The Machine That Builds the Machine: AI, Agentic Automation & the Future of Marketing Analytics, with Alexander Buttler | Senior Account Manager & Adverity

Speaker A: I cashed out my entire 401k thinking someone stole my identity.

Speaker B: A fake email cost me my dream home. After I sent my personal information to a scammer, my AI agent wired thousands to an account I'd never seen.

Speaker C: When billions of people feel unsafe, that's no longer a security problem, it's an economic one. At Jenn, we're building the trust layer for a more fearless planet with products and technologies from our global brands, Norton, Lifelock, Avast and Money Lion. See it in action@gendigital.com hello and welcome

Speaker A: to the Marketing Times Analytics Podcast, Season five. I'm with Alex Butler. He's been a guest on a few previous episodes. Alex, would you like to reintroduce yourself?

Speaker B: Yes, sure. Really happy to be back, Alex. For those who haven't joined our previous podcast, I'm, um, Alex Butler. I've been in the marketing technology space for over a decade now and Alex and I, we've had some chats about AI and its potential in the past and obviously it's been, uh, a quite phenomenal year of AI progress. And I think that's going to be our main topic of conversation to talk about the things we were. Our visions of what could be are actually now happening in reality. So, really excited to kind of do a follow up podcast with you.

Speaker A: Yeah, absolutely. I hate to say we were right all along, but I think right, we were right, we called it. So, you know, I think one thing I've always enjoyed is sort of a, uh, director's cut type of view. So I'm going to pull up our previous episode and you know, we'll sort of play a couple of the predictions and, you know, do some commentary on

Speaker D: it, because data infrastructure is a big challenge for businesses big and small. And it's tough to say if it's harder if it's a big company or a small company, because with the big company you have the complexity challenge and then with a small company you have a resource challenge. There's different ways to approach that problem, but ideas. I think there's a lot of value in the marketing and analytics world of creating a solution that really makes it easy for your business to manage all of that marketing data. And then I also believe I saw there's like solutions or at least some more complex data capabilities that Adverity offers on top of that marketing data that's being fed into the platform. So I think that's really valuable too that, you know, Adverity takes that next step as well of not just aggregating but also processing that data and allowing us to extract insights and make decisions from it.

Speaker B: Yeah, I always try to bring empathy to my clients where I can and especially in this position where I. Before I joined the technology space I had a short time in media buying and so I was the guy, you know, over a decade ago putting together all the Excel sheets and it was my job to make sure that all the media planning was reconciled and it was my job to provide performance reporting. And what do we.

Speaker A: Let's pause there. How's that changed, Alex, over the last

Speaker B: year the AI is doing it. That's the biggest takeaway from overhearing that. And for uh, those listening, I am still with adverity. What's exciting for what to talk about is obviously the, you know, the data pipeline infrastructure that we've been powering. But you know, the, the AI solutions that will sit on top of that data are really the main topics of conversations happening across my client base this year and even outside of the Verity client base. This is just something that's super top of mind for businesses is, you know, that the AI is available, but it's about how can we leverage it. And the biggest challenge right now is data readiness of AI projects. Sadly, there's been a lot of public AI project failures because organizations, super large enterprises were just pushing and just kind of assumed it would work and it was intelligent enough to be able to make business decisions on their behalf. But it didn't have, it didn't really have the data. You know, if we're going to start like at a really basic level, either the data wasn't available or the data was available but it was unstructured or even if the data was structured. The AI lacks context. It doesn't understand your business unless you train it over an immense period of time with m, with immense amount of users training that AI. It takes a really long time for, for an AI, an LLM, whatever an organization might use, it takes a really long time for, for it to really understand your business.

Speaker A: Yeah, absolutely. And I'll take a moment from sharing to comment on that because it took me several months to uh, compose. You could call it a context library or a canonicalized context library, which is good for people and agents alike. You know, every business should have some kind of a reference, especially for an analytics Org where consistency is paramount. You can't have different people or different areas of the org pulling metrics that should have a consistent meaning in different ways because it's going to result in, you know, slightly different numbers in some Cases or, or in other cases large differences in numbers. And that's just going to cause a lot of issues in terms of decision making because it reduces Trust. I used GitHub as the center or the warehouse for this library. And so anybody can clone the repository and then appoint their agent or their LLM, um, at it and say build me, forecast, uh, or kick off a model build for this type of campaign and the agent will have the process mapped out. It greatly reduces the amount of start from scratch that the LLMs like to do where they'll just make up their own process if uh, they don't have a clearly defined one. And this works almost too well. I think the bottleneck I didn't expect that I would never have called out last year was that the LLMs will do such a good job of the core work that the bottleneck will actually become reviewing the work.

