Webinar: The Death of Static Segmentations

Okay, excellent. Well, we will get started now. It's very good to see everyone today. I'm very excited to have this conversation with you around the quote unquote death of static segmentations. Before we get into that though, maybe I'll just do a quick introduction. I'm Geoff Gibbons. I'm a managing director at Board of Innovation, an innovation consulting firm based out of New York and have been working with clients over the last year or so, particularly on how we actually reinvent the way that they look at innovation and marketing and the entire product lifecycle using AI powered approaches. And we're very excited to have this conversation today with you around how we're seeing companies evolve from a situation where they're reliant on what we think of static as static segmentation decks in PowerPoints and PDFs that might sit on a shelf for a long time. And how we're seeing new approaches actually improve the value of segmentations, whether existing or new, for companies as they look to better engage with consumers and customers. So we will dig into it now. We're going to go through a few things. First of all, we'll start with how we see the disruption of static segmentations and what that means. We're going to get into what we call living audiences, which is a new approach that we've been starting to deploy with a number of our clients at Board of Innovation and talk about what they are, how they work, what are the benefits of them, and what's the actual way that you can actually put them into practice, whether you have an existing segmentation or not. Then talk a little bit about what might be coming next. Future applications. Please feel free to chime in with any questions in the chat. I will try to really address them as we go through, but we'll also have some time set aside at the end to actually go through all the questions in detail. Okay, excellent. So we'll start out with just the overall disruption of static segmentations. So this really is a view of how we see the market evolving currently based on the last year or so of the deployment of generative AI tools in our business. So maybe just to take a step back first before we look at segmentations, the emergence of generative AI we see is actually marking a historic development in the way that companies can innovate and the way that they think about marketing and customer insights and new product ideas, and also how they really even think about the role of marketing. So if we just take a step back, you think about in the 2000s, everyone was thinking about Design thinking and human centered design methods really focus on empathy and creativity. And the segmentation really took off on the basis of that of how do you actually have a really detailed understanding of different types of consumers or customers and what are their real needs and how can you address them differently and be more human centric in the way you run your business. In the 2010s, you really saw the emergence of the lean startup methodology in terms of how companies looked at innovations. Everyone was trying to be entrepreneurial, trying to behave like a startup. Not that people aren't doing that now, but there was a big focus on how can you actually have more experimentation and validation around things. And so if you think about what that meant for segmentations, there was much more focus around how can you use them for targeting, for running experiments, and also like getting a much more data driven view of your customers. And where we're actually looking at now is in the 2020s, the emergence of what we're calling AI powered innovation. So this is actually the AI powered imagination, creation and scaling of new products and services and really the reimagining of all of the aspects of how you think about innovation, insight marketing and the distribution of a product even. And so we're going to dive into what that really means for segmentations in particular. But this is a general topic that as a company at Board of Innovation, we've been really looking at in a lot of depth. And one of the things we want to want to flag too is we have an upcoming virtual conference called the Autonomous Innovation Summit about a month from now, where we cover all of the different aspects of that. But this, this webinar is really a deep dive into how it impacts the disruption of segmentations. So if we think about like a consumer or customer, customer segmentation, I'm going to, I'm going to assume that most people are generally familiar with the concept, but maybe dive a little bit into how they're typically used. So typically a segmentation study, what you will have is a large study that is conducted that involves surveys and interviews, potentially whether in person interviews or video interviews, or what it might be to actually develop and identify distinct segments of consumers or customers. They could be done as Personas and some kind of way of visualizing and describing them, and there's an underlying data set. And basically what those segmentations are used to do is to actually figure out how can you actually understand the needs of different types of consumers and customers, so their needs, attitudes and behaviors, how can you use that to sort of test ideas with them? Too, how can you think about from an innovation standpoint, how can you generate ideas that are targeted to different types of consumers or customers within that segmentation? From a marketing standpoint, how can you actually personalize the messaging or the channel selection that you might approach different customers with? Retention strategies? How could you identify which types of segments more of a retention risk, or pursue different specific loyalty tactics for that segment? Figure out price sensitivity, how you might price differently for