Chang: We met quite a few vendors during the selection phase who wanted to optimize our residual rate at season end and thus improve our margin. Some proposed plug-and-play solutions: you give the data, thresholds are ready, etc.
What won us over with Lokad is the custom-built approach: we co-construct around the needs specific to our brands. When you say “custom,” you might worry about timing: if it’s not framed properly, you announce a six-month project and it stretches to a year or more. The key to success here is that the Lokad team relies on its supply chain expertise: they can frame the need with the business, steer it in the right direction, and in the end we stick to the timeline announced initially. For a custom solution, that’s a key success factor in my view.
Swann Bareilhe: Thank you. Joannes, how do you present Lokad when you speak to a new CEO?
Joannes Vermorel: A few years ago I used to point out that most of our prospects have half a dozen failures under their belt over the past twenty years. If half a dozen vendors failed, perhaps the problem isn’t “this vendor versus that one.” There’s a problem of method, fundamentals, approach.
For those interested, at our booth we have a book I published a week ago—“Introduction to supply chain”—which compiles why, in my view, classical supply chain theories malfunction, leading to operational failures, and what other viewpoints give you a chance to make things work.
Very concretely, at Lokad we try to be the operational partner of the supply chain. Through supply chain scientists, we want to take responsibility—this may surprise you—a personal responsibility for decision quality. When you write a numerical recipe, in a way, that’s you. It’s not “a system” generating decisions: you wrote a numerical recipe you understand, and if something goes wrong you must reverse-engineer what happened and understand it.
This assumption of responsibility is very important, and fairly atypical. We don’t see ourselves as a software vendor saying: “Here’s a list of features,” ticking 600 boxes in an RFP, and then letting you shoot yourself in the foot with those boxes. We assume the technicality is our burden, so that you can—as I appreciated in the comments—reason as merchants, without technicalities becoming an obstacle with technocratic jargon.
The role of supply chain scientists is to carry that responsibility so that people who are merchants, specialists in luxury pieces, can execute their strategy without getting bogged down in details. We talk about probabilistic algorithms but—and I’m speaking under your control—I don’t think your teams need to handle them in detail. It’s abstract for them.
Carole Thomazeau: Completely. We look at the final decision. When we ask, “Why was the piece sent to this store rather than another?,” it’s generally justified, tangible, and measurable. And yet we dissected everything. I gave Swann plenty of scenarios, telling him “I don’t agree—why is it doing this?” and in the end it’s logical.
Yuting Chang: What matters is the explainability provided by the supply chain data scientists. We work very closely with Swann and his team; they are the ones who code. If we’re asked, “How do you explain the decision proposed by the Lokad tool: why not send to the Champs-Élysées and instead send to Provence?,” we always have an explanation based on the figures shown in the tool. Carole’s team is convinced; we trust these decisions; and over time we don’t even need to check each point because we know there’s a reason behind it.
Swann Bareilhe: On the responsibility side, the support from the supply chain scientist is central. Personally, there’s even an emotional aspect with the clients I work with—SMCP in particular. We’re interested not only in delivering features, but in delivering decisions that work.
I regularly look at sales performance; I also go to Sandro or Fursac stores myself—even if my wife goes to Maje and Sandro. There’s a human side and a personal commitment we try to cultivate and bring through the deployment of a solution.
Yuting Chang: It’s a very interesting and solid partnership we’ve built between SMCP and Lokad. We’ve even planned desk spaces at our headquarters.
Swann Bareilhe: Time flies, so let’s start to wrap up. Joannes, we’re about a year into the rollout. What would you describe as a reasonable impact to expect from a Lokad project after, say, 6 to 12 months—a medium-term horizon?
Joannes Vermorel: We essentially aim for a payback of at least ×10 relative to what Lokad costs. That may sound high, but for enterprise software—where there are risks—that seems reasonable. If you commit to a somewhat complex technology and, on paper, you don’t see a potential ×10 payback, there are probably other priorities. That’s important.
A heuristic success criterion: when legacy teams manage to let go of their Excel sheets. As soon as we hit that—regardless of measurements—I know the initiative is on track for success. When people drop Excel, it means we’ve solved all those issues that were poisoning them—often a series of small problems, not necessarily big ones. Then we move to continuous improvement, and the trajectory is very good.
I’d point out that, for the vast majority of supply chain initiatives, there is no continuous improvement. Companies have a system, live with it for ten years, and then say, “We’re fed up, we throw it out and start over.” That’s a shame. The supply chain shouldn’t improve once a decade via a big bang. If we can progress a little every week, that’s much better. A few years later the balance clearly favors continuous improvement compared with those who stagnate for a decade before making another leap.
The idea then is to expand scope, always with a very strong payback. But the most measurable effect is also the quality of discussions and analyses that improves. We can pursue single-digit percentage gains in revenue—it takes a few years, but the orders of magnitude are significant. We’re not talking 0.01%; we’re literally adding a full point of margin in absolute terms—and several points in the very best cases.
For that, once in production, teams must be able to tackle very difficult questions. For example: what does “service quality” mean in a luxury store? It’s not just a service level metric. People who say “97% service level” are missing the point—that’s a tough topic.
How do we valorize the luxury vision: defending a very high-end price positioning, maintaining a very favorable customer perception with a decade-long view? That’s a difficult exercise. What’s very interesting is when, once you’ve cleared the operational issues, you can move into these discussions, where the operational teams challenge the supply chain scientists: how do we refine this long-term vision while having an automated mechanism to manage the day-to-day? That’s the paradox: looking very far ahead, even though every day you must output a whole set of micro-decisions.
Swann Bareilhe: Last question before we open it up to the audience. If a peer came to you and asked for one lesson from the journey, what would you say?
