Using AI to capture signals of change - AUTONOMOUS session

Hi, everyone. Welcome to the second half of the summit. Today we're going to give it a couple more seconds for everyone to filter through and then we'll kick things off. Awesome. For those that have just joined, hi, and welcome to the second half of the summit. Today we're just going to kick right off. My name is Natasha and I'm joined by Leo. And we're going to get super, super tactical in today's session. We're going to be talking about kind of how AI can make consumer insights and consumer data super, super actionable and really impactful. So like I mentioned, I'm Natasha. I am an associate director here at boi. I lead AI strategy, AI product and AI innovation projects. And I'm joined by my super awesome colleague Leo, who is a specialist in data science and AI. So I am sure everyone maybe has heard the words autonomous before, maybe have heard the words AI. But if you haven't, we do know that this world is becoming more and more autonomous. So if we shift to this slide, it's happening kind of all around us, from assistance to products. In fact, I drove in a Waymo, an autonomous vehicle the other day and it was super, super awesome. But there's many other things that are happening, including kind of agents are becoming more autonomous. They're making decisions for us and on our behalf. But what we really want to do today is we want to kind of take a peek under the hood. So what actually makes an autonomous world? How is an autonomous world actually powered? So autonomous world actually means a constant stream of data, a constant stream of inputs that it can learn from that and then make decisions and optimize for the future. In this specific session, we'll be focusing really on kind of consumer data and consumer data meaning real consumers, not synthetic consumers. And the way in which we kind of get consumer data is through social media and social listening. But social listening today kind of leaves us with insights that aren't as actionable as we might want them to be. So who of you have seen a word cloud? And I just, like, I don't know what to take from this, but all I know is that currently, according to the social word cloud, social media is important. Or has anyone else seen a sentiment analysis where it just says X percentage is positive, y percentage is negative, and the rest is all neutral? And we don't kind of know why. We're kind of left really understanding just kind of the basics, but not really in depth. And so this is really hard for us to kind of power this autonomous world. Kind of static Social listening data and so, and social listening data that isn't as actionable. And so we at by decided to do something about this and so we developed something called agentic social listening. Now I am assuming that a lot of you have heard about what an agent actually is, but I just want to kind of level the playing field for those that are online. So what is an agent? An agent is an algorithm. An agent has a set of goals and it works on those goals in an autonomous fashion. But what's really awesome about agents is that it learns from the interactions that it's making in order to kind of maximize outcomes as it continues to achieve that goal. So that's kind of how we will define an agent. And so when we talk about kind of agentic social listening, we're really talking about a system of agents basically understanding the context, the sentiment, the emotion within social media. So for example, you can have a emotion agent within kind of social listening, just focused on the task of understanding what is the emotion of this consumer or this user or a context and understanding what is the context behind this consumer. Why are they saying what they're saying based on the time of day, based on where they're located and other factors. So I am not going to kind of bore you with any further detail. I'm going to kind of hand it over to Leo to really get into the meat and bones about what is agentic social listening.

