Didier Girard & Aude Defretière - IA Générative : passer du wow à la création de valeur

[00:00:05] Hello everyone, we're going to talk to you about generative AI.
[00:00:11] We're going to avoid certain topics that we could discuss for hours. We've included a small disclaimer on all the topics we won't be covering in this conference. My name is Didier Girard, I am the CEO of the Sphère group. And we support around twenty companies in the deployment of generative AI, from small structures of 200 people to very large deployments with over 200,000 users of our solutions. And so I am joined by Aude.
[00:00:37] So, Aude Frétière, I am a consultant on the advisory side of the Sphère group, namely We Envision. And so I have a product-focused role and, more recently, Gen AI.
[00:00:49] Before we start, I had two or three quick questions for you. Who has already used generative AI? I think everyone is raising their hand, we're good. Who uses it daily?
[00:01:03] Not bad. And who uses it daily in a professional setting? Wow! We have an expert audience. That's good. So, most of you have already adopted it.
[00:01:19] AI in the broad sense is something I'm quite happy, by the way, that this topic has come into the news in recent years. In fact, I defended a thesis in AI that I started almost 30 years ago, a little over 30 years. It's roughly the cycle we see between the time I did my thesis and the time when all this finally goes into production. And when I was doing my thesis, there were a number of things where I thought, this, we'll never be able to do with AI. I had a lot of hope in what we would be able to do with AI, and more specifically in machine learning since I was doing a thesis in the field of machine learning. There were topics I thought we would never crack. And at one point, we cracked a topic—well, researchers cracked a topic—and then I thought, wow, something is happening. And it was, so I don't know if this resonates with you, at one point, there was a Go player who was beaten by an AI. I think we talked about it a lot. Honestly, it was something I thought was impossible. And ultimately, when I saw what happened, I'm also a Go player, rather a very amateur one, In any case, I watched some games live and saw what was happening, and it was truly incredible to see those games. And then I thought to myself, we are changing eras. There was a second moment for me that was very important, which was when Google released its paper, where in fact, the RIA in 2017, so it's actually quite recent, it won a competition in the field of translation. And it was following this 2017 paper, so not that long ago after all, that this entire wave around generative AI started. So we're talking about something that was born not long ago,
[00:02:53] and that is moving very fast. Another moment that was very important for me was all the hype around OpenAI and what they were doing. They had started communicating several years ago, well, several years, maybe 4 years, something like that, about the fact that they had developed an AI and didn't want to release it because they considered it too dangerous, that it would be a game changer in certain professions, including journalism. From that moment on, they created a kind of appetite around this solution. I was lucky enough to be able to use this API before ChatGPT was released, about three years ago, it was in April 2022. And when I saw what it could do, we weren't yet at the level of LLMs we have today, but it was already extremely promising. And so at Sphère, we bootstrapped an activity around generative AI. And then, finally,
[00:03:48] December, late November, early December 2022, OpenAI released ChatGPT. And then it really went very, very fast. And when at Christmas, during the Christmas meal, my family was asking me questions about what is ChatGPT, how it works, etc. I thought to myself, 'Wow.' Ultimately, we're on a train that is much faster than anything we've seen so far. And I want to say that this train is only accelerating. Meaning, I was already on something I considered very fast. It's in the realm of the arrival of technologies. I saw the arrival of the web, the cloud, data, and the smartphone. I got a lot of companies on board with the arrival of these new technologies. And ultimately, generative AI is something I've never seen arrive this quickly. Also, one of the big differences is that, rather than tech topics and innovations often coming from the ground up—where we engineers were pushing these innovations—this time it's the opposite: decision-makers are actually pushing the adoption of these innovations within companies. So it's absolutely necessary to get on board.
[00:04:45] We could have done a talk on the technology. In the end, we decided that if we talk about tech, we'll present the current state of the art in generative AI. In fact, it quickly becomes obsolete because what wasn't possible yesterday becomes possible today. And ultimately, it would be a presentation with a rather short lifespan. We chose another option, which will unfold in two parts. I did an exercise I like to do called 'Remember the Future.' I told myself, it's 2030 and we've succeeded. We've transformed our information system thanks to AI.
[00:05:19] I did this exercise and I'll walk you through it. And then, to make that happen, we had to move the company forward and get the company and the people on board. So, Aude will handle this second part.
[00:05:30] Based on my experience from a short but still quite substantial year in terms of lessons learned.
[00:05:36] So, here we are in 2030.
