AI Agents Are Replacing Entire Marketing Teams

Iliya Valchanov: Our team used to be our growth marketing team. We used to be seven people. So we had seven people in sales and marketing. Now we have one. And we're doing more work than before.Intro: Built This Week, breaking it down Built This Week, we show you how A fresh idea, a clever tweak you locked in. Built this week.Sam Nadler: Hey, everyone, and welcome to Built This Week, the podcast where we share what we're building, how we're building it, and what it means for the world of AI and startups. I'm Sam Nabler, cofounder here at Rise Labs, and each and every week, I'm joined by my friend, cohost, business partner, Jordan Messner. How are doing today, Jordan?Jordan Metzner: Hey, Sam. How's it going? Another crazy, crazy week in AI. I feel like I say this every week, but there's been some really awesome new developments this week as well. I'm sure we'll get to talk about.But, yeah, every week seems more exciting than the one previous.Sam Nadler: Absolutely. And we have a special guest today, Ilya. Ilya is from Joomma. Ilya, introduce yourselves and tell us a little bit about Joomma.Iliya Valchanov: Hi, everyone. Thanks so much for inviting me. My name is Ilya. I'm cofounder and CEO of Joomma. Joomma is an agent or as you call it super agent for marketing tasks and which can autonomously plan, execute, and complete tasks which are in in the realm of marketing, whether that's content creation, performance marketing, ads, creating images or carousels, all of these things we we optimize for.So far so far yeah. I was just gonna say, so far, we're about we have about a 100 k users coming from all over the world, and we are thrilled to be building in AI.Sam Nadler: Amazing. Well, you know, the docket today, we're actually taking a little bit of a pivot. Typically, we demo something for our guest. However, today, Ilya is gonna demo for us how Joomma works. I'm really excited about that.Then we're gonna just jump into Joomma as a whole, and then we're gonna cover some latest and greatest news in AI. Ilya, before I kick it over to you, please don't forget to like and subscribe. We have new episodes out every week, and hit that hit that button and support Built This Week. So, Ilya, take it away. I would love to see how Joomma works, and this is, I think, really top of mind.I feel like, you know, every TikTok or YouTube video I've seen lately has something to do with leveraging AI to super drive your marketing process. So I'm really excited to see what you built for HipTrain.Iliya Valchanov: Yeah. Absolutely. So HipTrain, it's is basically an an app which helps you helps you with personal one on one training. And I was going through the website, and I'm like, okay. How can I help Heap Train with distribution?So I was thinking, okay. The website looks good. It has some good stuff on the blog. The message is good, so I cannot help too much with this. But maybe maybe we can help with the social media strategy.So what I did was I I gave Druma a clear social media calendar for April, covering LinkedIn, x, and Instagram, and that that was all that was all. So it goes through through this process. It goes, finds the what HipTrain is. We have a bunch of different skills. So in this case, we have the social media calendar, which is a skill that we optimize for.We have hundreds of these for marketing use cases that we're we're writing ourselves, basically, not with AI. We it goes to website and then creates this plan. Let's go research, do competitive do competitive scan, then present the findings, and so on. Dave goes through a more detailed process. You can you can check each of the steps.He's doing the research, and then it goes into question mode. Basically, it asks you a question. So very similar to how ClotCode would ask you a question, we also like to, you know, make it very interactive so that just with one prompt, you can go through a full workflow. In this case, I I I clicked, looks great. Let's build it, and and the tool continued to execute the workflow.So the final result of this is basically a social media calendar for April with an overview day by day. Here is the full calendar. Maybe I can zoom in, and, you know, it has a couple of pillars, corporate and wellness, fitness education, motivation. It goes through different concepts, so on and so forth. Also, it has the the the breakdowns and how many posts per view would go into each of these pillars that are very important for the social media strategy of of any company.And finally, there's some next steps that you can go through. So all of this was with a very, very simple prompt, which was just create social media calendar, and under the hood, we optimized for everything else that you wanna do. From there, you can just go to the next step, and you can say, in this case, I said, take this social media calendar and create an Instagram carousel about about it. And it, again, went through the process and created the the social media calendar and the the carousel. I still haven't seen it.I hope it will were looks great. So let's see.Sam Nadler: Even better.Iliya Valchanov: So this is basically a visualization, first first one, second