All right, thank you, Natasha. So we bring a case study that we consider that could be a really interesting way to explore agentic social listening. This is a fintech company that wants to connect with gensys. It's a real case, but we did some modifications. So please consider this as a hypothetical case as we are not, let's say, capable of sharing all the details, but there are so many elements of these that will help you understand better. What exactly is Agentix social listening? So long story short, this is a major fintech company. They are launching a new credit card for under 25, I mean Gensys and they want to understand better this audience to design a good strategy to position this product within this segment. So that was the request. I think this is one of the first project in which we start to use Agentix social listening. So it has evolved a lot. But it's an excellent case because we found the opportunity to create amazing visuals and amazing content to share with you. We are going to see how the process work and then you will have time for some questions in case you want to zoom in or zoom out in specific Aspects. So as any other social listening project, this start with data. In this case we collected 150k mentions of Gensys in Belgium. So the idea is that we basically collect these TikTok Reddit Instagram mentions and create a data set that then we analyze using of course agents. Let's see an example. We isolate one of the observations. As you probably already know, these observations contain multiple data signals. You have what is happening in the comments, you have different visual elements in the video, you have the transcript of the video where you can isolate pain points, competitors, et cetera. What we do is that we generate a couple of agents, like a layer of agents, and each of those agents is capturing those signals and, and this is happening multiple times. So once they have completed those tasks, we enrich our understanding of each element of that specific golden nugget. And then we repeat this process multiple times with each of the different observations that we have. So you see that this is helping us to understand better what exactly is happening out there. The good thing with agents is that they don't you are sleeping or focusing something more important. So we bulk the agents labor so they start to work with this large to clean up the noise, but also extracting those elements that are important to us. So in this case that's exactly what we did. This is basically giving you a data set that is, let's say cleaner, but also you have a better understanding of the different variables that you see in all that social media mentions. And that will help you to do something that we as researchers really appreciate, which is doing advanced analytics. For example, doing the famous clusterings. Clustering is grouping or segmenting soup segments within this big sample. This is a quick and easy example. You see that we create something that we import of innovation called visual clustering, which is getting different elements and cluster them. For example, you see cluster number one is related to sports. Cluster number two is related to fast sports like motorcycles, so on, so forth. And in the center we have the noise. So this is interesting because it's an excellent way to get to know better those sub population within this specific segment. For this project we detected 13, you see 6 examples. Excellent moment because this is where agentic social listening starts to make sense and to unleash the potential because we are connecting this with a new layer of agents that we call creative agents layer. Those agents are different because they produce content or they generate assets for your brand or for whatever you are doing in this case. Remember this is our research question. We want to know how to optimize visual Appeal to engage with the cluster seed. This is an example C2, which is the motorcycles and speed sports. And this is with what these agents came with, which is super interesting. You see that it really isolated multiple elements and this is again happening multiple times depending on the number of clusters that you have. So you see this iteration several times. You see the foam neuro, this is which is sustainable fashion. This is which is more related to sports, this is related to anime and cosplay. One of my favorites, if not my favorite, then the good thing is that for this project in specific, our agents were already really capable of understanding the different dimensions of the consumers. What we did was again to keep iterating and generate multiple experiments. For example, for the C2, C3, C5 clusters, we start to generate ideas autonomously of how exactly the communication of this product should be for that specific target. You see here some of the examples. The good thing about the agents is that they really get really well versed when it comes to understand the user's behavior. So that is not only about visual data, but also for example, pain points and solution analysis here at Kickout clustering, which is also super interesting for matching learning sometimes when we can have data to apply that. In this moment we are testing volumetrics, this is in testing phase, but also is a possibility. And just this is an example, but this is something that I really like about this methodology is that we are now testing the way to make this always on so always capturing insights from the market. It's an excellent example of how exactly this agentix system works. And I will pass it on to you, Natasha, to talk a little bit more about the benefits and downsides that we have detected and then open the conversation.

Awesome. Thank you, Leo. We've already had a couple of folks in the chat and also in the Q and A kind of start thinking about the applications in politics or in marketing. I think our goal is really more on the marketing side of things, but just to kind of call out a couple of different benefits. First and foremost, social media is filled with so much noise. And when we say social media, we're talking about the Reddit, the Tiktoks, the Instagrams and the likes, right? So it's filled with so much data and so much noise. So applying agents to social listening allows you to actually cut through all of that noise and actually targets you to uncover the question that you that you're asking sort of cuts through kind of all that kind of noise. The other kind of benefit that we really see with agentic social listening is that it allows you to kind of analyze multiple different types of media. So as Leo kind of showed, there is kind of a visual agent, for example, kind of looking at the visual elements around. It's looking at emotions. How is this person kind of talking, is there a kind of inflection in their voices, etc. So kind of the emotional stuff, it's looking at videos, it's looking at text, it's looking at images. And so it's really kind of multimodal in that sense, which is really helpful to then understand the kind of key drivers, the needs, the desires of a consumer versus looking at a post in a very static manner. The other benefit that I would like to maybe call out that you would have seen, and that is a kind of a stark difference from what we see in social listening today, is instead of going from social media and social listening data to kind of insights that you have to interpret and then create into an output, we actually see us going directly from raw data. You input a research question that you have and then you basically come out with kind of actionable outputs. And so on the topic of outputs, we just wanted to kind of run through a couple use cases with you and just understand from your perspective where is the heat coming from. This is by no means an exhaustive list, but some of the use cases that we can think about in terms of agentic social listening is creating kind of communication assets for your consumers. It's thinking about kind of campaign development, your branding assets. You can also think about it as kind of dynamic segmentation. As Leo showed before, there's kind of numerous ways to segment your audiences now. And so you can always think about kind of how do you dynamically segment your audience in an always on manner, packaging design, product development. So I appreciate, Konstantino, that all of them feel really, really exciting to you. I'm curious to know what are the key use cases that you think is the most important for you and for your business insights. Are all of these used in a company now? Yes, David, we've actually created these engines and kind of agent of social listening for our clients and for ourselves. So. So there are definitely stuff that's happening today. Great. So what we can do, we have about five minutes left, is potentially Leo. The two of us can kind of run through some questions before we, before we wrap it up. So let's take a look at what we feel might be interesting. How do you mitigate Leo, this is the question for you. How do you mitigate kind of irony or sarcasm in post to prevent Kind of false signals. How do you think about that?

