He Manages $100M in Ad Spend With a Trick Nobody Tried Yet

Sean Weisbrot: Everyone is using tools like Cursor and Claude to be able to vibe code, and you have a very unique function for it that you told me about previously. Why don't you share about it now? What are you doing? Because it's really unique.

Seva Ustinov: Sure. So I've been by, by calling myself, and I noticed that you can actually use cruiser and cloud code and other tools, not just for code, but also for documents, emails, presentations, project management, sales, uh, emails and, and everything. And like that was the. Aha moment. Uh, for me, when I realized how AI can actually help me automate like half of my work and half of my company's work. And, uh, now I'm like extremely excited about it and I'm, I want to show it to more people because it's cool.

Sean Weisbrot: Hey, business leaders and marketers. What if your brand could be featured right here? This ad spot could be yours. This channel is watched by a dedicated audience of ambitious founders, executives, and professionals who are actively looking for tools and services to help their business grow. If you wanna put your brand in front of this highly dedicated audience, that's difficult to reach. I'm currently looking for a few strategic partners for the channel. To learn more about sponsorship opportunities, click the link in the description. Let's go together. Yeah, I, I saw it in our last call and I thought it was really interesting because a lot of people are saying, oh, let me add a chat bot to Gmail. Let me add a chat bot to my calendar. Let me add a chat bot. But what you've done is essentially brought everything into cursor, into a GitHub repository so that you can just ask Claude to tell you about the documents that you've created. And I just think that's so cool because I don't think anyone else has really thought to do that.

Seva Ustinov: I think the best way to is just to show it. Do you want me to share screen?

Sean Weisbrot: Sure.

Seva Ustinov: So this is how my cursor looks like with files on the left document in the center and my chat on the right. And let's ask it to do something. Um, Hey, go to the strategy folder and use strategy context, and. Read company info, company story, product overview, uh, product vision, and also the last, uh, pitch deck. And come up, um, with a new idea how to pitch our product as AI first marketing operating system. So I'm talking to my cursor, so I, I don't have to type anything. I'm just think out loud what I would ask a real person, for example, and, um. Then I press enter and my agent goes, reads my rules about the company, the structure of my files, the processes we follow, rules not to hallucinate and not to create things I didn't ask for. Uh, then it read all those files, like product overview, per company story and everything. So it immediately knows everything I needed to know and created a plan how to solve my task. So with chat, CBT, that would take like at least five minutes to find those documents and upload them. Uh, here I just ask what I want and it comes up with, uh, ideas how to solve my problem. So the next step is to create actual documents and this is what's already happening here. And, um. This is a basic workflow. We will unpack it step by step to explain what's going on and, uh, how to use it. But yeah, this is, this is the base, the basic workflow.

Sean Weisbrot: And this has nothing to do with the product your company is making. This is just how you manage your company internally.

Seva Ustinov: Exactly, yeah. So we have. AI first product. We'll probably talk about it, uh, later in this video. And, uh, we have internal processes. So this is how we work. Right now. I'm showing you my example. I could use, I could build it my, just, just for myself. Um, yeah, here is the file. I can preview it and yeah, and this is new, new pitch strategy I ask for. Um, and it has a beautiful structure too. Yeah, just, just from one prompt, clean and easy to read.

Sean Weisbrot: Uh, it's, yeah, nice and clean and easy to read.

Seva Ustinov: It is the way it is because I have special rules and it took some time to, um, create those rules. Um, but yeah, now it works great for me. And, uh, let me show you a few more examples. Um. Um, hey, uh, this is a very lengthy document. Um, create another one, uh, with just one paragraph of text and five bullet points. Um, for example, in a way I could send it as a, in an intro email to an investor. Um. So what's happening here is on top of existing files that I've just created with ai, I can already reuse them, uh, to make like next layers of, uh, knowledge or documents or emails or something. And yeah, here it is. You can do the same in, uh, chat GPT, but only within one like chat thread. Um, so if you want to get back to this later in chat, GPT, that would be like hard or impractical, uh, especially with, uh, like chats you had like long time ago. But here, everything is saved to the same file structure. So whenever I want to return some, some, uh, work or things we've already done and continue from there or reuse it for other tasks, uh, it's. It sits in my context, in my file structure and I can do that.

