Powered by AI #:12 AI and People – From Knowledge to Organizational Intelligence

Hi everyone and thank you for being here this evening. I'm Muhammad Shaji. I'm in charge of digital innovation acceleration within NG. Today I'll be very pleased to introduce and ask questions to our panelists for this new edition of Powered by AI held by the NG and MIT Club France. But before starting our panel, I'll ask Claire to introduce the YPN program and YPN network because we are holding this session with YPN.

Good evening everyone. Welcome. I'm Claire Lua, part of the Young Professional Network. YPN is a community to connect young professionals at NG, to share ideas, foster innovation and better understand the vision of NG. Each year we organize an event, Creative Way Lab, for young professionals and it's in this partner area that we are welcoming the Powered by AI session.

Thank you very much, Claire. Very happy to be part of this YPN program. So to introduce the session today, we will of course talk about AI and people and how it's shifting the way we are interacting with machines today and how the adoption of these new AIs or new machines are developing a new way of trust and planning a new or designing a new future of work. So before starting and asking further, maybe I let the moderator finish the introduction and then give the floor to Enes and Florian to introduce themselves and talk about what they're doing in their current activities.

Thank you, Muhammad. Hello to our speakers Florian and Enes and thank you to everyone. Sorry for the late arrival. It's one of those days when things don't always go as planned, but I'm glad I'm here. Thanks for being here as well online and physically. We have an amazing roster of speakers today. I'm really glad to welcome Enes and Florian. We discussed a bit at Adopt AI that took place last week at the Grand Palais on new topics surrounding the intersection between machine and humans, and the short discussions we had led to organizing this session because I felt that our speakers had good insight on both work that has been done in the past and how their solutions are being implemented in industry, and what their vision is for the future of human-AI interactions. So maybe a quick word of introduction from both of you and then we can kick off.

Hi everyone. My name is Enes Besp. I'm the founder of Cedex. Cedex is an AI assistant that is specialized in speech to text. So basically we do transcriptions, we do summaries and then we have a RAG where you can find every information within the system.

We can start. So let's start with the first question. We see the movement today that goes from the chatbots and what OpenAI launched three years ago now and we're moving step by step to AIs and machines that are acting. From your perspective, what's moving today from this chatbot interaction with the machine to agents that are taking actions in an organization, in processes within an organization?

From my own perspective, I see that we went from being passive to being proactive. Before, we would just interact and ask for an output from a system. We ask a question and then it searches for information and then it can analyze information for us. Today it can anticipate a need. For example, if we ask a question, not only do we receive an answer, but we receive a request to, for example, prepare an email or draft something and even go more into the action of embedding it within the system that we're using right now. For example, I'm using Microsoft. The system knows that it's Microsoft and it can send an email directly from Outlook. So right now we're at the beginning of agentic. It's still quite blurry what we can concretely do and what is still a utopia for us. But what we can concretely do is prepare actions. What still is a utopia for us is the fact that we embed it perfectly within the system so that it's frictionless for the user, so that it doesn't even anticipate or connect anything and that any user today can use it and just create any action whatsoever. But obviously this depends on every use case that we have in the company.

So Florian, you've long argued that the chatbot was maybe the first entry point to this new generation of AI interaction. But if we pull the thread, we go towards whole process automation. How do you think about that journey from a user's perspective?

To react to what you're saying, Florian, does it mean that we will move from doers to managers of things doing or agents doing things?

I think we'll also have two different types of people: the makers and the managers. The makers would be the brain behind how we construct to evolve with AI and how it can actually fit a real need. But on the other side, you will have the people that are not interacting at all with AI because it's still manual work, and we will still have manual work. So I think there will be a gap between the workers that would just still do what they're doing today but at an enhanced pace without having a large component of AI disrupting what they do, and then you will have the managers and the makers constantly trying to innovate and improve not only their efficiency but also the way they think and share knowledge. So I think this is where we are in the industry as well. The industrial sector and the energy sector, I'm sure there are a lot of manual work that is still very far from how we're using AI in the day-to-day job that we're seeing today. And maybe their work is going to change in a different way and not totally in phase with what we're seeing with managers and makers.

