Rob Key of Converseon: Making Unstructured Data AI-Ready

As AI models become more powerful and increasingly commoditized, the real competitive advantage may lie in the quality and context of the data feeding them. In this episode of the Greenbook CEO Series, Lenny Murphy speaks with Rob Key, founder and CEO of Converseon, about the growing importance of structuring unstructured data for AI and research.

Rob explains why traditional sentiment analysis falls short, how context engineering can improve accuracy while reducing AI costs, and why brands need stronger governance around the data flowing into AI agents. They also explore the risks of AI-generated content contaminating future datasets, the evolving role of social and consumer intelligence platforms, and why domain-specific models could become critical infrastructure for enterprise AI.

Key Discussion Points

• Why data quality matters more in the GenAI era: Rob explains how small inaccuracies in unstructured data can become significantly amplified when AI systems consume and reason across that information.
• Moving beyond basic sentiment analysis: The conversation explores entity- and aspect-level sentiment, attitudes, intensity, trust, and other dimensions that can provide a much richer view of consumer opinion.
• Context engineering as AI infrastructure: Rob argues that organizations need to structure data according to their own brands, industries, language, and definitions before passing it into general-purpose AI agents.
• The emerging AI ecosystem: Lenny and Rob discuss a future where frontier models provide foundational intelligence while smaller, purpose-built models, platforms, and agents create differentiation through specialized context and workflows.
• AI-generated content, trust, and ethics: They examine the growing challenge of AI-generated content training other AI systems, the importance of distinguishing meaningful human expression from automated noise, and the responsibility organizations have when influencing AI-generated perceptions.

Resources and Links

• Converseon

You can reach out to Rob Key on LinkedIn

Many thanks to Rob Key for being our guest. Thanks also to our production team and our editor at Big Bad Audio.

Transcript

[Lenny] Hello everybody. And welcome to another edition of the CEO series. I am your host, Lenny Murphy. And today I'm joined by Rob Key, founder and CEO of Converseon. We have been trying to do this for, well, I think years, Rob. Right. So, um, so thanks for finally helping to make it happen.

[Rob] No, thanks for the invitation. I always enjoy our chats, even when they're quick and in passing. So glad to sit with you for a little more time.

[Lenny] Yeah, I, the pleasure is mine and our listeners. So I think that you guys are going to get a lot out of this. So Rob, for those who don't know you, why don't you give a little bit of your origin story and then we can dive into the fun stuff.

[Rob] Yeah, for sure. Well, I mean, Converseon, for those of you who don't know us, we actually transform unstructured data. So we work primarily with social media, voice of customer, other forms of texts primarily. Uh, we structure it for accuracy and meaning. And in this day and age, we transform it for AI and research readiness, right? Because unstructured data is so messy.

90% of all the world's data is unstructured. Less than half of it is actually usable because it's so messy and complex. And so we do that critical structuring of data. How this started though, it's a very winding kind of road. You know, I'd say my origins were, I was starting Converseon, we were, I came out of the world of WPP. I was on the WPP.com board. I was head of the innovations group at one of the communications agencies there. And Converseon started as a little bit of an agency. So we were really focused back in the early days of, of, uh, it, this Shell Holtz said to us, um, that, um, we were probably the first social media agency.

So, you know, at that point in time, we won lots of awards, but in 2008, you know, we realized that we needed a better way to kind of listen to what the conversations were that were happening in the marketplace. And we happened to come across two wonderful, amazing researchers, rock stars in AI, uh, Dr. Phillip Resnick at the University of Maryland and Dr. Jason Baldridge who's at UT Austin. Now he's with Google and they helped us in 2008, believe it or not, build our first machine learning model. Uh, we had a lot of training data, humans sitting around coding things and they realized, wow, we have this massive, incredible resources to start to train models on. And so we were one of the pioneers in doing machine learning on this data. Um, and we, through the years we've kind of really launched in many ways, the great innovations in our space.

And that included back around 2014, being able to, uh, introduce not just sentiment analysis, but things like kind of emotional analysis in this data. And back in 2019, we started to work with our first, uh, transformer models, so, you know, early forms of LLMs. And today I humbly and modestly think that we are the leader in, uh, doing advanced natural language processing and structuring of this data context engineering. So that the world of AI works effectively. And I would just simply say, you know, it sounds like, you know, a lot of transformation over the time. And I think one of the reasons we've been successful is that, you know, we were largely bootstrapped in many ways. So we've been able to kind of adjust with what's happening in the marketplace and be very agile.