Sean Weisbrot: Um, this is really cool. I wish that when I had my tech company five years ago, that this was optional or available. 'cause we would spend months putting stuff like this together. Oh yeah, we were thinking about every little word we were gonna use and making sure that it was just right, you know, on every webpage or in every, uh, you know, user story or just every time we had to write anything or do anything, it had to be perfect and it was like so slow.

Seva Ustinov: Yeah. To be clear, uh, I'm still responsible for every word here. So if I would send it to someone, I would definitely read it, adjust minor things manually or, or, or with ai. Uh, so you cannot trust it completely. But the more, the more I work with, uh, cursor, the more I update my context and my role and my rules, uh, the better it becomes. Uh, for example, like just last week, I had an urgent request from our, uh, PR team to give like comments to, to someone. Uh, and we had like three hours and the team was asleep. Uh, so it took me. 10 minutes to come up with, uh, the right answer. It was, and well, that was really straight to the point, specifically because it was based on everything we already have. Uh, here, if I would ask Chad GPT, it would come up with some generic fluff. Uh, but uh, here it's all based on the ground truth and there are special rules to not, uh, hallucinate or, um, put stuff that's, it's, is not in our knowledge base.

Sean Weisbrot: I'm sure that took a long time to make. 'cause I remember when I first started using lovable, they had this knowledge base system and you could put information in it, but you had no guarantee that it was actually pulling from the knowledge base when it was coding and, and prompt, you know, responding to your prompts.

Seva Ustinov: Yeah. Uh, it's actually like. Cumulatively it takes a long, uh, a lot of time and effort. Uh, but in reality it works like this, I think. Okay. Serving search clients with high hundred, millions in, uh, um, in cumulative ad spend. Uh, hey, actually we manage $100 million per year in ad spend. Uh, save it to the relevant file. Um. Probably to company info and update the speech deck and email this. Okay. It's not perfect.

Sean Weisbrot: Yeah, I saw it didn't get your, words exactly right. And that happens all the time.

Seva Ustinov: So this is like a real life example, I just work with it. Sometimes it make me, it makes mistakes, sometimes it doesn't do what I want, but, uh, instead of correcting it every time, I can just, uh, fix it once and this problem won't return. Because, because now it has the, the right knowledge and it marked it as canonical. So this is the accurate, confirmed information. Uh, and if you see, like I I, I, I ask it to mark everything as canonical or re-referred from somewhere else. And if it doesn't know what to write, it could put, could put a placeholder. So like we want to have this block of the text, uh, but we don't know it yet. So it's a placeholder to update it in the future.

Sean Weisbrot: I feel like you could go crazy just trying to make everything canonical. because there's so many files, you know?

Seva Ustinov: Um. Yes. And it's actually like not really a big deal. Like for example, this strategy folder, this is where I'm the owner of this folder and I work with it, um, a lot. But there are actually like courser rules here. Um, company story, company info, product overview and vision. And typically I just ask to load these file five files. Uh, and that's enough. Everything else can be loaded, like. When it's needed. So, um, typically like with the right structure, it's uh, it's, it is pretty convenient and not, not overwhelming. And also I can update rules so it loads these files automatically when I ask any questions or give any tasks, uh, related to things in this folder about strategy plans, road roadmap, fundraising, and, uh, things like that.

Sean Weisbrot: So how do you use it for like your, your roadmap? Because when, so like I was the product owner for the first few years until we hired a product manager and I had a document on Confluence. Mm-hmm. I had the, there's like another file that the QA team had, and then we had the design files in Figma. And so they wanted me to provide them with the relevant details from the right file along with the design files, so they knew what they were coding and how, what it looked like and how it's supposed to function. But even then, sometimes we would have mistakes where Al where something didn't align with the rest of, they'd be like, oh, you say that it does this, but the design makes it look like that doesn't work. Like, or. It, it couldn't possibly do the thing you're saying it needs to do. Um, or, you know, the designs work, but the feature is missing some steps that need to happen for this feature to actually work. So your, your user story or your requirements just aren't correct or they're not filled out correctly. Uh, completely. So I had all sorts of these issues, um, going through it. So how do you manage that? Or do you just have a person and, and you just leave it to them?

