AI in Marketing

I started testing AI quietly in late 2022, when OpenAI released ChatGPT to the public. I used it for internal work, personal projects, and side experiments. I did not talk about it much. I simply worked in the background, learning what worked and what did not. I built automations, refined prompts, and studied the machine learning behind large language models.

Now something interesting keeps happening: when I have an in-depth conversation with someone about AI, it sometimes turns into an impromptu consultation. People want to talk with me about AI. They even reach out because of AI, not Meta Ads.

That makes me wonder: why is there such a gap between what I see in practice and what people say about AI at conferences? The surveys look optimistic. McKinsey says 88% of companies “use AI.” But how many are actually scaling it? One-third. At most of the online retailers I work with, using AI means that someone opens ChatGPT from time to time. That is perfectly fine, but it is a different level from embedding AI into the decision-making process.

This article is about what those years have taught me: where I ran into problems, what works and why, and—most importantly—why most AI problems have nothing to do with AI.

Why AI produces nonsense

If I had to explain it in one word: context.

Imagine telling someone, “Drive to Brno,” when you actually want them to go to Karlovy Vary. Give the wrong directions and you will get the wrong outcome. That is not the driver’s fault, and it is not the AI’s fault. It is an input problem.

Garbage in, garbage out. It is the same principle we address with clients when we work on measurement. If I cannot provide reliable data at the start, the algorithm will not save me. Neither will the creative that everyone talks about so much today.

There is a second problem that receives less attention: the context window. Every language model has a limited amount of information it can work with at one time. Exceed that limit and the model starts to hallucinate—not because it is stupid, but because you have overwhelmed it. Give it too little and it fills in the gaps because it has nothing to draw from. Learning how to work within a context window—how much information to provide, how to structure it, and what to leave out—is a skill most marketers have never had to develop. Yet it determines whether the output is useful or merely polished-looking waste.

The way a model weights individual tokens can also make it fixate on the opposite of what you want. Write, “I do not want the copy to sound salesy,” and it may latch onto “salesy” and produce exactly that. This is why examples work better than instructions. Show the model one good output and one bad one, and you can replace ten paragraphs of explanation.

I did not learn this from a single article. I owe much of my understanding of machine learning—the field that includes the transformer-based systems behind LLMs—to Jirka Matern from Machine Learning College. Honza Barášek helped me understand the technical foundations: how a computer interprets code, the different types of databases, why SQL is so powerful for working with data, how React works (it was created at Facebook and underpins much of the modern web), and how backend security is handled. He did not gently introduce me to the subject. He threw me into it. Literally. Then he helped me stay afloat.

It is not magic. It is math and probability. Once you understand that, you work with AI differently. You stop blindly trusting outputs that sound convincing but do not make sense.

I see courses that promise to have AI build a landing page for you. Of course AI can make a landing page. But the page still needs a database and a backend. The form submissions have to go somewhere. API keys cannot be exposed in the frontend. What about SEO? Do you want the page indexed? What about measurement, the data layer, the favicon, and Core Web Vitals? A digital business card is not the same as a landing page with a form, even if you build it with Lovable or Bolt.new. Without context and an understanding of what is happening under the hood, an AI generator produces the average. The average looks great in a conference demo, but it falls apart in real work.

How I actually work with AI

Most people experience AI as a chat window: type a question, get an answer. To me, that is like driving a car without shifting out of first gear.

Today I do most of my work in Claude Code. At 12, I ran my own Warcraft 3 servers over a LAN and tried to become a scripter for Ultima Online. I could not program, but I reverse-engineered enough of a POL server to run a local environment. That approach—breaking a problem down into atoms even when I do not understand the whole—has stayed with me. Today I work with servers, RAG systems, Trigger.dev automations, and Hugging Face models.

Agentic workflows handle context very differently from chat. In a chat, context gets lost, the model forgets, and you repeat yourself. In an agentic workflow, AI can work with the entire project. It reads files, reviews documents, and returns to earlier steps. It is the difference between talking to someone on the phone and sitting them next to you so they can see your whole screen.

One of the simplest techniques is to tell the AI who it should be: “You are a web designer with 13 years of experience specializing in e-commerce.” That anchors the model in a specific field of knowledge. But that is only the beginning. I work with skills: structured knowledge files and references that the model loads for a task. I have skills for GTM implementation, analytics, Meta Ads account analysis, design, and UGC video scripts. Each one anchors the AI in a context and discipline. The result is a different world from a generic prompt. And you can create a skill like that yourself.

I also use AI to help write the prompt itself. I describe what I need, let the model propose a prompt, and then use that version. LLMs are often better at writing instructions for themselves than I am by hand. Prompt Cowboy is another tool built for this. Time spent preparing the input pays for itself many times over.