But the one true kind of North star that we've had since the very beginning is about how do you find accuracy and truth in this unstructured data set? The techniques have changed. The algorithms have changed. The approaches have changed, but that has been our mission for all those years. And for many years, I would say we were a little bit in the wilderness. You know, people were like, Oh, you have these conversations about what is your F1 score?

You're at a 6.5 F1 score. Well, you know, good enough is maybe good enough. I mean, it was kind of the work, even though we had, you know, Forrester, right. This is, uh, you know, a leader and Gartner group was giving us recognitions, but, you know, I think we were a little ahead of our time. With the advent of gen AI, of course, like that's changed everything. Where small little errors and data sets become extrapolated into massive problems. Instead of the garbage in garbage out, you know, it's garbage in, chaos out. And so at this point in time, you know, at this point in time, um, we work with a lot of leading platforms and others to help structure the data pre gen AI consumption. Uh, but also for research like gen AI is kind of where it's going, but you still need this really good data for insights or predictive insights and other things as well.

So, um, yeah, that's been our journey to date.

[Lenny] It’s a hell of a journey. Um, and the longevity proves the, uh, proves the validity of the mission in my mind. So, right. You have been focused on those challenges and the market has kind of caught up with you, uh, to an extent, because I know the past few years, you've been very active and focusing on data quality initiatives around unstructured data, because to your point, I hadn't known how to put that way that it's not garbage. It's garbage in, chaos out. I love that.

Um, because absolutely the risk of contamination of bad data now is infinitely bigger than it ever was before. Um, so let's talk a little bit about that effort, how you participate to like, look, we need to focus on the highest quality data period, regardless of source, right?

[Rob] Yeah. So that all came about in part because, you know, we've long, again, we've long had this mission that we have kind of a moral right and obligation of any way to capture the express opinions of wants, needs, hopes, et cetera, in a clear way. Right.

So that's, and it's also to be clear that was a differentiator and it remains a differentiator for us in the marketplace. Um, the average sentiment analysis, for example, standard document level sentiment analysis, that's pervasive is still about 65% accurate. That is a technique that goes back to 2002 and in 27 years, it hasn't really morphed a lot. And it's not that the models are actually wrong. It’s asking the wrong question. It looks at an entire document and it basically makes a judgment about the entire document that is this positive, negative, or neutral.

Those aren't the questions that people are asking. You know, if I'm a large CPG brand, I want to know, you know, what is the express opinion towards me and not just me, but what aspect is about price or taste or food or what, you know, so that's where the market has been. So, you know, the global data quality initiative, I got involved in that, I think about three years ago, I'm a director in an organization called AMAC, um, which is really kind of sets measurement standards for the communications professional world.

And so as the global data quality initiative is really rightly focused on all the problems in survey based worlds. You know, what was missing was the 90% of data that actually exists. And increasingly important, you know, the fact that we have these real strain, real time, uh, streaming data of what people are saying about you, whether, you know, it's X or Reddit or other forums that, you know, this is increasingly important, but the respondent rates, I don't need to talk to your audience about what the challenges are there, but there is kind of a greater need for, you know, at more passive, but, uh, access to unstructured data and opinions.

But the state of the market was a problem. So anyhow, I joined up with them and, um, along the way we have developed a set of principles based on interpreting kind of the global data quality principles but for the world of unstructured data. And we released them last year. You know, there’s seven core principles and we can run through what some of them are, but it really boils down to just a few things. One is number one is, you know, you have to take you as an organization need to make a statement that this is important to you that, you know, we all can say in the abstract, yes, of course, accuracy is important, but to make it a primary requirement in a world of competing needs is really important. The second point is you can't manage what you can't measure, right?

You can't. So, you know, the problem with that is if you ask a lot of folks who are consuming this data today, how accurate their data is, they'll give you blank stares. They have no idea of what it looks like. Well, it kind of looks good, but I don't really know. Well, that's not good enough. So, you know, instilling whatever form that in an end to end process so that you can evaluate the performance of this data across often it's an F1 score, which is precision and recall, and there's other scoring methodologies, but just installing whatever form that is that you actually start to do that.