Speaker D: Right?

Speaker A: It's very true because at the end of the day we're still responsible for what we produce. And so as the sort of volume of production goes up, you need to review it. And you would think it should be pretty straightforward to review it if you have, you know, the queries laid out the process, laid out, the decisions. But no, even that it starts to become quite overwhelming. And so I've come up with a few optimizations to speed that up. One, and uh, I would say the primary time saving one is using a local text to speech model. So this doesn't consume API credits and it basically, and then I ask for a transcript to be produced with the API credits, but the transcript being pretty, pretty quick and easy to produce of the whole decision making process. What was the approach? What was basically what's the proof? And then I have the local model produce it audibly and then I have a little HTML report created and basically it's like a uh, narrated briefing and I have that produced for the deliverables so that it speeds up my ability to download personally the, the process that was taken, decision making. So this makes reviewing the work a lot faster and it's really the only way that I've learned to keep up. And I think we're all different, right? So we, some people prefer like visual processing. I definitely am an auditory learner. I prefer to listen to the full logic mapped out. And so I think it's important to think through what is the fastest way for any person to digest information for themselves and build a process around that. Because once you solve the biggest hurdle, which is getting the process down of producing the work, you need to get the review bottlenecks figured out and that becomes quite, quite the next challenge that I don't think any of us saw coming.

Speaker B: Yeah, that's a really, a really impressive product that you put together. Alex and I always love taking these conceptual conversations and putting them into practice, you know, just to kind of recap that. Right. It's like we can't, it's, it's hard, it's, it's difficult to trust these LLMs, these agents to, to deliver on accurate insights because people have been burned in the past. Right. The LLMs are sometimes confidently wrong and so it's really important that we implement guardrails. This is what we like to call at a Verity and it's, it's funny, you built a very similar product that a verity has built which is around being that a governance layer between an LLM and an organization's business ready data. And just plopping an AI on top of that doesn't necessarily mean you're going to get the outputs that you need. So building that knowledge layer, that semantic layer, any business, even if we're going to go to now go kind of maybe a decade back, right to just business intelligence solutions, they're useless without a semantic layer because even those, you know, even those code based systems still needed to have a rule base on um, the logic of how the data should be interpreted, the business knowledge. Right. But now it's really the same exact concept. We're just executing so much faster with LLMs being able to do that work. A big topic that we're having at a very, and across our client base and generally this is where I see this going towards the end of the how the language will potentially change in the AI space I'd imagine by the end of this year is this idea of generative AI is the buzzword of 2020, 26. Right. But I really think marketers especially kind of giving the extremely objective line of work that they're in, especially performance marketers is, you know, it's really not about generative AI, about building something new. It's more so about having the AI do the brunt work of uh, manual tasks in order for these marketers to be doing more strategic work, more strategic planning, more strategic analysis. So the language that I'm speaking with my clients about is not so much about generative AI but more about agentic automation. And it's exactly the framework that you outlined in your product. Whereas let the agent should have a rule set, we can't just let it go rogue and decide how it should do this process right, call it a weekly ROAS report that a marketer needs to run every Monday. By design. If you prompt an LLM to pull a ROAS report, generally by its design, it's going to take a different approach each time. But that's not how businesses operate. Businesses need to operate on consistency and accuracy. So by implementing the framework of how that task should be executed, it doesn't become generative AI anymore, It becomes agentic automation.

Speaker A: Yeah, and I've been obsessed with this phrase recently. The machine that builds the machine, the

Speaker B: machine that builds the machine.

