“If we want to create artificial intelligence that delivers real impact, we can’t leave it at the level of just an analyst or assistant. For it to be truly valuable, AI must be embedded into the business process,” says Yevheniya Bohdanova, Sales Director at the distribution company Euromix and a graduate of the "AI Workshop for Teams" program at kmbs.

During the event “AI Demo Day: The Road to Systemic Work with Artificial Intelligence” at kmbs, the program’s instructor, Yevhen Sarantsov, spoke with Yevheniya Bohdanova and Ivan Moskalenko, IT Director at Euromix, about how their team is integrating AI into business processes.

We’re sharing a conversation in which they explain how AI has changed team roles and their approach to developing solutions — and what kind of management logic this transformation requires.

"If we want to create artificial intelligence that has real impact, we can’t just leave it at the level of an analyst or an assistant. In order for it to have value, the intelligence must be implemented into the business process," says Yevheniia Bohdanova, Sales Director at the distribution company "Euromix" and a graduate of the "AI Workshop for Teams" program at kmbs.

During the event “AI Demo Day: The Path to Systematic Work with Artificial Intelligence” at kmbs, program instructor Yevhen Sarantsov spoke with Yevheniia Bohdanova and Ivan Moskalenko, IT Director at "Euromix", about how their team is integrating AI into business processes.

Here’s the conversation where they share how AI changed team roles and their approach to solution design, as well as what kind of managerial logic this kind of transformation requires.

Yevhen Sarantsov: Had you tried working with AI before joining the “AI Workshop for Teams”?

Yevheniia Bohdanova: Our team had varying levels of involvement. At that time, I was more of an enthusiastic user — I used artificial intelligence for personal needs: doing research, testing things, gathering data through regular ChatGPT.

Ivan Moskalenko: At that point, we had already completed an IT project related to EDI — a platform that allows electronic document exchange and automatic uploading into the system. But problems remained with clients who didn’t use EDI. They would send documents in arbitrary and unstructured formats: some in Excel, some in PDF, others as photos. Sometimes — just an image sent via Viber. That information went to our sales reps, who either manually entered the order into the system or passed it on to operators — and then operators processed everything manually. It took a lot of time and resources.

So we decided to integrate ChatGPT via API and built a service that automatically processed these documents. If it was an image, OCR technology would recognize the text. If an Excel file — an algorithm identified the structure and extracted the needed data. On output, we would end up with clean, structured information that could be directly input into the system without manual handling.

We also implemented product processing even without barcodes. If a client sent in a product entry without a code, the system would search our database for the closest matching product by description, find the corresponding barcode, automatically apply it — and create a complete record. Clients could send information in whatever format was convenient for them — no unified formats or standards needed.

As a result, we significantly offloaded both the operators and the sales agents. It wasn't just about saving time — there was a significant reduction in errors. Finally, we had a stable and controllable process.

Yevhen Sarantsov: Everything starts with an idea. The better the idea, the better the result.

Throughout the “AI Workshop for Teams”, participants go through different stages. Tell us about your experience.

Yevheniia Bohdanova: I was convinced that everything would work easily: there’s a large corpus of quality data — we just need to formulate a prompt, load the data into the model — and the agent will do the work: provide analytics, highlight deviations, and summarize conclusions. But the actual result was far from expected. Then I tried a different approach — I started creating structured tables. With the first such table, accuracy reached about 70%. That already felt like progress — we began to feel like we were moving toward something real. Eventually, we reached about 95% reliability in the results.

Even though that was a breakthrough, it also brought the first disappointment. We truly believed that it would be enough to “throw everything we have into the model” — and we’d get an answer. But once you start working with language models seriously, you realize: at the heart of everything is mathematical problem formulation and precise prompt logic.

Yevhen Sarantsov: During the workshop, we often said that we belong to the big data generation. And we’re used to thinking: the more data, the better. Just load everything into the system — and it will generate results. But working with LLMs [large language models] is not about big data. It’s a completely different paradigm.

Ivan Moskalenko: Here, what matters most is not the quantity, but the quality of the data — and more important — the context. This is a fundamentally different approach.

We uploaded Excel files, exported data from the ERP system, tried to generate a report — but either the results didn’t come, or they didn’t meet our expectations. For two or three days, we kept adjusting prompt wording, clarifying, explaining context to the model. Eventually, we turned to our technical team for help.

That’s when we got a key piece of advice: “Work with smaller chunks of data. Break large arrays into pieces. Only then — analyze them.” From that moment on, we started experiencing a kind of “swing” — first excitement from breakthroughs and insightful outputs, then disappointment when the system didn’t give predictable results. This went on basically throughout the entire accelerator.