Um, and then part three is transparency. Um, you know, we look at kind of four criteria in the world of kind of social intelligence, a little bit of, if you're doing a good job, one is like, is the data from the people who you think it's actually from, right? And that's not an easy question to ask with bots and all types of other forms of data. We know within certain conversations, like 20% of the conversation is from humans. The rest of it is just, you know, noise and manipulation. Yeah.

Part two is, you know, once you have that data, is the metadata associated with it actually reflecting the underlying opinions? So that's your sentiment analysis, but not just sentiment these days, you know, attitudes, intensity, all kinds of different ways you can dimensionalize this data, if you're still stuck at document level sentiment, like you're way behind. Number three, though, does this data have any type of anchoring into the real world? Right. And that's because so much of what's happened in the world of social intelligence and media has often been vanity metrics, you know, so that, you know, your sentiment would have 5%. What, what does that really mean? And so there’s actually really interesting work that we're doing to kind of doing econometric modeling on top of this data so that we can say, and this is a real example for a big QSR brand, instead of telling them that their sentiment went up 5% last month, we can go in and now say to them, because of this clean data and AI and other things that if you were to improve your perceptions in these three areas of the environment by X, Y, Z, you will with 85% confidence generate $90 million of incremental revenue, quarterly.

We’re really tying the behavioral data to the perceptual data so that we can see that there's an anchoring in the world. And the fourth piece of this though, is even if you have the best data quality in the world, and if it's not trusted by your stakeholders, you're not going to make any progress. So this is where, you know, what we do is, you know, for all models that are being processed, there's a governance, uh, ability, you can do kind of sampling to understand the exact precision and recall and accuracy scores and area under the curve and all these other metrics. So you know exactly how well your data is performing before you consume it. And, you know, this is the framework and, you know, we've introduced it to, you know, the world of corporate communications and chief communications officers and, you know, and others in market research, I'm on the steering committee for the social intelligence lab, which is trying to advance capabilities in this area and we're making progress. Um, but, uh, you know, we have a lot of work to do.

[Lenny] So, you know, it's, uh, when you and I first connected and I was part of Dakota, I remember, point being, we have both been in this game, you deep, me on the periphery, for quite some time and seeing, you know, these changes happening. I remember a time for myself. Well, about when AI started to really emerge when generative AI and LLMs, thinking, well, gosh, what happens to the legacy companies that have built, you know, around text analytics and all of its, you know, broad permutations.

Um, and my first thought was, oh crap, you know, I don't know if this is good, right. They may just get swallowed up. Um, then after following you, right. And a lot of the work that you've been doing around this refocusing on the idea of the quality, recognize that no, actually this is vital foundational currency because trust is the currency now. It always has been, but it really is now. Uh, and you know, being the metric to understand the, you know, trust of content of a corpus is invaluable.

So that's my view. Is the market embracing that now they realize, oh wait, it's not yet. Sure. Can we analyze unstructured data at scale really quickly with AI? Yes, but that was never the game. It wasn't about the ability to analyze the content. I mean, there's a time where it was, but it's about, is it done right? And then how is it used? You know, are you hearing that in the conversations?

[Rob] Yes, I mean, what I would say is we see a few confluence of things happening. We have, we work with a lot of big enterprise organizations. There is a big drive to use AI agents, Claude co-work and others. Obviously we all, we all know that, right. That's happening for lots and lots of reasons, but you know, there's at the same time, you know, there is this recognition of the risk associated with using these things in the raw. And if you take another step back, the traditional SAAS platforms who have kind of played that job, the media monitoring plan platforms and the social listening platforms of which we have deep integrations with the majority of those leaders, um, they're kind of scrambling to see what do we look like in this world too?

Right. Because we know that the future is AI agents, right. But how does, how are those configured in a way that actually are workable? And we actually just did a study and I can release it to your readers and release it to you that, you know, everybody talks about the world context engineering, right. I think Gartner said context is like the word of the year and it's become, you know, it's getting overdone. It was like, how it was with AI a month ago. It's like, everybody's a context engineer.

[Lenny] Context and inference, right? Those are the words, right.