[00:05:40] Here in 2030, if you had to remember two things, it is conversational. That is to say, this 'here,' we communicate with it. And we communicate through conversation. That's very important. We ask it questions, and it answers us. The second thing you need to remember is that it is actually operated by two types of personas: human personas, but also a new type of persona that is emerging—AI personas. So those are really the two things to remember, which I think will completely disrupt what is happening. And I'm going to unfold this. We are in a real evolution, and this evolution, I have tried to symbolize it, and it is still quite difficult to step back from everything that is happening. And I share a little bit of my feelings with you. So, it's true that when I started, I didn't experience the era of punch cards, but overall, when I began in computing, what we handled was very mechanical, very hardware-oriented. We had keyboards; I can still hear the sound of those keyboards when we typed on them. We were on green screens, and there was this aspect... We entered character by character. What we handled was characters; we input characters. There was an evolution. It was in the 1990s, with the arrival of these new operating systems, where in fact, the way of communicating with the system, with the information system, was much more visual. It's the arrival of the mouse and windows. And in fact, before that, a visually impaired person could be very efficient in handling the information system; with the arrival of Windows, I'm not saying that visually impaired people cannot be efficient, but ultimately, they are missing something. And ultimately, communicating with the information system became extremely visual and also very guided. It's the era of menus and forms that arrived in the mid-1990s. A very visual and very guided era. Now, when I discuss with these AIs, with this information system, for me, it's much more cognitive. So, I understand that, for now, the AIs I handle are mainly assistants. We'll see that I differentiate between assistants and agents. And in fact, when I interact with my assistant, there is a whole phase of reflection where I try to convey something to it. I will make requests to it, and in fact, in my way of interacting and communicating with the information system, via this AI, it is much more cognitive than before. In the era we are leaving, I was very guided in the way I discussed with the information system. I had menus and forms. In fact, I was quite limited in the things I could do. Often now, we will be faced with a form that has a single field, which is the chat field, and then we will ask it questions. It's a space that is very open from a communication perspective, but as a result, we are... is much more cognitive. We sometimes even ask ourselves the question, what am I actually allowed to ask this information system? So I think that, we'll see, the chat is not necessarily the final version of the tool we will use to communicate with the information system. I think there is still a lot to see from the UX perspective. But keep this in mind, we are entering a... In a moment that is much more cognitive for the tool, to communicate with the information system. In this cognitive era, the main topic is that we will communicate with the information system, we will open conversations, we will dialogue. I will give you two examples. There are several ways to dialogue. I do mean dialogue. We can dialogue with... When we are on our smartphone, it will mainly be oral. But when we are on our computer, it will likely be via the keyboard, even though we can probably communicate orally with our computer, but depending on where we are, we may or may not be able to do so. But overall, we will have conversations. I will give you two concrete examples. Concrete in the sense that these are things we have... One thing we have already done for a client, and one thing we are currently doing for a client. The first thing is,
[00:09:24] We have someone arriving at a water treatment plant. Until now, to understand this water treatment plant, to know its status, how it works, whether it is functioning well, where there are malfunctions, they would take out their device and start navigating through menus and forms to try to understand how their water station is doing. What we have done now is that we are able to convert to interact with their water stations. And when they arrive on site, to start diagnosing, they simply ask a question, 'How are you?' It will respond, 'I'm fine' or 'I'm not fine,' and if it's not fine, it will start to say, 'Here, I have problems here, here, and here.' And so, we are now able to do this—that is to say, it is no longer a technological issue. to fine-tune these conversations, we have the technologies that allow us to do it. Another example, we are looking at how we are going to do it. We have a farmer who is on his farm, he is doing the morning inspection of his farm, he is going around his livestock and sees a little bit among
[00:10:27] these different animals, which ones are in good health, or if there are any that are in poor health, or if he detects problems or changes in behavior, he must note it. Up until now, he has to note it with a pencil, on paper. What we are looking at is to what extent he cannot dialogue with the application that will allow him to have an oral report on his livestock. Just now, I met Marguerite. Marguerite, I don't find her in very good shape. On the other hand, Tammy, what I saw yesterday, she has regained a bit of her joy of living. So, these are concrete things again, where we are in the field and we converse through our smartphone. Obviously, we can also converse with the information system from our keyboard, with a chat or with other tools that are undoubtedly yet to be imagined. Personally, I find that the chat interface is very intuitive and very easy to use. I often say that the... The innovation in ChatGPT is not GPT, but Chat. That's where they were very strong, because the APIs had existed for some time, and where they were clever was in bringing this notion of chat. So, it made things very accessible. But as an expert user, we often find things missing. And there are people working on what the UX will be that will help us dialogue with the information system. It is undoubtedly not the chat as it is today, but we need to imagine other solutions. So, the first thing is, we converse. To converse, in fact, we need APIs. If we don't have APIs, we won't be able to converse. And so, we used to think of the IS mainly through its interface and its forms. We will increasingly need to think of the IS through its APIs. So, that is an old dream of all engineers.
[00:12:02] We always say, yes, APIs are absolutely necessary. But in the end, what do we actually produce? We create UIs. In this case, if we want the AI to be useful, it won't go through—or it could, but it won't be very effective.
[00:12:14] It will go through APIs to act as an intermediary in the conversation between the physical person and the information system. It will use translation. The physical person will speak, and the AI will transform this conversation into a stream that we can then transmit to the APIs. This is already concrete. We are already capable of doing this. When we talk about tools, I don’t know if you’re familiar—LLMs now have tools. In fact, these tools are already capable of transforming language into something that can be used to request... This is already concrete; again, it’s not a technological issue, it’s simply a change in mindset. And then we need to prepare the information system to receive these conversations, so increasingly thinking API-first before thinking forms, even though this is something we’ve wanted to do for many years.