Seva Ustinov: Um. I am most, I'm mostly involved in, uh, deciding on priorities because there are so many things to do in, uh, marketing automation as automation. Um, and for that I need research. I need to talk to customers. I need to do sales myself. And I need to analyze all of those calls and everything and also do market research. So I'm not that involved in like pixel perfect, uh, things or like exact, uh, implementation. But it's my job, uh, to make right bets and to explain what we want to build and, um, how it will work altogether. Let me show you a couple more examples. Um, so I've asked it to do a deep research on what. People already automate around Google ads. Um. Google ads, meta ads. Where is it pulling

Sean Weisbrot: from that information? Like you, so you made this, you asked it this question. I, I How? Mm-hmm. Do you know that its information is correct for you?

Seva Ustinov: It's based on the deep research like this. This one specifically, I just manually went to charge GPT and run deep research. Now, I would just, uh, ask it to connect to parallel io. Use their API to to do a deep search if it's important stuff. If it's not important, I will just ask it straight away. Just use your search tools for the internet. Um, let me, let, let me show a few examples. Um, go to early analytics.com to the testimonials page and, um, add. Uh, a list of our customers and their testimonials to the user research folder as customer testimonials are great to actually extract what they consider valuable. Uh, and it's sometimes it's different from, uh, what we as like product managers or founders, uh, think so. Like in reality, we would have a folder in the sales, in our marketing team, uh, with full versions. Right now, just for the sake of this demonstration, I would ask it to pull it from the website so we could do most of day-to-day, uh, searches just from our agent, uh, agent in cursor provide, correct URL. Maybe it'll work, work better with direct URL. Um, same with, uh, deep researchers. We can do it here, we can do there, uh, uh, uh, outside, or we can use API for that. Uh, but once we have this data, uh, we could ask, um, cos to analyze. All of this, all of our testimonials and, uh, user stories and product interview transcripts and everything, and ask different questions like, Hey, what kind of, hmm. Like pain points, customers actually share with us, like themselves not what we found, not what we think, but what they tell us. Uh, let's prioritize this in, in this way or that way. Let's filter only those things that are related to Google Ads or meta ads or TikTok. Um, and hmm, let's create a rank list. So like if you do that in spreadsheets. Most people do that in spread, uh, in spreadsheets. Nowadays, whatever columns you choose will stay there forever. But with ai, with full context of your research files, you can rebuild these things on the fly. And again, you're still responsible for everything, for any decisions, analysis, highlights, everything, but it just takes like seconds to get a new perspective to do another line of analysis. Uh, yeah. Here are our customer testimonials.

Sean Weisbrot: Yeah, I, I started my, I started my tech company way too early. 'cause all of these things would've been massive undertakings. You would need people, you need to hire like a person whose full-time job was to do that. And you know, that's not to say that you wouldn't still need that now, but as a founder of a company that wants to remain lean while working towards becoming massively profitable and doing the next fundraising round, et cetera. There's so much more that a founder is capable of doing while learning how to build up to the point where you finally need a person. Where in the past we just had things that we just couldn't get done, and we just needed to hire someone to get that thing done because we didn't know how to do it. We just didn't have the expertise and we couldn't afford consultants and. I, so we started the company in 2018 and we folded in 2022, but it was, uh, yeah, there, I, I love ai. There's so much value that it provides now for such a low cost. It's like, you know, you could spend $200 a month and get the expertise of 20 different team members without needing to have them. Be involved, you can run much leaner teams.

Seva Ustinov: Absolutely. Uh, we haven't hired at least four or five people to date that we were considering hiring like a new person for this new role, and we decided not to because it was faster and easier to automate some process and workflows instead of having a set, a dedicated person for this or that.

Sean Weisbrot: Yeah. Thinking like an agency owner more than like a B2B SaaS, you know, tech founder,

Seva Ustinov: you could say. So yeah. Um, let's go through a few more, more use cases. So like till now we worked with just one folder and I'm the owner of this folder and I could do this myself without anyone else. But, uh, same thing. Does everyone on the team, so we have. Our potential clients list here. We have, uh, client meetings and transcripts here. Uh, we have everything from the marketing team. Like this is a demo folder, so not really everything is here. Uh, but uh, the structure is the same. Um, so like last week. Uh, I needed to answer to one of our, um, potential customers, and they asked a really tricky question about like, they had an experience with one of our competitors in the past and they didn't like it, so they were like checking how would their experience be different with us compared to that, compared to their specifically. So I asked my. Courser to go read all of our emails with them, all of our, um, call transcripts with them, go update that competitor analysis and come up with the answer. It did like a lengthy answer and I, I looked at it myself. I said, okay, uh, let's do a short version with just five bullet points. Prioritize this, this, and this, and also add this point. And the whole process took like 10 minutes. I didn't have to write to anyone, have calls, have discussions, set tasks to do a competi competitor analysis. Um, it's, it was faster to answer myself with courser, with the right concept context than to delegate the tasks to someone else.