To your point, in the industrial sector we're seeing embedded AI in drones and robots that do maintenance inspections instead of teams going out on the field, so we're seeing a shift in those kinds of profiles. One thing that I've argued for is when data science or programming students ask me what's the future of their field, I think there's a convergence in terms of individual profiles on domain knowledge and problem-solving capacities. By domain knowledge, I mean you need to know what you're applying your skills towards, because AI agents are now helping write code and solve problems, but you need to be able to identify the problem you need to solve and describe it in a way that is precise enough and also be able to evaluate the output of the model. So that's one of the use cases in terms of what's evolving in human-machine interactions. Development is one we've seen a big peak of usage and then it sort of plateaued and went down in terms of percentage of downloads and usage of tools like Lovable and Copilot. What's your feeling on that? In your companies, do you ask your teams to use AI co-development?

I do, but I also see that there are nuances. I don't force them to use any specific tools. I ask them to explore, to test, and also to see what their input is and how they can value their time. For example, we tried Lovable many times, but it doesn't really fit into our processes today. We found a few flaws in how we can create with Lovable, so we decided to use other tools. Obviously they use GitHub Copilot and other coding tools, but I feel that today is an important time to really do our benchmark and also try to see how the value of each engineer and developer can not only improve but, as we said, with just managing the tool, how they can manage at their best the tool with the knowledge they have and how they can also improve as a developer. So 10 times productivity is important, but not only productivity—quality. How can we improve the quality, not just using any tools? We had a peak, we saw many different tools, we spent a lot of time doing benchmarks, and from that benchmark we realized that some tools can give us a bit more time but they don't make our platform better. They don't increase specific value. The value that we manage to get as a developer is mostly the quality of embedding our code into different workflows and integrating it into a system that is connected with every single system from our clients. For that, we didn't need a specific tool; we needed specific knowledge in a specific environment like Microsoft. We needed specific developers for that, and we didn't find any tool that could replace that knowledge specifically. So I think it depends on which components and which type of feature we need to improve. But of course, the job of a developer is going to evolve day after day, year after year, and I'm sure in the next years we're going to see education change completely in that sense.

And Florian, I'm curious about how you see it in-house, developing that AI platform, but also for the users of the platform. Are those features that you're pushing for?

Let me go back to a question or to what Enes said about the process, saying that these new tools are not fitting in our process either in your own teams or organizations or with the organizations you're working with. Did you see how AI or the future agentic AI are changing processes? Do you have real feedback on processes that completely changed because we're introducing AI within them?

100%. For our clients, we see a before and after that is quite clear. The only problem before was that knowledge was really scattered—in emails, Microsoft, whatever we said with a colleague once, or in a notebook. Now we can capture all the data in one point. For us, especially in speech to text, we're capturing all the conversations we have—any type of conversation, physical face-to-face, video calling—we capture it. It's live data that is automatically updated every second. So we have a live knowledge management board that is constantly updated. Then we clean this data and make sure it is structured and usable for a specific use case. For example, if I go with an HR within a company, HR is going to make a lot of meetings with potential candidates. They're going to interact with 100 candidates for one big position. First, they call a number of candidates for 5 to 10 minutes, then they have a recap. Out of those 100, they meet probably 30. Out of those 30, the HR takes one hour to have a conversation with them via video call or face-to-face. We capture this data, put it in a structured note, and it can be sent to every manager in the process. The manager can ask different questions, find data that the HR probably didn't capture. Then once the person is onboarded, we have every conversation of the onboarding. The manager gets all the information already structured; they don't need to go back to the HR and ask the same questions. At the end, when the candidate is selected, they can have an extremely personalized notebook with everything discussed in the interview, without any friction, without going back and forth, without having data everywhere. Then once this candidate does 10 years of career and leaves, we have captured every single piece of data all along. Obviously we're not at 10 years yet, but we can already see the progress from a company we onboarded two years ago to today. Once the knowledge is gone, we can still keep this knowledge for the next person that is going to be onboarded with the same tool. So it is usable. I try to make the example as simple as possible, but this is a concrete example of how we can already concretely use AI effectively, have it within our processes, and make it work seamlessly for any type of company. This can work in every single different job in specific sectors. We have this for inventories in factories, for example, when they do live inventory within a factory. We have so many different examples.

The process you describe is interesting and it's an incremental improvement on existing processes. What I'm interested in thinking about if we pull this further into the future is how do we rethink the process end to end from scratch? If we were to use AI as a basis for everything, especially in the context of HR, one thought that comes to mind is that currently interviewing is screening CVs, interviewing, etc. Eventually, what might be interesting is to look at performance on the job of people who have been interviewed with certain parameters and metrics for them being selected. Can you use the outcome of annual evaluations, for example, as something that would train a model that would tell you that these are the criteria that made sense during the interviewing process, what are the misses, and what worked well? Both in terms of your experience and maybe Florian as well.

