Full transcript
Conor Doherty: This is Supply Chain Breakdown, and for the next 30 minutes we’ll be breaking down the generative AI value gap. Now, my name is Conor. I’m Communications Director here at Lokad, and to my right, as always, Lokad founder Joannes Vermorel.
Now, to be clear, when we say “generative AI value gap” we mean, and I quote, “the overall lack of clear data supporting positive return on investment (ROI) for many enterprise AI projects.” Before we get into that, comment below what supply chain problems you’re hoping GenAI will solve, and also get your questions in as soon as possible. Joannes and I will discuss those in about 20 minutes. Now let’s push on.
Joannes, the background to this conversation came about when I found some surveys from big consulting firms and from some public institutions basically talking about the questionable return on investment for these AI projects we’re talking about. Now, for anyone who missed that, a very quick primer. According to BCG, that’s Boston Consulting Group, approximately 75% of companies can’t scale or monetize their AI projects. According to McKinsey, about 80% see zero EBIT impact. According to PwC, at least half of companies have not even run AI risk checks. And according to KPMG, approximately one out of ten major execs at multi-billion-dollar companies are, quote, highly confident in ever seeing ROI for their projects.
I hear that and I raise you: Harvard, who said between 0.5% and 3.5% of work hours actually use generative AI, and that translates to approximately a 0.5% productivity boost. The National Bureau of Economic Research in 2025 said AI chatbots have had no significant impact on earnings. And lastly, according to Reuters, more than 50% of major organizations are not even tracking the ROI of these projects. Now, Joannes, I could go on, genuinely, for the full 30 minutes, but I come to you and I say: do any of those figures surprise you?
Joannes Vermorel: No. It’s the exact same patterns that the industry observes with all transformative technologies. You see, the web 25 years ago was exactly the same. If you go 45 years ago, that was the introduction of computing in companies—same thing. More recently, let’s say cloud computing—again, same.
So, the bottom line is that when you have transformative technologies, it’s very interesting. A lot of companies see the technology, they correctly identify that it’s massive, that it’s transformative, and they do something—and it fails miserably. That happens over and over and over.
It may sound as if 90% failure is absolutely terrible, but the reality is that the 10% that will succeed will change the industry forever. So you can have both at the same time: something where it’s 90% failure, and the last 10% changes the industry. If you go back to the web, the reality is that the quasi-totality of the investment made for web stuff in the 2000s went extremely poorly, and yet here we are 25 years later, and the web is everything. E-commerce is super massive; 80-plus percent of people find their date online, etc. The impact is absolutely enormous. Videos and movies are now sold through the web—Netflix. The impact is enormous. But the web portals that were in vogue in the 2000s? They’re gone. That was among many similar bad ideas of the time. And I think GenAI is on a course that is very similar. There are tons of gadget-y takes that will fail.
My take is not so much that we see a lack of returns because, again, early stage for transformative technologies—we should not be fooled by the lack of metrics. The question is more the lack of substance: are you doing something that is truly meaningful with it? At Lokad we have more than half a dozen projects, and for some segments where we are using those technologies, they have completely revolutionized our practice. We are not going back. There was a world before and a world after. It’s deeply different—for the better.
Conor Doherty: Okay, well, I don’t want to bury the lead at all. Some people, including me, raised the point that there’s a huge amount of spending. Again, Gartner—I think—said something like for 2025 there’s an estimated $644 billion, if I recall the data point correctly, estimated for this year. Everyone can’t be in the top 10%. So that’s a lot of money and a lot of organizational restructuring around this technology. Are there any parallels to be drawn between, let’s say, the tulip hysteria in Holland in the 17th century, the dot-com bubble, the GFC? Is that hysterical, or is there some substance?
Joannes Vermorel: The foundations are genuine. I mean, anybody who spends 30 minutes playing with ChatGPT—the shock is just incredible. Those LLMs are doing something that is mind-boggling. Same for image generation; same for all those classes of generative technology. The shock is real. There is really something here. We are not talking about pure speculation; the technology is real.
Now, we can discuss whether the valuation of the companies who are at the genesis of those technologies is justified. That’s another question that depends on whether the technology will remain limited to a small number of companies as providers, or be extensively commoditized. But the technical innovation is real.