Sean Weisbrot: Yeah, I hear that be because let's say you have a hundred customers that all ask you different things about different competitors. You now need one person or two people to manage research on a hundred different competitors or, you know, a hundred different situations. Maybe it's 10 different competitors, whatever. It's,

Seva Ustinov: so we have, uh, script that we basically wipe coded for ourselves and I can, uh, ask it like, go download my. Uh, emails with that person. Uh, it'll run a script and have files with emails in the right folder. Uh, or I can say, go download the, I had a meeting yesterday with Sean. Can you, um, go find it and put it in the right folder and then like, we'll create a follow up based on that. Uh. I cannot show here because it's like, it's a demo, folders without actual connection, but it literally works like that. Um, I think it's, uh, I, I should give a little bit more con before I show the next case. I need to give a little bit more context of, uh, what we do as a company because it's relevant. Um, so basically we work with. People and companies, well with companies that spend hundreds of dollars per month on digital campaigns, on Google ads, meta ads, tiktoks and everything like FinTech, health tech, um, software, a consumer software, AI companies, education. I think like you get the point, it's consumer software and consumer consumer services, not e-commerce companies, and one of their biggest. Issues with data is that to make decisions about ads, they need to centralize data from their backend, their payment systems, their CRM, their call tracking systems, like compared to ad uh, compared to e-commerce, it's not enough to just have ad platform data or web analytics like Google Analytics. You need all of that data stitched together, uh, to get reports like this. Uh, where for every. Um, channel, campaign and creative. You have full funnel data from ad spend. Yeah. Facebook, Google, Bing, TikTok, Reddit, Twitter, influencers, affiliates, SEO and so on with details up to ad sets and creatives with data from ad spend to. Web events to new paying customers to their lifetime value, to their lifetime value forecasts with different attribution models, like based on this data, you can make better decisions about the performance and scale what's working, um, compared to what you see on Google Ads account or Facebook account. Um, but to make all of this work, we have a team of, um. Customer success managers and like data engineers who actually connect to all the data sources and make sure that the data is accurate and everything is working correctly, and you can actually trust this data to make those decisions. Um, yes. So let's get back to That's the, the

Sean Weisbrot: biggest, the biggest thing. That's the biggest thing.

Seva Ustinov: Um, and imagine like. Running like 15 integration projects in parallel and how to manage all of those and all of those team members and project statuses and everything. So typically for customer success manager, it takes like, uh, an hour per day to update everything after calls. With clients and with internal teams and to track everything and not to forget every anything and to follow the right process, uh, and so on. So like the use case that really impress them is when they can ask something like this, Hey, go to the projects folder, uh, read the latest meeting transcript from, um. August 11th and update. Um, client beauty one status card, like in reality, it would, it would go to fireflies to download the latest, uh, transcript, uh, not to this folder. Um, and the status card is simple here, but it's follows pretty much the same structure as a real one. And we have special rules here. What to check, what to follow, what to avoid, um, which scripts to use, how to name files and everything. Um, and yeah, it read, it, read the transcript, it read the, um, card. And so I found that we fixed influencers at attribution, attribution promo codes, clarified, uh, the list. Um, and there are some actions required. Hmm. Regarding manual expenses that couldn't be pulled via API, um, I wouldn't save it right now because it's like my demo example. Uh, but think about like one hour per day per manager.

Sean Weisbrot: Yeah. And

Seva Ustinov: multiply that, uh, by like a year or something.

Sean Weisbrot: That manager still has to go and check the update to make sure that it's legitimate before keeping all of the safe. Maybe it doesn't take them an hour, but maybe it takes, I mean, they, it might still take 'em an hour depending on how many clients they're managing.

Seva Ustinov: Um, it's typically like in like this integration phase, it's, it is just a few per person, but still they're like very like, intense, uh, to track like two 20 moving pieces, uh, within each pro, each project. So yes, but instead of manually writing things and trying them to remember and writing them down during the meeting and then trying to remember what we were talking about, uh. They can like immediately do 90% of work and actually focus on what's important and say, okay, like you missed this point. We should actually focus, focus on this and at this point, uh, but like we can talk to it. You don't have to to, to type it yourself. So it's actually saves like 90% of time on manual work and it leaves more time to actually think like, what's, what should we do? What should we. How, how to like, uh, actually get things done in time accurately and not to forget anything. And that was impossible to do with any kind of project management software before.