[Rob] So, but the real challenge, and I think the sophisticated brands get this is like, you know, Gen AI is obviously as good as the data that it works from, you know, Gen AI understands kind of broad concepts. It does not understand the structured meaning of your particular organization and the way that you define and you see the world, which is kind of unique, right? It's a generalist and the linguist would say it doesn't understand language at all. It's purely statistical. So there's a lot of, in the world of computational linguists, like they don't love these LLMs for some very good reasons, right? They're a little bit of a different animal. So what we see happening is this really kind of interesting new kind of ecosystem emerge. So what we did is we did a study and we took the 4 million records of like the airline industry and food delivery, and we took the same data set and we ran it through, we gave it to Claude, Fable 5, and said, you natively enriched this data, you know, for sentiment and all these other things. Like, what does it look like? Right. We actually then took what we call a standard enrichment, which is the standard classifications that come out of your traditional media or social listening platform, which is like document level sentiment, which we were talking about earlier, and then fed it to them and said, let's look at the results.

And then we did a third level and we said, let's do an advanced enrichment. Let's structure this data specific to the definitions through advanced natural language processing to the specific brand and industry, because language differs by domain, you know, small is good for smartphones, bad for hotel rooms, and let's structure it. And we're not going to just do sentiment, we're going to do, you know, sentiment, the sentiment will be entity aspects are very, very specific. And then, but also attitudes. So problem with sentiment analysis is that, you know, disappointment, frustration, and outrage are all negative, but they're different. So, you know, that we now have the ability to classify for attitude. So there's 15 different attitudes and let's look at intensity of how strong the expression is and trust, which is very different than sentiment. The who, what, when, where, why. And we then put it into Claude. And we said, let's calculate what happens downstream and what we found, you know, probably not a huge surprise in many ways that, you know, the advanced enrichment is 92% accuracy across the human gold standard. The native enrichment of Claude was about 65%. Importantly, it also could not figure out things like trust and attitudes.

Like it just, you know, it was kind of invisible in the dataset. And the third thing that's really interesting was that because it was using a cleaner dataset, that the inference, there’s the other word, inference and token costs were reduced by 90% at the end. So what you're starting to see intelligent conversation is, you know, you don't, somebody said the other day, you know, you don't drive an F1 car to your office every day, right? You don't use an… an agent for this type of thing. What you need to do is figure out what's the model size for what you're trying to do. So we use small language models, purpose-built transformers understand slang, sarcasm, context, do all that enrichment. And then MCPs kind of connect into these AI agents. And then they do the stuff that they're magical at. And then you have what I see is going to be the ecosystem of the future in our particular space 

The SAS platforms then are kind of like, where do we fit? Well, because you also see things like X have their own MCPs and, you know, people were saying, well, just go direct. The problem is MCPs are dumb pipes. What matters is the context that flows through those pipes. So what happens when you structure upstream of those MCPs is what really matters. And at the end of the day, the best context wins whether it's, and it's probably not going to happen at the X level. You know, you can't really go into X and kind of structure that data to my precise meaning and then get the data through. Right. It could happen downstream at my data lake level for the big brands that are building data lakes and they have ontologies and knowledge graphs and stuff, and you could do some structuring down there, but you still have to tame the beast of all this data and do everything else you need to do to prepare the prep for it to do it. So I see actually a big opportunity for the current platforms, whether it's like decisions, meltwaters, brandwatch all those, like we work with them to be context as a service, to really play that critical role of structuring this data to your specific meaning, to your specific accuracy, to tell the accuracy gateway so that this data can efficiently be transferred into an AI agent. So it works well.

And that's a really exciting vision, right. And, you know, I see all the time, people are like, I just go to agents now, well, you know, first of all, the cost is outrageous. Because every time you prompt it, by the way, it has to recreate that classification that had before, which will probably not be the same as what you had previously when you do it upstream, they’re deterministic. So when it gets there, you can repurpose that all the time. And so the costs are massive costs of reductions. And, um, so I think that's what it looks like in this day and age, and I think for the foreseeable future.

[Lenny] I agree. I mean, even as we’re recording this on August 13th, right. The last few weeks, the change in the broader tech narrative, I would say something I couldn't imagine saying a year ago, the intelligence is commodity now, the models, the big frontier models, they're all roughly a parody. Um, and they may, you know, they're, they're still in the arms race. There may still be some leapfrogging from a capability standpoint, but the duration between that is shorter and shorter and the difference between the models, especially when you throw in the open source models that are coming out, is negligible. So they're all operating at roughly the same level at this point. So the economics of deploying Fable versus Grok 4.6, why the hell would I pay for Fable when Grok 4.6 is so much cheaper, or why would I pay for Grok 4.6 when I can use Kimi, right? So that, I think the entire shift is recognizing there's a foundational layer of intelligence. Fantastic. They're all operating at basically the same level.