At the time, we didn’t yet realize one critical point: the quality of the input data determines the quality of the model’s output. If the data is incomplete, poorly structured, or lacks context — the system can’t produce a valid conclusion, regardless of its computational power.

Yevhen Sarantsov: During the training, we discussed the importance of changing perspectives. That’s when you decided not to build another oversight tool, but to focus on creating a solution that adds real value to the sales rep and supports them during their customer visits. Tell us more about that experience.

Ivan Moskalenko: We began developing a system that allows a sales rep to receive personalized recommendations — specific suggestions about what to offer the customer during a visit. These recommendations are based on data: order history, client profile, current inventory, seasonality, product category effectiveness at that location, etc.

For example, the system might suggest which products from a particular manufacturer have the highest sales potential at this specific outlet, which to avoid because they typically don’t perform well in similar formats, or which product categories are currently underrepresented on shelves. This way, the agent receives a clear, automatically generated action sequence and arguments for client dialogue.

Yevhen Sarantsov: How did you apply what you learned during the program? Did it impact internal decision-making in your company?

Yevheniia Bohdanova: Absolutely. I realized that basic analytics can be delegated to an AI agent. So I created an agent that evaluates the sales rep’s route, analyzes the number and frequency of visits, and then formulates recommendations for the manager — where it makes sense to change visit patterns, where weekly visits are needed, and where monthly visits suffice.

It automatically classified information by sales channel and flagged which points had the highest revenue — and which needed more frequent visits. It also identified cities where sales reps were underperforming.

The most valuable feature turned out to be route analysis and optimization. For example: “These locations are currently visited four times a month — twice would be enough due to low efficiency.” Analyzing dozens or hundreds of stops manually is a complex task, even for seasoned professionals. Meanwhile, the agent does it in seconds.

That’s when we realized something fundamental:

"If we want to create artificial intelligence that delivers real impact, we can’t leave it just at the level of analytics or assistive tools. To have true value — the intelligence must be embedded in the business process."

That was a real breakthrough for us.

Yevhen Sarantsov: We've grown used to thinking that a manager has ten agents and makes all the decisions manually. In reality, the AI model cannot exist independently. It only works as part of a broader system. And that immediately reshapes process logic: roles change, work styles shift, people start doing things differently.

That was one of our core assumptions: we won’t succeed by merely building a smart assistant. If we don’t go further — that is, if we don’t integrate the agent into business processes and reimagine the role of the person working alongside it — it will ultimately fail.

Now we see the next horizon clearly. We can no longer afford the old approach: “Write a prompt, analyze the data, draw some conclusions about your routes — and do as you see fit.” It must be part of a systematic, coordinated approach.

Yevhen Sarantsov: What changed in your company’s approaches or processes after participating in the program?

Ivan Moskalenko: During the program, we developed a prototype that is still functioning. However, over time we encountered a problem — data quality. So the next step for our team was a deeper dive into the data structure and efforts to improve it. Now, the team’s main focus is on modifying and enriching the data set — making it more in-depth and accurate.

Yevhen Sarantsov: I remember during the program we started analyzing data at a national level and realized that context gets lost at that scale. As soon as we moved to regional data, things got better — we began to find more relevant insights. But the most precise and meaningful context comes from working with data from a single point of sale.

That became a turning point: it changed not only our analytical perspective but also our approach to developing AI solutions. We decided to work at the level of a single store and a single sales rep — and to generate personalized, relevant recommendations just for them.

Ivan Moskalenko: Technically, we addressed this by limiting the amount of context used, working within the capabilities of the language model. We narrowed down the input data, selecting only what had the greatest impact on decision-making. The ERP system pre-filtered the data, and only then was it sent to the AI model for processing.

Over time, we realized we don’t need dozens of departments — just a dozen or so ambassadors who can drive systemic products inside the company. Then a new goal emerged: to ensure that AI approaches are no longer isolated initiatives, but part of a systematic process.

Yevhen Sarantsov:

AI must become a built-in part of business processes.

It’s extremely valuable when a company has AI ambassadors or architects.

There are employees who can independently test prompts, evaluate models, and expertly explain how exactly they should work. Then there are others — those who use AI without even realizing it. For them, the updated system is like Google Maps — just a tool, nothing more. Your case with photos and receipts illustrates this well — the tool operates automatically, but most users are unaware it's powered by AI behind the scenes. Gradually, more and more people in the company find parts of their work governed, at least partially, by AI — even if they don’t consciously realize it.