100%. If I can answer quickly, I talked about being proactive earlier. But what if the candidate said they can do a specific thing about engineering, and then once the person is in the job, leading a project, and the manager comes into it and the candidate finds difficulties, we can have an AI go through every single document, every single thing that the person said and go back to it and say, 'Oh, but this is an example. You said that you worked on that case previously at that time with this company and it has the same components. Maybe you can use that directly.' If the AI can be ongoing, we're working on proactive AI where when we speak, the AI processes every single thing as a real transcription and goes into all the documents and finds all the documents that are related to what we're saying. We can do the same thing for a candidate with the HR process. They talked about so many different projects, made a case study, or even worked on a project in the beginning, and then two years later we can go back to it constantly and find the similar points that have been talked about.

Maybe just a last question on this topic. You were mentioning earlier local developers or citizen developers developing something very specific versus the idea that we can develop generic models that are applicable at scale throughout the organization. How do you see in practice, with customers you work with, the balance between both? That's a question we ask ourselves: decentralization versus centralization and how do we go at scale?

What I got from discussions with other colleagues and counterparts in other big organizations is that we asked people from business to define what are the best use cases. Don't you see in your ecosystems that these people from business define just processes where we ask AI to change their human way of doing things and not go further into the finer part of how AI is working and the real capabilities of AI to go further and change the entire process?

I always talk to different businesses about what they think would be next for them after giving them AI assistants, agents, whatever they have in mind. It reminds me of one thing I saw in a documentary about AI: back then, 150 years ago, if you asked people how they would see the future of transportation, they would say faster horses. That was the question from Ford to his customers. They wouldn't see a car because they didn't redefine the entire concept of going from point A to point B. That's concretely what we're doing right now: how can we go from point A to point B and redefine the entire process between A and B? What I found is that most people, when asked how they can use AI, say 'go faster, make the process easier, shorter.' It's always the same terms. But when we go back to the first question about the managers and the makers of AI, how are the makers going to rethink the entire jobs and different businesses for tomorrow? In the next 25 to 50 years, I think the giants are going to be completely different, and typologies of companies are going to change their entire strategy and also their entire way of defining productivity. You can do a lot more with a lot less. So how do you make business more viable in the long term with the amount of people we have today? How is education going to improve and create the new systems, the new jobs, and probably the new subjects? If coding or engineering is not going to be the core topic of the future, it's probably going to be about psychology or other aspects that are still not completely understandable by AI. So maybe one of the future jobs would be mixing psychology with current systems like coding and improving the entire area and the way people work together in a specific task. That's a lot of thought process, and I think we can debate over this and try to redefine every single job all night, but there are many different areas.

So if we talked a lot about workflows and processes, if we look at the interface between humans and machines, like we were saying in the beginning, chat and text have been the primary interface that shifted the mindset. You're working a lot with voice through meetings. There's multimodality in new models with vision included in the mix, also quantitative data. So what you were mentioning might be the right tool for the right usage—not LLMs and large models for everything, but specialized models that are more frugal, statistics-based, or have some level of expertise to do the right kind of application. How do you see the evolutions? You started with voice through a tool that listens to meetings. Are you also thinking about what the different modalities are that you can include in using AI for use cases?

I would say it's between two things. The first thing that any company asks us because we're working on voice and sensitive conversations is security. How can we innovate? We still have to be compliant, regulated, and follow many different rules within Europe, within France, and within the organization itself. So the first futuristic thing I see is that innovation cannot be slowed down by security processes. Today we have to go through so many different compliance and regulatory systems, and I feel like even the government is still trying to find the right cursor of how much to regulate and how much to leave innovation to continue to foster. Our goal is to create the most secure AI assistant that can be created and used by the organization in the most simplistic way, embedded frictionlessly for any user without even seeing that it's an extra tool or agent, directly embedded in their entire process and the way they work. So that's the first thing: having a frictionless tool while still being compliant with security and regulation. The second thing I see as what's next is proactivity. Whereas before we had a passive AI system that could capture notes and transcription and then we could talk to the AI afterwards to find information, tomorrow we will see this assistant anticipate my needs in real time. When I talk to, for example, different business developers in my team and try to see what we have done wrong and what we have done right, at the same time that the business developer talks about it, I have the news from what happened lately with this company, the different talks we had with the customer, and the different data points from Salesforce of how it can perform overall. So what I'm seeing is that we will empower in real time our thought process so that we can be not only more productive in general but more proactive in real time and also in making decisions. That's what we're working on, and hopefully in the next years we can have augmented not only meetings but augmented exchanges with different people, using every single resource and every single data that exists on the market, whether external or internal, to make the right decisions. Maybe that information will appear in virtual reality glasses in front of you during the meeting.