different folks, and then also just where you think overall you might want to grow your business. So where do you think you're actually less penetrating in different segments? Where do you see the most potential? So there's all sorts of ways that segmentations can typically be used within a company, you know, and often, you know, companies will spend a six or seven figure sum to actually develop these segmentations. And then there's a whole process of how do you actually get people to use them, understand them? Sometimes this might take a year or two, but typically what ends up happening I think in most situations is that you have, you know, a large amount of money and a large amount of time was put into the creation of a very well put together structured set of PowerPoint slides and PDFs and there's a lot of presentations that have to come after that. And then you might have a team that manages that. Right. And so they might field questions around that from folks in the business. How would this segment look at this? Like there's all sorts of follow up questions and things that come up once you have a segmentation. So it's really what matters is not just the actual segmentation, but how do you use it and how do you make use of it. The problem is though that there's a lot of limitations with that. So some of the most obvious ones I think that stand out, I think is that you have very static data. So the actual data you have in the segmentation is usually when the segmentation is done, is at least a few months old and it just kind of stays there. And so the segmentation slowly gets a little bit more out of date each month, each year. It's a very expensive process to gather and analyze all those findings. It takes a long time to actually generate the organizational understanding of the different segments, get everyone to talk about them, use them in their day to day work. Really also segments, if you have five or six segments, basically they will lack nuance and reinforce stereotypes. That is a function of what happens when you go from millions of people down to five things. You are going to have to make generalizations and you will reinforce stereotypes. You often find too that there's limited actionability of those segmentations when you're querying or asking questions about topics where there wasn't data captured. So, you know, you might, let's say, you know, let's say, for example, you're working in a, you know, you're working for a pharmaceutical company and you, you know, gathered all sorts of information about how people perceive a certain symptom. If you're looking at a new symptom that wasn't studied in that segmentation, you might not know really how to actually figure out what that segment thinks of that. And then generally there's sort of a lack of digital data sources, whether that's from social media, from Amazon, from whatever kind of external data sets that might exist. They're typically sort of hard to integrate into these types of segmentations. So a lot of problems. But I think that it's a commonly established way of doing things that companies still find value from. But I think what we found is that actually there are some new approaches that are possible now just in the last year or so, which can build on segmentations and actually be much more valuable for companies. And so I want to talk to you a little bit about what we call a living audience, which is an evolution from a static segmentation which we have found in our initial deployments with clients to be very effective. So what do we mean by a living audience? A living audience. Imagine instead of having, say, five or six segments, you have a very diverse panel of AI generated synthetic agents. And we'll talk a little bit about what that means that mirror the distinct demographic, psychographic and behavioral profiles within those segments that can be used to generate what we call synthetic predictions about those segments and the nuances within those segments and also be kept updated over time. So when we say a synthetic agent, we really mean it's a digitally constructed Persona of an individual person within a segment. So imagine for an individual segment, you might have 50 or 100 different agents that represent all the different types of people that actually might fit in that segment, so that you can actually really model out and understand all the different variants and ways those people might be different. Rather than just sort of relying on the sort of stereotypical view of a segment, they can be used to generate synthetic predictions. So synthetic predictions, basically this is a technology, that functionality from generative AI which enables you to predict based on an understanding of what that synthetic Persona does or is how they would respond to a certain thing. So what would be a need that they would have in a certain situation, where would they buy something? How would they respond to a certain type of question? And what we found from research, academic research, as well as our own actual experimentation with working with clients on real case study examples in real life business contexts, is that actually the synthetic predictions are very closely able to mirror what real people would actually say or do. So it gives you the ability to answer questions and make predictions around how a specific segment would approach something in a way that actually helps you to make better decisions moving forward, rather than you just inferring something from a PowerPoint deck that isn't quite addressing that topic. And then the last piece is the living piece of this. So the idea is that a living audience can actually be updated over time. So whereas a static segmentation, you have a data capture stage and then you develop the segmentation and then the deck and all of the materials, what you would actually do