Yevheniia Bohdanova: This raises an important point: a person has to want to improve their own productivity — that's the first thing. The second is that this person, especially in sales, has to truly understand the business context.

At some point, we may need to introduce a specific role — for example, an AI supervisor or head of the AI track. This person understands their operational context and is capable of working with business models.

Yevhen Sarantsov: AI today demands a new group of people within companies — those who consistently ask, “How can we do this better?” “What kind of model should we build?” “Where can we optimize?”

Ivan Moskalenko: And these aren't necessarily technical people. Quite the opposite — these are folks who didn’t even exist in the org chart before. Previously, effective automation required programming knowledge and was always associated with IT. Automation was perceived as a "tech spec": someone wrote it, passed it to development — and something would get made at some point. Now the paradigm is changing: roles can be redistributed.

Business teams now include people who deeply understand their own processes — the ones they deal with daily. These individuals can configure their own agents. IT, in turn, no longer does everything "turnkey" but supports the process, helps choose the right tools, advises on the tech stack — and becomes a technological partner. Not the “agent developer,” but more like the “co-pilot.”

Yevhen Sarantsov: This approach naturally creates at least three roles:

1. Those who can properly define the task — formulate the need, test the model, refine it, and pass it on.
2. AI architects — those who create or configure the solutions.
3. IT teams — supporting finished processes, ensuring technical sustainability, consulting, and integrating solutions.

Then there’s also a sort of “fourth role”: people who don’t engage with AI consciously but are already using tools that have embedded AI under the hood — without realizing it. They just click a button and get the result, unaware of what’s powering it.

Yevheniia Bohdanova: In our company, the culture of working with AI is just beginning to take shape. So for now, only a few people are involved in the process. But I believe that the more people engage, the more success cases we’ll start seeing. That’s how we can significantly boost the company’s productivity.

Ivan Moskalenko: Take our case with promotions. Since we are a distribution company, we don’t usually create our own marketing promos. Most often, we receive them from manufacturers and relay them to the market. Manufacturers send us these briefs in various formats — some by email, others through messengers. Then our marketing team gets involved: each brief is manually processed and entered into our internal system — a promo tool that allows us to log the campaign, track its efficiency, and forecast future outcomes.

Yevheniia Bohdanova: Later, we calculated that on average, around 900 promotions are entered into the system each month. Each one takes about 10–15 minutes to process. In total, that’s around 225 hours per month. In monetary terms, that came to about 30,000 UAH (Ukrainian hryvnias).

From my perspective as Sales Director, there’s additional complexity. If a manufacturer delays sending a promotion and instead of the 25th sends it on a later date, we end up launching it late — say, on the 2nd or 3rd of the next month. Sometimes, if the responsible employee is busy or has left the company, we have to wait for someone else to enter the promo into the system.

We analyzed how this process could work using AI. According to our estimates, the same number of promotions could be processed in 15 hours instead of 225. And the cost to the company would be around 4,000 UAH — instead of 30,000.

This way, we can operate much more efficiently. At the same time, our employees have the opportunity to focus on other tasks — which means a shift in roles.

Yevhen Sarantsov: It’s not just the roles that change, it’s the way of thinking. When people start working with AI, they gradually move from automatic task execution to seeking out unconventional solutions. They gain the ability to spot non-obvious patterns and indirect effects.

What’s most interesting is that these changes happen with the people who were initially skeptical, but still decided to try using the new tools. Those who overcame internal resistance discover a new way of thinking — more holistic and geared toward solving complex challenges.

All of this needs to be integrated into our operational systems. A sales agent can't be working in one system and receiving recommendations in another. It has to be one unified process with cohesive architecture.

Yevhen Sarantsov: Have you ever considered stopping your work with AI?

Ivan Moskalenko: This is definitely not a technology we should abandon. What can now be easily automated with AI used to require significant effort — or had no solution at all using standard tools. Today, for me, it’s a very powerful analytical instrument.

Yevheniia Bohdanova: It’s a new style of management. Thanks to the use of artificial intelligence, every person in the company — regardless of their role or position, including managers — gains the ability to make decisions much faster.

As the number of AI ambassadors and architects in the team grows, this capability will increase: decisions will be made more efficiently and their quality will improve. This happens because the data we receive with AI support is becoming increasingly accurate — assuming the prompts are well-formulated and the agents are properly configured.

If the model is tested, adapted to the business context, and trained on high-quality queries, its conclusions can be trusted. And it’s these very conclusions that can then move to the next level — becoming fully integrated into business processes.