However, the problem is that many companies approach those transformational technologies the wrong way. They identify correctly the potential, and then they just allocate budget to it. That was the same for the web: “The web is skyrocketing—okay, let’s have a $50 million web project. It’s going to be a web portal, and we’re going to consolidate the ideas of everybody and see what flies.” And guess what? It was wasted money at the time. GenAI is very frequently approached the same way, and it will be wasted money now.
The problem is not something you solve by pouring more money on the case. You need to look inward in your business and what needs to change to take advantage of those technologies—which are quite cheap, in fact. Again, that was the same for the web: the web is cheap; generative AI is cheap; cloud computing adoption is cheap. If you use it at massive scale, yes, it will be expensive. But when you want to get started, there is no reason to think that leveraging GenAI starts at millions of dollars. It starts at $20 per month with a Google Gemini subscription or a GPT subscription.
For me, the real case is that there is a lack of substance because, very frequently, people venturing like that don’t have a mechanical business empathy. They’re lacking empathy for their own business on how they are going to use it—empathy for the use case: “How am I going to…”—that would be empathy with the customer; and then mechanical sympathy to understand the potential and limitation of the technology so that what you try has a 90% chance to succeed. I mean, at Lokad—small numbers—but my series of attempts at generative AI for various stuff is over 90% success because I have, I think, a good idea of what things have a very decent chance of working, and I don’t even try the stuff where I think it’s not going to work.
Conor Doherty: Again, that is a key point because what you just said there—when you use it—again rough numbers, approximately 90% success rate. When I posted about this on LinkedIn, I got a lot of private feedback of people saying, “Yeah, for me, on an individual level, it is transformative. I went from performing at this level to this level. I can parallelize a lot of different tasks. My individual productivity has skyrocketed.” So then the question becomes: what is missing? On the individual level, people can be incredibly productive, but then when you have a company, which is a collection of individuals, suddenly, according to a wide array of data sources, that productivity just collapses inwardly.
Joannes Vermorel: Yes. Because, again, if you start a GenAI project—for the vast majority of companies—approaching the transformative technology through the lens of the technology itself is wrong. You need to think: what am I going to do for my customers that makes my company better, more efficient, serves my customers better? And then it turns out that things that used to be impossible have now become possible thanks to GenAI. Not everything; a lot of things that were impossible still remain impossible. But we have a slightly wider range of options, and that’s where there is massive opportunity.
You cannot just allocate money to GenAI and expect that things will come out. That’s exactly the attitude companies had with their web portals 25 years ago: “Oh, the web is great—let’s put dozens of millions on something web.” No. But if you have the idea, “I’m going to have an e-commerce store,” then very good, and you execute brilliantly. It’s very different from saying, “I invest $10 million on the web,” or “I build an e-commerce store and an e-commerce experience for my customer base.” That’s a very different proposition.
Conor Doherty: Then, again, this comes back to a key point. If we’re not—if, okay, I’m paraphrasing; you can correct me where I’m wrong—it sounded like you said you’re not surprised that there’s not a great ROI yet because it’s too early to tell. Okay, we can play with that hypothesis, but then you do still have to show some impact. So if you’re not measuring return on investment, what, in general, are signs that people are doing the right stuff or they’re having a good impact?
Joannes Vermorel: If you try to use those technologies and you don’t end up in a position where you say, “Okay, the world I live in today is now radically different from the one before,” then you missed the turn. Just to give a few examples: a few years ago, we decided to go for automated translation. Our website is automatically translated; nowadays it has been for years. We went from managing more than half a dozen professional translators to zero, and it’s completely automated. Now, whenever we publish a page in English, within hours we’ll have the page translated in more than half a dozen languages.
You see, the before and after are just different worlds. One case was we were managing a team—we even had an app to manage the workflow of those translators. Nowadays it’s purely automated. Another case: RFPs—Requests for Proposal. Some companies send us hundreds of questions, and we now answer them automatically. Again, that’s the sort of thing where we went from hundreds of questions to “let’s spend 20 minutes to have those same 400 questions answered,” with then a few hours spent on the most important ones to have the human touch and improve the key answers.