page, third page of the carousel, fourth, there they are, and so on and so forth. So this is fully fully autonomously built. If I was to if I could actually, you know, give it your branding or your colors, it would be fully ready to be published on Instagram immediately. So, basically, that's that's what we are optimizing for. We have connectors to Google Analytics, HubSpot, Google Search Console, Google Ads, Meta Ads.We are doing all of these ourselves so they're not the so we can expand them with whatever you need. And currently, most of our clients, they just connect all of their tools together, and we make the magic happen, basically.Sam Nadler: Amazing. And does it learn, like, for instance, that Instagram carousel that you posted and, you know, over the course of the month, it does, you know, 20 carousels. Does it begin to learn which kinda carousel format had the most engagement, the most views, the most comments, whatever, and then begin to, you know, if you run the second month or the third month and begin to improve upon itself?Iliya Valchanov: Yeah. So the first step to do this is you need to create a project, and in this case, I can create a HipTrain HipTrain project. And once it is in the project, it starts, you know, gaining memory about the stuff it has done. It can go and over time, you can involve it. The this thing with that you mentioned from his for Instagram, we have just released this Instagram connector, and this is exactly the way we're thinking about it.Once you start posting the stuff, the project should know what you have posted already, how it performed, and whether it should do more of this or less.Sam Nadler: Very cool. Very cool. And what types of businesses are using it today? Is it mainly consumer business, b to b businesses, smaller teams, bigger enterprise? I'm sure it's a little bit of everything, but where where do you think what seems to be what client seems to be gaining the most value and traction?Iliya Valchanov: Yeah. So we we call it, you know, an agent for marketing tasks. So whoever is doing marketing tasks, and these are mostly seen in marketing agencies where everybody does marketing tasks. Second biggest group is marketing teams where, you know, these are smaller, but, you know, in marketing teams, you have two, three people doing the work of 50. And also solopreneurs, founders, people who they have a project out, but they don't know how to market it.They come to us to kind of do this work for them.Sam Nadler: And have you ever run, even if it's like back of the napkin math, like, you know, running a project through, like, the full kind of marketing workflow to, you know, really deploy everything possible. Like, what a normal team of five would spend, you know, six weeks doing can now be done literally by one person in, you know, eight hours? Or when can you walk me through some of that time savings and obvious probably cost savings in as well?Iliya Valchanov: Yeah. Absolutely. So so our team used to be our go to market team. We used to be seven people. So we had seven people in sales and marketing.Now we have one. And we're doing more work than before. So, basically, we started by trying to automate each and every activity that we do in marketing, and I would say that in some of the cases, we're much faster. Let's say, we integrate with Webflow, and now you can write and publish an article in three minutes from beginning to end with with no supervision. So it takes three minutes to do this, but it's not only about the speed.It's about stuff that we were not able to do before. We're hearing these full reports from Google Analytics, from Google Ads. I had never had a a report as good as as this one in the sense of AI is giving me the full overview of what's happening, who's converting, what's working, what's not working. No agency is actually doing this in a good way, not to mention a marketing team of two or three people. So, basically, I think it's more about these extra stuff that you can do that that were not possible before work.And, yeah, and I'm most excited about about these, unlocking these.Jordan Metzner: Can you tell us Ilya, hey. Thanks all for the demo. That was really cool. Can you tell us a little about the impact on IRR and ROI and ROAS and kind of it seems like the biggest opportunity is in the ads, particularly in the ads space less than in the organics in the organic space. But, yeah, tell us a little bit more about kind of like how the impact is to ROAS and whatnot.Iliya Valchanov: Yeah. So the it really depends on the client. And if you take a very big agency or a very big company, they're thinking about it like this. Can you save us 10% of our billable hours? If you can save us 10%, we're gonna make shit on shit on money.Right? And and, basically, in this case, for for we for them, we are trying to remove all the repetitive work, and and, basically, it's measured in terms of billable hours. Every billable hour we save, this they they make this much more money. Let's say they bill a $100 and we save them twenty hours. It's very easy to see to to how much we have saved saved them.Now when you talk about a small marketing team, the it's not so much