is have a live model where over time, based on new data that emerges, whether that's from research studies that you conduct looking at social listening, integration, point of sale data, e commerce data, whatever it might be, to actually figure out how you can actually update those Personas and make them more relevant around how they actually evolve over time. I see a question in the chat from Ann around, have we actually conducted AB testing on this research? The answer is yes. We've actually worked with clients in projects where we've actually done things like, for example, let's take 12 product concept ideas and let's run interviews with real people and then let's actually run synthetic interviews with these synthetic Personas and figure out which of the 12 they prefer most. And so you actually would run the interviews with real people. Usually in this instance, I'm thinking of, we actually did with a CPG client, we actually ran 20 interviews with two different segments via video interview. And then we constructed these synthetic Personas of the different segments and ran the same study synthetically. And the prioritization that we actually got was exactly the same. So if the goal was to really prioritize from 12 down to 3 or 4, you got the same result from running 20 interviews conducted by people is actually conducting this synthetic evaluation of what the Personas would actually would say. And so with that, you actually get massive time saving. And also you're able to not just test 12 ideas, you can test 300 or 3,000 at the same time. And so the benefits really that we see around this is that you have, as I mentioned, you have real time data that evolves, you have faster, cheaper data gathering so you can analysis, so you can actually use AI to really speed up the analysis process. And you can also have these kind of low cost updates, right? So you could actually develop an integration with a social listening tool or whatever it might be to make updates to the data over time. You can have intuitive interfaces to actually build organizational buy in and relevance. So I'll show some examples of what that might look like. But imagine that instead of just having like a PowerPoint deck that has the results of the segmentation, you actually have a tool that people can ask questions from, generate ideas, generate insights, needs, whatever it might be. And then you actually also have more diverse and inclusive panels, not one sided segments. So you actually have let's say 50 or 100 Personas representing each segment. So you understand all the different facets of people within that segment. You can actually think about ease of inference in topics that are not covered in a study. So even if a certain type of question or query you had was not covered in the study, you can still make a prediction around based on what it does know about these people. How would it actually work? As I said, in these AB experiments we've done, it's actually very effective and very cost efficient. And then it's built for digital integration. So you could really imagine. And we'll. I see there's a question in the chat in terms of how you actually generate the synthetic agents. We'll get to that in a second actually and talk a little bit about the process by which you build these and also how you actually can keep these things updated and think about the right kind of data sources that you have. So those are the general benefits. And so what we see is this paradigm shift that emerges. So you go from static data to real time data, a significant upfront cost to faster, cheaper data gathering and analysis, PowerPoint slides to intuitive interfaces, stereotypes and lack of nuance to these diverse and inclusive panels. You have limited actionability beyond the specific data captured a year ago or two years ago to easy inference on topics that weren't covered in the study. And then you have this real ease of integration of digital data sources from your customers, From Amazon, from TikTok, whatever it might be, as you think about this. So we're actually going to do a poll now just to talk a little bit about which of these benefits are most compelling to you. So we have a, a poll running. You'll see on the right hand side of the screen. Take a vote. There are six different benefits. We'd love to see what people think are the most compelling Benefits and then we'll have a little discussion around and if anybody wants to drop in the chat why they think that's the most compelling benefits, that would be. That would be great. Okay, I see from what we're looking at so far, the most compelling benefit is real time data that evolves. Maybe anybody who voted on the real time data that evolves, if you could drop into the chat why you think that's the most compelling benefit? And then if anybody wants to comment on which of these benefits they think are not that relevant. Okay, so Kevin has a comment that he's interested in reducing the cost so that you can actually have smaller businesses take a bunch of segmentations. Yes. You can sort of democratize access to this. Scott is saying that the real time data is compelling. It's so useful when you've done initial segmentation. And I'd say, Kevin, to your point, actually, you know, we, we had some discussions with, with one client actually around. They, they were looking at doing a new segmentation and we were actually discussing the possibility of a much cheaper approach to getting a new segmentation rather than doing a big expensive study, actually just using the Amazon API and actually developing the segmentation just from Amazon data alone. So basically integrating and reading Amazon reviews from their product category and using that to actually develop segmentation as a starting point rather than doing a big expensive study. And I imagine that could be an interesting