Bear in mind that in RFPs you have hundreds of very pedestrian questions like “Will you accept to have an NDA for our data?” and whatnot. Tons of pedestrian questions that need to be answered but don’t necessarily deserve the attention of a human in this process. So I’m talking about something where, with or without GenAI, the impact is not 1%. No—it’s day and night. The impact is absolutely massive. The new process is completely unlike the previous one, and it is perceptibly vastly better, even if you can’t really measure it precisely.
It would be very difficult for a company like Lokad to say what is the exact ROI of automating the super tedious problem of answering RFPs. You can measure in work hours how long it took before—yes—but that would vastly underestimate the case because the reality is you don’t necessarily have that many people in your team capable of answering that, and the very few people who are capable find it super tedious. They don’t want to do it; it’s punishment for them. Thanks to this new process, you can retain your best people in your sales team longer. It’s super difficult to assess, but the impact is massive.
Conor Doherty: Again, this is a good point because it transitions to the next topic, which is: you’ve outlined very well the marketing applications, the administrative applications. When we talk squarely in the domain of the supply chain and supply chain optimization, are you familiar with specific GenAI use cases that are in production? The reason I ask is an IDC study from earlier this year found that 88%—let’s say nine out of ten—GenAI pilots in this domain, actual supply chain decision-making processes, never make it into production. They refer to it as a proof-of-concept dead zone. So are you familiar with specific supply chain use cases that have actually graduated?
Joannes Vermorel: Yes. But supply chain is very quantitative—at least the way Lokad does it. It’s about resource allocation. You want to allocate your resources—inventory, production capacity, transport capacity, shelf capacity, etc. If I invest $1 of capacity in anything—inventory capacity, etc.—what is the option that gives me the highest return once you take into account all possible futures, to have a risk-adjusted optimized decision?
Now, this is a highly quantitative problem, and here LLMs are not a good fit to approach that directly. LLMs can approach that indirectly, potentially to help you generate the numerical recipe that governs your decisions—if your decisions are governed by a numerical recipe. At Lokad this is the case for clients, but for most companies this is not the case, so the LLM is powerless.
Then you have a few ancillary use cases—an add-on. For example, catalog data cleaning: you want to improve your product labels; you want to enrich your catalog with categories that were not existing. That’s the sort of thing where LLMs can really help you, but it’s not solving the core supply chain problem. It is just making your life much nicer when solving some of the sub-problems of supply chain. So it’s good; it’s useful—very useful—but if you start front and center saying, “I’m going to do a GenAI project,” it’s not going to solve your supply chain problems. As far as supply chain is concerned, GenAI is more like an optional technology that, for sub-problems, will make your life much easier.
Conor Doherty: But again, we add—“Oh, I’m just going to plug this in and everything’s gravy”—as opposed to, “Here are useful sub-applications.” They’re not as sexy as an entire transformation, but they help.
Joannes Vermorel: Yes. And when you have transformational technologies, you need to rethink your business as well. Think of the web: you have a website, but in itself it’s kind of useless. Those web portals were mostly useless. What is, for example, very useful is to have e-commerce. But then, in e-commerce, that means you need to have a distribution center that can do fulfillment. So it’s not just the web technology. You realize that if you want to make money on the web, then you need to reorganize your entire business so you can operate this e-commerce segment. It’s a much more demanding transformation compared to just “invest and get a website.”
That’s where I say those investments are typically misguided, because they invest in web portals—and GenAI is the same—on the naked technology instead of considering the transformation of the business that goes along. Buzzword-driven investment just never flies, and it has been the case for the last, probably, 50 years of software-driven innovation.
Conor Doherty: Are there any subtle signs that people can internally discern—like, okay, I can’t measure ROI yet, but here are some indications that maybe we want to pause the spending a bit, or conversely we should increase the spending?
Joannes Vermorel: The reality is: don’t think of it as an investment. It’s too early. Investment is a capital allocation problem once I know the terrain—once I know where I should be investing. Here again, it’s a technology that is very cheap. The question is more: do you have people who identified something that would make a lot of sense and that can be prototyped—even at the level of one employee? Do you have something that works, and then you say, “Okay, I have this thing; it really makes sense. I can see it makes sense; it resonates. I don’t even need fancy metrics—I know it’s good.”