[00:13:04] An example that is really emerging—I don’t know if you’ve already heard of vibe coding, maybe some of you, quite a few. It’s a term that has emerged very recently, February 2025. It’s really a very recent expression. What is vibe coding? The idea is that we no longer code, but we describe the code we want to have. We describe the intention, and it’s the AI that will generate the code. So even if some people are skeptical about this—myself, a year ago, I was; six months ago, a little less; three months ago, almost not at all; and now, I’m certain we’ll move toward this type of solution. Now, I’m not saying that all problems will be solved with Vibe Coding; I’m simply saying that Vibe Coding is coming and will allow people who are not developers to produce applications that will impact the information system. That’s the idea behind it, in my opinion. So it’s happening now, and again, we’re really in this natural aspect—meaning that in Vibe Coding, we’ll transform something in natural language, and somewhere, it will generate code that will interact with the information system. So this is really happening now. And the latest generation of tools, you may have little... Heard of Cursor, Windsurf, I don't know if that means anything to you. The most recent one I used less than a month ago called CloudCode, it's the so-called development environment from Anthropic, the company that develops Cloud. It's really quite astonishing what we're capable of doing with this tool, and when you use it as a developer, you realize we're truly changing the landscape. So, coding vibe, the cognitive air. That is to say, in the vibe coding, we will transform something into natural language, and somewhere it will generate code that will attack the information system. So this is really happening now, and the latest generation of tools—you may have heard of Cursor,
[00:14:20] Windsurf, I don’t know if that rings a bell. The most recent one I used less than a month ago is called CloudCode; it’s the so-called development environment from Anthropic, the company that develops Cloud. It’s really quite staggering what we’re able to do with this tool, and when you use it as a developer, you realize we’re truly changing the air. So, vibe coding, the cognitive air.
[00:14:48] The second very important point is the arrival of personas. Personas—everyone knows what they are—and until now, we often thought of personas as being associated with physical people, with humans.
[00:14:59] We’re going to have to think differently. We talk a lot about assistants and agents. In fact, for me, when we start to... Deploy AIs in production that will interact with the information system, we need to start thinking—and especially if they intervene in a business process— we must, before moving to the coding phase, go through the persona and clearly define what the AI we’re putting into production will do.
[00:15:27] For me, there’s one thing an AI will never be responsible for what it does. For me, the accountable aspect will be very important.
[00:15:39] The accountable aspect will remain with physical people, humans, but on the other hand, we will have AI personas going into production, though they will never be accountable. In some cases, it’s not a problem, and in others, it’s more troublesome. Another thing, two other points: we’ll need to work on collaboration between human personas and AI personas, or even between AI personas themselves. We’ll have this issue to resolve, which hasn’t been cracked yet; we’ll need to think about how all of this collaborates and communicates.
[00:16:09] Another point is that currently, in AI Personas, there are very few. Generally, companies have one assistant or one platform or a chat available in production that everyone in the company interacts with. We can imagine that in the very near future, actually quite quickly, we’ll have a number of AI personas going into production, to the point of having more AI personas in production than human personas in the company. That’s what I think, the direction we’re heading. So here, I’m projecting you into 2030. I think that by 2030, we will have many AI-driven entities in production, and undoubtedly more than we have employees in the company.
[00:16:46] I distinguish between two types of people going into it. That is also an important concept. I often say, assistants assist and agents act. I had a lot of trouble with...
[00:17:01] What is simple, and sometimes it takes a long time to express it, I had a lot of trouble with it—everyone was asking me what an agent is, what the difference is between an agent and an assistant. In fact, I had almost the etymology right in front of me and I didn’t see it, and I always went into technical considerations to explain the difference between an assistant and an agent. In the end, I arrived at this thing that is very simple and basic, but for me, it sums up the situation well. An assistant, it assists—what does that mean? It means that no matter what happens, it will not act on the information system.
[00:17:32] And mechanically, it is the human staff who use this assistant who will be accountable for what happens. We can ask questions of the assistant, it will respond, but afterward, it will be the human staff who will act. That is the assistant. And overall, everything we currently see in terms of AI are assistant-type personas. There is a new generation of personas that are emerging, being built. These are personas that can manipulate and change the state of the information system. And again, when an agent—so these are entities that act—manipulates the information system, who is accountable for this manipulation? When defining these agent-type personas, it will be necessary to clearly define that there will be an action taken and an owner of this modification. When defining its character sheet, it is important to remember that this agent will act, and behind it, there will still need to be someone who is accountable for what happens.
[00:18:36] To achieve all this, a certain number of things will be needed. First condition—there are many—I’m trying to summarize in a few conditions what will happen. First topic: obviously, data will be needed. Again, this is a somewhat unrealistic topic: companies need to become data companies. We will need to go much further than that, and everything we hear about data governance, thinking of data as a product, will be somewhat of a necessity. We will really need to work on this topic, and AIs in companies will only reach their full potential if we have data, but also if this data is governed. For example, knowing that a piece of data is an expiration date. A piece of data can be what we call lineage, meaning that a piece of data can be consolidated from other data. And if one of the primary data expires, the entire lineage chain derived from that data must also expire. So, this entire data lifecycle will need to be addressed. And if we don't do this correctly, we will end up with AIs that are ultimately quite poor because they won't be able to rely on the data we've collected.
[00:19:43] The second topic is that we will need to work on the company's digital twin. So, it's not enough to just have data; we will need a comprehensive way to represent the company through this data. This is what I call the digital twin. Now, the digital twin is not a recent term; we've been talking about digital twins for some time, particularly in industry. The idea here is to say that my information system actually represents my company. And everything that happens in the company is ultimately captured through data. But from the moment I have digitized my company and it is represented through data, AIs will be able to manipulate and play their role within this information system. A replica of the company, a digital replica that captures the dynamics of the information system.
[00:20:36] The third condition is that we need an AI platform. What is an AI platform? Simply put, it is the platform that will host all the assistants and agents we will put into production.