Sean Weisbrot: Yeah. I'm a, a big fan of Cursor in Claude. I use them every day if I can.

Seva Ustinov: What's your experience? Like? What, what have you tried already?

Sean Weisbrot: I vibe coded a software from scratch to production Oh, wow. By myself with it. Yeah. Yeah. Um, I came across one giant bug that I haven't been able to fix, and I kind of left it alone, but I spent nearly four months on the, and I had four companies that were ready to pay me a few thousand dollars each to use it. Mm-hmm. But this, this bug was so big. It was the only bug that's left, but it's the only thing that's preventing them from, from me satisfactorily letting them use it. And so I decided it wasn't worth taking their money, and I was just gonna leave it because I, I, it was at this point, I would need a developer to fix the problem, and it could cost several thousand dollars to fix this problem. And if I pay the few thousand dollars to fix the problem, then I have to go and actually launch it and get the customers onboarded and serve them. And like, I wasn't sure I wanted to have an actual SaaS company because my last SaaS company was, is a very painful memory for, for me. And, but the difference being, we raised a lot of money and it took years to do where this, like, it's already done and I have people ready to pay. So it's like a different, I'm not gonna raise for it. Um, so yeah.

Seva Ustinov: Got it. Uh, it's actually like a typical thing to, to, to, to, to, like, to do prototypes fast, to experiment, to make, to like see what works, what's not, how it looks, how it's, how to implement it. But then like the final production version actually like still needs, uh, real developers to take care of everything.

Sean Weisbrot: Yeah. And I'm fortunate enough that I have the skills to be able to do that, but. I'm not a developer, so there's only so much, like, the only reason I was able to get it to where it was, where it was basically ready for production was, uh, my experience working with A CTO and frontend developers and backend developers. You know, I, I constantly communicated with 'em about the database, like I was for a non-technical founder. I was extremely involved in the tech because I wanted to understand what we were building and why we were doing it that way. I felt I was, you know, I know some, some CTOs might hate this, but like I wanted to learn, right? Like, I'm investing hundreds of thousands of dollars of my own money into this thing. I think I have the right to ask questions to understand how the damn thing's working. So I was basically buying a tech degree, you know, I paid a hell of a lot more than, than a typical tech degree would cost, um, you know, to learn all those things. And as a result, I can take an idea and I can get it to production. I don't have the skills to serve the customers and you know, so vibe coating allows me to get further, but I'm also very cognizant of the fact that I can't tell you how clean the code is. I can't tell you if it's scalable. Hmm, I can look at the code and I can kind of read the code, but I can't tell you if like the architecture is, is, is scalable. I can just tell you, you know, from the front end point of view, it looks the way it's supposed to look. It acts the way it's supposed to act and on the back end it's fine. But you know, maybe if I have five customers it breaks, you know, me using it by myself, maybe it's fine, but. Like, what's the breakpoint? And I don't know. And I need technical people for that.

Seva Ustinov: By the way. There is a trick there with courser and with charge PT and with anything, uh, typically, like in that case previously I would ask something like, Hey, what do you think about this architecture or this library or this framework we're using or this approach to solve this, all this issue? And it'll say something, uh, so I can go Google things and come up like with the. More grounded, uh, answer about like our architectural, technical, technical decisions. It would go read the manuals and say it would notice like stupid mistakes, but it wouldn't tell you like if it's scalable or uh, is it the right solution. But there is a trick you can ask it. Hmm. Let me even, um, hey, I want to create a new connector to fireflies. Go search for actual developers and users experience and tell me, um, um, what actually works for them. Um, so find successful cases and tell me what's the right approach to do that. We are API official and official MCP or whatever other options, but, uh, give me your answer based on real people's experience. So that one prompt saves like 50% of wrong decisions, maybe 90% of wrong decisions.