Speaker A: That's how I see AI now, after all of the blood, sweat and tears over the last year. The idea comes from historically how computing has been applied to business. The goal is not automation in itself, but rather enhancement would be one word for it. Basically enablement, enablement of faster processes for people. This doesn't always look like automation. You don't necessarily want to take the tools away from people, but you want to make it easier for people to produce things and faster and more intuitive. The goal of like. I heard a story recently about like the computer mouse and how the goal of that wasn't to automate the clicks that people were doing or uh, the selections on the computer. It was just to make it more intuitive and easier. So I tried to think through what I'm building with, with an, with an agent. In that framework, the agent isn't necessarily building a process that you don't touch. It's building a process that, you know, takes your time to build that report down significantly. And you know, one thing with a, uh, with a ROAS report, for instance, is you need to be able to sort of tweak it, especially over time, or answer questions inside. And so having a black box process where, you know, it's producing the same thing consistently is good. But then the next level is, uh, how do we put that in our hands so that we can see the machinery or the processes on that weekly basis and, and be able to explain it, be able to tweak it, change it to daily if we want to change the formula. And so it looks a lot more like traditional software than, you know, something that is in, you know, Claude desktop and that you're, you know, sort of configuring, but that, that's always sort of at arm's reach. And so that's what I've been focusing on recently is like, what is the machine that we're building with, with this machine that's AI. So it's uh, it's quite a Lot of fun. And it's, that's another thing that I would have never guessed a year ago.

Speaker B: You know, I'd agree the, the speed to market on it has been really impressive. And uh, just to speak to the software piece, because that's been my expertise in the last decade, is it's automation. And people might think it's okay, it's just lines of code executing, but it's agentic automation. And so to really dig into the language there and why agentic is the really important piece is because generally software is just, it's code and it's going to do a deliverable, but it's not thinking right when you integrate LLMs with a software approach. So, meaning is that there is a, there's a set of rules to follow, but it's what's making this powerful for marketers specifically is that it's also assisting with the insight generation because okay, yes, it's tasked with pulling the ROAS report every Monday. Right. But the agentic automation piece, the agentic piece of that is that it's pulling the ROAS report and it's not just spitting out CSV file, uh, you. It's designed to be bringing you fresh insights every Monday. And what the power of AI is doing is that what would take hours for call it a data scientist or an advanced marketer to dig into this, to this report and extract and report insights. And LLM is doing this in minutes. Right. And so that is really one of the most powerful things about this, is that it's this, is this the power of AI on top of business operations, consistent business operations, but just bringing more brains to the table whether they're, whether they're human or simulated.

Speaker D: Sure, yeah.

Speaker A: And one thing, especially for internal use cases that I found is like a burgeoning skill set that I've been building is, I would call it like a data ecosystem awareness. So knowing when I see a result or whether it's AI, uh, generated or not, and being able to understand. Does this make sense? What are the possible failure points along the data pipeline that may result in an outcome that what I'm looking at versus data?

Speaker B: Sorry to interrupt, but data quality always is one of the main.

Speaker A: I cashed out my entire 401k thinking someone stole my identity.

Speaker B: A fake email cost me my dream home. After I sent my personal information to a scammer, my AI, uh agent wired thousands to an account I'd never seen.

Speaker C: When billions of people feel unsafe, that's no longer a security problem. It's an economic One At Gen, we're building the trust layer for a more fearless planet with products and technologies from our global brands, Norton, Lifelock, Avast and Moneylion. See it in action@gendigital.com Pain points from

Speaker B: my clients around where the process stops or where the process breaks. And it always falls into a data quality issue, whether it is a campaign where a naming convention was not met or a, uh, key ID or a UID is missing. Right. Any of that core foundational data pieces that connect the dots are. Right. Those tend to be one of the most common areas of where just downstream reporting and then agentic automation is breaking down and failing.

Speaker A: Yeah. And that's also a plus side of using LLMs is they are phenomenal at debugging.

Speaker B: Oh yes.

Speaker A: It saves me so much time. And they go to levels of detail that we would never go to. I had it.

Speaker B: And quickly. Very, very quickly.

Speaker A: Yeah, because it's just too tedious. Like it would not be worth the turn for a person to go into that level of detail. I had it doing a reconciliation, uh, of data that we were trying to understand, you know, what went wrong. It was another company that was analyzing data and there was a discrepancy between how we had analyzed it in the company and that other company had had issues in their matchbacking process. And these were serious issues because they hadn't traditionally analyzed direct mail. And so there were things that they didn't catch that we would have caught instantly. For instance, there was 20 touch points reported for a single direct mail recipient. And what I really appreciated was the LLM said for an online brand doing display, 20 touch points is nothing. They wouldn't have processes to catch that. But because it's direct mail and that represents, you know, $10 or $15 worth of spend for that particular recipient in a direct mail company would need to catch this because it makes a lot less sense in a direct mail scenario than an online scenario. I thought that was such a, such a great connection to make of, of the, why this mistake might have happened. And, and just doing all of the reconciliation processes to figure, uh, out. Exactly. It ran a simulation to identify what, what was the mistake that was made provably so saying like, we can prove that this is how they reached that scenario and this is the sort of thing that manually we would never do. We would sort of edit and say, you know, here's the problem area. This is what we recommend to do to fix it. It's another story entirely to say we've replicated what went wrong. And we know exactly what it is. That is just a new capability I think uh, that is probably underappreciated right now in, in the AI world.