[00:20:50] This AI platform, I often say, will be the nervous system of the company. It hosts, executes, manages, and monitors all the AI characters we have put into production.
[00:21:00] And I think we will increasingly focus on, ultimately, when we score or determine the value of a company, we can imagine that by 2030, part of the company's value will be linked to the quality of its nervous system, thus to the number of AI agents and assistants in production, and ultimately to the tool that allows all of this to be executed. What I also say is that this nervous system will be very important in the future of companies, and what I advise my clients is not to outsource in the sense of relying on publishers or SaaS solutions for this nervous system. It is so important to the future of the company that the company must have confidence in the future of this nervous system and must control its evolution, rather than being dependent on the lifecycle of a publisher or the releases of the various solutions that may come along to run this AI platform. So, that is also an important topic.
[00:21:59] And then the last topic, which will gradually allow me to hand over to you, is that the company will need to transform.
[00:22:07] We started deploying generative AI within Sphère very early on. And at first, we saw this topic as a technological one. So, we made our AI platform accessible internally, but in the end, there was very little usage. And so, we had to work on change management so that people would adopt this technology. So, this is about getting the company to transform, and it's not just a technological issue, as I was saying. We are truly in a change of era, moving into the cognitive era, and I think you will say it better than me, We're talking about changing people's mental processes; you put it much better than I do.
[00:22:52] So, what I often say is the path and the goal. Well, now we have the snowflake, so I crossed out path and put flow and goal. But basically, for all this to happen, we need an information system that is adaptive and designed in an environment that is in perpetual evolution. There will be a very rapid wave of innovation in the field of AI and generative AI in particular. We need to be able to have an information system that can accommodate them. First topic. Second topic: we need to shift to a culture of experimentation. With what I mentioned earlier, vibe coding, It's almost certain that quite soon, many people will be able to create applications that will go into production. There are already players offering this, where you create specs and send code into production, and if you want to modify the code in production, you don't modify the code, you modify the spec. Natural language becomes the programming language. Again, I'm saying that not everyone will do this. I don't think the job of developer will disappear; it's just that we will increase the number of people who are able to bring code into production. A bit like the arrival of Excel in its time, which changed and impacted the way we imagined information systems, for the happiness of some and the misfortune of others, but regardless, we will have to embrace this. Another major topic is that this information system, which will change, which will transform, will need to be stabilized. I'll give you an example, and I think this will happen. We have an agent, for example, that modifies a piece of data in the information system, and there is another agent that says, 'No, I don't want this data to be like that,' and it modifies the data back the other way. And then finally, the first one says, 'No, but I just modified this and I'm modifying it again.' So if we're not careful, if we don't operate, if we don't audit, if we don't modify what's happening, we can end up with an information system that, little by little, becomes unstable. So we're entering an era that is constantly changing. For me, the issue is more technological. Technology is here and it's advancing very quickly, and even if every day there are innovations and things that weren't possible yesterday are now possible today and even more so tomorrow, what we need now is to manage the organization's transformation to accompany this change.
[00:25:13] 100% agree.
[00:25:15] I had just put up this slide. So we are the Sphère group, which is about 1,000 people. So Sphère is the engineering part. So we build information systems. And We Envision is the consulting part we started about 3 years ago, which is there to help imagine the future and guide our clients toward that future. And I’ll pass the floor to you.
[00:25:35] Alright.
[00:25:38] Well, my background is very different from Didier's, since I haven’t been working on generative AI for 30 years—I’ve only very recently started with the subject. Originally, I have a pure product background, and in quite traditional companies at that, legal publishing to name one. So, I can tell you that I had to take charge of product management, and I knew why I was doing it because in these old companies, there’s still a lot of sense in implementing some methodology around it. So they told me, 'You’re going to help traditional companies,' and I thought, 'Okay, I know that well,' and 'You’re going to install a generative AI platform.' I thought, 'Okay, it’s a product. Generative AI isn’t what I know best, but I’ll learn, and in any case, my product foundation is solid—I’ll manage.' Except I quickly realized that I wasn’t just a product manager for a generative AI platform and that something a little bigger was happening than what I had potentially managed in the past. So at first, I went in with high hopes—I set my product vision, I was going to break down my roadmap, my ROI would appear quickly, and then I’d be able to measure my AI. Adoption, something fairly classic, I think, that resonates with a number of people in the room. And in the end, I hit a bit of a wall. Meaning that, in the end, adoption wasn’t necessarily there. Measuring ROI on a generative AI platform—not simple. So I clearly saw that there was something a bit off.
[00:27:04] And very quickly, I told myself, okay, alright, the product is just a pretext, behind it there's something much bigger, it's generative AI, Didi just talked about it, and so my product, in a way, is anecdotal in the story. And so once I accepted that fact, I revised my approach a bit, and I told myself, okay, maybe take things in a different direction, rather than waiting for my product to be in the hands of users and just check adoption, maybe I should start with adoption of generative AI, try to encourage experimentation, and then give rise to use cases. And maybe at that point, we can talk about ROI.
[00:27:37] And when you do that, are you only addressing developers or not at all developers?