about how much time you save. It's much more about how many more things you can do, and it's about the opportunity to explore more and more things and build projects that you would never do and so on. So long story short, for marketing agencies, the return is is is in hours. They're usually saving, you know, 10 x to 20 x of time on each task. And for smaller teams, it's much more about chasing every op every idea they have in their head, they can chase it now, and ultimately, they they have, you know, better results.Jordan Metzner: Awesome. I guess this is like allowing marketers to not have as much need for a like design team and kind of creative team so that they can run more creative opportunities themselves, a more a b testing kind of more ability to understand what content works by creating it themselves.Iliya Valchanov: As well as well. But even even let's say for it's just redefining the the workflows. In the past, let's say you want to to to do SEO or GEO. You don't know how to do keyword research. You know, that's a specialized activity that a specialist has been doing.Right now, we connect to, let's say, Ahrefs or Semrush MCP, and it does it for you. You know? It is kind of crunches this whole 10 activity in in one minute. So so it's very, very redefining in every way possible for the teams. So SEO, ads, you you mentioned ads.Ads is an activity where companies spend a lot of money, so each tiny optimization you can make matters a lot more than, let's say, content. So this is why we decided to double down on on data analysis and bring the data from Google Analytics, from Google Ads, and so Meta Ads so that you can have better intelligence about these systems. But, yeah, over there, sometimes the savings could be, you know, millions if your budgets are in millions. Right?Jordan Metzner: Yeah. Totally. And tell us a little bit about just, you know, as you guys have been running these agents and whatnot, what are some general marketing learnings that you have taken away that like maybe maybe apply to all brands and businesses looking to market online in today's day and age?Iliya Valchanov: Yeah. So the first thing is it's not about doing stuff faster. It's about doing stuff better. And I think this is a very big misconception. Everybody's like, save time, save time, but but actually, save time to increase quality.This is kind of the the main learning. And the best users of AI, this is what they do. Another learning that we that we have is that after coders, marketers are very, very, you know, excited about technology. People in ops are quite excited as well, and other professionals are a bit a bit less excited, I would say. So that's why kind of coding and mark coders and marketers are leading the wave in terms of what gets built and how it gets built.Final thing that I'm that I'm maybe interested in learning is every company is going through this transformation right now, and the most important thing they need to do is to allocate one person to be leading this project. Companies that do not have a dedicated person to lead AI transformation, whether it's a five people company or 50 or 500, they fail because they think, oh, I'm gonna get the tools, and it's gonna work. No. No. No.It's about complete change management and rethinking of the workflows.Sam Nadler: Super interesting.Jordan Metzner: Yeah. Really interesting. I think kind of your comment resonates about kind of, you know, saving time isn't really the key key variable here, but, you know, improving the quality of the product that you're outputting is is the output. Because it seems like, you know, with AI we get time back sometimes. Right?These agents can run all night long, and you know, for me sometimes like I don't even have work to give the agent because I I'm just like I've ran out of ideas for the day maybe. But you know, to your point, I think it's not necessarily about time, but quality, and I think that that kind of resonates. So it seems like it raises your your quality boat of your marketing and increase your brand in that sense, and that's that's obviously pretty interesting.Iliya Valchanov: Yeah. Yeah. Definitely. When somebody tells me, oh, I'm running 50 agents, and I'm like, this is this is ridiculous because, you know, you cannot process the output of 50 agents. And, you know, I can run no more than three agents at the same time.I don't know about others. But they're so agents are so fast at getting the job done that, you know, by the time I run the third agent and I give it the context, the first one is ready. So Yeah.Jordan Metzner: I feel similar to you. I think, like I mean, it seems like we're now starting to get at the like multi level agent orchestration level, you know, kind of running this this symphony as OpenAI calls it, you know, kind of, you know, being the master of ceremony with all these sub agents. But, yeah, I I kinda feel similar, you know, I've seen people who have like an, an insane amount of agents running, and I don't understand as my agents like not busy enough, and I wish I could give him more work. So I don't understand why I need to multiply that, but I don't know. Maybe it's, you know, it's just an early early days, and