That's not the alpha. That's not the differentiator. You know, it is going to be the use cases. And that's where agents come in. And you're right that I hadn't thought about that way of, you know, the SAAS platforms, all the platforms that I like that idea of, you know, context as a service, I would even say that effectively that's context as an agent, fitting within a workflow, so doing the pieces that they do within the stack to optimize the outputs. And I think that cost is going to continue to come down. It's just going to shift infrastructure. So just compute, right? That's where all the money is going to flow because that's what's necessary while everything else gets somewhat commoditized, but not the expertise of a model like you have built that really does create valuable impact, you know, rather than the general use case stuff. So I think we're going to see a series of connections from different platforms. Yeah. We'll use that analogy for lack of a better term, you know, where they're connecting as part of a workflow to take data, make it useful, now apply it around the business issue and it's still up for grabs on who owns the chunks of that workflow, right?

And that's still, there's still lots of opportunity there. There's no clean winner within that. Um, but again, a few weeks ago, I would not have bet against it being Anthropic or OpenAI or SpaceX owning the big layer, now I will bet against it, I don't think they do. I think they own the foundation. But the value contributions come from companies like you and others that are really building off of that to deliver what's really necessary is how does it help to make a decision? So, and what is the value of that decision? Cause that's what the brand's most interested in. The buyer isn't interested in all the pipes in the flow. This one is the product. Does the product give me competitive advantage? Right. Period. So I don't know. What do you think? Is that, am I?

[Rob] No, I think, I think you're right. I mean, you know, the challenges we hear about token maxing and, you know, and everything else associated with it are serious. And I think that, you know, that we're getting to this, this stage where you're, we're going to see the right sizing of the intelligence needs with the right sizing of the model fit for what they do, and there's going to be a lot of room underneath, you know, the Anthropics and OpenAIs of the world for more domain specific models that are really smart, but very purpose built for what they do and I think, and how do they get configured together becomes, becomes really interesting. And I think that, yeah, there's a lot of opportunities there.

I did a presentation not long ago for a bunch of chief communications officers and the title of it was from the manufacturing of consent, which was kind of an old term within communications to the manufacturing of context. And the idea being that who defines meaning for your organization is up for grabs within an organization, right? You have folks in BI, you were doing the ontologies and stuff. They kind of want to have it. You might have a chief data officer here overdoing it, then like it's up for grabs. But I would argue that, you know, the communications function in many ways should own it. They're the ones dealing on the front lines of this data all the time. They're imbued in language. They're kind of, in terms of nuance, they understand issues and how it imports the company. And so how do you, how do they become more involved in structuring the meaning of an organization and, you know, becoming kind of a meaning architect? And you see a lot of this work where they're trying to influence how AI talks about a company and other things as well. There's a dark side to all of this, by the way, which is, you know, and it's incentives to kind of, you know, there's good manipulation for precision, there's bad manipulation for promotion.

And, you know, how do we have the ethical guidelines of how to approach this the right way? And understanding that every piece of content that you were creating is training data for a model. And when you're doing that, by the way, that comes with a lot of responsibility. If you really want to look at things like the AIEU Act, if you're contributing to the training of the model, it is about precision, it is about removing bias. It is, well, communications and marketing in many ways.

[Lenny] And provenance, improving provenance, which is increasingly a big deal.

[Rob] Yes. Absolutely. And don't forget that, you know, in many ways, communications and marketing inherent in what we do, and I throw myself in there, I was from that world, is their job is to instill bias into the market towards your products, right? So there's going to be push and pull happening here. Um, and I'm not sure exactly how it's going to play out. I do really have concern, and I think we're seeing it right now about, uh, this AI slop that's training other AI, it's creating more AI slop in this kind of endless cycle, um, that can, you know, if we don't intercede, uh, soon, um, you know, we're going to have real problems.

