AI blew the entire budget: An interview with Lyaskovets and Patsan

“The AI system has blown the entire budget into the red”—this is a classic scenario, and there’s nothing you can do about it. Most media buyers who are just starting to integrate AI into their workflow face this frustrating situation. But we’re here to tell you that this is just the tip of the iceberg, because AI is capable of much more.

That’s exactly why the AFFHUB YouTube channel invited Myroslav Lyaskovets, Founder and Lead of Marketing, and AI expert Mykhailo Patsan. “Short and sweet”—that’s exactly how you can describe this episode, which is entirely dedicated to working with AI and all the nuances you need to know.

How do you use AI in your work and daily life?

Myroslav: My typical use case with AI is usually simple: create new creatives, rewrite some text, or redraw a design. And often, it’s Excel. I pretty much run all the AI processes myself because it’s interesting—I’m just curious to see what else it can do. You ask it a question about creatives, and it says, “Oh, maybe let’s make you a creative in this style”—you don’t even think about it sometimes. And you’re like, “Sure.” It spits it out. You’re like, “Okay, let’s test it.” It gives you the result, and you’re like, “That’s cool.” And based on that, you couldn’t even have imagined that it could be done that way. And it often works, converts better—it’s really awesome.

Mykhailo: When I wake up, my AI agent immediately tells me: “Look, today this happened once, twice, three times, four times—that’s interesting, think about it, write it down in a note.” Seriously, I wake up, and I have all this information. I do all of this with Claude. So Claude is basically my main tool for most of these kinds of tasks. I have separate projects everywhere, each one tailored to a specific purpose. And I’ve trained it to gather information. It’s just a foundation—you always have to keep training it; it always needs to have up-to-date data.

Where to start—what kind of subscription do you need?

Mykhailo: It’s best to forget about the free ones altogether—there’s nothing there. But you can start off comfortably with $20 and then see how much you’re using. Gradually, you’ll scale up. For example, my developers are currently spending around $2,000 a month—which means they haven’t cut any corners. Because we have this fully automated system: you just give it a description—everything you need—and it delivers the finished product. Because of this, we let half the team go. The most interesting thing is that nothing has changed: we deliver just as we did before, and even much faster.

Is it true that AI is free labor?

Mykhailo: Let’s debunk this myth: artificial intelligence is not a free workforce. You pay the same way, depending on usage volume. No less than for an employee—but it can handle more and scale up further.

Is it possible to launch an AI agent right now to manage Facebook ad accounts?

Miroslav: Facebook and Claude recently opened their APIs, and there are already colleagues who have implemented this integration—they say it’s a big deal. It integrates fully into the dashboard. You just type into Claude: “Which of my ad campaigns are really taking off?”—and it shows you. It also says: “Look, this campaign isn’t performing well right now, but I have a hunch it’ll take off tomorrow”—and suggests: “Keep it running,” or: “Shut it down; it won’t convert anymore.”

The only catch is that it can’t handle conversion metrics. If you’re running a gambling campaign, we have a funnel: ad, CTR, CPM, CPC, installation cost, and then registration and deposit. And here’s the catch: we’re already looking at deposits, because that’s exactly what we need to get. But they aren’t immediately available on Facebook—you’re looking at derived metrics. It’s still a bit behind; it needs to be trained further—fed more information.

Can AI completely replace salespeople? Wasn’t there some kind of experiment?

Myroslav: There are people who ran an experiment: they took ChatGPT and a bunch of other AIs, set them to work for a month, and allocated a budget to each one. They fed them automated traffic. And in the end—after a certain period, like a month—they all went into the red. Meaning those agents that were feeding traffic drove everyone into the red.

How can an AI agent close sales through Telegram chats?

Myroslav: One of the CRM systems we use for messaging—where Telegram is connected and all managers work through it—has implemented an AI agent. It saved all the dialogues with the managers: it pulled up the entire history of all the conversations, analyzed the logic of the exchanges—they wrote this to you, you replied with that—and based on that, it built a chain of dialogues on its own.

We’re turning it on. Obviously, it still has some kinks to work out, but the fact is: when someone writes to it aggressively or hypothetically, it gives a more or less adequate response—one that managers have already provided to a similar customer before—it pulls from that and formulates a reply. Or the customer writes: “Yes, I want to buy,”—and it replies: “Glad to hear it”—not in a templated way, but based on dialogues it’s picked up from some manager somewhere. One guy added a smiley face—it’s already changing its tone a bit. He sends a link and reminds them: “You came to register—have you registered yet?” I was sitting there when they showed me this, thinking: what on earth is going on?

And he doesn’t burn out. A manager typically works for six months to a year, and then something doesn’t sit right with him—say, his salary. But an agent doesn’t have those emotional lulls. He just does his job every day. And every day he learns—he gets smarter and smarter. And if he gets worse, you can reset his memory and train him all over again. You can’t do that with a person.

What does the ideal AI team for an arbitrage team look like?

Mykhailo: For a full-fledged team, you need at least three people. The first is an AI expert—someone who optimizes and configures the system. The second is a techie who understands security: basic security—you don’t need top-secret knowledge here, but you do need to know the basics. The third is someone who writes automation scripts. And if it’s a full team, a fourth person to collect data.

Ideally, you should look for such a person, bring them on board—let them observe and write code. Just like hiring a developer before, only now an AI developer.

How do you tell during an interview that an AI specialist isn’t just blowing smoke?

Mykhailo: Understanding the stack. That is, what agents are currently in use, what they’re based on, and how they were built. I asked this question during an interview. Most people last year built on n8n—it’s a platform on which various AI agents are built. I simply asked, “How did you build this?” And the person admitted they didn’t know. There are many like that.

How much does it cost and how long does it take to develop a full-fledged AI agent?

Mykhailo: The average cost for one such AI automation with an agent is approximately $20,000–$25,000. That’s about 3–4 months. And just like that, you have a ready-to-go, live production system.

But there’s a catch: nowadays, you can automate a lot yourself using Cursor or Lovable—simply by describing the logic or going through a bit of training to understand how it all works. And then you can build many of these small automations on your own. It only makes sense to do all this once you fully understand the entire process.

What will happen to AI agents in 2–3 years?

Mykhailo: In about 2–3 years, something even cooler will come along: it will predict and make decisions on its own—whether to do the work or not. Imagine, it’s sitting there, looking at you and saying, “Look, these ad campaigns of yours aren’t effective, but these ones are really great. Let’s scale them up for Canada, because I’ve tested them there, and my agent friends who’ve already launched them say they’re a hit.” And they already have a shared database where they exchange information.

Instead of a conclusion

AI has long since ceased to be a simple chatbot that answers basic questions and often makes mistakes. Technology is advancing, and now it’s on par with a full-fledged employee capable of performing a range of different tasks.

Of course, it still has room to grow. Think about it. Believing that AI will replace humans is a big mistake. Ignoring its existence and not integrating it into your workflow isn’t the best idea either. The ideal lies in balance. AI doesn’t take over all your responsibilities, but it significantly lightens your workload and is an excellent assistant.

Well, in this article, we’ve covered only the key points. You can find much more useful information right now in the new episode on the AFFHUB YouTube channel! 👇🏻