hopefully, we'll all learn how to be better working with agents and sub agents and orchestrators of agents, etcetera.Sam Nadler: Absolutely. And speaking of agents, I actually wanna transition us to kind of the First News article in the day, which is kind of the high level, you know, this ongoing competition between Cloud Code and Codex. You know, similar to you, Ilya, I can really only run three agents at the same time. I I I I feel like I forget one as I, like, engage with the other two, and then I come back to it and forget the other one. But, you know, we're kinda I would argue we're power users of both here.I know Jordan shifted a little bit more towards Codex, and and I still use both Cloud Code and Codex about 5050. But just, you know, a, do you have a preference? B, what are you using? Have you used both? Give me give me some understanding of of of your stack.Iliya Valchanov: Yeah. So everybody on our team was using Cursor before, then we all moved to ClothCode, and basically last week, we canceled Cursor, and we haven't been nobody wants to move to Codecs. We test it from time to time, but for our use cases, Cloud Code is working so so well that we do not see why we should move to codecs anymore, while the difference between cursor and Cloud Code was so big that we actually intentionally moved away from Cursor. So actually, I'm very interested to ask Jordan, why are you using Codex more?Jordan Metzner: Yeah. So we went through kind of like a similar set of iterations. Think like at one time, we were all using Cursor, and we've kind of moved from Cursor over to Cloud Code. Especially, you know, when Opus 4.6 came out, I think that was the big jump, you know, in mid December. And one thing I noticed is like, I I really enjoy using Cloud Code.It's very fun and they do a really good job of kind of keeping you entertained while the agent is working and kind of, you know, those words and all those things. And it's actually like a pretty like enjoyable experience. I was a little reluctant to try Codecs, but once I went to Codecs 5.3, I noticed significant improvement in code quality. One thing I would notice like repetitively with with Claude code is that it would make mistakes, but be confident that there were no mistakes. So it was I I guess you can call like buggy code for lack of a better term.And one thing I've noticed with Codecs is that it's it's much more performant. So as a user experience, it's like a little less enjoyable. It's a little bit slower. It's like not as fun. But the output of the code, both with 5.3 and now especially with 5.4 on extra high, I found to be significantly better quality.I think like to each their own in the sense that, you know, I I will kind of allow my developers to kind of pick what their favorite tool is. Like, they still have access to all of these tools and kind of trying to coalesce everyone to use the same thing. But, you know, switching costs are high. You know, it's not that high, but, you know, everyone kind like switch and move to a new tool, but we still have some developers preferring Cloud Code over Codecs, but for me personally and some of our other developers, we found that Codecs has been much better. I think I think the big unlock for us is we're we're working on trying to get, you know, these PRs done in the background, kind of like whether it's Codex or Cloud Code kind of working in the background.So I think as that gets more improved, that will probably push us towards like one ecosystem over another. But I think what's most interesting is that like, you know, back to the switching costs. Yeah. Maybe for a developer like to learn a new system picks a little bit and like the idea desire to like close one and open another. So that seems like honestly not very high, but people get custom are accustomed to working in a certain way.I mean like, you know, developers who like JetBrains, like, have been using the same IDE for, you know, fifteen years or something like that. So it's sometimes hard to switch, you know, the form and operation which you work. But to me, you know, it's like whatever's the best model of the of the day. And you know, like, we'll just keep switching to find like where the best outcome is for that. It's, you know, it's crazy that no one's looking at the code anymore.I mean, you know, both Cloud Code and cursor excuse me. Both Cloud Code and Codecs are built as applications that don't have the code easily available to show you. And that's an interesting paradigm, but it seems like more and more that nobody's reading the code anymore.Iliya Valchanov: We have a coefficient at work. So how much of the code do you read? And each each person on the team has has the coefficient. And some people read 90% of the code. Others read 30% of the code.It's very, very interesting because this varies all the time, but it's kind of clear that we're all converging to reading less and less of the code. Actually, what we did was we spun up four more environments of our software, and we're just putting the PRs on these environments so we can test them. So it was much more important for us to to have more places to see what the agents are doing to the product rather