Speaker B: Yeah, I couldn't agree more. And what is that forcing businesses to do? Historically? It's forcing them to make concessions. It's like we're just going to have to accept a less than 100% accuracy on these things. And when there are gaps like that, that's lost revenue, that's lost potential revenue. Right. That's, you know, when you think about cost per acquisition. Right. That makes cost per acquisition higher. Right. It reduces efficiency. What I love about AI powered data auditing use cases is that businesses no longer have to make those concessions because they don't have to do that resource allocation to outcome analysis anymore. Because again, businesses are always trying to decide, okay, where do we put our resources? Because that is one of the most finite things is resources from an employee perspective. And it's like one person can only be doing so much and not to get too corporate, but that's what, you know, that is what global corporate is thinking about right now is how can one person do more with less or how can one organization, one department, one team do more with less. And that's where these types of use cases come into play. And really into that word you used before, enablement, it really is moving the needle.

Speaker A: Yeah, I uh, enjoy the process of making a lot of our processes more accessible. So you take really complex processes that previously you needed analyst time on and you can codify them and turn them into tools and make them usable by the people who were requesting the analytics work. And you can't do that for everything because sometimes the analyst uh, needs to decide if something is worth pursuing given the context of the request to the client, internal or external, what, you know, what's the value? And that stuff really does require subject matter expertise. But other things that the stuff that we don't like to do, you know, but that we have to do, that sort of thing is perfect for codification and self service and that I think AI has done a tremendous job at.

Speaker B: Yeah, the one piece of our previous conversation definitely want to re reference again I think it was maybe towards the end of our conversation where we were talking about autonomous marketing. Right. And so think of the performance marketer life cycle is okay, it's time to make an optimization. We need to make a decision, um, to maximize our performance metrics on our paid media. So what is that process? Okay, we need to get the data, we need to harmonize it so it's all aggregated for it to become business ready. And then once it's business ready, then we need to create business intelligence on top of that business ready data. And then we need to now, okay, now we have the reporting, now we need to generate the insight. And once we've generated the insight, then we need to come to a consensus on what that means for paid media optimizations. Then we need to go into the paid media platforms and execute whether, you know, it is, you know, whatever, whatever levers need to be pulled on each native platform. And you could think of, you know, how many there are right now, right. We talked about in our last podcast around autonomous automation. I think we compared it like a stick shift car to a Tesla, right? And it's in the world of mcp. This is such an amazing, such amazing progress on how the agentic AI can operate cross platform because you know, every single, you know, every single SaaS platform has its agent, right? But what it's, but what makes this agentic automation so powerful is it being able to execute cross platform and having these agents communicate with each other through MCP in order to fully to create more of an autonomous workflow of that life cycle that I just outlined. So I'd be curious, I could talk about how it's actually being put into action.

Speaker A: Yeah, let's do it. Let me pull it up.

Speaker D: But still not autonomous by any means. It was assisting the human, right? Driving cars early, very manual is everything.

Speaker B: We're auditing it now to make sure the data quality measurement stays high. And then now we could start having our departments take that extra time for collaboration and then get to that strategic revenue driving activity where they can start to scale data scalability, creating templates, right? Being able to quickly copy and paste infrastructures that you have built for one department for another. Or each time there's a new marketing campaign that provides a new brand initiative. How can we quickly copy and paste what we built for the previous one for the new one? And rather than having to rebuild any pipelines or rebuild any kind of procedure, scalability comes into play to ultimately create that data democratization where any department who needs to answer a business question can go to a singular source of truth and answer that question to the business.

Speaker D: I like that a lot. I think it's similar to if we were to look at cars, for example. So starting with the sort of horse drawn carriage, very early, very manual breaks down a lot, actually superior in some ways to the early versions of cars. But then cars came in, replaced the horse and bogd those progressed for a long time but still very manual driving is very manual for the vast majority. But you know, tools came along that made.