Think of when I went from managing six translators to completely automated, end-to-end translation. I didn’t do a case study. It was very obvious. What’s more, we actually got better translations. You could think, “If it’s going to be a machine, it’s going to be worse.” It turned out it was higher quality. Why? Because Lokad has so much to translate that, at the price point we were negotiating, our translators could not spend infinite time on every page. It had to be done swiftly, and quality would sometimes suffer. GenAI solved that.
You can first do it on a limited scale: “I use a change to translate—does it work? Can I provide extra contextual instruction to make the translation better?” Yes—and it works. Once I have validated that, I can enter the investment phase where I robotize that—not for one page, that’s a demo, but for a thousand pages—and put some IT plumbing to make it convenient. When I engaged in the robotization of the initiative, I was already 100% convinced that it would work. I had already stress-tested the thing by manually doing a few pages.
It’s the same with the web. If you do a first experiment—does it really make sense? E-commerce: if I start to sell a few products online, does it make sense? Is there anybody willing to do that? If I cannot do a first experiment where it resonates with what I’m trying to deliver for my clients, it’s probably nonsense and you should stop the project before pouring more money on it.
Conor Doherty: You’re describing very, let’s say, text-based applications, and those are certainly transformative—we know that because we do that in-house. There are also other applications, even at the initial stages of forecasting, where you can deploy GenAI—and we do that as well, using embeddings as part of initial forecasting, etc. There are applications of LLMs in this.
Joannes Vermorel: Yes, but again, it’s an add-on. The substance is doing probabilistic forecasting, and if you use embeddings, then you can get slightly better probabilities in some circumstances. But this is an element—and here it’s very incremental. I believe that’s probably not what people are looking for when they think GenAI. They’re not thinking about something that is just going to make one subprocess slightly better. They want something super transformative and very visible immediately.
You do have such situations. For example, can you automate the relationship with your suppliers? They write emails; you write emails back. Can that be extensively automated? You don’t need an IT company and multiple millions to figure that out. You can do a first experiment: can I compose a prompt that would give a reasonable answer to a question raised by a supplier? Can I compose a prompt that would compose the email I want to send to this supplier? Does it work? Can I have the toy prototype working? Once I have that, then businesses can start thinking about automating big portions of what they do with this sort of technology.
Conor Doherty: This topic is so interesting because there are multiple strands to this. We’ve covered potential applications, but the actual process of applying that is a big problem. Let me add some context here: many—too many for me to go through; I’m just going to group them as “many studies” this year—say there is a lack of robust change management for these GenAI projects. Read between the lines: they’re just being forced into place. Do you think better change management would help raise the numbers in terms of ROI—training, skilling?
Joannes Vermorel: That’s exactly what I was talking about: capital allocation. The buzzword could be cloud computing, big data, blockchain, and now it’s GenAI. “We allocate this many millions on that.” This is just wrong. Then you wonder about the change management—no. It starts with “There is something I understand on how I can serve my clients better,” and it entails a transformation. This transformation is made possible at the endgame just because GenAI is there, but you’re thinking it completely differently.
You start by focusing on crafting the right problem. Ninety percent of the effort is crafting the right problem, not the technical execution. Back to the web: the web portal was thinking about creating a website and web technologies—that was the problem. It turns out it’s absolutely not. The web is easy as far as technologies go. Having a website up and running is the easy part. Having an e-commerce business that is profitable is the difficult part.
So, first experiment: “I want a very thriving e-commerce segment.” Maybe you realize that the web part of this investment is not really related to web technologies. You have many other problems you want to solve if you want a viable e-commerce business. Same for GenAI: if you think tech first, you invest in that, but for the transformation you’re looking for, 95% of the effort will have nothing to do with GenAI. GenAI will be the one component that made this whole transformation possible; before that it would have been impossible. It just makes something possible, but the rest of the transformation is your focus. That’s your starting point. GenAI is just plugged at the right place—critical, but fundamentally a technicality.
Conor Doherty: Well, the way you’re describing— I don’t know, because I know you well at this point. I’m not sure even you believe that—no, let me rephrase. When you say that what big companies need to do is rethink the problem, and then, once they’ve rethought the problem, they can work backwards and identify the nodes on the map where they can plug in AI—how likely is that, on the scale we’re discussing, when we’re talking about half a trillion dollars?