Oh, maybe not.

Don't you see in this user experience a difference? I don't know if there is a word in English, but in French there is a difference between 'usage' and 'utilisation.' You don't need to know how the subway works to go from point A to point B. Don't you see in the future of usage of this AI people that don't need and don't want to know how it works to get what they want from it? As you said, the new way of displaying information, people will take the easiest way to get what they expect, going through confirmation bias and the way our brain works in this interaction, just following what the machine is saying.

About the user experience, one way to think about it is that the software as a service model has dynamic data but a static interface. What you were describing as digging deeper and deeper is maybe a dynamic interface that is on demand for accessing dynamic data based on your needs. Is that something you are exploring? How do you see the evolution? For me, chat interfaces are like the console level, ground zero of interactions with machines from 30-40 years ago. Then there were graphical user interfaces that improved from there. Where do you see the next step?

I just want to go back to what you said about the fact that we're not maybe there yet and we're not ready. I think there are two things. First, there are people that are techno-addicts; they will try everything out, be very enthusiastic, connect and create their own agents. And you have an entire population that we cater to today, and we have to do change management for them to make sure they understand how the technology is going to be integrated with their workflows, but also to reassure them because they are scared. They are actually really scared of what's coming next, scared about their own job, the way things work. They think things are going extremely fast. The internet took almost a decade to be implemented and really used and integrated into the system. Today, it's been what, two or three years since it started to explode? So people are still trying to accommodate. That's why we still have this delay and we're not there yet, because we still need to have this entire change management process within the organization and try to make people at the same level of understanding and integration before we go to the next step. I don't know how long it's going to take, but I guess every organization has to take this into account before going faster.

Maybe that's a good segue for the last few minutes of the discussion around knowledge and expertise, things that are inside our minds as workers, specialists, experts, professionals. How do we get that out of our heads so that we can leverage it, capitalize it, make sense of it, share it with others, make others grow with it, but also use it for machines to collaborate with us? Florian, maybe.

Just to quickly finish on that, on the most simple basis, I think it's very important to just talk to the users. Don't just give an app and tell them to log in and find it out by themselves. It seems very basic, but trust me, the number of companies that were so happy that our team was really talking to them, onboarding them, taking them step by step through the process and reassuring them—I think for them it was change-making. I didn't see that in the beginning, but then little by little we found this as a very big difference in the way people actually take innovation into their own businesses and their own jobs and make it better.

So on these wise words, maybe we can end the conversation. Maybe we'll take a question or two. I know that Florian has a hard stop in a few minutes, but let's take a question from the audience here. Are there questions in the audience or you can ask also in the chat?

Hello, many thanks for the talk. Going back to the citizen developer concept. I'm lucky enough to be a developer and vibe coding is a blast. You can go 10x, 20x productivity, and that's a huge game changer for developers. The makers you were talking about that I can see in every organization are starting to develop—they can output Python, they can output with Lovable full systems at scale. I'm really wondering, maybe at NG or for the customers you are discussing with, how do you prevent this from getting completely out of control? You have two forces going against each other: one is you give productivity for the 15% of people who are actually managing the organization, and I'm wondering what the other 85% are doing. On the other way, you have IT professionals who are completely afraid of agents going everywhere and messing with data, maybe on the big processes you were talking about.

To be able to keep control and not make things go around, today we also try different types of applications. We give them the freedom to try and test out every single thing. But at some point, even in our size of company, we have to regulate and try to define specific processes that really work within our system and how we can embed them with our organization. Same for our clients. We have clients, large corporations, that are extremely reluctant with us working with other subcontractors or other providers, even if it's just to help us code a specific part of our platform or to help us with an API. They are extremely thorough on this because we have to go through every single process for them to validate it. We have to really think through: is this really worth it? Is it really going to go through the entire process? Even for us as a small organization, we have a lot of regulation from our customers on the first base. So I don't know how that works at scale, but for us we have a lot of boundaries. We cannot do anything we want. We would love to, but once we find something that really works for us, we go to the customers and say, 'Okay, that's what we're using. Can you just validate it or not?' But we cannot just change it every single month.

I think at the end of the day, it's about culture and mindset. The AI transformation is really a transformation of people and processes rather than a technological transformation as such. So thank you very much for this amazing conversation. For those of you who are here, we have a networking cocktail at the end of the day, like usual, and see you next year for more Powered by AI. Thank you very much.