Speaker A: I want to pause there for a second because I mentioned something and I didn't explain it. I mentioned that horse drawn carriages were superior to cars in certain ways. And I think the difference, I think that explains the last year in a lot of ways.

Speaker B: Where are you going with this?

Speaker A: Yeah, why is a horse drawn carriage better than a car? First of all, because the car, because it's a brand new concept of a product, it's inherently buggy, it has a little pun there, it has some issues that needed to be worked out. Now a Tesla is better than a horse drawn carriage in every way. But that took a hundred years. The horse drawn carriage, if you were to fall asleep, the horses actually know where you live there. There's stories of horses taking you uh, know, drunk drivers for instance, back home. You know, they were passed out in the carriage. There are, you know, there's things like uh, you know, if the horse, the, the horse drawn carriage survives on the fuel of like grass or whatever, they feed the horses and that's like pretty available there, there isn't any issue. You know, imagine what the fuel infrastructure looked like when, when cars first came out. Probably not very extensive. And so running out of fuel is probably much more common than running out of whatever people would feed their horses because everybody had horses or at least they were very commonplace in comparison to gas stations I presume. So the initial technology adoption was probably uh, quite, quite rough. But I think people saw the vision and were willing to struggle through it. And I think that's kind of what we saw over the last year was certain things, it would actually have been better. There were benefits to doing them manually. And it's only if you keep in mind that eventually we're building to a place that will far exceed the old state. But we're only in the first phase. We're in the Model T phase right now and there's going to be things that are underdeveloped and there's going to be things that the previous way won out on. So yeah, I just wanted to mention that because I didn't explain that and that's actually I, and I didn't realize at the time that would actually be a really solid example of what we were about to face. I didn't expect it. I thought we were just going to jump straight to the really good car. But now like you have to suffer through the first phase.

Speaker B: Right? And the horse and carriage discussion, you know, really, what, what makes it so great? Right. I was thinking, you know, because it's, it's simple, right. It's simple and reliable. Right. And so that conversation, it, what we're seeing is that exponential growth and it's, that curve is way steeper than we thought it would be. But for those who are not, uh, able to climb that steep cliff right now, they're the ones falling back and they have, it's that reminder, you got to go back to basics. Right. And that's where we talked about the beginning of our conversation, where, you know, you got to go, you have to make sure that your data foundation is ready to go. Got to go back to the basics. You got to hop back on the horse and buggy and make sure that that wheel is spinning. Correct. Because then once you throw a motor in there, uh, you know, you throw a computer on top of it, but you got a broken wheel. Car is never going to work.

Speaker D: Yeah.

Speaker A: And even to take that a layer further, if we are city planners and we build roads that are poorly built and they're sort of all lumpy and jagged, people's wheel, the wheels are breaking, then it doesn't matter if you have the horse and buggy or the, none of them will work if that foundation and those roads aren't built properly. And that is sort of having a semantic layer or having standardized processes across the business if you don't have that. And it doesn't matter if you want to introduce all the autonomous processes or none of them and just bring in manual analysts to do everything without a. Neither of those will work with inconsistent processes and data. It's not just data that's incorrect sometimes. This is the very commonly cited issue with data. It's also not having a consistent approach to, um, how things are defined and queried. The data could be perfect. But if two people and two different sides of the business are using different methodologies to pull the same numbers, it's going to cause tremendous confusion in the business because they both could be correct, but because there is no aligned definition, it's going to slow everybody down.

Speaker B: The, uh, business knowledge needs to be predefined.

Speaker A: Yeah. Okay, let's keep going it a little bit easier.

Speaker D: You had the check engine light. Instead of checking your engine, you had all these indicators that told you if something was broken. So this is not an automated state by any means, but it's almost an assisted manual. You still have to do everything yourself. And then you get like the parking brake sensor, uh, or the sensor when you get too close to things, they

Speaker A: started to add little features.