Sean Weisbrot: I think I'm gonna test this on, on my software because I, I'm highly confident I know what the problem is, but I also don't know this how to fix it. And I actually had gotten it working. But then there was some like little thing about it that I didn't like the way it had fixed it. And so I had like several prompts further that went to try to like fix that problem and it broke it again and I was unable to revert correctly so that I could go back to it working fine. And that's where I, my problem is, is that it basically got stuck in one of the like last five commits. Basically it was working and then it wasn't working again. And I, that's when I was like, you know what, like I tried for days to fix it and revert it to correct commit, and it just wasn't working again. And I was like, well, I just don't know what to do from here. I broke it, I, I fixed it and then broke it more than it was before. Um, and, and at some point you, you struggled to trust Claude to accurately tell you how it screwed up.

Seva Ustinov: Hmm. That's, that's true.

Sean Weisbrot: And this happened to me as well, actually, I, with the software, I had developed something, I had developed a feature set and then I developed another feature set on top of it because it was meant to kind of expand it and it broke the entire app. And so where I was in the app development, I said I spent like three or four days on it and I said, you know what? Forget this. And I just said. Destroy everything from the code base related to both of these features. And I started from scratch with the initial feature set. Made sure it worked perfectly well, and then I started to add it again. The, uh, expanded features set again, uh, and then it worked.

Seva Ustinov: Yeah, that's, uh, the way this look, it found, it's specifically mentioned that this like hoo approach, one of the default ones. It wouldn't work based on the real user experience. But where does it get its sources though? Uh, Reddit, GitHub people were places where people actually discuss such things

Sean Weisbrot: and, but you didn't tell it to go to Reddit. So how did it know to go to Reddit?

Seva Ustinov: Um, it either knows by default that those are the right places for this, or I have it somewhere in my course of rules. Uh, I'm not like 100% sure like what exactly is there because like maybe last time you didn't search for the right sources and I just ask it like, Hey, go to save to the rules that when we ask questions like this, you should research, uh, like GitHub and Reddit first. And by the way, interesting. Let, let, let, let's like get back from like coding to managers and CEOs and founders work. Um, because, um, I feel like this is a completely new approach to how things work, uh, how people work. Like right now, 99% of people I know they would use AI via chat, GPT or just like integrated AI to like Google Docs or Notion or something. Um, and. Basically use them as a magic answering machine, like better version of Google. Uh, and here, like it's a real life example. Um, how to make any kind of work in a new way. Uh, and I strongly believe that in the future, most of the work with any kind of systems will, will, will look like this. Of course, it wouldn't be a cursor, it would be like something nicer in terms of interface and everything. Uh, but the process of building up your context, bringing everything in one place, uh, updating rules and making your personal and your company agents, uh, more and more knowledgeable and smart. And following your personal routines like this is like the future of work, similar to how we all of us learned to use spreadsheets and docs and PowerPoint and then to Google and then chat GPT. This is how knowledge work will look like in one way or or another. What do you think about this?

Sean Weisbrot: I think it's a really interesting way of doing work. Like I said, uh, when, you know, when we spoke last time, I think it's very interesting. I think it's very clean. It makes your life easier. You probably save a lot of money on tokens, having to give it context every single time. So I, I think more founders will be pushed to work like this for sure. Yeah.

Seva Ustinov: Let me, I actually, I'd love to tell you a little bit more of about what we're building. So I already showed you how we work with, uh, cursor as a smart agent with all the relevant context. But if you want, wanted to work with like analytics or ads. That would be a more challenging thing because it's not just text. Uh, you have to work with databases and SQL, curious and scripts and API connections, uh, and lot of that. So we're basically building a course for performance marketers. Like we noticed that our. Best customers. They have multiple marketers on their teams, and what they do is they spend three, four hours per day looking at these dashboards and saying, okay, like last week we launched this creative. And it initially went well. So we decided to scale it, and then the metrics went slightly down because it started to burn out. So we decided to downscale it and then reload again. And when you have. Like two products, five regions, two different sales funnels, three campaign types in each, uh, group, 10 creatives in each, in each campaign that easily gets to hundreds or thousands of creatives running at the same time. So yeah, we, we decided to build an AI agent that can actually automate that, all of that with a similar approach so we can. I just start talking to it like, Hey, what, like who are you? Uh, what can you do? What kind of metrics do you have access to? And build me a simple report with seven key metrics for the last seven days, split by days. Um, so it's. Already knows like the basic details about performance marketing and analytics and all of that. Um, so Ellie, senior Performance marketing and data analysis assistant, it already has access to all those, uh, data sources and metrics. Uh, I mentioned before it can build. Sql, curious, uh, and execute them to get the right data on the fly. Like we can ask it basically anything we want. Um, yeah, it got the right data. It format months. It is a nice, nice report. Bring some high level observations. And that's just level one. Um, level. Hmm.