And that's why, again, data quality really matters when you're structuring your data, when you're listening to your data, when you're analyzing your data, you’ve got to use AI to be able to cut through the stuff that really matters and is meaningful or you will be kind of caught in this kind of swirling mess of meaninglessness that you can't trust. And it's been interesting to see though, recently, I know LinkedIn has done this and Reddit, you know, and you see even what the, with anthropic this week, kind of watermarking some of their content with the idea that we can start to really identify what is human created versus what AI is created.

Um, you know, Reddit has cleaned up a lot of AI slop in their data. Um, and you know, LinkedIn, you can, there's also helping to identify it. So that's going to be a big, big move here because these signals only work if number one, if people can express their opinions as they want to express, which is a whole other matter, but once the, but you know, that we're actually capturing, you know, the real human voice, uh, in the noise.

[Lenny] So that is, I had not thought about that till right this second. So I obviously, uh, anybody who follows me you know, particularly on the substack, know that I am a copious user of AI from a production standpoint. My personal view has been, I am the orchestrator, I'm the originator. It's my thinking, it's my context, it's my et cetera, et cetera. And I really don't give a shit about the words. So that's just a delivery mechanism. So my point is I use AI to write a lot. It's a better writer than I am. It's not nearly, you know, it's, it's still painful.

What I want to do is to get my thinking out. Now, here's the point. I'm also keenly aware that it is devoid of emotion. So the, now this is business writing, so maybe that's okay. There's not much personality to it. I mean, maybe it's somewhat trained off of my previous years of writing, but if the context is trying to understand my feelings about things, not going to get that. So you'll get my thoughts on something. But not my feelings. I didn't think about until right this minute is, as you said that that's, there's a trade-off there.

Um, and I struggle with it. My kids, my teenage kids hate any AI generated content. They won't even touch it. Now that's wonderful because they're still in school and I don't want them to depend on AI. I want them to create their own content and do those things themselves. However, and sorry, I'm trying to form this in my head as we're talking, Rob, where does the balance get to? Uh, cause I reject the idea that, and maybe this is totally self-serving, that what I generate from personally is AI slop. Bullshit.

No, it's not, it is my thoughts, my contexts, my inspiration, my orchestration, et cetera, et cetera. And AI just generates the words because it's faster and easier. Fine. So, but the core, it’s not slop. So I don't like that bucket terminology. However, to your point, the recursion issue and the lack of depth from an emotional standpoint, as if more people think like I do around that, that can become a real challenge. Um, and I hadn't thought about it. So what you can say, Lenny, you're, you know, it's AI slop, man. You need to quit that crap and whatever.

[Rob] Well, look, I would just say what you just described isn't kind of an AI slop. It's like having a really good ghostwriter sitting along with you and helping you just kind of, you know, express your opinions, you know, and structuring it in a way. And I, you know, I use it a lot too. I mean, for sure. Um, but I do think that there is a danger that, you know, we, as people do, if this is broadly adopted, we lose our unique expressions that, um, you know, come across in a way that, you know, systems one systems two, kind of that emotional kind of like structuring within data. Now we find, we have, as I mentioned, we have this attitude classification, so we can pick up on frustration or laudatory or disappointment and others. And so we can pre-filter data to clean out some of the extraneous noise and then really kind of isolate that stuff. But that stuff is I think increasingly important, you know?

Um, and you know, I haven't seen any decline kind of when we look at the classifications that we're finding less structured attitude in data. It'd be an interesting kind of exploration to tell you the truth, like to kind of start to apply these models to more specific and see kind of what we are losing kind of in translation. But for now, I think, you know, what we just really need to do is kind of hone in on getting the obvious slop out, right. And kind of delineating it from, you know, the assisted communication slop. Yeah. Because it's meaningful, like, you know, I mean, let's be real, like your opinion and what you say has influence and has an impact in the industry.

And, you know, that is important to capture that, right. For sure. But, you know, I think that, and I guess I go back to why we go back to why organizations have such responsibility when they're, you know, trying to either influence AI perceptions or other things as well, that they focus on this concept of precision versus promotion. That you were really trying to, you know, understand, make the world more understandable and more clear and create, you know, greater intelligence in data and not just creating more noise with the idea that, you know, somehow that will kind of rise up to the top because we saw this all with the world of search engine optimization in the early days is that, you know, people were creating doorway pages and doing dark arts and doing spamming and Google got smart and they started to chase the algorithm and then BMW got thrown out of Google at one point for these techniques. And so we haven't quite seen that happen in this world of AI, but I think it's coming. We will have that moment where organizations overstep what is considered ethical and you know, there will be pushback and, and then we'll see where, where it goes from there, but I'm waiting for that moments for the wall street journal to say one day, you know, the communications industry is, you know, manipulating AI for X, Y, Z, right.