Carole Thomazeau: Everyone talks about AI, automation, and so on. The key is combining technology with operational intelligence. The teams doing replenishment have operational business intelligence. We’ve managed to combine their know-how—they didn’t have time to do everything—with a tool that helps them go faster and be more effective.
Yuting Chang: In the same spirit: we talk a lot about AI. But what exactly is AI? What can it really improve in operational efficiency and, ultimately, in margin? Often, AI is a black box: we don’t really know what happens inside. With Lokad, I come back to explainability: every decision made is explained and understood.
If the sales director asks us, “Why did you decide to send these pieces to that bigger, higher-revenue store instead of another?,” we can answer. This explainability builds confidence not only internally in the supply chain department—Carole’s team—but also externally: the commercial department, etc. That’s what I find interesting in Lokad’s approach.
Swann Bareilhe: Thank you again for your contributions, Carole, Yuting, and Joannes. Let’s move to questions. We won’t have much time, but I have one on explainability: is there also a need for explainability towards the stores? You mentioned the Champs-Élysées or elsewhere; at some point won’t “elsewhere” be tempted to say, “I never receive the pieces with the best potential”?
Carole Thomazeau: First, it’s not “never.” Smaller stores, because they have a narrower offer, may receive more than before. Previously, we automatically cut the smaller store and supplied the bigger ones—something the tool no longer does. And yes, there’s a real change-management topic with teams: explaining to digital teams why they have less reserved stock; explaining to retail that we’ll supply them on a one-week lead time instead of two.
But, as Yuting said, since everything is explainable and measurable, it’s not an issue.
Yuting Chang: I’d even say the opposite. With the previous tool, we managed by clusters. Cluster A means stores that generate more revenue, are larger, etc. Now we look at stock-management prioritization based on the sales of that particular store on that particular SKU. That actually favors smaller stores that do sell, but were previously drowned out by “bigger” stores selling three per week. In reality, the smaller store sells; we just didn’t have time to review that SKU and that point of sale previously.
Audience: Hello. You mentioned co-construction, and we can feel the enthusiasm and the pleasure you had doing this. You’re now in an ongoing collaboration—I understood you even plan potential office space for the Lokad team—and in both directions. How long does this last? At what point do you plan to have supply chain scientists in your organization, or is the model to rely on Lokad’s organization?
Yuting Chang: Thank you, that’s a very good question. To be transparent, we discussed this internally, and even with the Lokad team. What’s interesting is the transparent discussion with the partner: if one day, in SMCP’s internal roadmap, we plan for the supply chain data scientist capability, Lokad is open to us bringing those capabilities in-house.
The question is whether SMCP has that ambition. For now, we don’t have the answer. In any case, within the collaboration with Lokad, it’s an option; it’s not taboo.
Joannes Vermorel: I’ll add that all the code of the numerical recipe Lokad co-builds with its clients is the client’s property. The code is already in the client’s hands. Lokad isn’t a technology hidden in a packaged piece of software. What remains is a question of ambition: do they want to build that capability internally? We’re entirely in favor of training people if they’re interested.
Carole Thomazeau: I confirm the code is truly accessible—I went in myself to modify certain tables.
Audience: And at the same time, if we follow your reasoning, you probably add more value by staying focused on elaborating the criteria and staying customer-focused, with a team by your side. If you integrate data scientists, there’s turnover, you’ll have to retrain them, keep best practices… There’s surely great value in continuing to collaborate.
Beyond the teams, you explained that you take your best expertise, which spends more time adding intelligence to the criteria to be defined. Do these criteria evolve regularly? Do you change them? When you take a new range, do you have new criteria? Do you carry learnings from one range to another? How do you make that choice? It’s not just revenue that determines your choice in the end.
Carole Thomazeau: The criteria: before, “Ship-from-Store” can become a new criterion. Stores eligible for Ship-from-Store enter an allocation criterion we didn’t have two years ago when not everyone was Ship-from-Store. Tourist stores: stores don’t change all the time, but when we open a new store—will it be touristy or not—we’ll prioritize accordingly.
Rather than “fixed” criteria, it’s more about where we are in the season: at the beginning of the season, the cost of sending a piece to a store where it won’t be useful is lower than at the end of the season. So we make lead-time and risk-taking adjustments throughout the season and depending on the quality of stock we have in the warehouse.
Audience: Right. And that’s where you put the intelligence. The rest you’ve robotized, as you said.
Swann Bareilhe: Thank you. We’re really at the end now; we can continue the discussion afterward. I didn’t see who wanted to ask a question. Thank you for the presentation. I have a short question for Lokad, to better understand the quantitative approach to the supply chain. I understand there are mathematical models that take in data and compute an allocation proposal—possibly for SMCP’s case. So, is there a training time for these algorithms? Is it included in the project, or is it instantaneous?
Joannes Vermorel: The long answer: I invite you to read “Introduction to supply chain,” available on Amazon, which details what we do. Our algorithms are public. Generally speaking, there are training times, but my approach is to have things that can typically be trained in under 60 minutes, using potentially very distributed cloud resources, to stay very agile.
Is there training? Yes. Do we engage in approaches that take weeks for a model to converge? No. From our point of view, it’s very important to be able to retrain these models potentially several times a day. It’s not that the supply chain changes that fast; it’s that when the business raises an objection or a strategic reorientation, we don’t want to wait three days for the grinder to finish.
Someone says: “What if we thought about the topic this way?” You need to relaunch training, and we want the result to come quickly—at least to know: “If we do it differently, what does it look like?” So yes, our models—both learning and optimization—have computation times, but we try to always keep them under 60 minutes, regardless of company size, to remain very agile operationally.
Maxime Rabillet: Thank you to all four of you for this stimulating session. I have no doubt it will also spark post-stage conversations. The Lokad booth is right over there. Thank you, everyone.