[00:27:41] Not at all. And that's actually my problem. Because when we asked the question at the beginning about who has tried generative AI, there were a lot of hands raised. But my daily context with the people I support on the platform is not that at all. I'll talk to you about that a bit later. But there you go. So, I wanted to share a number of lessons with you. It's worth what it's worth, it's drawn from my experience. But it's starting to have a bit of history. And I think it's still interesting to keep all this in mind. So the first message is: humans first, technology second. As Didi said. I think that with technology, even if there's still a lot of uncertainty, we roughly know where we're headed. However, the human aspect and its impact, and the impact of technology on humans, well, that's still a bit less clear for now. So the lessons I've learned from this so far. The first is that we're experiencing a major cognitive shift. Didier mentioned it earlier, but we've had a few big shifts like this in history, typically with the arrival of writing, when our brain was trained, formatted to memorize, we freed up significant bandwidth and space in our brain, since we no longer needed to memorize, we could read, and so we used it to do other things. We can think that the same thing will happen with generative AI and that in a way, we will free ourselves from a whole range of repetitive tasks for which our brain is currently programmed. And so perhaps we will develop new abilities. I'll put this here, I don't have a crystal ball to see, but in any case, there's a strong chance that this will impact our cognitive system in its entirety.
[00:29:12] Second topic, as I was telling you, when I ask—because, as a product manager, I do a lot of acculturation and awareness work around AI in traditional companies—so the average age is not a start-up environment, I'm not dealing with thirty-somethings, I have a bit of all age groups, let's say, in the room. When I ask who has used generative AI, I never get what I saw here today. If half the room raises their hand, and even then, I know some raise their hands because they think, if I say I haven't used it, I'll be stigmatized. So there you go. The reality is that when you dig a little deeper and start doing exercises, you realize that it's not widespread at all. There's a lot of media hype, but the people who have taken the step to use it aren't as widespread as you might think. And on top of that, I think many people feel like they've already missed the train, that the step is already too high for them, and so they hold themselves back in a way. So there's a lot of work to be done around this to lower the pressure a bit. It's not a problem if you haven't used generative AI. You might just be six months behind the others. There's plenty of time to jump on the bandwagon and catch up. There's no issue at all.
[00:29:33] Today, if half the room raises their hand, and even then, I know some raise their hands because they tell themselves, 'If I say I haven't done it, I'll be stigmatized.' So, the reality is that when you push a little further and start doing exercises, you realize that it's not massive at all; there's a lot of media hype, but few people have actually taken the step to go for it, It's not as widespread as that. And on top of that, I think many people already feel like they've missed the train, that the step is already too high for them, and as a result, it somehow prevents them from moving forward. So, there's a lot of work to be done around this to lower the pressure a bit. It's not a problem if you haven't used generative AI. You may just be six months behind the others. It's a great time to jump on the bandwagon and catch up on that delay. There's no issue at all.
[00:30:19] Second topic, and this is a bit of my favorite, I admit, ROI and Gen AI. I've given up on that, except we still have the systematic pressure from leaders who absolutely want us to put a number on the ROI they'll get from their generative AI platform. That's not the point, actually. We can't presume the ROI we'll get from something.
[00:30:41] The issue is that people haven't taken hold of it. It's very difficult to project an ROI at this stage.
[00:30:47] There's real work to be done in terms of acculturation and sponsorship with leaders to calm this ROI obsession a little. ROI is great; it has certainly changed our lives on many topics, but with Gen AI, please, we need to stop—it's not at all the element that should be measured.
[00:31:03] If you'll allow me to intervene, I had the case as a leader at Sphère, as soon as we decided to deploy the platform, we naturally brought up the cost, and quite quickly, my co-leaders asked me, 'But Didier, ultimately, what's the ROI? How much is it going to cost us?' I didn't know what it would bring to the company. And in fact, I reversed the topic—not in the same way you did—I simply said, 'I don't care about the cost.' So to speak. What does that mean? It means that what you pay when you develop your own platform is the number of tokens. Now, I don't know if the number of tokens means anything to people in general. I will summarize it: it's the number of words that the generative AI will generate. And what I'm saying is that a generated word is a word that is not written. And in fact, in a way, the more words my platform generates, the less, in a way, all the people in the company write words. And, so to speak, the happier I am because I've won.
[00:32:03] There may be debates, but in terms of speed, quality, and ease of conversation with the information system. So, my focus, actually, what interests me in the end is the more consumption I have, the happier I am. Where everyone is afraid of how much it will ultimately cost, I say on the contrary, the more I spend, in a way,
[00:32:22] in tokens, and, so to speak, the more satisfied I am because it means people are really using it. Of course, one could say to me, yes, but they're using it to generate images and things that are useless. It's a bit like the fear of the arrival of the Internet. in companies or the fear of the arrival of the photocopier. We have always had fears about the fact that employees would misuse the tool we made available to them. I start from the principle that people will use it for the benefit of the company. And the more money I spend on the use of these tools, the happier I am, so to speak, because I think we are ultimately improving. The company.