Super, super good questions. We created an agent that was capable of detecting sarcasm and irony and you give an score. So when you create your crew of agents, there are some of them that are literally specified to detect noise. Not only. I wouldn't say sarcasm is the worst, like the worst part of it, for example, when you have a spam content or when you have, for example, commercial ads that are coming from a user, like what people call user marketing, something like that, that's somehow super hard to detect. But agents, we create a specific, in this case, we create a specific agent that was detecting the sarcasm and we calculated a score. It was not a big part of the sample though, so I wouldn't be that much afraid of that one. Unless that your topic is sarcasm. Sensible, like politics, for example.

Right. Another question we had a couple of comments in this, in the chat as well is can you clarify how the clusters were defined? With such inherent messiness in the underlying data, it feels harder than just asking AI, make me a cluster.

Yes, it's important to mention that this is not only generative AI. So we are using generative AI to enrich our understanding of the data and what exactly is happening out there. But then we are using some other clustering methods that are quite popular in this case, I think if I'm not mistaken, this was something similar to a KNN which is nearest neighborhoods. But I will need to check exactly what was the method. But it was not generated. In this case we use IT matching learning to do the clustering. But it's important that the data that we use it to give to the matching learning models was generated via generative AI so that it's a complementary system.

Another question that came up, which I'd happy to take and Leo, feel free to jump in as well, is the study looked like it's based in Belgium. Is there a risk that the approach would not work for a global company with cultural biases? Super, super, super good question. Good news is, is that kind of we have access to social listening around the globe. And so I think there are a couple of things. One is in your initial kind of data set, by that I mean kind of whatever social listening data that you're kind of getting. We want to make sure that we're getting kind of data from across the world and in your countries or cultures of interest. The other thing is you can actually think about creating a cultural agent, right, that actually uncovers the different nuances between the different cultures. You can even break that down Even further, let's say you're playing with three different kinds of cultures or demographics or whatever it might be. Imagine you had an agent for each of those that represents each of those and their needs and their desires. So if you had those three agents that then work in tandem, you're able to then kind of uncover the nuances and needs of various folks from around the world with different backgrounds, with different perspectives. So anything to add, Leo to that one?

I think you already covered quite well the points. I think that sometimes when we need to change the geography, it happens. Like any other research that you are doing, you have an impact. So you need to recreate the scope. Just, for example, language will require also a different methodology because we are in many cases we were using, for example, Dutch agents and translated to English, sometimes because of the need of the consultant understanding. But it will require a nuance methodology indeed.

Awesome. We have about 45 seconds left. There has been one question in the chat and in the Q and A function that every single person has been asking. And Leo, this is a question for you. How did you make your slides? What tool did you use to make your slides? Because it was awesome.

Canva. Canva? Yeah, Canva. And a lot of passion. I know we wanted to animate because we wanted to make it simple to understand, so that's why we wanted the animations. But Canva, super easy to use.

Awesome. Well, folks, feel free to reach out to both Leo and myself. Leo, for the technical and Canva stuff and we'll see you next time. Cheers.