way in for smaller businesses. Alessandra says that, you know, with the current pace of change, when you end a segmentation, your data are already obsolete. Did we put the poll out to synthetic agents as well? That is a really good question. We should have done what we didn't. But actually, you know what we will do? We will actually do that and share that as a LinkedIn post to see what the relative results are. I love that suggestion. Picking up on the SEI do gap. Very interesting question, Alistair. We'll get to that in a second actually. And then just, yes, a few other sentiments around sort of the importance of the diverse views and making sure that that's not neglected as it is, as our fun is. Awesome. Okay, cool. So interesting, yes. General sense of like agreement around maybe which are the most compelling. And so we'll get into some more like, specifics around how this works. Okay, so in terms of the actual process, there was a question around this before. So the general steps of it would be, would be that somehow you would gather the data for this living audience. Right. And so that could be an existing segmentation, that could be a new segmentation. Or it could be a new type of data set. So for example, as I mentioned, we were talking to one client, they produce consumer durable devices and a lot of their sales are on Amazon. And they were actually looking at how could they actually create a segmentation just on the basis of all of the data they get from Amazon reviews around their product category. So actually Hackathead has used that as a non traditional data source for creating the segmentation. But imagine you have all that data, you have the actual underlying data from an existing segmentation and the PowerPoint slides, all that kind of stuff. So imagine you have that. If you don't have a segmentation, you would need to go through that process of creating one from that data. But then once you have those segments, you'd actually then identify from the data within those segments what are the factors that vary within those segments. And also identify the future data needs that you might have that might be that you would actually identify effectively from the underlying data in the segmentation what are the biggest variables in terms of people's attitudes, behaviors or background. And then, and that would actually to the question here around the confidence of it being representative, what you'd actually have to do is really analyze what are the different factors that actually vary within the segments in order to figure that out. And then you would actually basically go through a process of creating these AI synthetic agents that represent each of the segments. There are different approaches to doing this. There are some third party tools that do this. We found most success in actually working with clients in their own existing Azure environment. For example, using in most instances using OpenAI. Actually we basically use what's called RAG is a training approach with AI. We actually basically upload custom data sets and use that to create these agents. And then there's a series of prompts that are created to actually create the, to create the panels and identify the sort of the variants within those segments to create sort of 100 or so or 50 or so different Personas within each segment. You then go through a process of actually testing and validating and updating the audience panels based on querying it. So basically asking questions. So usually we would start out with creating a very simple kind of chatbot proof of concept interface. Right? We can just ask questions, get results, and then you can basically use that to update the Personas and the agents and actually figure out if they're working or not. This actually is a time too where you can actually start to assess that SEIDOU gap. There was a question around that before. So one thing we haven't really done this yet, but we see potential for, is that you could actually train the model to actually counteract the patterns you see in that seidou gap. So really what that would require is actually taking the predictions that you get from the model, taking the actual behavior you see in the world, and then also the research findings and figuring out what is the pattern of the sevugap and using that to train the model so that any predictions eliminate that pattern, encounter that pattern. And then the last piece and I'll get to just a couple of questions. So synthetic users, we have used them sometimes. There are others that we've used as well, but I think primarily we've used our clients own environments rather than a third party tool. And then the last step is to create a usable interface to what we might call queries or ways to create content out of these panels. And maybe I'll just skip to that first. So I just got some sort of example screen, screenshots. Really what you would actually do is get to, is get to the stage where you might have a way of actually conducting analyses of existing data. So you might be able to ask questions, you might be able to generate predictions. So once you've figured out the types of questions you might want to do, you'd have either a simple chatbot where you can actually just ask it a question, or you could actually have the development of specific features. In this example, these examples that we have here, it might actually be that you would actually use it to generate specific insights or needs or actually identify specific questions that it might develop options around. So somehow you create some kind of interface that people can use. There's a question from James around. Can anyone set this up easily without much expense? I think the simplest way would be that you could actually do this very manually using a like, using sort of like an OpenAI GPT tool. You could do that. The