Speaker D: Driving started to get more advanced, but still not autonomous by any means. It was assisting the human. Right Driving. Cars continued to improve in how well they could assist. Assist the human. And now we are getting to the point where cars can replace the driver. And basically now we have this autonomous driving an autonomous car. I would postulate we are moving towards autonomous marketing as well. And I'm not sure what the timeline is, but all you have to realize to know that that's coming is if you ask a marketer to create a viral clip, they cannot do it. Okay. If they can't do it, then a computer can definitely do whatever the human being is doing. I mean, if you, if a human being could just create a viral clip, you know, the average marketer could just create a viral clip. I would say, okay, there's no, like, I don't think computers can replicate it, but the vast majority of marketers would not be able to beat an autonomous marketing machine, especially if it like.

Speaker A: So that, that's an interesting take. I don't think it's the strongest way I could have phrased it, but the point was that the sort of magic, uh, the Mad Men side of marketing is largely out of reach. It's not really the boots on the ground thing that marketers are doing day to day. We're really doing a lot of analysis, a lot of basically synthesis of what are we saying to the market, how do we say it, connecting that to, uh, what we can execute and what we can, you know, who we can reach, how do we reach them? Um, it's a lot of analytics and that is why we can probably use AI for a lot of the processes. It's not that genius creative inspiration that creates the super bowl commercial. That is genuinely really hard. I mean, I don't think I've ever laughed at an AI generated joke ever. Because humor is one of those really higher order, like, ah, one of the highest order forms of thinking. And uh, yeah, that's far, far away. We may have to dumb ourselves down to get AI to that level.

Speaker B: This is true. Like people, people point out AI content versus human content all the time. You know, that's the thing is, is it, you know, it's like, uh, it's. It's not a human brain, it's a simulated brain like you mentioned. Right. So the thing is, it's. And as I in the last podcast too, uh, where I talk about, you know, empathy, right? Like AI cannot empathize with the human person. The human brain, the human psyche, because it cannot, it will never have that. Um, we'll see what Elon Musk does in the future. But you know, at its, at its current state, right, that's why there's always that disconnect and why things like AI generated content always gets pointed out by people. You know, it's like the perfect example you mentioned. It's like, you know, LLMs, you know, really, it really is just like this genius child and it's like if you don't really give it the required, like, if you really don't guide it into making right decisions, it's. It will make wild decisions. And we can end today with a funny story with some comedic relief shout, if I may, but a really, a really funny one. A Reddit thread of this. And this is when I think, uh, Anthropic released Claude and they basically did the opposite, right? So we've been talking about this whole time about how it's so important to implement guardrails and to give AI business context because then you can't trust it, right? So this was kind of like an experiment this, this one startup company did where they were just basically kind of really testing. It's like, okay, what will CLAUDE do? And so what they did is they ordered a vending machine and they implemented CLAUDE within their Slack and they tasked it with saying, hey, you're in charge of filling the vending machine with whatever everyone wants. Uh, but then it gave it performance targets. Like it had to, you know, reach a certain amount of revenue by an X by a certain date. I think they gave it like a week or a month or two, I can't remember. Uh, but it's super funny Reddit thread. But in, in short, you know, because they go on as a super funny story. But the, the consensus here is it kind of went crazy. It was ordering some snacks, but it wasn't happy with the margins and it kept like it had full autonomy, right? So it was like annoying people on Slack, like asking everyone doing at here's at channels, like asking about, you know, can you imagine like in these major business channels, the, the Claude. The vending machine Claude bot is like at channel to the entire company is like, hey, what snacks do you want? Then it went on like, then it started like, uh, autonomously thinking about how could it could improve its margins. And so I think that if you Google like tungsten cube vending machine Claude, right? Like this, this Reddit thread will come up and then eventually made the decision that because it was running out of time to meet its revenue target, that if it sold one tungsten cube, that it would achieve the revenue that it needed to attain, uh, by that. So they actually went out and ordered

Speaker A: a tungsten cube, a tungsten cube came

Speaker B: to their office, and then the agent went around the entire company begging people to buy the tungsten cubes. Really funny story, but that's the kind of craziness that could happen if we're not governing these AI, these LLMs. And uh, it's just so interesting. You look at the case studies of you can learn. The end of the day, I'm a big believer in you got to fail in order to learn. I think some of these large enterprises in the first half of this year have had those hard learning lessons. But end of the day, that's only going to get us to the next step with really, uh, making AI as productive as possible.

Speaker D: Couldn't agree more.

Speaker A: Thank you, Alex. It's always a pleasure to have you.

Speaker B: Likewise, my friend. Thanks for having me again.

Speaker A: Awesome. Thanks everyone for listening. We'll talk to you soon.