than read more code.Jordan Metzner: Yeah. I I don't know. We don't have like a coefficient like that, but it's you know, I think it it has to do with like developer trust, like how well they trust the model and how well they think of their output. And you know, I think I still think, Glenn, that the 10x developers are still like maintaining themselves in position of being 10x. Like they're still providing a lot more value than like a standard developer because they just can, to your point, like abstract more agents and get more complicated work done.But yeah, it seems like we're driving towards a path of like nobody is going to read it. It's just a bunch of, you know, gibberish per se. Right? And what's even more I think interesting is like how do people learn computer science in today's day and age? I mean, if you were to go to, you know, university and take a, you know, take a computer science degree, you know, is learning how to write Python and functions and for loops like even even critical or important?And you know, is that type of base knowledge important now or in the future? Or you know, is it something that like we don't ever have to think about ever again and it's maybe a wasted talent? I don't know. I mean like programming in general has always been this kind of thing that's like very difficult to do for most people and for many people. Like only certain people are like have been able to do it really well while others struggle with it.Plus you do the abstraction of it and just the learning of any specific language itself like, you know, that's why kind of TypeScript was so performant because it was able to take away so many of the annoying things of JavaScript. But, you know, now the next thing is just script. Right? And that's where we're we're at right now. It's just just write in plain English and get back what you want.Yeah. Don't know. It's really fun and interesting and exciting and like I mentioned in the intro, this week has been crazy and so was last week and so was the week before and, you know, I mean, I think these these in the frontier models coming out every day with new product features just as a developer, it makes our lives super exciting.Sam Nadler: Well, speaking of crazy, let's cover our last story, which Jordan, I think may be helpful for you to give us an overview of this new open source auto research that lets you do hundreds of AI experiments. Ilya, you may know a bit more than me about this, but I know Jordan has some context and, you know, just give us the the download.Jordan Metzner: Yeah. So Andre Karpathy, one of the founders of kind of AI overall, and he's the originator of the word vibe coding. You know, he worked at OpenAI and Tesla and has been building AI tools for the public for the last few years. And this week, he launched kind of two new pieces of technology both on his GitHub. This this auto agent tool, and then this other tool that we're mentioning here about this kind of build your own model.And I've just seen some incredible feedback of what's going on in the market of people saying that like this has been incredibly helpful. I think this morning, Toby from Shopify said that this auto research tool works really well on even code bases and other non kind of research based products. So it is a really exciting time of this, like, kind of mini model of training the model and kind of then having the model get better and you know, one thing that's pretty interesting, I don't know if you read about this Sam, but the model training works in five minute increments and it's specifically timed so that it limits the amount of training it does to a maximum of five minutes. And I think what makes that interesting is that, you know, almost all training of these models takes hours and days and weeks and months and all of that. And I think that it was a really interesting interesting function like selection of like code choice that he he selected to have this model trained on only five minutes at a time.And it gives this ability of, you know, back to what Ilya said, you know, I can only do so many agents that are only so busy, etcetera. It's like, you know, you can have your agent trained for five minutes, go work on another task or so, and then five minutes later you're gonna get results. And then, you know, seeing those results and then iterate again and again. And so I really like this idea of kind of like these small mini experiments, but, you know, you can do that obviously so many times an hour and so many times a day, etcetera, etcetera. Wow.Sam Nadler: Yeah. Ilya, did we did you get a chance to read the article?Iliya Valchanov: Yeah. Yeah. I actually read the the GitHub repo when we came out, and it was very interesting because it's it's it's it's very small and stripped to to the essence, and only someone like Karpazi can can do it so, you know, to the point. So it was very, very cool to see. In my background, we were teaching people how to train these models, and, you know, one of the first things that that we did so so imagine, we've created more than a 100 courses, AI and how to train models.And then Gypsy comes out, and everything that we've been teaching people is, like, in