Joannes Vermorel: It’s going to be fairly rare. Based on history, how many companies 15 years ago were massively investing in data centers? Tons of them. Who had the guts to do what Jeff Bezos did—“We are investing so much in our data centers that we are going to open that to the world”? People would say, “Amazon?” Think how many banks had super-large data centers; they could have done this move ten years earlier. Amazon didn’t have the largest data centers in the world at the time. There were very large banking corporations with far larger data centers. But the guts was this: this emerging technology that we have come to call cloud computing. Bezos decided, “You know what? We are selling books, but we are also going to rent servers.” That’s the sort of transformation I’m talking about.
This is very challenging because it requires very deep transformation in your business. Statistically, if you observe the last century of business, very few large companies managed to do that truly. Obviously tons of brilliant management teams will manage to be the outliers that outperform the market and do the turn where their peers failed. But they are going to be the outliers.
Conor Doherty: Then we circle back to the original point about essentially a bubble. If a lot of people are investing a lot of money in a thing that they’re not equipped to deal with or execute properly, doesn’t that mean at some point a critical mass will be reached and—
Joannes Vermorel: A bubble is as if it was truly exceptional. What I see in GenAI is that we have very high valuations for a series of companies—that could be a bubble, yes. But for businesses spending money, my take is the baseline is probably 80% of the money spent on enterprise software is wasted. That’s the baseline. Maybe GenAI is 90%, but the baseline is that 80% is wasted.
So, for me, there isn’t a very specific bubble on GenAI. It’s more like the buzzword of the day. A few years ago it would have been wasted on blockchain; before that, wasted on some random big data initiative; before that, a random Web 2.0 initiative, etc. The bulk of the money spent by large companies on software projects is wasted—80% would be my baseline. Here it’s not far outside the norm; it’s just a little bit bigger. I would not see a bubble as significant on this front compared to the really mind-blowing valuations you have for GenAI companies broadly speaking.
Conor Doherty: There is another dimension to that consideration, though, which is: a lot of companies—let’s take one, not one of our clients—Shopify and them introducing GenAI skills as an actual requirement for hiring and for evaluation purposes because they’re going all-in on this technology. There’s another consideration, which is the actual employment restructuring effect of this technology, which in the near term might actually be for naught.
Joannes Vermorel: Again, that’s where I would challenge that a little. My take is that Shopify has a very specific business model that is not overly complex. If we talk about our clients—aviation, for example—it’s a thousand different trades, super arcane, extremely difficult. It takes months to even understand what is going on exactly in this segment of the business because it’s technically very challenging and complicated.
Conor Doherty: But they’re just an example of the trend.
Joannes Vermorel: Yes, and my point is that for something like Shopify—yes, it’s very good if the corporate culture is leaning into those emerging trends—but I think it will mostly fall on the shoulders of the top management to identify the key things where GenAI will really transform the experience for the customers and partners of Shopify. I do not see this business as so diffused that it’s going to be a bottom-up transformation.
I would say the same for Apple. When you have a super-massive business built around one hyper-successful product—the iPhone—it’s not really having 100,000 employees familiar with GPT that will transform your company. It’s more the very top who understand very clearly what it means for Apple, for the iPhone, and then roll out something that really makes sense—making the right choices. That’s where I’d say, yes, there is a culture shift, but the challenge for most companies will lie in the management more than at the bottom of the pyramid.
Conor Doherty: I’m going to push on a little, and ask you—because it’s something that had been messaged to me privately on LinkedIn. We’re on YouTube right now because of technical difficulties with LinkedIn, but it sets up this next question. Gartner—whatever you think—according to their Top 10 Strategic Technology Trends of 2025, that is the source, they put agentic AI—autonomous software agents that act as virtual workers—at the head of its list. So the question is: chatbots, AI agents—do you see these as the game-changer that Gartner—
Joannes Vermorel: What Lokad has been doing for a decade—yes, it is absolutely game-changing. But I would push aside the buzzwords. What we want is unattended decision-making. That is game-changing. Then what are the technologies to make that possible? I would say LLMs are a tiny, optional part to make that happen.