[Lenny] Of course we are now it's, you know, AEO. So, I mean, of course that's already happening, especially when there actually was an article in wall street journal today about publishers pushing back on Google and realizing, you know, the game has changed. We're feeding the beast now that is actually disintermediating us. So they're not referring traffic to us anymore. That was a great synergistic relationship, right? We're going to optimize our content and Google's going to send traffic to us. You know, that's a win-win. Now it's we're optimizing our content and we don't even get the traffic. So Google's just, so there's a real tension there. Uh, that I think we're already down that path where the role of AI both as a content generation tool, but also as a disintermediation to the content itself becomes a real challenge for the ecosystem, for the web ecosystem entirely.

Especially a company like green book, we're a publisher. We sell access to eyeballs, right? That is the fundamental business model of every publisher in the world is traffic. And when you do all this work to create good quality content, but you're not getting a payoff out of it. That creates a sign to pay the piper, right? We have to find a way through that overall. So I think we're already there.

[Rob] Yeah, I agree with that. And you know, we're speaking about, you know, all the challenges and what may happen. I mean, the only thing that I would just emphasize on that, you know, by going through this technique of, you know, cleaning, cleansing, making accurate, classifying, you know, with breadth for multidimensional analysis and not these kind of flat, shallow sentiment, you know, applying economic econometric modeling and doing other things that you can do, you know, the power of what happens downstream is pretty mind blowing. I mean, you know, we, so whether it's trend detection of, you know, or it's, you know, it's reputation intelligence or brand intelligence, you know, we're running a, we're doing a big program for one of the biggest CPG companies who, you know, they're running a, their traditional tracker and they want to know, they want to answer why are these scores changing?

Right. And so we can answer that. And by the way, this data is, if you get it there it can be predictive. So it can kind of predict the outcome of your tracker before the tracker is there. So, you know, there's incredible value in that particular space. And when you mash that up with the power of a Claude or a co-pilot with this data, you know, my mind gets blown. I mean, I see things for one of the world's biggest, you know, international finance companies. Um, we were able to put the data into a Claude instance and they were able to identify a false narrative and the source and the information in seconds, patterns that a human being would not be able to see. Yeah. Um, and so that combination is massively important. I mean, we're going through kind of a messy middle right now, you know, where, you know, these policies have to kind of evolve, you know, organizations need to kind of change the way they talk about these things, you know, these things like accuracy and, you know, context structuring, you know, may seem foreign to kind of what people do, but these have to become mainstreamed in the organizations.

You know, the agents themselves, I think have to do more. I would like to see them do more too, and doing what Google did in the early days of search, and they kind of created a structure to be able to engage with the community more about the kind of data they would like to see and things you should be doing, and so that we can retrieve things better and we can make things more precise. There has to be a conversation about this stuff. There has to be an adoption of principles. But what I see happening on the other end of this stuff is incredible, you know, and I don't have to echo it here. You said all the time, things that would take weeks are now hours, sometimes minutes.

And you know, we're going to see the most exciting. I've been, yeah, I've been doing this for a long time. This is the most exciting transformation that I've seen. And we're only partway through it.

[Lenny] Yeah. Yeah. Well, and happening so quickly, right? Yeah. I mean, the pace of, of change is breathtaking. But yeah, for you, I'm sure that it's a piece of the, you're foundational, right? So it's like the world is changing really rapidly around this, but you’re in the catbird seat, man. Like, Hey, we're in the business of making data useful. And that's not going to change. It's only going to get better.

[Rob] I mean, look, I had a couple existentialist moments too, just like you did, you know, when the first emergence of open AI was and said, what does this really mean to us? You know? And, you know, and as we've gotten deeper into it, you know, the, again, what I realized is that the type of custom NLP and structuring that you do with this data is infrastructure for those. And so I don't, I think it's more important than it's ever been in the past. And, you know, I think the legacy models are kind of struggling to kind of keep up. I think Scott Brinker, I don't know if you know, he’s written a lot about this space. I mean, I really think this concept of headless delivery, delivering the intelligence wherever it's needed using MCPs, using, you know, accuracy gateways in terms of model governance and is it, is really exciting. So how do these organizations kind of evolve into that? You know, so the SAAS platforms have to allow more context into their systems.