challenge would actually be in actually creating a real panel of lots of different types of synthetic agents for each segment. You could probably easily create a reasonable version of one segment, but it would be hard to go much beyond that just using a sort of open source tool like that. And then in terms of what you actually do with these over time, basically imagine that you have these living audiences set up. You would have real data from real people that would emerge over time. So whether you are maybe you do interviews or follow up surveys with real people that represent these segments, maybe there's external data sets you get, you can then use those to actually check the accuracy of your predictions. So let's say you did some research on topic X. You could actually get the model to predict what the findings would be for that same topic and then actually check and then cross validate those findings and say basically use the findings to actually using that kind of rag process I talked about to actually update the underlying data set and then improve the prediction basically by training it to get better, by saying you predicted this would be the response. Here was the actual response. Here's the difference that you should be thinking about. And so over time you can see that there's sort of like a narrowing of the gap between the prediction and the reality as, as you actually evolve and use this further. So there's a question here around what are the actual kind of use cases? So obviously it varies greatly by different types of industries. There's also a question here around how does this work in B2B versus B2C? So I'll get to those questions in a second, but maybe just sort of run through a few examples. Right, so first of all, so you could imagine that you could do things in a very real time way. So you could imagine saying a query that you could pose to this interface would be of the top things that are trending on TikTok this week, which are the most relevant for our prime consumer. And then can you come up with some ideas of what we could do about them across the product lifecycle? So maybe there are some ideas in terms of how you might come up with new product ideas from that. Maybe there are ideas around how you might come up with a marketing campaign, whatever it might be, but you can actually figure out from a real time situation that was never studied in the segmentation data, what is the different response that you might get from these different Personas, what's relevant? And then can you come up with an idea on the basis of that? Let's say you have some ideas for new products. They could be within a B2C context, B2B context, you could actually figure out which of those product is will be most relevant. So sort of a prioritization, but then also relying on the fact that you have lots of different Personas within, you have a panel of different Personas within that segment. You could actually figure out also which might be the most polarizing. So where would there be the highest level of disagreement within that segment? And actually where would be the, what would be the things around people around which people might disagree of different retention tactics, let's say, which are actually most likely to succeed? And then how would they actually need to be personalized to individual consumers or customers. So you could imagine a situation whereby you identify for a certain segment that you actually might. A certain type of message at a certain time is going to be most relevant, this type of segment. But maybe there are ways to actually personalize that message to different people within that segment. So same kind of tactic, different message that's slightly personalized. And so it would enable you to actually do all of those pieces at once, and not just the tactic, but how do you actually personalize it? But then let's not forget it's not just about like a chatbot, right? So you could actually do other things. If you think about AI actually working in the process of, you know, generating and analyzing text and videos and images all at the same time, you could imagine, you know, here's a photo of a shelf of our products in a store alongside a competitor. You know, how do these products stack up in the eyes of our consumer or our customer? And what could we do to stand out? So it could actually analyze whatever you might input and provide some kind of prediction around it. And so just take a pause there and come to a couple of the questions. Okay, so there's a question around how confident are you with the living audience segmentation being representative of the market? I think it really. It comes down to really the quality of the data being fed in. So I think if you're. If you're confident that the actual segmentation study itself that was fed in was representative, then I think there's no reason to assume that the audiences wouldn't be. Where you might have challenges is if you, if, for example, I mentioned the example of, like, pulling in and relying on Amazon data or TikTok data, if that underlying data set is not representative, that would be where you have an issue rather than related to the sort of AI technology. Can you imagine a scenario in which, based on the data you accumulate, that the original segmentation becomes invalid? And how do you design for this? Yeah, I think that could actually be interesting and almost. I could imagine actually that what you might do is be able to test the ongoing validity of your segmentation and figure out when it's actually out of date. So you could actually have a feature you could develop in a tool that would actually identify when you need to replace your segmentation, in theory, because you could actually see when there's a growing gap between the data and the segmentation, and the data you're seeing outside in the world