MCP servers—the Model Context Protocol—connected AI directly to the tools I use: Notion, GA4, Meta Ads, Figma, and Nano Banana. It took time for me to understand where the real value was. Here is one example: when I refine an SOP or methodology, I need to challenge the structure and find blind spots. An LLM helps, and the output can go directly into Notion. That has improved speed, but more importantly, it lets me go deeper. When AI takes over routine work, I can spend more time below the surface, looking for the correlations, patterns, and inconsistencies that a standard report will not reveal.

I use Perplexity for initial research when we start with a new client: a quick view of the category, pain points, and competitive landscape. No analysis is better than a bad one, but I still need an initial perspective so that I can ask the client better questions. Those questions often have nothing to do with the campaign: What email automations do you have? How well do you know your customer? Where do you lose people between the first and second purchase? AI provides a starting point. The client provides reality.

And Cowork? Have you ever tried to configure policies for a Google Cloud project? That interface can be painful. Seriously. With Claude Cowork or Google Shell, I can set up what I need without clicking through five levels of menus where every page looks different.

Reporting went from 4 hours to 30 minutes. Meeting summaries. Anomaly detection. If CPA jumps by 40%, I want to know before the client does. Why before the client? Because the moment a client comes to you with a problem you should have seen first, they no longer need you. An agency’s value is not in managing campaigns. It is in seeing things earlier and understanding them better. If you cannot do that, you are only a pair of hands on a keyboard—and AI can handle that on its own.

The answer is not black and white. One A/B test from PPC Hero caught my attention more than the large case studies: AI copy with human editing produced the lowest cost per lead. AI-only copy had twice the CPL. AI and a person together beat either approach on its own. AI generates variations quickly, while a person adds brand context, cultural relevance, and strategic intent. I see that pattern consistently.

Meta Ads: The offer matters more than the creative

Meta launched Andromeda in late 2024. The algorithm no longer asks, “Who should see this?” It asks, “What does your content tell me about the person you are trying to reach?”

I see it clearly in client accounts: narrow targeting generally does not work very well. A greater mix of formats and media helps. But the largest difference does not come from the creative alone. It comes from the offer. I spend more time analyzing the entry product—the first thing customers buy from a client, the way they enter the client’s world, if you will forgive the poetic phrasing.

You can have brilliant creative for the wrong product, and the algorithm will not solve that. CPMs in Advantage+ increased 94% year over year. It is an interesting correlation: we do things faster, but we pay more for them. Where does that equation converge?

After more than 30 tests, independent measurement company Measured.com found that the platform-reported ROAS for Advantage+ overstates reality because ASC allocates heavily toward existing customers. It is like a speedometer showing 180 when you are traveling at 90. Most agencies overlook it because platform reporting looks strong, and no one asks how much of the result is genuinely incremental.

Creative diversity is a competitive lever: 12–20 concepts, refreshed every 1–3 weeks. AI helps by accelerating the iteration of variants. But without a clear offer and an understanding of the customer’s entry point, even the best creative is only a nice-looking image that does not sell.

What this means—for me, for the Czech market, and for you

Almost no one in marketing talks about AI security. API tokens exposed in the frontend, conversations shared to train models—sometimes it takes one settings toggle to prevent it, but how many people know that? We manage campaigns worth millions, yet few people ask where client data goes. That should be basic hygiene.

There is plenty of discussion about AI in the Czech Republic, but until recently, the data was missing. APEK launched the first benchmark for AI in Czech e-commerce in January 2026. Rohlik, Alza, and Smartsupp are further ahead than many people might expect. But the average online retailer? Sixty-nine percent of Czech companies have no data strategy. A company that wants “AI personalization” without even having customer segmentation in place is solving the wrong problem. Technology is not the barrier. The foundations are the barrier. It is a leaky bucket.

The same dynamic is visible in agencies: 27% have already been asked to reduce their prices because of AI. AI is not eroding creativity. It is eroding the pricing power of good ideas. I am curious to see where that dynamic goes over the next year.

Clients sometimes ask whether AI is a threat or an opportunity. I think that is the wrong question. AI is a tool, and a tool changes nothing on its own. What matters is who holds it, why they are using it, and what context they bring.

AI accelerates what already works. It does not rescue what does not. If your data is messy, you will get more confident wrong answers. If you do not know who you are selling to, no model will figure it out for you.

I began by quietly testing AI on side projects. Today it runs a substantial part of my internal infrastructure. What has not changed—and will not change—is the need to understand what is happening under the hood, whether it is a campaign, a website, a dataset, or a language model.

The algorithm answers questions. Someone still has to ask the right ones.

---

I owe the technical foundations to Honza Barášek and the machine-learning foundations to Jirka Matern from Machine Learning College. The complete data research for this article, including citations from McKinsey, Gartner, the Czech Statistical Office, the EIB, Measured.com, and other sources, is available on request.