And I would say for our partners we work with, you know, they used to have like, you buy our sentiment analysis and this is what you get. They now allow, you know, for our models to be customized for their customers, you know, sophisticated customers within their environment. So it's all happening in there. That's the change for them, you know, like how do we create more intelligence or, or it's going to all happen downstream. And so I see some really positive moves in the right direction.

[Lenny] Yeah. Yeah. I agree. I want to be conscious of your time as well as the listeners. So there's another question that I want to bring up and we'll, we'll head here. So came from WPP, you're in New York, you know, you used to, you had a place over, left the Playland Studio and now you're wearing flip-flops and shorts in Costa Rica. So just what that's been like as you're in the middle of tech, right. Fundamentally, obviously you're incredibly well-versed right in the middle of that, but yet there's this whole other part of your life where, you know, you're living in the jungle, uh, or at least kind of the jungle. So give us a glimpse into the, uh, the strange bifurcation or dichotomy of Rob's life.

[Rob] Oh my goodness. Well, that's a long story that I wouldn't want to put anybody through. But, um, yeah, no, as you pointed out, I was fortunate to live in, uh, Jimi Hendrix, a little old apartment down behind Electric Lady Studio, the studio he bought in 1970 for a very long time. And, uh, the studio to this day is cranking out incredible, you know, music. But, um, you know, at the end of COVID, I did find myself longing for something a little different. So I do have a home in Costa Rica overlooking the ocean with holler monkeys in the morning. And, you know, a greenhouse and living off the land like you do. And I think that this combination is really important because I think that we get so sucked into what's happening. And I, like all of you, I get sucked down rabbit holes with whether it's Claude or these other things and how fast this is moving and, you know, what I find out is that when I'm there, I can, you know, get the perspective and clarity necessary for the wisdom on how to use this stuff.

Right. And I, you know, I think what I do, all my friends, you know, and I see all of many of them come down and visit and it's like, you know, three days of kind of clarity. Uh, you know, and, um, so yeah, I'm kind of, I'm kind of straddling both these days, but, um, I'm in New York right now. And people say, where are you? I live in the cloud, so I'm everywhere and nowhere. And, uh, yeah, and I'm fortunate. I feel very fortunate to be able to do that.

[Lenny] It's very cool. So you're the ghost in the machine, Rob, is that what you're saying?

[Rob] Some type of spirit or ghost or something in the machine. I'm happy to be, uh, I'm happy to be a subprocessor within other platforms. So I'm a subprocessor and that's what we are. Yeah.

[Lenny] Now we're going to get a simulation theory. We should probably stop. So at least, uh, at least recording, we can have that conversation offline. I'm so glad we finally made this happen. And I don't think it'll be the last time. Where can people find you?

[Rob] Oh, well, they can always go to the Converseon.com website and, you know, click through that. You can also reach out to me on LinkedIn. I'm pretty visible. My email is our [email protected]. Um, again, I'm pretty much everywhere and nowhere. So yeah, reach out anytime. I mean, I would encourage anybody who's interested in that study that was really kind of about how do we kind of value, you know, put real tangible numbers around the value of context enrichment for our particular space. Consumer insights, media and others. I'm happy to send the study.

We're also making it available on our LinkedIn and others too. It's the first in a series of, it's not to be all and end all conversation on this, but, you know, I felt like, you know, this lack of guidance of how people should be adopting, you know, AI agents for enterprise readiness is something that we need to contribute more literature to, you know, more information to, and why those by themselves do represent real risk, you know, in isolation. And so, but then again, I know I'm happy to also partner with other folks in the industry who are trying to solve these same issues and, you know, we have a lot of work in front of us, but it's an exciting time.

[Lenny] Yeah, it is. Why don't you, make sure to send us the link to that study and Emma, let's put it in the description below the post and give people access to that, but Rob, thank you so much, really appreciate your time. Uh, enjoy your time in New York and your time when you go back to Costa Rica. Thank you to our listeners. Thank you to our producer Emma, to big bad audio. Or, our production assistants, to our sponsors. And most of all, again, to you, our listeners, that's it. Take care. Bye-bye.