[00:32:57] Exactly. And a second obsession of leaders, in general, is use cases. And it's a bit of a hobbyhorse on LinkedIn, it's a bit of a race for use cases. How many use cases do we have? Have we cracked any use cases? Except that the problem, once again, if the population and all the professions have not embraced generative AI, in general, employees want to please their leaders, meet their annual objectives, so they will create use cases. But they will create use cases potentially without the real end users. We all know what that leads to: a mismatch, and also the potential to create a gap between the professions and the people who created this use case. I have faced this exercise myself. I am currently on assignment at France Télévisions. I had to work with a person who works in internal communications and who writes press releases. She told me, 'Yeah, I spend hours on it, I finish at midnight every night.' Okay, no problem, I'll create an assistant for you, you'll see, it will help you write press releases, you'll have a first step, it will be great and everything. I had an interview with her, I showed her my assistant, she told me, 'It's great, it's amazing, the exercise is top-notch, the result is great.' Except that it was me who did the demo, and afterward, I followed up with her ten times to find out if just... She had opened my product, she had never opened it. Because actually, I think I created even more fractures between generative AI and her, since in fact, I showed her it was magical, I showed her it was great, but in reality, it seemed complicated, since she had never even prompted once in her life. So in fact, this risk of pushing use cases at all costs means that if we go too fast in adoption, instead of bringing people along with us because the result is splendid, we risk losing them simply because they felt it was going too fast for them and that they weren't part of the subject.
[00:34:32] So for me, if there's only one metric to follow, I think you've understood, it's adoption. But in this case, it must be done well. There's only one, but we must aim for 100%. The real idea of a successful company is one that has brought all its employees along with it on generative AI, without putting pressure and with this truly collective dimension on the subject.
[00:34:51] It's just to give you some feedback. We deployed this at Sphère. It's been almost two years now since we had our first tools. And we really follow this indicator. I think in daily use, among the 1,000 employees, we might be at 75%.
[00:35:12] While I think everyone could ultimately use it on a daily basis.
[00:35:17] Our challenge is to manage to bring everyone on board and understand why there are still people in the company who ultimately don't use it.
[00:35:23] And you can imagine that if it's 75% at Sphère, it's not 75% at France Télévisions. We're not dealing with the same population, the same addiction to technology. So there are still a lot of people to reach. retrieved from slightly more traditional companies. And in my opinion, this is how we will achieve the use case, that is, by generalizing use, and also by showing what types of use we can have, okay, it makes a report, but I can also create an assistant that is my career coach, oh yes, okay, it also has this brainstorming capacity, and the more we show the range of possibilities, open up the field of possibilities, that's when the professions, the end users will be able to project themselves onto 'Ah, but okay, I could maybe use it in what case,' in such parts of my daily life, particularly what bores me to death. And there you have it, in any case, it's by taking this tool in hand and having this experimentation that we will push for ideation, for creativity around use cases.
[00:36:18] Next, we must accept the question of the long term. Didier spoke about 2030. So, there are certainly things to do before 2030. However, here we are really in a logic of adaptation, of profound transformation, almost Darwinian, where quite a few things will still fall into place. And so, we can't just say, OK, that's it, I gave you a masterclass, generative AI has no more secrets for you. We are much more in something a bit more daily, like we had with coaches in Agile, why not Gen AI coaches, or even why not champions. We easily identify in departments people who have taken hold of the technology. Thus, this proximity, this accessible knowledge can be extremely valuable. A small caveat, be careful about representativeness among these champions. We know that in general, women tend to exclude themselves, that sort of thing. So, interesting. still aiming for a certain parity in representativeness. But there you have it, accepting this long-term perspective, also accepting that it's a logic of 'we tame each other, we don't impose,' there's no point in setting annual goals on the number of assistants created, I think that's not necessarily the right method to get everyone on board.
[00:37:23] There is a real question of inclusivity, and this time, it may be somewhat generational. Here, I've shown you the figures from the digital barometer. We see that on average, 33% of French people use it, which seems more realistic to me in my daily life than what I saw in the room. But of these 33%, we still have 77% of 18-24 year-olds who are obviously included in the 33% who use it. That makes you think a little about what it looks like across a population rather than a generic case or other. So there is still a societal challenge in saying that we need to bring everyone along and especially not start creating a generational divide around this tool.
[00:37:59] Obviously, and from experiencing it daily in seminars on generative AI, there is a lot of fear, a lot of anxiety. It ranges from 'will humans be replaced by machines?' to 'what is the ecological impact?' and 'what are the biases within the tool?' » There are a lot, a lot of questions, and that's normal, it's new, we hear about it in the media. My technique is to lay everything out flat, to be hyper-transparent about everything—what we know, what we don’t know, what it does, what it doesn’t do, what it does well, what it does less well—to try as much as possible to calm the emotional side a little and try to rationalize things.
[00:38:37] Responsible use includes ecological questions. I believe there’s a conference on Green AI tomorrow, so it’s a real topic. Now, it’s also a real opportunity to ask ourselves questions a little differently—that is, AI is here, what do I actually do with it? At what point is it appropriate for me to use it? If I just need to rephrase some language for a colleague sitting next to me, whom I talk to every day, I’d be faster writing my email than iterating six times with an AI to get the right tone. So there are situations where AI is not at all relevant. Then, afterward, when I’ve generated things that seem to be of good quality—typically images or others—why not consider something a bit collective where we’ll store AI-generated material instead of generating it a second time. Meeting minutes—if everyone does their own little summary with their own transcript, it’s not very eco-friendly. We can imagine that only one person uses an assistant to get the collective minutes. In any case, there are plenty of things that can also give rise to ideas on how to have responsible use of this AI. Then, biases. Biases are a real question around AI. My point of view on this is that biases can really be a deterrent. Eventually, I might really tell myself, 'Oh dear, I don’t want that, it doesn’t represent me.' Every time I ask Midjourney to create a secretary in English, I don’t specify whether it’s a man or a woman. It gives me a 30-year-old girl. with rather pretty glasses, I don’t relate to that. Okay, but in fact, if I know in advance that Midjourney will give me this pretty girl, I can also, in my prompt, act against it and tell it I want a 60-year-old male secretary who’s a bit paunchy. Why not? At least I’m in control of my prompt and I know what I’m going to do to avoid being disappointed. Again, it’s all about the knowledge and awareness of what I can or cannot do and how far it can go.