If by “agentic AI” you mean the outcome—unattended decision-making—so we decide what to buy, what to build, where to stock, the price we put on display; all of that changes daily, automatically, in a way that is unattended—yes, it is absolutely massive. So if Gartner by “agentic AI” means this outcome, then I agree. If by “agentic AI” they mean putting an LLM inside the loop, then I disagree.
Conor Doherty: I’m out of my questions. I will transition to the questions that were submitted. Some of these came in through LinkedIn even though we are actually on YouTube, so thanks for that—thank you for making the transition. So, Joannes—yeah, Joannes—I’m going to read this verbatim; it’s quite long: “What publicly available metrics can we use to gauge industry-wide progress in adopting generative AI and extracting financial value from it? Should we look at GenAI API spend, large-scale layoffs, or other signals?” I can repeat that.
Joannes Vermorel: No, it’s good. It’s a long question. My suggestion is: do not look at metrics. Metrics will be lagging, and when the metrics are visible, it will be too late. Again, think of Amazon with e-commerce. Amazon was nothing, nothing, nothing—it was a nothing-burger for companies like Walmart—until suddenly they were unstoppable and too large. The same thing happened for digital cameras versus the old chemical cameras: digital cameras were nothing for a long time, and then they were suddenly dominant. That’s the thing with most technological transformations.
The same happened in many industries—for example, fly-by-wire for aviation. It was nothing, and then Airbus did it, and then it was the norm, and anything not fly-by-wire was pretty much toast. The problem is that metrics will be lagging. Yes, you will see layoffs, but those layoffs will be done years after. Companies can automate and have massive productivity savings, but they won’t necessarily trigger layoffs immediately—they want to preserve morale, be nice, give people the opportunity to move somewhere else—and then there will be an economic downturn, possibly a decade after, and then there will be layoffs. So you can have effects that are extremely lagging.
For API spend—yes, but it will be very difficult because you will have AI specialists or companies like Lokad that can completely distort the market because they are spending a lot. When you see that a lot of people are spending money, is it your average company, or is 90% of the spending done by very specific companies—say, the video gaming industry? That will be difficult.
My take: do not pay too much attention to metrics. They are irrelevant in cases of technological transformation. You’re back in 2000; you’ve not experienced shopping online; try to project: “Is it going to change the life of my customers if they can buy online?” That’s an example of something transformative. Think: can I do something truly transformative for my customers with those technologies? If yes, then go for it. Do not wait. If you wait, you will be facing giants that have popped out of nowhere when you finally decide to go in this direction.
Conor Doherty: Perhaps this is something—based on what you’ve said—Lokad obviously has an ROI-first perspective when it comes to decisions. So to the CFOs watching—and we both know many in our networks—to the CFOs who are watching and who have got in touch privately to say, “Yeah, you know…” time to time about, okay: what is the return, in whatever way you want to express it? What is the return? What is the impact? What advice would you give to them specifically—the CFOs—when it comes to navigating these projects and their role in all of this?
Joannes Vermorel: For CFOs, the key question is: what is the value-add of your white-collar workforce? If we have to have a serious assessment, then you have to ask very tough questions and maybe play with ChatGPT to think: is this thing going to be automated, or is it really beyond what the technology can do? You don’t need expensive consultants, expensive IT companies to answer those questions. You can play with GPT and do small experiments to answer this question.
By doing that, you can answer the gist of the message: do we think those technologies are automating 10% of our workforce, 20%, or 90%? For Lokad, there are entire classes of tasks that have been automated to 100%—we go from half a dozen people involved to zero—for that task.
So I would say: start building a deep assessment of your white-collar workforce and what is exactly at stake, and then set a trajectory. Obviously that’s just to see the sort of return on investment, but it means very deep transformation. It’s not going to be CFO-led. Those transformations are so deep. Think back about Amazon: you’re an online book retailer and you want to become a cloud computing provider. It’s obviously at the very top that the decisions will be made—probably CEO level; the CEO will have to convince the board, considering the magnitude of the transformations we’re talking about.
Conor Doherty: Well, Joannes, we’re out of questions and we’re definitely out of time. To the people who did attend on YouTube, and to the people who watch this later, thank you for your attention. By the way, if you’re not already connected with Joannes and me on LinkedIn—why not? We’re lovely. Reach out; we’ll talk. But with that, Joannes, thank you very much for joining me and for your answers. And to everyone else, I say: get back to work.