the tip of your hands, and you you get the scripts. They're ready to use, and then the question is, what do you do with them? And then as the hardware becomes faster and faster, and you can do all these experiments, one thing that stuck stuck in my brain was a friend of mine came and he said, I saw your website, and it said that you're doing, let's say, AI for marketing, And then I refreshed the page, and it said the same thing. Why is this the case? You should be AB testing every single moment of every single minute, and you you should be, you know, constantly optimizing on every page, every word.And I'm like, well, this this sounds like an overkill. And he's like, but but now it's cheaper than ever to do it, and, you know, I can't imagine why people are not doing it yet. And this article was basically saying, okay. Here's kind of a blueprint to start running these experiments in a scale never seen before and at a cost never seen before. So I was I was very excited about this spin of of it with the experiments.So I don't know. This makes every website a living being. Right? So websites are they're no longer static.Jordan Metzner: Yeah. This is a huge unlock. I I think I was talking with someone yesterday just about this idea that, you know, historically, you kind of like build a website, you put it up, and then you focus most on your web application. If you build a new feature, you go back to your website and update it. But for the most part, kind of they stay static for a pretty long time until kind of you do a refresh or whatnot.And, you know, obviously there's always been tools to like a b test buttons and copy and things like that. But they've been, you know, one not of high value. You know, changing a button, you know, from blue to green, I I think is like an over optimization where people find that, you know, the results results, you know, even if they come in statistically significant, like at the end of the day, like if you didn't change it, you'd end up getting the same type of results. And I learnedSam Nadler: a lot ofJordan Metzner: that in Amazon. I think like they tried so many things, so many tests of like changing the shape of you know, the design of the star ratings and whether that would have an impact on sales. And you know, just so much kind of historical like web lab type testing where you know, you run this experiment, you give it its time, you get enough, you know, you get a coefficient, you know, and then you know, and then you kind of decide like, okay, now we decided we're gonna go with b or whatever. Right? And so this idea of this like constant experiment where like you don't even know what's happening on your website because it's constantly testing every single variable and all these other variables.So, you know, even customer that comes in today sees a different website than one that comes in an hour from now, even one that comes in tomorrow, you know, I think that's obviously incredibly cool. The real question is like does that actually mean something for the like the net impact to your business? You know, like does optimizing your website over and over again have, you know, significant lift over time? You know, that probably has an impact on what is the business. Right?You know, I mean Google's probably the classic of this, but like they haven't changed their homepage in forever. Right? I mean, almost like since the beginning of time. And you know, they make small changes and add a microphone here or there, but you know, they've been very delicate on kind of changing that core kind of Google launch page experience and obviously have done incredibly success been incredibly successful. It does seem like e commerce is probably where it's like this opportunity lies the most kind of like because, you know, e commerce like people check out and so you're able to check like and get a real feedback loop of whether or not like this has an impact on customer behavior versus just like, you know, in a b to b setting to your point, Ilya, like on your software, like customers might visit your website, but it might take them, you know, a longer period of time to make a purchasing decision, etcetera.So but it just thinks I I just think this is like more democratization of AI and machine learning, and overall just giving more capabilities into like developers hands, and I think that's that's incredible.Sam Nadler: Cool, guys. Great episode. Ilya, thanks for joining. Where can they find you? I know it's juma.ai, but, you know, if if someone, you know, is desperate for this product, what's what's the best way to reach out?Iliya Valchanov: Yeah. Juma.ai, they can start free trial. They can check they can check the product. It's a bit and they can check the product. They can see if it if it's useful to them.Or on LinkedIn, I'm posting regular LinkedIn, checking my inbox there. So if they wanna reach me, Ilya Volchow on LinkedIn is the best place.Sam Nadler: Amazing. Well, thanks for joining, and, Jorda, we'll see you next week.Jordan Metzner: Thanks, everyone. Thanks, Ilya. Thanks, Sam. Great episode. And, yeah, exciting times in AI, looking forward to what's coming out next week.See you. See you this week.