[00:40:35] We saw it earlier, but there’s also a collective aspect to develop. We’re always telling ourselves that digital progress will make us even more individualistic. As we saw earlier, there is still an opportunity to somewhat reconcile generational differences by perhaps leveraging the strength of potentially younger individuals in the workforce. There is also this idea of sharing, as we saw with resources. We can also imagine, and this is the case in the platform we are setting up, assistants that are shared within the team. And thus bring this idea of collaboration to find these use cases, these AI individuals that will serve the entire team.
[00:41:10] Finally, as you will have understood, there is truly a balance to be found between these AIs that are arriving and humans to know, for us today, how we define ourselves as human. Some time ago, we would have said that a human is someone who reasons. Tomorrow, will it still be someone who reasons, knowing that AI, we say it does not yet reason completely, but to what extent will we say it does not reason completely? There is this whole question of balance to be found to know where our place is, where the place of AI is, how it can augment us, how it can serve us, and in what ways humans remain key. Didier spoke earlier about the question of accountability. This is one of the points that we foresee as still being in the domain and the prerogative of human beings.
[00:41:53] On this topic, we have been asking questions about artificial intelligence for some time. And what I like to say is that artificial intelligence has nothing intelligent about it. It's a debate, we might be getting into things that are a bit outside the mainstream, but for me, as soon as a machine can do something, it means it's not intelligent. I give the definition of artificial intelligence in reverse. That is to say, if we can do it with a machine, it means there is no intelligence behind it. Typically, if we had arrived with a calculator 200 years ago, we would have said, 'Wow, it's super intelligent.' And ultimately, who says a calculator is intelligent? It is not intelligent. Translating a text, 30 years ago, Frankly, you had to be super intelligent to translate a text. Now, who would say that DeepL is intelligent? No one would say that DeepL is intelligent. Because now, machines can do it at scale. So there is this notion of intelligence somewhere; it constantly shifts with what we observe in what we are capable of doing algorithmically. And I think there are a number of things that we consider to be part of
[00:43:05] of human knowledge and thus of human intelligence, which will gradually emerge because we realize that after all, AIs can do it, and we managed to do it without necessarily the notion of intelligence.
[00:43:18] I come to my last slide, just to say that today, in my opinion, forget ROI, traditional KPIs, invest in humans, as Didier did at Sphère, meaning recognizing our limits a bit, and having full hope in reaping the value tomorrow from this investment of the moment. And there you go, the sooner you start, the better, because as we've seen, it's going very, very fast, and the companies of tomorrow that have adopted AI in time will certainly be the most successful.
[00:43:48] I might add two economic points regarding this: invest in humans and provide the tools. So for our generative AI platform, we didn't go to an operator, and we don't pay per number of users per month. In fact, we connect directly to the LLM, to the APIs offered by these different players. As I was telling you, we have about 75% daily users in a group of 1000 people. Currently, our bill... Related to AI, is 2000 euros per month. After all, it's not that expensive. Second topic, I was telling you about the arrival of CloudCode, which changes things a bit for me—we are still mainly a company of developers—and which changes the way we will produce code, the speed at which we will produce it, the simplicity as well. And we must consider that Cloud Code will cost... the company about 10 euros per day.
[00:44:51] And yet, I believe it is absolutely essential that we deploy this tool within the company and that people use it. Because I know that through the use of this tool, which will significantly increase our AI bill
[00:45:05] —since we're now talking about 200 euros per developer—I believe it will be beneficial for the company.
[00:45:12] It's 10 euros per day per user, meaning 10 euros per day.
[00:45:15] 10 euros per day per developer. That's roughly the cost of using these modern tools, because they consume a lot of tokens, but in the end, what is produced is also...
[00:45:29] of fairly high quality. I don't know if you attended the previous talk on code review. What we notice is that if we use code generation tools a lot on development platforms, code review is something that is quite tiring. It requires a lot because you have to dive into someone else's code and concentrate. It's a task that, cognitively, leads to a lot of fatigue. I think we're going to have a bottleneck because developers will increase their productivity through these tools. However, code review, I think it will still remain human for some time. As a result, we still have humans who will review the code produced by people who have increased their productivity.
[00:46:18] Thank you.
[00:46:20] We can take questions.
[00:46:29] Oh yes, with pleasure.
[00:46:31] Hello, I was very interested in your presentation, particularly about AI agents.
[00:46:39] But I still wonder, what is the point of having an AI that does things compared to an algorithm where we control the... The different processes and thus have uncertainty about what it will produce.
[00:47:01] Okay, great question.