Sean Weisbrot: So this is kind of what I was building for my software. Hmm. So the, the original thing was having the expense data from the company, all of the expenditures. Mm-hmm. And pulling them from QuickBooks and ignoring all of the, the money going in. Right. So, ignore the income, only look at the expenses. And tangentially, I was building an agent that was connected to the data so that you could communicate it about your finances, about your expenses.

Seva Ustinov: That's an amazing use cases. I know how much people, um, spend time on this or ignore this and they have some problems, and it's like a perfect use case to apply AI because there are a lot of repetitive actions, a lot of analysis, a lot of, uh, things to check. And, uh, it's a perfect use case for ai.

Sean Weisbrot: Yeah, I was building it for, I was using to build it for businesses. Yeah,

Seva Ustinov: I think it's on our roadmap on the, like we have a tiny, tiny operations team, but still, uh, there are a lot of transactions with clients and with, uh, expenses. Um, so it's probably heading in that direction, but within courser, but having something like that already built and accurate and working and maintained and so on, that, that would be nice.

Sean Weisbrot: That's what everyone else I talked to said.

Seva Ustinov: Um, so next layer is to actually, uh, analyze performance. Uh, hey, show me ad sets and their performance over the last seven days with, uh, ad spend. Leads, cost per lead, oh no, not leads, uh, with, uh, orders and revenue and, uh, return on ad spend. Show me the most, the best a assets and the worst assets. So we can do like different kinds of analysis like this, and typically that, that would require you either right sql, curious or manually sorting and filtering things here. It's optimized for, for that kind of work with filters and grouping and like, there are tons of things here you, you can do with it. Uh, but. The problem is like whatever analysis you do in the end, you'll have to go and manually do things based on this analysis here, it's already completed, this best performing ad assets and worst performing assets, uh, actually now create an automated rule that will scale. Best performing ad assets, um, with this 70 timeframe and stop worst performing a assets with raw less than, um, 50%. It'll take like 60 seconds, but I already have a rule like this here. Uh, no. This one. This one? Yeah. So this is a list of ad sets. This is a recommendation by the agent, uh, based on the exact rules. So it's like, uh, it's actually creates, uh, code for these rules and executes them. Uh, here is all the details about this ad sets, so we can check like why. Was the decision made, we can click here and like stop this assets, or increase budget or limits for this one, or we can put it in auto mode. With approval and without approval, or we can change it. Like actually I see like there is a problem with, uh, this one. I, I wanted to exclude from the rules or I want special rules for specific products, or I want different set of rules for brand campaigns, uh, or different targets and limits for different regions. So whatever my. Workflow as an ad manager looks like, uh, we can just talk to agent and update to it to a way we want it to work, confirm that it's working and put it on autopilot. And that's like the most boring work in the world when you walk us through the numbers day by day.

Sean Weisbrot: For sure. Yeah. I hate looking at the numbers like that,

Seva Ustinov: and that's just the beginning. this is

Sean Weisbrot: really cool.

Seva Ustinov: Uh, imagine just one last thing here. Um. I was running a digital marketing agency before with like a hundred plus employees, 15 million in revenue, and uh, I always wanted. People to actually log changes and analyze each change and what are were the results. And like create a knowledge base about what's working for this project, what's not working, what's the aggregated view from different projects, uh, what did we actually learn from all those changes and experiments. So when you have like multiple people working on your head account, running different things, ideally. You want to have all that knowledge and data compressed, combined and reused for all future experiments. But it's so, so hard for people to do it accurately and intentionally and um, like because it's hard. But with AI, we can actually automate the analysis of all the change logs, all the automatically create look books automatically. Extract insights and like check with people, automatically update playbooks for all your campaign management and then run new experiments and processes and creatives and targetings and everything based on competitor analysis and all our previous experience. Like everybody in the industry wanted to build something like this for performance marketing, but that was just impossible before AI agents came in. And similar to how I built context, layers of context, uh, about my company inside Corer, were going to build layers of context about ad management and ad performance for each. Client with AI agents and actually close the loop of running multiple like hundreds of hypothesis per month, extracting knowledge and scaling what's working.

Sean Weisbrot: Thanks for watching. If you liked this insight, I've handpicked another video for you right here on the screen. For more actionable strategies that get you real results, hit subscribe.