Revenue is rising, campaigns are working, PNO and margin are holding. Everything seems to be in order. This is exactly how the data of many clients looks at first glance, who come to us with the question of why, despite the growth, nothing remains in their account. The answer is always in the numbers, which few people know how to connect.
Case Study: When Everything Seems to Work
This is exactly how this case looked.
On the level of marketing dashboards, everything was fine. Campaign performance was stable, orders were growing, and revenue was increasing, and they even managed to maintain PNO. There was no apparent signal indicating a problem.
Until a simple question arose: Why isn't the absolute profit growing?
This is the moment when most companies start to backtrack. Reports are reviewed, periods are compared, and deviations are sought.
We could have spent hours, maybe even days, searching for the answer in data that isn't immediately connected. Despite each advertising system showing positive development, none of them had access to real orders in ERP and CRM. And that's where the answer was.
It turned out that the problem wasn't in communication or product quality. It was in one product detail that would likely be overlooked in a regular report - the clothing sizes didn't match standard expectations.
Thanks to the AI agent, we discovered this anomaly by looking at multiple sources simultaneously:
1. First, it focused on products that made up a significant portion of revenue and looked for anomalies among them. However, at this level, it found nothing that clearly indicated a problem.
2. Then it expanded the view to the entire category where these products were located. Even there, no significant signal appeared; the category as a whole seemed healthy.
3. The next step was to move to the brand level. And this is where an interesting pattern began to emerge. A brand that included multiple products across different categories made up a significant portion of revenue and simultaneously showed below-average low PNO. At first glance, everything seemed fine again. The brand is a sales driver.
4. The difference only became apparent when looking at return rates. At the brand level, it was more than five times higher than average. This signal did not appear at the level of individual products or categories - individual products separately had a relatively small share of revenue, and the categories were composed of other brands that averaged out the overall return rate.
5. Only at this point did it make sense to go deeper. The agent connected data from multiple sources - product details on the web, analytical data from Google Analytics and Google Search Console, data from advertising platforms like Meta Ads, real order and return data from ERP, and simultaneously analyzed thousands of customer reviews and feedback.
It was in these that the same problem repeatedly appeared: the product sizes did not meet expectations. And this "small mistake" led to increased return rates, which had a direct impact on profit.
How Much Does an Overlooked Detail Cost
At first glance, it's a small thing. Size labeling.
On the data level, however, it's a fundamental problem:
- Product price - €33
- Margin 35%,
- which means a profit of €11.55 per piece.
With a return rate of 28%, it meant that approximately 2,500 pieces were returned monthly. Just on lost margin, it represented more than €30,000 monthly, almost €360,000 annually.
And we're still only talking about sales margin. When logistics costs, handling, or restocking are included, the real loss can climb up to half a million euros annually. All this in a situation that at first glance looked completely healthy on dashboards.
The Problem of Today's Companies is Not a Lack of Data
This case is not an exception for us. It is a consequence of how companies work with data today.
We see 2 common scenarios with our clients:
1. They have a lot of data, too much to process and read
2. They don't have properly set up analytics and reporting, and therefore the data is not of quality, and situations like this one cannot be seen in them
Analytical tools, advertising platforms, CRM, web data, or internal systems, each provides different data reporting options.
The result is an environment where important signals appear in multiple places at once, but their significance must still be manually assembled by someone who can connect and understand them in context.
AI Agents Bring a Revolutionary Shift in This
Instead of working with one tool or human capacity to process data, a layer emerges that operates over the entire existing stack. It's not another system added to the infrastructure, but a way to work with what the company already has, across analytics, marketing, web, and internal systems.
That's why we took the path of our own solution and built a custom AI agent as a controlled system over our own stack. We focused on three things: control over data, auditability of outputs, and the ability to work in the context of a specific project.
Why Don't We Use an Open-Source AI Agent?
There are many open-source solutions for creating AI agents on the market today. They are fast, accessible, and allow you to have a functional prototype in a few hours. We consciously chose to go a different route.
Instead of ready-made frameworks, we develop agents ourselves, through our own code and direct connection to models (currently, for example, Anthropic). The reason is simple: we want to understand what we are doing and how it works. It's in the DNA of ui42. Not only to use new technology but to explore it, understand it, and then develop our own solution for the needs of clients and the market.
This technology is only a few months old, and we cannot deploy something that new into a client’s production environment without full control and understanding.
Šimon Zámečník, Software Architect at ui42
Open-source solutions are great for a quick start, but they abstract key things "under the hood." With our own development, we have full control over how the agent works with data, how it makes decisions, where it makes mistakes, and what its real limits are. We can also better optimize performance, costs, and behavior in specific use cases. Thanks to this, we know how it found the error, how it finally arrived at it, and we exchange contexts mutually.
For us, however, control is not the only key. Understanding is crucial.
AI is Not About Blind Trust
An AI agent is not an authority; it's a tool. If you don't understand it, you quickly find yourself in a situation where you accept its outputs without critical thinking. And that's a problem.
That's why we see "challenging" as a key part of working with AI, not to undermine the technology, but to be able to challenge, push, and improve it.
Without this layer, AI becomes a black box that you trust but don't understand.