[00:47:06] So actually, there are times when,
[00:47:10] If you will, there are indeed two types of AI agents. There are AI agents that will execute graphs, as you say, and so there is a well-known technology in this field called workflow graphs, maybe you've heard of it, where in fact we have predetermined the set of things that can be accomplished. But there are some cases where we can't represent the space of possibilities in the form of graphs. Because potentially, we will use several APIs to respond, one API will respond with something, and determine which is the next API we should use to continue... To move forward, in fact, it's done on the fly, and it's what the AI responds with that will allow us to determine which tool to use next. In fact, we are not able to represent what needs to be done in the form of a graph. At that point, we try to do it through AI. Someone said, I wanted to add, it's interesting your question,
[00:48:08] to add a point, it's that recently there was a very large conference in New York, about a month and a half ago, something like that, on agents, and there was a person from a company who said that in fact, regarding the deployment of AI in production as agents, an agent In production, it consumes a lot of tokens. As I was telling you, for a developer, it's roughly, you have to consider 10 euros of consumption per day when you have a somewhat powerful agent that is the developer. And what he was saying is that for him, with the current cost, which will necessarily change, we imagine it will go down, you shouldn't answer a question through
[00:48:50] an agent if you're not certain to create more than 1 euro of value with that answer, 1 dollar, sorry, 1 dollar of value with the answer that will be provided.
[00:48:59] So what I mean is that, to summarize regarding your question, if we can represent it in the form of a graph, we should do it. Sometimes, we can't represent it in the form of a graph. And actually, we will address it, the flow of responses will be determined by an AI, and
[00:49:21] the path the AI will take to answer will consume a lot of tokens, and therefore, we need to be certain that it will create a lot of value afterward. I don't know if that's clear or not. Moderately.
[00:49:36] If you want to experiment with agents, there's a technology that is very easy to use, called Small Agents. It's a framework pushed by Hugging Face. Small Agents, S-M-O, Small Agents. It's about 1000 lines of code. And when I want to test and tinker with things around agents, or even have things go into production, right now, I do it with Small Agents. It's really very simple to use.
[00:50:02] And as long as you haven't used it, it's quite abstract to understand what's happening. You really need to manipulate an agent and watch it execute and see its thought process unfold. I'll give you an example, I have one that amuses me a lot. I love Wikipedia. And I wondered, am I capable of creating a news agency based on Wikipedia? That would be great. Because in the end, there are lots of people who fact-check on Wikipedia, we have a fairly reliable data source, but ultimately, I'd like to have news bulletins coming out of Wikipedia. So what I did is I have a first tool, because Wikipedia publishes its traffic, daily, and so we're able to determine on Wikipedia which pages receive the most traffic. And from there, what I do is detect the pages that get the most traffic, I take the text of the page today, I compare it to the text of the page from a week ago, which allows me to detect what has been modified. in the page, so it's a series of lines that has been modified, I send this text to an LLM and I tell it, based on this text, produce a news dispatch for me. And so it outputs for me, I tell it in French, follow the style of an AFP news dispatch in French, and it really produces fairly accurate bulletins, and it works rather well. And in fact, I ask it questions in natural language, and it will determine, based on the question I asked, which tool it needs to use to answer my question. Will it search in the stats? Will it generate the news dispatch? So, this is really the agent that, depending on what I sent it in text form, will determine which tool it will use to answer my question.
[00:51:50] Okay, I'm a bit talkative. Alright, the question is for Didier regarding the digital twin. I'm here, sorry.
[00:51:55] Okay.
[00:51:57] Regarding the digital twin of the IS and the conversational aspect of the IS. From the moment the prerequisite is to have data, data governance, and also to have a documented IS, How do you think we're going to tackle these issues that are historically... It's been a long time that everyone has been trying to document their IS. And so, it's also been a long time that we've been discussing how to have a conversation with the IS. How do you think we're going to tackle these two prerequisites?
[00:52:28] One minute.
[00:52:32] First, I reverse the subject. I say, for me, if I'm in a Darwinian system, there are companies that will succeed and companies that will not succeed. I think that the companies we will have in 5 to 10 years are the ones that have managed to solve the problem you mentioned, which is an old problem and one that, until now, we haven't been able to crack. After that, there's the issue of data governance. An issue that is... When you go see a member of the executive committee, they say, 'I want money because I want to do data governance.' You get zero. Until now, these were topics that were very difficult to defend at the COMEX level because it's very abstract, people don't understand. There is a way to recover money. which is quite simple.
[00:53:17] It turns out that I often pitch to COMEX these days.
[00:53:22] What I explain to them is to say that it's simple. I tell them that we need to implement the governance of the DNA, and they don't quite understand what that means. I say that it's extremely simple. You will ask a question to your IS, as CEO, you will get an answer. An employee from the Landes region of the company will ask the same question and may potentially get a different answer.
[00:53:45] If you haven't implemented data governance, you can't do that.
[00:53:49] And so, if you want your test to be relevant and for there to be answers that are tailored to your question as CEO, who has privileged rights regarding data access, you absolutely must provide the governance teams with the means so that these AIs can answer you pertinently.
[00:54:07] And just to add something on this, it's also much more concrete for the business lines because we, for example, created assistants for the purchasing services on purchasing procedures. And the assistant responded very well based on fairly simple and high-quality documentation. Except we said stop, we can't put the assistant into production because, in fact, if the procedure changes, who is responsible for the assistant? How does it work? And that's when the concept of governance comes to the table and takes on its full meaning. It is much more explicit than indeed pushing governance for the sake of governance. So it's also a way to perhaps address issues that have been latent for a long time.
[00:54:44] Thank you very much.
[00:54:45] I have to interrupt you... We have a talk after this. Thank you.
[00:54:50] Thank you all.
[00:54:51] Thank you.

Transcript (Translated)