would be theoretically possible, I would imagine. How long does it take to create the Panel of Personas to the point where it's ready to use the first time. I would say the actual creation of the panel of the Personas, once you have the underlying data, I would say you could probably develop an initial usable thing within probably a sort of a week long, maybe two week time frame. But the real, the real goal is to actually validate that both the accuracy of the output and the relevance of that. So probably we're looking at sort of a few weeks of work to do that and have some rounds of feedback on that. How small can be a living audience. I mean, it could just be one agent per segment, obviously. But I think generally what we would imagine is you might look at sort of at least sort of 50 or 100 per segment in order to actually, to actually get the diversity of different types of people within that. Sort of thinking about the same way you might look at in terms of what would be quantitative scale within a cell, within a survey study. Yeah. And that granularity, the number of Personas you generate within the audience could be defined by the human. You could say, I want 50, I want a thousand, or whatever it might be. And then are there opportunities to use the Personas as an adversarial network to challenge and train the underlying model? Yeah, I think absolutely. I think absolutely you could do that. I guess the question would be how you actually sticker out the method of validating and figuring out what's true. I guess you'd have to have like a real, real world, real time data set in order to do that. But in theory, I think all of these things are possible and then you have to sort of validate and test that. A question from James around the components of the system and the platform we need to set up to achieve all of this. MVP setup. Yeah. So I think there's sort of a few levels of that. The simplest version is you could do this using Chat GPT and present GPT like that's the simplest version, as I mentioned. Typically what we'd assume though is that you basically create like the. The next level of that would be a chatbot interface where you use a rag method to create a simple chatbot within, within your company's closed OpenAI environment, for example, or Google environment. So it actually requires very little in terms of actual software development and engineering work. The actual more complex version of this that requires front end and back end development is actually generating an interface with features that actually help you to structure prompts and queries and generate specific outputs. So that's where it becomes More challenging. And then last question, I'll get to from Carolina before we move to the last step, what are the limitations of this method? As well, I would say like the, you know, the top three, I think that come to mind for me is obviously the quality of the data that you have going in. So there was a question around the representativeness of that. So if you don't have good data going in, obviously that's going to be a challenge for any kind of approach, even an existing segmentation. Number two is the quality of the validation of the outputs of this. So both in terms of accuracy and relevance. So are the predictions accurate? Are the predictions also relevant to the needs of the business and actually providing feedback to the model to actually improve that? And then I'd say that the third piece is really more about sort of how you actually operationalize this within a company. So it's great to have an amazing tool, a chat bot or a website where there is. But the real limitation, just as with, similar to how you would with a, you have limitations around operationalizing a PowerPoint deck in an organization, is that it's not just about understanding the technology, but it's also understanding how do you actually get people to use this differently and integrate this into their ways of working. And so I think the limitation is really more about forgetting that it's not just about the tool and the technology is about how do you actually make use of it in your business. And so in terms of where this could go next, I've talked a little bit about some examples of use cases we started to explore with clients and what we've done so far. I think we're seeing a few other ways this could evolve over time. And we'd be interested in if folks have other ideas to throw into living audiences. We've talked about how they could be consumers or customers. They could equally represent suppliers or competitors or buyers or chefs. If you're in the food business or regulators, whoever it might be, basically a human that might have an input on your business. So I think it's interesting to think about how you sort of build infrastructure of different types of people that you could represent. They could also be used to simulate market outcomes that involve different people. So it might not be that you're just testing and simulating what happens from one person, but how does what one person actually do influence what another person does? So we have actually done some exercises around what we call synthetic war gaming. So if we do X, what's the response of our competitor? And then what might a customer do from that. So how do you actually start to think about second order effects as well? Equally, I would think about what are the kind of integrations that are possible. So could you actually integrate this with personalization tools or CRMs to actually generate personalized experiences autonomously? Right. So you could imagine that actually you could manage this, to integrate this with your CRM to develop, customize email campaigns or whatever it might be, and also enable the feedback loop from the open rate on those emails, for example, to actually offer some kind of like real time retraining and validation of the model and the predictions. And so we see it evolving in lots of interesting ways, but it's still obviously an early, early process. We're about sort of six months into this in terms of like deploying this with our own clients, so still a lot of questions to be figured out. Now we got a little bit of time for Q and A, so feel free to put any questions in the chat and I'll see if there's anything else that I haven't addressed yet. Okay, so there's a question from Karen. If you only have certain times in a year when the audience changes, like attending events, would it be important to then look at the common parts of the segments related to parts of the event program that do not change? So we then have an audience segment that is relevant over different events? Yeah, I think that would be interesting to think about. What are the components of the. The segmentation that might be variable and might not change to. We've also seen with some clients they would actually conduct sort of quarterly events where they actually have an intake of data as well, which is, which is quite interesting. A comment from Eric that is more of an art than a science. Absolutely. And it's also about like how do you actually use it as people rather than just the sort of the technology itself. Any other questions that people want to throw in? Is this tool more useful for the early stages of. I think that might be a typo of sedation or for validating and prioritizing existing ideas. I'll assume maybe sort of like early stages of understanding needs and insights. I think that we found it can be, it can be used for both equally. I think you just have to be clear around what's your level of expectation. So it's not like the prediction you can get from this is going to be as good as doing an actual real life one hour interview with a person. But it's going to give you the broad strokes, same type of insights, just maybe not at the same level of Depth. The GPT model is based on probability. Sometimes in UX research we look for unique and non obvious things. The improbable can generate nice ideas as well. How do you address that? Yeah, it's a great question. I actually think there are sort of two ways. One is actually the way in which you prompt for those things. You can actually deliberately prompt and ask for none of those things. On the technical side, there are things you can actually do around what's called changing the temperature of the response within the model, which gives you. It basically gets you less accurate but more creative and non obvious answers. You can actually program for that. You can actually sort of. You can almost have a button that actually could be pressed to do that. Within each segment the agent profiles are kept unique. Yes. So to be clear, the basically you'd have an agent for each. So a panel would be made of lots of different agents. Yes, that is correct. And there's a question from David. Will there be a recording? Yes, there will be a recording link sent out to everyone who registered. Have you found anything unexpected about how internal customers are starting to use these new chatbots? I'd say, I'd say what we've found is that it enables you to do more than just ask questions. I think typically what we've found is that when you have a segmentation, people will reach out to the team that will be managing that segmentation and ask them questions like do we know this? Do we know that? What we found is actually most interesting is where you can kind of make multiple different types of predictions about different questions and then figure out the implication of those two things together. So which would be, let's say, one of the use cases we talked about earlier, which of these product ideas was most relevant and which would be most polarizing? If you ask those two different questions and then figure out what's the implication of those two things together, that actually helps you to get to a deeper understanding of which product you should prioritize rather than just asking one question at a time. Okay, in the interest of time, we're going to wrap up in the next few minutes, I guess how to get involved with this further. Obviously if you are somebody who is a buyer of segmentation, you have a segmentation you're interested in, how could you have another one or a new one? We'd love to get in touch and see how we can actually identify what would be the use cases for you and build a proof of concept. If you're actually a research and service provider, I saw on the signup list. There's a bunch of folks from the market research industry. Obviously companies that actually conduct all these segmentations would love to see if you'd be interested in actually approaching your clients together and also seeing how we could actually help you to reinvent your business model. So obviously there's a different kind of revenue model that could be associated with this that actually offers more of a sort of an ongoing recurring revenue stream. We'd be interested to talk to you about that too. If you're just an interested innovator, please follow us on LinkedIn and also just you can learn more about the upcoming Autonomous Innovation Summit. So this actually is something that Board of Innovation is running on the 5th and 6th of June. So we have about 40 speakers across a range of different topics around how the autonomous age is transforming the way that companies think about innovation. Sina has just put a link into the chat here around signing up for the summit. It'll be a great event. We had about 10,000 folks register for the last one in December. We've got very interesting speakers across different topics, everything from synthetic testing to how to actually reimagine sustainability through the lens of AI, and would be very interested to see folks join for that as well.